Crossroads the ACM Magazine for Students WINTER 2014 VOL.21 • NO.2 XRDS.ACM.ORG

XRDSCrossroads The ACM Magazine for Students WINTER 2014 VOL.21 • NO.2 XRDS.ACM.ORG

Health Informatics The Anatomy of a Human Disease Network Did I Take My Meds Today? Challenges in Personal Health Tracking ACM’s Career & Job Center Are you looking for your next IT job? Do you need Career Advice? The ACM Career & Job Center offers ACM members a host of career-enhancing benefits: • A highly targeted focus on job • Job Alert system that notifies you of opportunities in the computing industry new opportunities matching your criteria

• Access to hundreds of industry job postings • Career coaching and guidance available from trained experts dedicated to your • Resume posting keeping you connected success to the employment market while letting you maintain full control over your confidential • Free access to a content library of the best information career articles compiled from hundreds of sources, and much more!

Visit ACM’s Career & Job Center at: http://jobs.acm.org

The ACM Career & Job Center is the perfect place to begin searching for your next employment opportunity!

Visit today at http://jobs.acm.org

ACM_printadv1_final.indd 1 11/21/13 11:36 PM Inviting Young Scientists

Meet Great Minds in Computer Science and Mathematics As one of the founding organizations of the Heidelberg Laureate Forum http:// www.heidelberg-laureate-forum.org/, ACM invites young computer science and mathematics researchers to meet some of the preeminent scientists in their field. These may be the very pioneering researchers who sparked your passion for research in computer science and/or mathematics. These laureates include recipients of the ACM A.M. Turing Award, the Abel Prize, and the Fields Medal. The Heidelberg Laureate Forum is August 23–28, 2015 in Heidelberg, Germany. This week-long event features presentations, workshops, panel discussions, and social events focusing on scientific inspiration and exchange among laureates and young scientists.

Who can participate? New and recent Ph.Ds, doctoral candidates, other graduate students pursuing research, and undergraduate students with solid research experience and a commitment to computing research How to apply: Online: https://application.heidelberg-laureate-forum.org/ Materials to complete applications are listed on the site. What is the schedule? Application deadline—February 28, 2015. We reserve the right to close the application website early depending on the volume Successful applicants will be notified byApril 15, 2015.

More information available on Heidelberg social media

PHOTOS: ©HLFF/B. Kreutzer (2)

ACM_HLF_ad-2014_v4.indd 1 10/1/14 1:41 PM Crossroads The ACM Magazine for Students WINTER 2014 VOL.21 • NO.2

12 begin

5 LETTER FROM THE EDITORS

7 INBOX

8 INIT HEALTH 2.0: The Digital Health Revolution By Diana Lynn MacLean

9 ADVICE The Essentials of a Computer Scientist’s Toolkit By Numair Khan

10 UPDATES Staying In Touch By Claudia Schulz

11 MILESTONES Vital Technology By Jay Patel

12 CAREERS An Antidote to Impostor Syndrome By Dean Jackson and Taliver Heath

14 BLOGS My $300 Home Cloud Server: A Story of Blood, Sweat, and eBay By Wolfgang Richter Image by Valentina Photo / Shutterstock.com

2 XRDS • WINTER 2014 • VOL.21 • NO.2 Health Informatics

23 44 62 features end

18 FEATURE 38 FEATURE 56 LABZ Gathering People to Gather Data Did I Take My Meds Today? The Regenstrief Global Health By Diana Lynn MacLean By Matthew L. Lee Informatics Group By Suranga Nath Kasthurirathne 23 FEATURE 44 FEATURE Opportunities of Social Media Seeing Is Believing 57 BACK in Health and Well-Being By Megan Monroe Radiography By Munmun De Choudhury By Finn Kuusisto 48 FEATURE 28 FEATURE Wearable Technologies: One step 58 HELLO WORLD Here Comes the #Engagement: closer to gait rehabilitation in The Anatomy of A serious health initiative made Parkinson’s patients a Human Disease Network trendy By Sinziana Mazilu By Marinka Zitnik By Fay Cobb Payton and and Gerhard Tröster KaMar Galloway 61 POINTERS 54 PROFILE 32 FEATURE Trevor van Mierlo: The Story 61 ACRONYMS Challenges in Personal Health of Building a Startup Tracking: The data isn’t enough in Health Informatics 62 EVENTS By Matthew Kay By Adrian Scoica˘ 64 BEMUSEMENT

We would like to acknowledge former XRDS feature editor Virginie Lerays, who was not properly credited in the last issue. She worked alongside Daniel Bauer and Adrian Scoică on the Natural Language issue. Left Image by Everything Possible / Shutterstock.com; Right Image by Sean Pavone / Shutterstock.com

XRDS • WINTER 2014 • VOL.21 • NO.2 3 M.S. programs at

EDITORIAL BOARD Poornima Sundaram SUBSCRIBE Editors-in-Chief University of Southern Subscriptions ($19 Inbal Talgam-Cohen California, USA per year includes XRDS Stanford University, USA electronic subscription) ADVISORY BOARD are available Sean Follmer by becoming an MIT, USA Mark Allman, International Computer ACM Student Member Science Institute www.acm.org/ membership/student Departments Chief Bernard Chazelle, Princeton University Non-member Vaggelis Giannikas subscriptions: Your future career, secured. University of Cambridge, Laurie Faith Cranor, $80 per year UK Carnegie Mellon http://store.acm.org/ Alan Dix, acmstore Master the technologies, economics and policies Issue Editor Lancaster University ACM Member Services Diana MacLean David Harel, To renew your ACM of secure communication networks, systems and Stanford University, USA Weizmann Institute membership or XRDS of Science subscription, please send services a letter with your name, Issue Feature Editor Panagiotis Takis Metaxas, address, member number Hanieh Moshki Wellesley College and payment to: HEC Montreal , Canada Noam Nisan, Hebrew ACM General Post Office Examples of research areas: University Jerusalem P.O. Box 30777 New York, NY Feature Editors Bill Stevenson, big data analytics operating systems Apple, Inc. 10087-0777 USA Erin Claire Carson browser security sensor networks University of California Andrew Tuson, City University London Postal Information cloud computing storage and file systems Berkeley, USA XRDS (ISSN# 1528-4981) Richard Gomer Jeffrey D. Ullman, is published quarterly in Internet of Things wireless systems University of InfoLab, Stanford spring, winter, summer Southampton, UK University and fall by Association for machine learning network architecture Computing Machinery, Suranga Kasthurirathne Moshe Y. Vardi, 2 Penn Plaza, Suite 701, network security … and more Indiana University- Rice University New York, NY 10121. Purdue University, USA Application to mail at Periodical Postage rates Numair Khan EDITORIAL STAFF New York University, USA is paid at New York, NY Director, Group and additional mailing Full federal scholarships available for U.S. citizens Talia Kohen Publishing offices. Bar Ilan University,Israel Scott E. Delman www.ini.cmu.edu Hanieh Moshki XRDS Managing Editor & POSTMASTER: Send HEC Montreal , Canada Senior Editor at ACM HQ addresses change to: Denise Doig XRDS: Crossroads, Billy Rathje Association for University of Puget Production Manager Lynn D’Addessio Computing Machinery, Sound, USA 2 Penn Plaza, Suite 701, Art Direction New York, NY 10121. Andrij Borys Associates, Department Editors Andrij Borys, Offering# XRDS0171 Mia Balaquiot ISSN# 1528-4972 (print) Arka Bhattacharya ISSN# 1528-4980 Columbia University, USA Director of Media Sales (electronic) Somdip Dey Jennifer Ruzicka University of [email protected] Copyright ©2014 by the Distinguished Speakers Program Manchester, UK Copyright Permissions Association for Computing Machinery, Inc. Permission Rohit Goyal Deborah Cotton [email protected] to make digital or hard West Chester East High copies of part of this work School, USA Public Relations for personal or classroom Bryan Knowles Coordinator use is granted without fee Western Kentucky Virginia Gold provided that copies are http://dsp.acm.org University, USA not made or distributed for profit or commercial Finn Kuusisto ACM advantage and that copies University of Association for bear this notice and the Students and faculty can take advantage Wisconsin-Madison, USA Computing Machinery full citation on the first Jay Patel 2 Penn Plaza, page or initial screen of of ACM’s Distinguished Speakers Program University of California Suite 701 the document. Copyrights Berkeley, USA New York, NY for components of this 10121-0701 USA work owned by others than to invite renowned thought leaders in Ashok Rao +1 212-869-7440 ACM must be honored. University of Abstracting with credit Pennsylvania, USA CONTACT academia, industry and government to General feedback: is permitted. To copy Claudia Schulz [email protected] otherwise, republish, post deliver compelling and insightful talks Imperial College London, UK on servers, or redistribute For submission requires prior specific Adrian Scoica˘ guidelines, please see permission and a fee. on the most important topics in computing University of Cambridge, http://xrds.acm.org/ Permissions requests: and IT today. ACM covers the cost of UK authorguidelines.cfm [email protected]. Apoorvaa Singh transportation for the speaker to travel IPEC Ghaziabad, India PUBLICATIONS BOARD Marinka Zitnik Co-Chairs to your event. University of Ljubljana, Jack Davidson and Slovenia Joseph A. Konstan Digital Content Editor Board Members Pedro Lopes Ronald F. Boisvert, Hasso Plattner Institut, Marie-Paul Cani, Germany Nikil Dutt, Roch Guerrin, Carol Hutchins, Web Editors Patrick Madden, Shelby Solomon Darnell Catherine C. McGeoch, Clemson University, USA M. Tamer Ozsu, Mary Lou Soffa

4 XRDS • WINTER 2014 • VOL.21 • NO.2 LETTER FROM THE EDITORS

A Shrimp’s Tale: Why we need to fund research or me, it all started with a YouTube video1 of a shrimp on an underwater treadmill accompanied by the “Benny Hill” theme song. Why was this shrimp on an endless journey, running seemingly forever, and why was it so gosh darn funny? The humor question is easy—you don’t see shrimp hitting the gym every day. But why did this video Fexist in the first place? Of course there are thousands, if not millions, of humorous videos involving treadmills on YouTube, but only this one was being investigated by the U.S. House of Representatives Committee on Science, Space and Technology. It turns out this video was part

of research that was funded by the fessor Robert Full, a biologist at UC UPCOMING ISSUES United States National Science Berkeley who works closely with ro- Foundation (NSF). boticists, uses similar videos of vari- Summer 2015 You might start to wonder why put- ous animals (from crabs to cats) on [June issue] ting a shrimp on a treadmill is im- treadmills to understand the biome- portant research, but let me assure chanics of their locomotion. It turns Computational Biology you, it is. Professor David Scholnick at out these animals are extremely effi- Article deadline: February 27, 2015 the Pacific University in Oregon con- cient. By understanding how nature ducted this research to understand can achieve such highly efficient loco- Fall 2015 how water quality can affect shrimps’ motion, we may be able to build more [September issue] movements and the distances they advanced and smaller robots. New, can travel in a short time. And Pro- efficient mini robots could be de- Virtual Reality ployed to search and rescue missions Article deadline: July 17, 2015 1 https://www.youtube.com/watch?v=Qj-yAHTf- because they are smaller and could VeE run for longer on a battery.

XRDS • WINTER 2014 • VOL.21 • NO.2 5 about the Large Hadron Collider at social networks and understanding CERN, often gets asked about evaluat- how people interact with and through We need research ing the results of such a large project computers. We need research that not that not only in financial terms. His reply: “I have only points to new, more efficient al- no idea. We have no idea. When radio gorithms, but also to more efficient points to new, waves were discovered, they weren’t ways to use computing to stop the more efficient called radio waves, because there spread of disease or to help people were no radios. They were discovered lose weight. This issue covers health algorithms, as some sort of radiation.”2 and technology, and what we see is but also to To me this is at the heart of why ba- that so much of this research really sic research is so important, because focuses not only on computing, but more efficient ways undiscovered knowledge can impact also on understanding people and the to use computing the world in amazing, unintended, intersection between CS and social and unforeseeable ways. Often ad- science, or CS and health or biology. to stop the spread vanced research topics may sound ob- This kind of research is critical for CS of disease or scure or arcane, or even unnecessary and for CS students to have an impact (much like a shrimp on a treadmill) in the world today. However, it seems to help people to outsiders, but experts in the field we as a community need to do a bet- lose weight. may think these are the very projects ter job of conveying the importance that should get funded. Right now in and value of our work. In the age of the U.S. a witch-hunt, reminiscent of YouTube, TED, and crowd-sourced the Golden Fleece award of the 1970s funding on Kickstarter, we need to and ‘80s, is on to find government be thinking beyond our publication funded research that should not have record. been funded. Currently the House One could make an argument that Committee on Space and Science has CS research could be done by the been investigating a number of NSF large tech industry. Certainly there research grants across a wide variety is great innovation coming from in- of topics, including computing proj- dustry in the form of new technolo- ects that investigate crowd sourcing, gies, new applications, and also new ubiquitous sensing on mobile devic- research topics (Microsoft Research es, and creativity support tools and is one of the largest publishers of CS gaming.3 And the time-honored tra- research). But, often the constraints This formative basic research may dition of anonymous expert peer re- that help industry come up with new seem unimportant or even comical, view of research projects at the NSF is solutions can be burdensome for yet it has the potential to provide dis- coming under close scrutiny as well. pursuing research that has no clear coveries that could change the way Of course, there are many issues with financial benefit today. And the fast robotics works or have other unfore- peer review, and we should no doubt pace of industry means labs come seen applications. have more oversight into how the NSF and go,5 interest in research topics Why should we do this basic re- allocates funds, but my fear is that change quickly, and internal funding search and more importantly why these investigations are very politi- is often short term, making it hard to should anyone, especially the gov- cally motivated. have long-term research projects. We ernment, pay for it? And why should In many ways this is part of a larg- still need publicly funded research in computer science research take place er goal of the House Committee on computing. in academia if there are thousands Space and Science to defund social Social science and applied re- of companies in Silicon Valley and science research.4 While we as com- search matter. And they matter in a around the world creating new, in- puter science students may think this CS context. We student members of novative technologies and products? has little to do with us, it actually hits the computing community need to These questions are becoming in- much closer to home. Computing stand up and make it clear that we creasingly important for us as a re- research is increasingly focused on see a future in which computing is an search community to answer as inter- integral part of any research agenda, est in federal funding for the sciences ensuring that computing has a broad 2 http://sploid.gizmodo.com/heres-the- decreases. perfect-answer-to-why-we-must-invest-in- and diverse definition in the years to Other fields have long dealt with pure-1665820877/+ericlimer come. having to provide financial justifica- 3 http://news.sciencemag.org/policy/2014/10/ 5 http://www.zdnet.com/article/microsoft-to- tion for their research. Physicist Da- battle-between-nsf-and-house-science-committee-escalates-how-did-it-get-bad close-microsoft-research-lab-in-silicon-valley/ vid Kaplan, theoretical particle phys- 4 https://www.insidehighered.com/ ics expert from John Hopkins and news/2014/06/02/house-passes-nsf-funding- ——Sean Follmer and producer of the film “Particle Fever” bill-takes-slap-social-sciences Inbal Talgam-Cohen

6 XRDS • WINTER 2014 • VOL.21 • NO.2 begin

INBOX

ON THE NLP ISSUE For #NLP people: did anyone find the easter egg we hid on the cover of the current issue of @XRDS_ACM (which is about Natural Language)? —Daniel Bauer, Natural Language Processing Ph.D. candidate, Columbia University, Twitter (@ dnlbauer)

My article “From wax tablets to touchscreens: an introduction to text entry research” is out in latest @ XRDS_ACM: http://pokristensson.com/pubs/Kris- tenssonXRDS2014.pdf … —Per Ola Kristensson, University Lecturer Department of Engineering, University of Cambridge, Twitter (@ pokristensson)

NEW ON THE BLOG Why and How #Parallel# Programming? A great post by Fahad at @XRDS_ that it’s pulling an absurd Blogger’s Reply: by writing code. ACM http://xrds.acm.org/ amount of power. As a Totally true. But for our use They learned fast! blog/2014/09/parallel- general rule of thumb, each case, research workloads, we —Andrej Bauer, programming-through- watt is $1 to run 24/7. That only turn it on sporadically Mathematician, dependence-analysis- means if it’s pulling 300 (2-4 times per week). computer scientists, part-i/ … #computing watts as I suspect, you’re —Wolfgang Richter, Twitter (@andrejbauer) —Pedro Lopes, Hasso paying an extra $200/year Graduate Student, School Plattner Institut, HCI more than more energy of Computer Science, researcher, Twitter efficient hardware. It Carnegie Mellon University (@plopesresearch) may still be worth it, but I wanted to point out the hidden cost. Sometimes OTHER TWEETS A QUESTION it’s worth replacing old @XRDS_ACM How to contact XRDS: Send a letter to the ABOUT EFFICIENCY hardware just for the power @marinkazitnik editors or other feedback by email (xrds@ Have you connected savings from My friends became acm.org), on Facebook by posting on our group page (http://tinyurl.com/XRDS- something like a kill-a- a financial aspect. wizards on MUDs (http:// Facebook), via Twitter by using #xrds in any message, or by post to ACM Attn: watt to see what the power —Jason Young, en.wikipedia.org/wiki/ XRDS, 2 Penn Plaza, Suite 701, New York,

Image by KreativKolors by Image draw is? I think you’ll find Azure Developer MUD ) and had to do magic New York 10121, U.S.

XRDS • WINTER 2014 • VOL.21 • NO.1 7 In the U.S., although medical records contain patient information, medical providers—including hospitals—own all patient records, which can be shared with insurance companies, pharmacies, researchers, and employers.

INIT HEALTH 2.0: The Digital Health Revolution

t’s the crack of dawn, (e.g., the increasing afford- helped doctors make better and Jane is on her ability and sophistication treatment decisions? What morning run with her of sensor technology) have if, armed with sensors and smartphone. She uses only served to more deeply analytics, we could crack Ia running app to pace her- embed Health 2.0 into our to code to getting a good self through the full six everyday environments. night’s sleep, every night? miles without stopping. Beyond ubiquity, how- In this issue, we bring She notices her split times ever, Health 2.0 has dra- you a glimpse of cutting- are a bit slower than usual, matically reformed the edge research that begins and wonders if it has some- role patients play in their to answer some of these thing to do with her rest- own healthcare. Just like questions, and more. From less night; she had a tough Jane, many of us are active the inside scoop on online time falling asleep. When participants in managing health communities, to she gets home, Jane exam- our own health: We turn to the question of social me- ines the charts her wireless technology to help us lose dia’s role in diagnosing scale generates. Over the weight, improve long-term and tracking conditions, to past two weeks, her body fat memory capacity, figure out the form of plain text, sur- technological innovations percentage has dropped, what triggers our asthma vey responses, or even raw for “hacking” our well-be- and her weight has gone attacks, and to collaborate biometric data—is both ing, to visual representa- up: She’s built some mus- and share experiences with abundant and easier to col- tions of medical histories, cle. Looks like incorporat- people who have the same lect, store, and share than to designing and building ing those strength sessions health concerns as us. And ever before. a gait training system for into her morning workouts as a result of our participa- In the past few decades, people living with Par- is paying off! tion in Health 2.0, health- Health 2.0 has grown into kinson’s Disease, we have While it’s hard today to related data—whether in one of the most vibrant packed in a collection of ar- imagine a world in which and active topics in com- ticles that represents the vi- technology and health are puter science, and it is easy brance and excitingness of not enmeshed, the phenom- to intuit why. In a world in Health 2.0 itself. Of course, enon is somewhat recent. which almost anything can such excitement is accom- The tech + health (or Health be measured, we have all panied by thoughtful cau- 2.0) movement is rooted in The idea that become, to some degree, tion. Issues of privacy, data the mid ‘90s, with the ad- we can leverage direct beneficiaries of the quality, and health policy vent of the commercial web. Health 2.0 movement. The remain pertinent and in The Internet made medical technology idea that we can leverage tension with newer techno- information available to on novel data technology on novel data logical advances. We hope anyone behind a browser: to improve both our own to leave you with a piqued No longer did people have to improve well-being as well as that of interest, a balanced per- to overcome financial, phys- both our own those around us is compel- spective, and most impor- ical, or geographic barri- ling. What if, for example, tantly, enthusiasm about ers to learn about diabetes well-being as we could use social media the current and future out- treatments, for example, or well as that of data to preemptively identi- look of Health 2.0. to research Lyme Disease fy disease outbreaks? What symptoms. Since then, oth- those around us if visualizing patients’ med- ——Diana Lynn MacLean,

er technological advances is compelling. ical histories intuitively Issue Editor KatherinePhoto by Breeden

8 XRDS • WINTER 2014 • VOL.21 • NO.2 A “data silo” refers to records owned and stored by one provider, but not shared with others.

ADVICE The Essentials of a Computer Scientist’s Toolkit

ho wants to set break- points when they can pepper their code with far more insightful state- Wments like p r i n t f(’’a s d a f s’’)? While learning how to use a debugger is a skill most programmers love to hate, any- body who has spent long hours trying to make sense of a crash report—often couched in equally insightful terms by the operating system—knows better. They say a wise person learns from the experience of others and as I hold a very high opinion of the mental capacity of my readers, I have compiled the following list of skills that I was either thank- ful for possessing as a computer science student, or, lacking which, was often cessor performance and the trend is simple scripts to the mix and before left pulling my hair in anguish. expected to continue in the future. you know it, you’ll have a “Live Free Programming in C. Disregarding This means sequential programs can or Die” license plate. the popular maxim that “real men no longer hope for regular speedups LaTeX. A typesetting language, La- (and women, of course) program in and all future programs must be, TeX makes it extremely convenient to C” on the grounds that it, perhaps, to some degree, concurrent. Not so produce high-quality scientific docu- expresses a somewhat subjective and much a practical skill as a state of ments. Not only does it help you as an unscientific opinion, the reasons for mind, the ability to think parallel is, editor, but also being the de facto stan- learning this beautiful language are, nonetheless, indispensable to pro- dard for research publications, it gives nevertheless, numerous. Not only is grammers who wish to exploit the full each document a well-established aca- it the progenitor of most modern pro- potential of multicore processors. demic look. gramming languages, it provides the Knowledge of basic probability. As a The Make utility. This last choice perfect level of abstraction for under- computer scientist, expect to encoun- may seem a bit controversial. But, standing the architecture of the un- ter uncertainty often and be prepared believe me, if I had a nickel for ev- derlying machine. In my experience, to analyze and quantify it. Familiar- ery time my eager efforts to compile C programmers generally possess a ity with statistical inference methods downloaded source code have been better ability to think low level and is especially important if you plan on foiled by the appearance of a cursed have a more solid understanding of conducting research. makefile. For the unfamiliar, make core programming principles. More- Unix. Computer clusters at aca- is a utility for managing and build- over, contrary to popular belief, C is demic institutions and large corpo- ing source files, and most online co- not a dead language. Nor is it difficult rations commonly run Unix (or one debases include a makefile specify- to learn, being amongst the simplest of its variants) and to take advan- ing the order of compilation and any and most concise programming lan- tage of these resources you should build parameters. If you intend to guages around. feel at home logging in, perform- interact with a large codebase and Thinking parallel. For the past few ing basic file operations, using the maintain a good head of hair at the years, multiple cores rather than high compiler, and editing text files from same time, learn to use make.

Image by Slavoljub Pantelic clock-speeds have driven micropro- a Unix shell. Add the ability to write ——Numair Khan

XRDS • WINTER 2014 • VOL.21 • NO.2 9 begin I 10 media—a newsletter. media—a of print forms ness using traditional chapter aware raise to skills creative their flexing reach but students, also organize about nine events per year, events about nine organize it possible to making basis, untary ondevelopment. contribute avol All photography, IT, marketing, web and management, event as such task, responsible is and bers for a specific 10 and mem two ment between has heads. Each depart 12 department made of the Body aGoverning ter has Chap treasurer—the and secretary, president, ofing president, the vice Body—consist Executive its to tion is: addi In organization structural a good what Student Chapter knows ACM the SZABIST Pakistan, Karachi, (SZABIST) in Science Technology and of Bhutto Institute Ali heed Zulfiqar to perfect it . it perfect to effort of alot spent and newsletter the issuing about enthusiastic more even was committee year’s This further How print and digital media can Touch In Staying UPDATES UPDATES Established in 2007 at the Sha 2007 at in the Established platforms, such as social media, to media, such social as platforms, ter that ter astudent you to chap introduce to of XRDS edition n this

is not only using digital using digital notis only student chapter outreach chapter student

, I would like , Iwould like

------Twitter and Instagram, following a following Instagram, and Twitter on announced also are updates tion, addi In likes. 1,000 more than has page its members; group Facebook 483 are There platform. munication com most important their is book Face Department. Media aSocial has puter science. Chapter the why is This field the in of com advances sharing and activities, chapter upcoming about updates providing community, student the of keeping touch in with importance of the aware very ter is documented. and out, carried advertised, which are (right). newsletter the in covered topics of apreview and (left) committee executive the and department rial edito the of head the introducing newsletter, 2013/14 the of pages two first The The Student ACM Chap SZABIST - - - - - published the first issue at the endissue at published of first the which formed, thus was Department world the of computing. Editorial An in news chapter as well about as the updates share to publication annual an idea of the creating introduced chapter’s year, the Body Executive Fernandes. Michael president chapter emails,” explained for than us better much works media social so work; cial [emails] them use for but offi mostly do here “People information. spread to used rarely is email that surprising It’s platforms. media cial not very so new these toward trend current During the 2011-2012 the During academic XRDS

• WINTER 2014 • VOL.21 •NO.2 •VOL.21 2014 •WINTER - - -

Photo Credit TK The Cyborg Foundation is founded by Neil Harbisson, who is known as the world’s first cyborg. Its mission is to physically apply technology to the human 2010 body to not only enhance, but create, “new senses and perceptions.

MILESTONES Vital Technology

Vital signs are measures of the body’s basic functions and can give an assessment of overall physical health. The four major vital signs: are body temperature, blood pressure, the academic year. This 12 page “mag- heart rate, and respiratory rate. The technology utilized azine” not only introduced the mem- to measure these signs is constantly evolving to be more bers of the previous and new commit- comfortable and precise. Here are just a few milestones in tee and provided information about the history of vital technology: past and upcoming chapter events, but also informed readers about cur- Santorio Santorio designs the first rent trends in computer science. Un- 1612 clinical thermometer. It consists of a fortunately, the following year the thermoscope—a container filled with buoyant bulbs—and committee was faced with some inter- a numerical scale. It is designed to be placed in the patient’s nal problems and was unable to issue mouth. a newsletter at the end of the year. As a consequence, this year’s commit- The first sphygmomanometer, better known tee was even more enthusiastic about 1881 as the blood pressure meter, is invented in publishing the newsletter and spent a Austria and consists of a rubber bulb filled with water and lot of effort to perfect it. a mercury column. The device measure blood pressure in Inspired by their leader Sana Haid- millimeters of mercury. er, members of the Editorial Depart- ment diligently got to work. Their Researchers in Germany report using red and articles highlighted the Chapter’s 1939 infrared light in an “ear oxygen meter,” which 3-D printing workshop and student works by measuring changes in tissue thickness and light conference on programming mobile intensity. This device is an early predecessor to the pulse phone applications, as well as driver- oximeter, which is a common non-invasive device used today less cars and Windows 8. Supported to measure blood oxygen saturation. by the Creative Department, which did an excellent job in structuring the Smartwatches are gaining increasing newsletter and designing the layout, 2014 popularity and feasibility as a tool for a hard copy of this second newslet- measuring vital signs. For example, one of the most ter was issued to academics and an publicized devices, the Apple Watch, is equipped with a built- online version was published for stu- in heart rate sensor using infrared and photodiodes. dents. The feedback from both groups was very positive, encouraging the ——Jay Patel chapter to begin work on publishing another newsletter next year. Feeling inspired? Have a look at the online version of the most recent ACM SZABIST chapter newsletter: http://issuu.com/alishbachapsi/docs/ acm_newsletter_2013-2014, or visit the Chapter’s webpage for more information: http://szabist.acm.org/about/. —— Claudia Schulz

XRDS • WINTER 2014 • VOL.21 • NO.2 11 begin

According to trends in Twitter data, people tweet about “health” most often between the hours of 6 a.m. and 9 a.m.

CAREERS An Antidote to Impostor Syndrome

mpostor syndrome is defined as “a psychological phenomenon in which people are unable to inter- nalize their accomplishments” [1]. IIt triggers thoughts like: “I don’t belong; someone made a mistake when they hired me.” Or, “My ID badge didn’t work on the door. Oh my God! They caught me! I’m fired.” This affects people across all races, all genders, and all ages. It most commonly affects those who are new to their work, and those who don’t have several role models in their workplace. It leads to the burnout of strongly capable people. Left unaddressed, it reduc- es what people can do in their personal and professional lives. Impostor syndrome affects most people at some point during their careers. Online estimates put the number at 70 percent of engineers, while a quick poll of 30 personal engineering contacts put it above 90 percent. The first step to fighting it is easy enough: Simply acknowledge impostor syn- healthy system of validation. Feedback learn, so they often rate themselves as drome as a common feeling, built into from others is never fully informed, as competent. That’s beneficial, as it lets human psychology. If that doesn’t others will not have full knowledge on us keep learning new things without work, here is a five-step approach to what you have done and why. You have initial discouragement! prevent it from slowing you down. this missing context, which can be es- But the second half is the part that pecially useful later for evaluating neg- stalls our learning and progress. Once STRATEGY #1: TRACK ative feedback. Tracking can help you you start to gain enough skill to judge YOUR ACCOMPLISHMENTS hold an honest light to criticism. what a terrific skill would look like, you Track your accomplishments in a for- realize how far away that might be. You mat where you can review them over STRATEGY #2: COMPARE WITH CARE then see others who are considerably time. The format is up to you. The log- There’s a related cognitive bias to im- more skilled than you are, and discour- ging doesn’t have to be long form; some- postor syndrome, called the Dunning agement can set in. thing short and concise is likely better. Kruger effect. Quoting from the initial Everyone makes a mistake here: When in doubt, focus on the positive study, “the miscalibration of the in- comparing without context If I compare here, and make sure to document your competent stems from an error about my skill at Java to another engineer’s, I successes, even the small ones. When the self, whereas the miscalibration of don’t know what it took for them to get you record a success, also make sure to the highly competent stems from an to their current skill. I don’t know what pause and reflect. Dwelling on success- error about others” [2]. other efforts or distractions they had. es as they happen can help to build and The first half of this is easy to see: The fix here is simple. In the first step strengthen long-term motivation. People who are new at something do of this approach, you started logging

Logging also helps to build a not know how much more there is to your work and successes. That gives you Image by Valentina Photo / Shutterstock.com

12 XRDS • WINTER 2014 • VOL.21 • NO.2 When a hospital employs more empathetic nurses, its ratio of positive to negative sentiments on social media jumps from 2.4:1 to 3.3:1.

one person you can compare against, do not excuse yourself from account- other people around you will provide with full context, and in a healthy way. ability. Do not dwell on failures, but do ongoing support. You compare and compete against a quick retrospective analysis to help If you get tired or stuck trying to work yourself. This helps stop comparisons avoid the same failure going forward. on this, actively find help. Ask someone against others, which is a common re- The postmortem is a technique you respect and trust (a professional, a curring trigger for impostor syndrome. used at Google to learn from failure. mentor, a leader in your community, or The postmortem is a concise written a family member who knows you well). STRATEGY #3: SET GOALS document that explains the following: These people can offer perspective and When you compare against yourself it 1. What went wrong? Add enough tools for calibrating your judgment and helps to have set goals, so you can see detail to bring it back to mind later. for staying engaged. how you’re doing and what you set out 2. The cost of it going wrong. How to do. It is important to never lower bad was it? YOU ARE NOT ALONE your self expectations based on impos- 3. A triage list. What tasks do you By keeping track of our successes, tor syndrome. Set your goals high, even have to do to fix it? comparing against ourselves and not if you miss them from time to time. 4. A preventative list. What can you others, setting aggressive goals in a You might set goals yearly, quar- change so this will not happen again? reasonable way, avoiding self sabotage, terly, or even weekly. They should be There is no punishment or negativ- and seeking additional help, we can ef- easy to grade. Examples include: “learn ity in a postmortem. Punishing our- fectively and mitigate the harmful ef- enough ObjectiveC to show a demon- selves for failures does not help most fects of impostor syndrome. Thus we stration app,” “learn enough French people learn to succeed. can improve both our happiness and to order dinner,” or “get an A in data Google postmortems are public to productivity in our careers and every- structures.” any employee who could potentially day life. Each person and each project at repeat the mistake. For use in personal Because enough people are report- Google has quarterly goals, which they tracking, the document can remain ing impostor syndrome, it is a regular set themselves. At the end of each quar- private. topic at colleges, corporations, and ter, they score the results, and partial Failures have costs, which can in- conferences worldwide. However, it is success counts. Getting three quarters clude time, money, reputation, and/or something that you can work on and of the way there scores as a 75 percent opportunities. Since we have already work through. Give it some thought, for that goal. If an engineer always hits paid the cost, a postmortem helps to and discuss it with a good friend or a all of their goals (averages 100 percent), reinforce the lessons learned, which friendly peer; you may be surprised at they know they’re aiming too low. If can make the costs worthwhile. Make their response. they get half of the goals or less, they the postmortem’s preventative list part know they set the goals too high. An ide- of your future goals. References al success rate is 70-80 percent, which [1] Impostor Syndrome. Wikipedia, The Free Encyclopedia. Wikimedia Foundation, Inc. 27 Oct. means successful engineers miss 20-30 STRATEGY #5: FOLLOW UP 2014. Web. 28 Oct. 2014. percent of their goals, every quarter. AND GET HELP [2] Kruger, J., and Dunning, D. Unskilled and unaware Setting goals like this is healthy, There are a lot of talks, papers, and oth- of it: How difficulties in recognizing one’s own incompetence lead to inflated selfassessments. while still pushing strong perfor- er resources about impostor syndrome Journal of Personality and Social Psychology 77,6 mance. Failure after strong effort is a available online; a quick web search (1999), 112134. part of trying hard, and learning to fail will turn up quite a bit. Go do some re- is an important skill to build. search, as unblocking your success is Biographies worth direct effort on your part. Dean Jackson’s a member of the ACM and a technical program manager working at Google, focused on Search STRATEGY #4: Impostor syndrome is definitely and Security. AVOID SELF SABOTAGE something to discuss with friends and Taliver Heath works in the technical infrastructure area of Google at the Mountain View campus, and has been Your tracking needs to be honest. If peers. It does not have to be a deep facilitating the Life of an Engineer program to new hires something fails, do not be negative, but chat, but just confirming it affects for the past five years.

XRDS • WINTER 2014 • VOL.21 • NO.2 13 begin

The Google-founded Calico lab researches age-related illnesses, such as Parkinson’s, ALS, and depression, which together affect more than 40,530,000 U.S. citizens.

BLOGS

justify indulging a new hobby without good reason—I was My $300 Home Cloud stuck waiting for just the right impetus. I didn’t wait long. My fiancée’s wish became my command! Server: A Story of Blood, I immediately related the story of my co-worker to her: Sweat, and eBay A server for just $40! Granted, we would want a little- endian, x86-64 architecture. Plus, for her algorithms By Wolfgang Richter and my virtual machine research we’d probably want a lot of cores and as much RAM as possible. Oh yeah, did “I really wish I had a dedicated Linux computer to run com- I mention I also wanted a beefier machine at home so I puter vision algorithms on,” said my fiancée a couple of could manipulate large virtual machine images? Virtual weeks ago. If you were there you would have been blinded machines (VMs)! We’d need CPUs with VT-x or AMD-V so by the metaphorical light bulb that lit over my head. You we could run VMs with accelerated hardware support. see, just the week before, my friend and co-worker had or- VMs run slow as a snail without acceleration. That would dered an old, decommissioned (complete with “non-clas- make the machine useless to me. sified” stickers!) Apple Xserve off of eBay for merely $40. Focused on my quest, I started scanning eBay listings Like my fiancée, he wanted to have a machine for a special daily. My co-workers even began to notice and started purpose: test compilations of open source software on a asking me if I was looking for something specific. big-endian architecture. I was quite envious that he was I responded I was toying with the idea of trying to able to hack on such cool hardware for such a cheap price. grab some cheap data-center-class hardware (for the But, I wasn’t yet ready to bring out my wallet. I couldn’t astute, cheap data-center-class hardware should be

14 XRDS • WINTER 2014 • VOL.21 • NO.2 Apple has been working on wearable health sensors— for example, apps that detect heart attacks before they happen—since the original iPhone’s debut in 2007.

an oxymoron). I was worried my project would end in failure, and wasn’t quite ready to announce to the whole Figure 1. NewEgg shipped these within a day! world my larger plans. After several days of failed bid attempts—I always seemed to get sniped in the last few seconds—I finally found what appeared to be the perfect fit. There is a lot of conjecture on where the Dell 1U rackmount model CS24-SC came from [1]. Some people say Facebook data centers [2]. Others just say that it was mass-produced for “clouds” [3]. Whatever these servers were used for, they were all retired by the thousands and can now be found all over eBay. The general consensus is Dell never sold these to general customers; the CS24-SC was a special custom-designed server sold by the tens of thousands to certain large customers. Thus, the CS24-SC has no support from Dell [4]. I haven’t been able to find anything outside of what random other CS24-SC owners have found in the years since the great decommission event. The CS24-SC has a few variations, but they don’t deviate too significantly. The one I had in my sights came with two quad-core Xeon E5410 @ 2.33 GHz CPUs. OK, fairly beefy compute from a few years ago giving us eight total real cores. It had 8 GB of RAM installed, which felt a little wimpy. Articles from second-hand owners online were conflicted on the maximum amount of RAM supported by the CS24-SC. Some said 48 GB [3], others said 24 GB [2] was the max. Well, it didn’t matter, because the greatest amount of RAM I could find at a reasonable price was 24 GB of data-center-class ECC RAM (PC2-5300P for the interested). Cool, what were we still missing? Oh, most of these second-hand machines don’t come with hard drives or hard drive caddies. After a quick visit to Newegg, I identified a cheap Seagate 1 TB hard drive (ST1000DM003) to slot in. Meh, Figure 2. Dell CS24-SC sticker, it’s authentic! let’s do this right and add in an extra hard drive for RAID1 to protect our work. I also threw in two CAT6 Ethernet cables so we could use both of CS24-SCs gigabit network ports, and a power cable. Well, that’s about it right? The server on eBay had its own case, 400-watt stock power supply, motherboard, and other needed components. We had to wait 1.5 weeks, but finally the CS24-SC arrived. I anxiously picked it up at our local FedEx location just up the street. My fiancée and I unboxed it together and hooked up all the components together. She tried putting in some of the RAM herself, so this counts as a date right? We were both worried that it wouldn’t boot, and in a sense that became a self-fulfilled fear. After plugging in a VGA monitor, we just had a

XRDS • WINTER 2014 • VOL.21 • NO.2 15 begin

Figure 3. It’s alive! Figure 6. See, it really does work.

Figure 7. Always check for bad hardware.

Figure 4. The HDDs connect fine without caddies. Figure 8. Installing Ubuntu.

Figure 5. Cabling in the back.

16 XRDS • WINTER 2014 • VOL.21 • NO.2 The approximate number of people worldwide who have received Cochlear implants, a device that bypasses damaged inner-ear 400,000 organs and directly stimulates the auditory nerve.

black, blank screen. Uh oh, maybe the hardware is bad? What were the total costs? Or maybe the RAM is bad? COMPONENT QUANTITY COST SUBTOTAL I really racked my brain thinking of ways to check on this system. I plugged in the Baseboard Management Dell CS24-SC Rackmount Server 1 $120.00 $120.00 Console (BMC) port into my router. Based on its 24 GB PC2-5300P RAM 1 $64.98 $64.98 DHCP client table, I guessed a certain device on my Seagate 1 TB HDD 2 $54.99 $109.98 network was coming from the BMC port. My hunch was confirmed when I port-scanned and discovered port CAT6 Ethernet 2 $1.99 $3.98 81 open and running an Apache server. After going to BYTECC Power Cable 1 $4.99 $4.99 the server in my browser I was presented with a login prompt. I was getting desperate and worried. I thought Shipping 1 $2.99 $2.99 that even if the VGA port was bad for some reason, we’d Taxes 1 $3.90 $3.90 at least be able to get into the remote console. But how Total $310.82 to get past this login screen? I tried several username/password combinations, and luckily root/root worked. I found out later online that is There you have it, a[n] [old] data-center-worthy the default username/password combination [5]. Thank home cloud server for only $310.28. Just for fun I tried you to whomever left this at the default, or reset it! If you customizing a hypothetical order on Dell’s website for sell a server with such a management console, please new hardware configured the same way our CS24-CS is: reset it if you customized it at all. It turned out the VGA It came out to more than $3,300—with discounts it only port wasn’t bad, we just didn’t have the monitor plugged drops to $2,400. Simply adding a second CPU on the Dell in before the BIOS flashed its screen. The system went to a website costs more than $500, more than our entire setup! blank screen after failing to boot an OS. The more modern hardware is faster, but our little CS24- Okay, phew, things seemed to be working. I downloaded CS is almost 90 percent cheaper. Thanks for reading my Ubuntu 14.04 LTS Server and copied it onto a USB stick. story of how I built a $300 home cloud server. And now Our CS24-SC had no trouble booting into the Ubuntu maybe you can too with a little elbow grease and eBay. installer off of USB. We installed Ubuntu, named our tl;dr Quick Server Specifications: server “phoenix,” after the ever-reincarnating mythical bird, and started customizing our CS24-SC. The two hard TYPE COMPONENT drives, 24 GB RAM, and whole system were recognized CPU Two Quad-Core Intel Xeon E5410 @ 2.33 GHz perfectly by the BIOS and Ubuntu. The only lingering Memory 24 GB ECC RAM issue I have is that Ubuntu doesn’t seem to properly Disk Two 1 TB HDDs (RAID1) display through the VGA interface after it boots. Grub Network Two Gigabit Ethernet Ports displays fine, and so do the early-stage kernel messages. Perhaps this is just a driver issue I need to track down. References Also, the fans on the PSU don’t spin, but it doesn’t appear [1] Dell CS24-SC Server. http://www.tedunangst.com/flak/post/Dell-CS24-SC-server. to be going bad yet. [2] The Definitive Guide to the Dell CS24-SC Server: Drivers, Config & Tips. Hurtig Hardware virtualization seems to work, and we are Technologies. June 5, 2014. https://hurtigtechnologies.com/2014/06/the- setting up our own work environments within Vagrant- definitive-guide-to-the-dell-cs24-sc-server. managed VMs. I’m using this opportunity to experiment [3] On the Dell CS24-SC Server. Rambling Geek. Nov. 13, 2012. http://www. aramblinggeek.com/on-the-dell-cs24-sc-server. with some advanced Linux functionality I’ve never tried [4] Dell CS24-SC Drivers. Dell.com Cloud Services Forum. 2012. http://en.community. before. Our two hard drives are not a traditional RAID1. dell.com/support-forums/cloud/f/4715/t/19456940. I’m using the new btrfs file system to mirror our root [5] Willis, R. Dell CS24-SC BIOS & BMC v2.5 Firmware Download. May 5, 2014. http:// partitions. There would be some work involved in setting robwillis.info/2014/05/dell-cs24-sc-bios-bmc-v2-5-firmware-download. up the second hard drive to boot, but we won’t lose our data. I setup the dual gigabit ports into a single bonded Biography virtual device using the Linux kernel’s balance-alb Wolfgang Richter is a fifth year Ph.D. student in Carnegie Mellon University’s Computer Science Department. His research focus is in distributed systems and he works under algorithm to try and balance inbound and outbound TCP Mahadev Satyanarayanan. His current research thread is in developing technologies flows across both ports. leading to introspecting clouds. tl;dr: Cloud Computing Researcher.

XRDS • WINTER 2014 • VOL.21 • NO.2 17 feature Gathering People to Gather Data An interview with Paul Wicks, Vice President of Innovation at PatientsLikeMe, a patient network and real-time research platform.

By Diana Lynn MacLean DOI: 10.1145/2676566

atientsLikeMe (www.patientslikeme.com) is a patient network and a real-time research platform. Through the network, patients connect with others who have the same disease or condition and track and share their own experiences. In the process, they generate data about the real-world nature of disease, which helps Presearchers, pharmaceutical companies, regulators, providers, and nonprofits develop more effective products, services, and care.

This sharing of online medical data and what drove you to pursue these issues. For example, we’ve worked has led to more than 50 novel studies what you do now? with Dr. Max Little to prototype tools that including: a patient-led observational PAUL WICKS: My job at PatientsLikeMe gather voice data from a Parkinson’s trial of lithium in ALS (amyotrophic boils down to this: Doing things for the patient from a phone call. Dr. Little lateral sclerosis), new patient-reported first time. I’ve been lucky to be a part believes a 30-second call—recording outcome measures in neurology, a of the team here for eight years now, the signature of their voiceprint—could “dose-response” curve for the benefits first as a community moderator, then a be a more accurate, rapid, and objective of friendship between epilepsy pa- scientist, and then leader of our research way of measuring their disease than tients, and new methods for gaining and development team. During that existing methods. We’re using our Open patient input into clinical trial design. time we’ve expanded from a handful of Research Exchange to help researchers Specializing in the conduct of on- neurological conditions like amyotrophic crowdsource the development of new line clinical research, Paul Wicks— lateral sclerosis (ALS), multiple sclerosis patient-reported outcome measures. who is the Vice President of Innova- (MS), and Parkinson’s disease to more We’re also charting the use of wearable tion—is responsible for shaping the than 2,000 health conditions—meaning health sensors that passively and scientific validity of the Patients- that anyone can join. The main core of continuously monitor activity and vital LikeMe platform and generating in- the platform involves patient-reported signs. Perhaps the greatest opportunity sights from the personal health data data on conditions, treatments, and is in connecting patients to their medical shared by members. In this interview, symptoms. While there’s no better source Wicks illustrates the progress made for how a condition affects people than by PatientsLikeMe and new challeng- patients themselves, there are some es facing the organization. issues with patient-reported information The Internet allows that we are working through. These BACKGROUND include ensuring the data is kept up-to- you to reach patients DIANA LYNN MACLEAN: Paul, you date, is accurately recorded, and that it faster, cheaper, and are currently VP of Innovation at isn’t affected by psychological factors. PatientsLikeMe. Can you tell us The types of projects we’re looking to in a way that’s more a little bit about what this means develop are those that start to address convenient for users.

18 XRDS • WINTER 2014 • VOL.21 • NO.2 records so they can import the most valuable data that they want to share with other patients or for research. I originally got into this area back in 2002 through a Ph.D. [program] studying ALS at the Institute of Psychiatry in London. By day I would go out and visit patients to give them a battery of cognitive tests; by night I would moderate an online forum for people with the disease to connect to one another. The Internet was nowhere near as ubiquitous then, but the opportunity seemed ripe—why see one patient a day when you could gather data from hundreds or even thousands all in one go? As a trainee scientist, it felt like most of my time was spent doing the logistics of data collection rather than actually thinking of new hypotheses or analyzing data. If we could get vast amounts of data continuously, then we should be able to dedicate more of our thoughts to what we’re meant to be doing—unraveling diseases! In 2006 some of the users of the forum I was running became the beta testers for PatientsLikeMe, which was founded by a family affected by ALS over in Cambridge, Massachusetts. As soon as I saw the site I knew they had something really special—the same research tools I’d been administering were right there for patients to use. Even better, they provided visualizations to help patients understand their condition in context, and connect with other patients who had walked in their shoes. Although I trained in quite a traditional environment relative to the world of startups and technology, I’ve always tried to keep one foot grounded in academic medicine. One of the first things we did when I started working with PatientsLikeMe was to try and get dataset that would have taken me five TECHNOLOGY, PATIENTS our first peer-reviewed scientific paper years to gather through traditional AND HEALTHCARE published. During my Ph.D., I had come methods given the rarity of the disease. DLM: The practice of patients sharing across a handful of patients who yawned Not only did we find excessive yawning health information with each other uncontrollably, sometimes hundreds was relatively common (occurring as a online manifested as soon as the of times a day, so much that their jaws severe problem in about 9 percent of Internet became publicly available dislocated. But as they say, “the plural patients), it was more common in certain in the mid-’90s. How has sharing of anecdote is not data,” so I could subsets of the disease affecting the health information online changed the never get anywhere having only studied mouth and throat, although there was no landscape of medicine and healthcare about 90 patients in my research. When correlation with measures of breathing from the perspective of both patients I stumbled across a paper reporting severity. We hadn’t just identified an and healthcare professionals? uncontrolled yawning as a side effect of unmet need among patients; we probed PW: I think we’ve seen a shift in the anti-depressants, I saw my chance. We the underlying pathology of the condition, past decade or so from those early text- added “excessive yawning” as a symptom and had gone from concept to publication driven set of interactions, which were in our ALS community and within weeks in a couple of months. From that point on, anecdotal stories from text-based forums had data from 254 ALS patients, a I was hooked. and message boards, to a more data-

XRDS • WINTER 2014 • VOL.21 • NO.2 19 feature

driven approach where patients want to digital communities or apps or wearable the “active ingredient” in intervention. help develop new measures, improve the devices, but the vast majority of these lie We’re just finishing up a report of a study clinical trial protocols, read the peer- outside the traditional healthcare system. recently completed with the Department reviewed papers for themselves, and even Professionals are under pressure from all of Veteran Affairs in epilepsy, where we run their own studies. For both patients sides, including ensuring they are getting recruited veterans with epilepsy to see and healthcare providers we see huge reimbursed for their time. If the “fee how PatientsLikeMe might affect their variation between conditions, and there for service” world were instead a world self-management and self-efficacy. The are all sorts of different factors that go in which fees were paid for improved results should be published shortly, but into that. The parents of a child with a rare outcomes, it would likely be easier for it’s been heartening to see that other developmental condition want to learn as these professionals to engage with groups can detect these benefits too. It’s much as they can, and will rapidly get to innovators. not just us tooting our own horn! the stage where they know as much as the The major pitfall I see with nearly any small number of experts out there (which EFFICACY OF ONLINE health technology is they’re a clunky bolt- arguably isn’t much). In cancer the state HEALTH COMMUNITIES on to your daily life. When you’re feeling of medicine has advanced to the point DLM: What are some of the benefits better you don’t want to be reminded of where you’ve got very deep molecular that patients derive from participating your illness and devote time to entering genetics, personalized treatment in online health communities? data, slapping on a wearable device, or regimens, imaging, all of which are difficult Correspondingly, what are some of filling out a survey. You just want to get for a patient to get a hold of and which the pitfalls, and what technological on with living. That’s why we need to learn are changing all the time. An ALS patient advances (form factors, interface design how to harness the whole tapestry of big might be willing to self-experiment techniques, algorithms, etc.), if any, do data being collected, from smartphones because they might feel they have few you see addressing these in the future? to pharmacy loyalty cards—all with the other options, whereas a psoriasis patient PW: We’ve published several surveys appropriate permissions, of course— whose condition flares up every few reporting the benefits PatientsLikeMe with an emphasis on returning value months might be less willing to take risks members experience, such as learning to patients, not creating commercial with experimental treatments. about a new symptom (72 percent opportunities for exploitation. We see the same level of variability of survey respondents agreed), on the health professional side. For understanding the side effects of a THE VALUE OF instance in ALS, our most developed treatment (57 percent agreed), and PATIENT-GENERATED DATA condition where we have a decade of finding another patient who had taken a DLM: As more people turn online for experience, we work with a consortium treatment they were taking (42 percent medical advice, the quantity of medically of more than 80 clinical experts called agreed). We’ve even seen benefits that relevant, patient-contributed data “ALSUntangled,” who use the Internet might be clinically relevant like improved available online continues to grow. to engage with patients who want medication adherence in HIV, reduced What is the inherent value of this data, experts to investigate complementary self-harm in mood disorders, and greater and could you give an example of an and alternative medicines. The group seizure control in epilepsy. Intriguingly, interesting discovery or insight that reads the scientific literature, checks in a follow-up study in our epilepsy derives from patient-contributed data? their case files, and reviews data from community, we found the greatest PW: The Internet allows you to PatientsLikeMe, then publishes their predictor of benefits experienced by reach patients faster, cheaper, and in a findings “open access” in the main ALS patients was the number of friends with way that’s more convenient for users, journal [http://informahealthcare. their condition they had made on the site. and lowers the barriers to conducting com/loi/aml]. There, you can really see As it turns out, patients themselves are research. Of course you’ve still got to healthcare professionals committed design your study well and realize your to meeting patients where they are, audience is easily distracted: You’re and responding to their questions competing for their attention with respectfully. I think healthcare providers Facebook and “Candy Crush Saga.” are pleased when their patients take Many healthcare There are few different classes a more engaged and activated role in of patient data we see online. One is managing their own condition, but a few professionals are just a straightforward like-for-like can sometimes feel threatened when interested in the of traditional methods for capturing patients begin educating themselves. data such as postal questionnaires, In fact, PatientsLikeMe published potential of all these interviews, or telephone surveys. The data suggesting about 10 percent of technologies, but second type of data is information that PatientsLikeMe users change physicians used to exist in silos but is now being as a result of information they’ve learned the vast majority shared openly. So, for instance, patient- from the site. of these lie outside reported outcomes (PROs) had been used I think many healthcare professionals in clinical trials for 30 years, but it was are interested in the potential of all the traditional information taken from patients and then these technologies, whether they’re healthcare system. locked away. Now, we’re seeing patients

20 XRDS • WINTER 2014 • VOL.21 • NO.2 who can enter their PRO data, use it to halt their disease might be within as a tool for gathering insights about The fact that they reach. Given the slow pace of medical themselves, communicate with other have to share their research, most of the patients who opted patients, and contribute data that can be to take lithium would have been ineligible aggregated by multiple researchers. data online is an for the carefully controlled trials that The third type of data is entirely new. indictment of the would take years to enroll, only to have a It includes the semantic density of email 50/50 chance of receiving a placebo and messages, sensor data from the GPS in healthcare system— being kept in the dark until the study had smartphones, and biometric data from why can’t they access finished. We partnered with some of the wearable devices. These things would pioneering patients in that community to have been high-end R&D tools in a few their medical records upgrade our tools to allow the recording highly selected places on Earth just 10 or online in a way that’s of dosage levels, side effects, and, 20 years ago. Now, they’re at everyone’s crucially, the same patient-reported fingertips. There are studies and tools meaningful and outcome measures widely used in clinical that can harness all three types of data valuable to them? trials. By visualizing the data continuously and mash them up in new and powerful and in real time, we quickly got the ways. That’s why I’m so excited to get up sense that lithium was not a miracle for work every day. drug, and when we subjected the data to more rigorous analysis, using data from DLM: What are some of the main hundreds of historical matched controls, challenges in extracting useful insights communities to drive, design, and even we were able to say more conclusively from these data, and how do you see us run clinical trials. Could you describe how that lithium didn’t work. This finding was addressing these in the future? this would work, and what it would take subsequently replicated by a number of PW: We’ve been making incremental to meet standards of medical validity? clinical trials that followed in years to improvements since the beginning. PW: When we started operating as come. In terms of credibility and validity, Today PatientsLikeMe gets an upgrade a business, one of the ways we made one thing we’ve learned is when you’re about every two weeks. However to really money was by raising awareness about doing something new you’ll be held to catalyze this whole field, I think we need specific clinical trials. In one example even higher standards than the status to look at the consumers of this data, we helped Gilenya recruit MS patients quo. That’s why we published our findings the decision makers, and ask them “what for their clinical trial, and they stated in a high impact peer-reviewed journal— would it take for you to integrate patient publicly that we sped up recruitment by Nature Biotechnology—published all the data gathered online into the decisions several months. Later, we built a mash- details of our matching algorithm, made you make every day?” In healthcare up with ClinicalTrials.gov so any patient sure the paper was open access so that those are decisions about what drugs to could find out about any clinical trial for anyone could read it, and even uploaded approve, the value at which treatments which they might be eligible, regardless of a de-identified copy of our dataset so should be reimbursed, the definition of whether it was sponsored or not. But we others could replicate our findings. I’d like good quality care—big decisions. When knew just feeding the existing machine to say this made our position unassailable those decision-makers signal that yes, wasn’t enough, so we sought to find ways but, even today, the ALS field continues under the right circumstances they would patients could really change the system. to minimize the impact of this study and take patient-generated data seriously, In many ways, we followed the lead of claims because it was patient-reported then you’ve just aligned a whole group of patients themselves rebelling against the the finding was null and void. individual stakeholders—entrepreneurs, system and breaking the norms of clinical We’ve also taken another approach academics, patients, investors, trials. Back in 2008 a small Italian study to clinical trials, which is to, rather than advocates—to solve the challenges in was published. This study suggested condemning the whole system, ask front of us. lithium carbonate effectively delayed what can patients do to optimize it to The big challenges are the usual the progression of the normally fatal be more patient-centric. This year we’ve suspects: data quality, data density, disease ALS. Although that study was very undertaken a number of projects with validation, reliability, bias, and small, just 16 treated patients and 28 commercial sponsors to gain patient generalizability. But the good news is controls, it caused a big stir in the patient feedback on aspects of the trial protocol, we’re working on these issues every day. community and within six months we recruitment materials, and outcome And because this whole ecosystem is built had 10 times that number. One hundred measures so that trial designers can upon technology, we’re going to benefit and sixty patients, who had got hold of make trials more appealing to patients, from consumer technologies like tablets lithium off-label, were sharing their data less burdensome, and ultimately more or smartwatches as they become more on PatientsLikeMe to see if the drug was aligned to what matters to patients. In widespread too. working. Some clinicians thought this was this way we hope we can bridge a gulf dangerous and might harm recruitment that risks developing between trial CHANGING MEDICAL RESEARCH for the formal randomized control trials designers focused on their scientific DLM: Some of your recent work has to come, but patients weren’t satisfied to and commercial goals versus patients explored the idea of using online sit on the sidelines when an opportunity trying to improve their outcomes. It’s still

XRDS • WINTER 2014 • VOL.21 • NO.2 21 feature

early, and only a tiny handful of trials are experience they are vulnerable, but they designed with systematically gathered We need to learn also realize they are among friends and and robust data about patients’ views, how to harness the hopefully they benefit enormously from but I hope we can help to change that. what others like them have shared in a Unlike the scientific community we find whole tapestry of big vulnerable moment, so the virtuous cycle businesses can be more responsive. data being collected of sharing data becomes apparent. If we can show we help avoid a costly In part, the fact that they have to commercial failure or an expensive with an emphasis share their data online is an indictment protocol deviation that slows down their on returning value of the healthcare system—why can’t path to regulatory approval, then they’re they access their medical records online willing to accept evidence that it works. to patients, not in a way that’s meaningful and valuable creating commercial to them like their personal health record TACKLING A VARIETY is on PatientsLikeMe? When you look at OF CONDITIONS opportunities for the emphasis of most health IT systems DLM: Are certain medical conditions exploitation. though its billing, locking the data down in more amenable to being the focus of a silos, and driving lower cost care through successful online health community? physician workflows. I’m afraid to say If so, are there particular attributes patient benefit is either low on the priority (rarity/prevalence, stigma, etc.) that list of most electronic-record systems or correlate with this? transplant patients, and when they’re on absent altogether. PW: Absolutely, we even have the waiting list or awaiting surgery then I’ve been doing this for eight years an internal term for it; we call it the their engagement is very high—they’re now and helped to build communities “PLMability” of a condition. Where we looking for tips, sharing information, and of more than 250,000 patients and seem to do best is in chronic conditions looking for support. But once they’ve had I’m pleased to say I have yet to hear that have a major impact on patients’ the transplant, assuming it’s successful, about a major breach of medical privacy lives, where they feel directly affected then they just want to get on with the rest resulting from sharing data online, but by the condition (unlike, say, high blood of their lives. That means we do probably it’s certainly a concern. Then again pressure), and where their own knowledge have unknowable biases in terms of who when online banking first appeared, the and behavior is likely to have an impact keeps coming back, and again that’s why world seemed terrified of identify theft, on their outcomes. For instance we often we’re interested in exploring other ways of and while that’s certainly been a valid say if you break your leg in an accident, receiving information from patients that concern it’s also the case that we went you wouldn’t necessarily need a system might be useful to them if they need it, ahead and rolled out online banking to like PatientsLikeMe. You have an acute, but that doesn’t require they keep coming the whole world to great benefit. very curable condition that will be getting back to submit more data. Privacy is certainly a very real better soon, and the medical system concern, but I think it’s time we gave knows how to treat it. Contrast that with PRIVACY more credit to patients in making their a neurological condition, like multiple DLM: Data privacy is a growing concern own decisions—people can drink or sclerosis, where you’re going to live with in today’s technological landscape, and smoke or drive cars or go skydiving or it for the rest of your life, we really don’t legislation around safeguarding medical buy a handgun, but for some reason when know what causes it or how to treat its information is particularly strong. Yet they become ill or want to share medical many varied symptoms, and your needs patients seem willing to share detailed information about themselves, the health may change dramatically over time. health information with each other in system suddenly become very concerned Even within those conditions for which public online forums. Why do you think that they don’t know how to act in we’re highly suitable, we see variations in this is, and how do you see it playing out? their own best interest. It’s certainly a when patients come to us in the course PW: There are a few factors at play rapidly evolving area and we are always of their disease. We see a lot of people here. On our site, patients are making conscious of the trust patients put in who’ve been recently diagnosed who are an informed decision about the risks us with their health data. It’s up to us just trying to find out everything they and benefits that sharing their health to honor that trust and always try to be can, and we see another group of relative data online might bring to them. (See one step ahead of those that would take veterans who have been managing their our privacy policy here: http://www. advantage of their generosity. condition well for years but are now patientslikeme.com/about/privacy.) experiencing something new they don’t They’re only sharing as much information Biography know how to deal with. They come back to as they feel comfortable with. For Diana Lynn MacLean is a Ph.D. candidate in the Computer Science Department at Stanford University, where she is the community to draw upon the wisdom instance if they’re on the site reporting advised by Jeffrey Heer. Her thesis comprises cross- of the crowd. data about lung cancer, they might not disciplinary research focused on extracting medically relevant data from patient-authored text on the topic of Sometimes there’s a conflict between be reporting another condition they have substance abuse. Out of the office, Diana is an avid rock the research needs of a scientist and such as a mental health diagnosis, and climber and Tomb Raider fan. Diana received her B.A. from Harvard University in 2009 in computer science. the lived experience of a patient. For that’s fine. They’re putting in and getting example we have a community of organ out what they want to. By offering their © 2014 ACM 1528-4972/14/12 $15.00

22 XRDS • WINTER 2014 • VOL.21 • NO.2 Opportunities of Social Media in Health and Well-Being Intelligently leveraging data from millions of social media posts is a modern public health approach that has the potential to save many lives.

By Munmun De Choudhury DOI: 10.1145/2676570

eople are increasingly using social media platforms, such as Twitter and Facebook, to share their thoughts and opinions with their contacts. One in six people in the world today is a user of Facebook [1]. In a way, social media has transformed traditional methods of communication by allowing instantaneous and interactive Psharing of information created and controlled by individuals, groups, and organizations. An important attribute of social media is that postings on these sites are made in a naturalistic setting and in the course of daily activities and happenings. As such, social media provides a means for capturing behavioral attributes that are relevant to an individual’s thinking, mood, communication, activities, and socialization. Moreover,

this real-time data stream of social in- of research is how computational nuanced predictions of flu infections formation is often annotated with con- techniques may be applied to natural- based on online search queries. Paul text including location information, istic data that people share on today’s and Dredze developed a disease-specif- cues about one’s social environment, online platforms in order to make ic topic model based on Twitter’s posts and rich collections of multimodal in- sense of their health behaviors and in order to model behavior around a va- formation beyond text, such as images related experiences. In the context of riety of diseases of importance in pub- and videos. this surge in research interest, this lic health [5]. Through language mod- With the increasing uptake of article highlights the potential oppor- eling of Twitter posts, Culotta found social sites, there has been a corre- tunities and challenges in the use of evidence of high correlation between sponding surge of interest in utilizing social media as a novel stream of in- social media signals and diagnostic continuing streams of evidence from formation for augmenting traditional influenza case data [6]. Sadelik et al. social media on posting activity to re- approaches of health assessment. developed statistical models that pre- flect on people’s psyches and social dicted infectious disease spread in in- milieus. In fact, the ubiquitous use ROBUST, LARGE-SCALE SOCIAL dividuals based on geotagged postings of social media, as well as the abun- MEDIA MARKERS OF HEALTH STATES made on Twitter [4] (also see Moreno et dance and growing repository of such Leveraging Internet data for model- al. [7]). data, has been found to provide a new ing and analyzing health behaviors In the behavioral health domain, type of “lens” for inferring health- has been a ripe area of research in the Park et al. found initial evidence that related behaviors and mechanisms recent past. Google Flu Trends (http:// people do post about their depression [2–4]. A common thread in this body www.google.org/flutrends/) provides and even their treatment for depres-

XRDS • WINTER 2014 • VOL.21 • NO.2 23 feature

sion on Twitter [3]. Kotikalapudi et NEW ASSESSMENT TOOLS experimenter and typically comprise al. analyzed patterns of web activity Due to the richness of data generated recollection of (sometimes subjective) of college students that could signal on online social platforms, we are pro- health facts. Social media measure- emotional concerns [8]. In other relat- vided new opportunities for under- ment of behavior captures social activ- ed work, our past research examined standing health behaviors of individu- ity and language expression in a natu- linguistic and emotional correlates als, in a scope and scale not possible ralistic setting. Such activity is real for the postnatal course of new moth- before. For instance, surveys and wear- time, and happens in the course of a ers, and thereafter built a model to able sensing tools that are often used person’s day-to-day life. Hence it is less predict extreme behavioral changes extensively for personal health moni- vulnerable to memory bias or experi- in new mothers [9]. Our findings in- toring can indicate one’s geographic menter demand effects, and can help dicated behavioral concerns such as location, as well as other physiological track health concerns at a fine-grained post-partum depression may be re- and affective responses [10, 11]. How- temporal scale. flected in mothers’ social media use: ever, they generally cannot capture the including lowered positive affect and context and content of these reactions. NOVEL PLATFORMS OF raised negativity, and use of greater Conversely, SNSs provide a rich PSYCHOSOCIAL SUPPORT first-person pronouns indicatingecosystem where the context and Self-disclosure is an important thera- higher self-attentional focus. In fact, content of one’s affective, behavioral, peutic ingredient [12], and is linked to the behavioral changes of mothers and cognitive reactions, as well as improved physical and psychological could be predicted by leveraging their social interactions, can be observed well-being. In fact, self-disclosure has activity from simply the prenatal pe- over extended periods of time [9]. The received a great deal of attention in riod. characteristics of such context and counseling research because of its hy- A common thread in this body of content can be learned for thousands pothesized benefits for the client dur- work is that these methods, which in- or even millions of people. These so- ing the course of therapy, such as an intelligently leverage data from millions cial factors are known to be key in crease in positive affect and a decrease of social networking site (SNS) posts the detection and assessment of a in distressing symptoms [13]. Jourard and users, have demonstrated effi- number of health conditions and out- reported the process of self-disclosure ciency in performance and accuracy comes, and can be made to work in was a basic element in the attainment when applied to a number of health do- a complementary fashion alongside of improved mental health [12]. Ellis mains. Together, they point to the po- traditional approaches. reported discourse on emotionally lad- tential of social media as a signal to le- Moreover, in typical behavioral sur- en traumatic experiences can be a safe

verage in the study of health concerns. veys, responses are prompted by the way of confronting mental illness [14]. Image by Everything Possible / Shutterstock.com

24 XRDS • WINTER 2014 • VOL.21 • NO.2 On similar lines, seminal work by Pen- REACHING LARGE, seeking advice, connecting with ex- nebaker et al. found participants as- DIVERSE POPULATIONS perts and individuals with similar signed to a trauma-writing condition Social media and SNSs are being in- experiences, sharing questions and (where they wrote about a traumatic creasingly adopted across different concerns around treatment options, and upsetting experience) showed im- walks of life. According to a Pew In- and understanding professional di- mune system benefits [15] (see also ternet study [1], currently 73 percent agnoses. Online health content can Ramirez-Esparza et al. [16]). Disclo- of online users use at least one online enhance coping and self-efficacy [21], sure in this form has also been asso- social platform, with Facebook being affect health-related decisions and ciated with reduced visits to medical the most popular. Beyond the everyday the behavior of users and their friends centers and psychological benefits in use of sharing details about the mun- and family, enable better manage- the form of improved affective states dane goings on of life, 59 percent of ment of chronic health conditions, [17, 18]. Rodriguez similarly found re- U.S. adults have used online resources and fuel discussions with healthcare vealing personal secrets to an accept- to obtain health information in the providers [22]. ing confidant could reduce the feeling past year. In the context of health Hence there is an opportunity for of alienation and, as a consequence, and well-being, social media use can healthcare and clinical profession- can also lead to health benefits [19]. serve a range of purposes, including als to leverage such online presence Social media platforms are known to reach a variety of audiences and to allow increased self-disclosure [20], provide health help and resources for allowing individuals to discuss sensi- timely intervention and management tive or otherwise considered stigmatic of a condition. For instance, women health topics with communities they Proactively using tend to use Pinterest and Tumblr more identify with. Because users can be social media to [1], while use of Instagram and Twit- essentially anonymous or pseudony- ter is notably high among African- mous on social media, and therefore increase public American and Hispanic audiences. are not bothered by self-presentation awareness of and Potentially, such affordances can be or concerns related to tracking their nurtured to target public-health mes- history on the site, these services can education on health saging for these demographic groups. facilitate fruitful connections among issues is a logical peers with similar stigmatic experi- NEW OPPORTUNITIES FOR PUBLIC ences and provide an open and honest modern public HEALTH SERVICES platform of discourse. health approach. SNSs provide a unique avenue for

XRDS • WINTER 2014 • VOL.21 • NO.2 25 feature

public health organizations to reach out, monitor, and support large and Due to the richness diverse populations. Clinicians and of data generated caregivers can tune in to social media conversations in real time, listen and on online social collect feedback, identify information platforms, we AACCMM Confeerreennccee gaps, and quell misconceptions about health needs of individuals. Addition- are provided new PPrrooceediinnggss ally, due to the multi-way, interactive opportunities for functionality that is inherent to these NNooww AAvailaabbllee v viiaa platforms, social media can allow understanding PPrriinntt--oon-Deemmaanndd!! these agencies to increase direct en- health behaviors gagement to maintain and increase trust and credibility about the variety of individuals. Did you know that you can of health information that surfaces on now order many popular these platforms. There is also potential for public ACM conference proceedings health organizations to be able to en- via print-on-demand? gage with opinion leaders and influ- to communication between people, encers on health-related topics in so- the emergence of pro-self-harm social cial media, and their conversations. media sites and content may present Institutions, libraries and Influencers can be both organizations a new risk to vulnerable people who individuals can choose and individuals and exhibit the char- might otherwise not have been ex- acteristics of credibility, persistence posed to these hazards. from more than 100 titles in convincing others, and ability to on a continually updated drive conversations so that others take PRIVACY AND ETHICS OF UTILIZING list through Amazon, Barnes notice of the topic or idea and show SOCIAL MEDIA DATA support. Potentially, by engaging with The ability to illustrate and model in- & Noble, Baker & Taylor, such influencers, public health of- dividual behavior using social media Ingram and NACSCORP: ficials can discuss ways to promote data shows promise in the design and CHI, KDD, Multimedia, messaging on shared communication deployment of next-generation well- goals to increase the reach of public ness-facilitating technologies. Privacy- SIGIR, SIGCOMM, SIGCSE, health communications. preserving software applications and SIGMOD/PODS, services can serve as early warning and many more. RISK TO VULNERABLE POPULATIONS systems, providing personalized alerts Despite these positive benefits in and information to individuals. health assessment, social media may Beyond monitoring behavioral For available titles and pose a hazard to vulnerable popula- trends in real time, social media-based tions through the formation and in- measures—such as degrees of activity, ordering info, visit: fluence of “extreme communities” emotional expression, etc.—can serve librarians.acm.org/pod —groups on SNSs that promote and as a personal diary-type narrative provide support for beliefs, attitudes, resource logging “behavioral finger- and behaviors considered typically prints” over extended periods of time. harmful or unacceptable by the social The application might even assign mainstream. Examples include pro- health-risk scores to individuals based anorexic behavior, pro-suicide tenden- on predictions made about forthcom- cies, deliberate amputation, or other ing changes manifested in their behav- forms of self-harm. Essentially, for ior. In operation, if the inferred likeli- these populations, social media may hoods of forthcoming changes surpass provide them with an environment a threshold, the individual could be and avenue to seek and provide sup- warned or engaged, and information port and acceptance that is difficult might be provided about professional to obtain through offline means. Al- assistance and/or the value of social though these online groups may pro- support from friends and family. In vide the benefit of support, they may short, we hope analytic approaches present a risk to the public by encour- based on social media data can play a aging vulnerable individuals to hurt role in helping individuals find timely themselves. Additionally, because the and appropriate support from health web eliminates geographic barriers care professionals and others.

26 XRDS • WINTER 2014 • VOL.21 • NO.2 Concerns regarding individual pri- rious repercussions (e.g., higher insur- [6] Culotta, A. Towards detecting influenza epidemics by analyzing Twitter messages. In Proceedings vacy, including certain ethical consid- ance rates, denial of employment, etc.). of the First Workshop on Social Media Analytics erations, may arise with this form of Developing interfaces that remind us- (Washington DC). ACM, New York, 2010, 115–122. analyses of social media, as they ulti- ers of these risks (e.g., triggering an [7] Moreno, M., Jelenchick, L., Egan, K., Cox, E., et al. Feeling bad on Facebook: Depression disclosures mately leverage information that may “are you sure?” dialogue when detect- by college students on a social networking site. be considered sensitive given their fo- ing sensitive information being en- Depression and Anxiety 28, 6 (2011), 447–455. cus on behavior and health. I envision tered into a post) is an important area [8] Kotikalapudi, R., Chellappan, S., Montgomery, F., Wun- sch, D., and Lutzen, K. 2012. Associating depressive the systems described to be designed of future exploration by the social com- symptoms in college students with internet usage as privacy-preserving applications that puting research community. using real Internet data. IEEE Technology and Society Magazine 31, 4 (2012), 73–80. are deployed by and for individuals, [9] De Choudhury, M., Counts, S., and Horvitz, E. thereby honoring the sensitive aspect A NEW AND EVOLVING FIELD Predicting postpartum changes in emotion and behavior via social media. In Proceedings of the of revealing different types of health- To summarize, it is important to SIGCHI Conference on Human Factors in Computing related information to them. bring the potential of social media to Systems, (Paris). ACM, New York, 2013, 3267-3276. Closely intertwined with this pri- the fore, so as to leverage the benefits [10] Ertin, E., Stohs, N., Kumar, S., Raij, A., al’Absi, M., and Shah, S. AutoSense: Unobtrusively wearable vacy issue is the challenge of inter- of this new data source in order to en- sensor suite for inferring the onset, causality, and ventions. Can we design effective in- hance the quality of life for people. consequences of stress in the field. In Proceedings of the Ninth ACM Conference on Embedded terventions for people whom we have Further, this will stimulate discus- Networked Sensor Systems (Seattle). ACM, New York, inferred to be vulnerable to a certain sion and awareness of the potential 2011, 274–287. illness in a way that is private, yet still role policies could play in support- [11] Shiffman, S., Stone, A. A., and Hufford, M. R. Ecological momentary assessment. Annual Review raises awareness of this vulnerability ing the identities and practices that of Clinical Psychology 4 (2008), 1–32. to themselves and trusted others (doc- individuals suffering from certain [12] Jourard, S. M. Healthy personality and self- tors, family, friends)? In extreme situ- illnesses develop in the face of social disclosure. Mental Hygiene 43 (1959), 499–507. ations, when an individual’s inferred disadvantage. [13] Vogel, D. L., and Wester, S. R. To seek help or not to seek help: The risks of self-disclosure. Journal of vulnerability to an illness with risk- The role of social media and its po- Counseling Psychology 50, 3 (2003), 351. taking attitudes is alarmingly high tential in understanding health be- [14] Ellis, D., and Cromby, J. Emotional inhibition: A discourse analysis of disclosure. Psychology & (e.g., self-harm-prone individuals), haviors is a relatively new and evolving Health 27, 5 (2012), 515–532. what should be our responsibility as phenomenon, one which society is only [15] Pennebaker, J. W., and Chung, C. K. Expressive a research community? For instance, beginning to assess and understand. writing, emotional upheavals, and health. should there be other kinds of spe- Because social media are mostly cre- Foundations of Health Psychology, 2007, 263–284. [16] Ramirez-Esparza, N., Chung, C. K., Kacewicz, E., and cial interventions where appropriate ated and controlled by end users, the Pennebaker, J. W. The psychology of word use in counseling communities or organiza- opportunity for surveillance and pre- depression forums in English and in Spanish: Testing two text analytic approaches. In Proceedings of the tions are engaged? In short, finding vention can be extended to all users. Second International AAAI Conference on Weblogs the right types of interventions that One way to do this could be the public and Social Media (Seattle). AAAI, 2008, 102–108. can actually make a positive impact promotion of direct and easy avenues [17] Joinson, A. N. Self-disclosure in computer-mediated communication: The role of self-awareness and on people’s behavioral state while for people to access help through social visual anonymity. European Journal of Social abiding by adequate privacy and ethi- media sites. Lastly, proactively using Psychology 31, 2 (2001), 177–192. cal norms is a research question on social media to increase public aware- [18] Smyth, J. M. Written emotional expression: effect sizes, outcome types, and moderating variables. its own. We hope this article triggers ness of and education on health issues Journal of Consulting and Clinical Psychology 66, 1 conversations and involvement with is a logical modern public health ap- (1998), 174. the ethics and clinician community to proach that has the potential to save [19] Rodriguez, R. R., and Kelly, A. E. Health effects of disclosing secrets to imagined accepting versus non investigate opportunities and caution many lives. accepting confidants. Journal of Social and Clinical in this regard. Psychology 25, 9 (2006), 1023–1047. [20] Johnson, G.J. and Ambrose, P.J. Neo-tribes: The References Beyond interventions, there is need power and potential of online communities in for work on educating users about the [1] Social Media Update, 2013. Retrieved from http:// healthcare. Communications of the ACM 49, 1 www.pewinternet.org/2013/12/30/social-media- (2006), 107–113. privacy risks of sharing sensitive infor- update-2013/. [21] Eysenbach, G., Powell, J., Englesakis, M., et al. Health mation online that can potentially be [2] De Choudhury, M., Gamon, M., Counts, S., and related virtual communities and electronic support linked to their health. Participants’ so- Horvitz, E. Predicting depression via social media. groups: Systematic review of the effects of online peer In Proceedings of the Seventh International AAAI to peer interactions. BMJ 328, 7449 (2004), 1166. cial media use suggests they might not Conference on Weblogs and Social Media (Boston). [22] Sillence, E., Briggs, P., Harris, P., et al. How do AAAI, 2013, 128–137. be aware of the implications of some of patients evaluate and make use of online health these sharing practices, indicating they [3] Park, M., Cha, C., and Cha, M. Depressive moods information? Social Science and Medicine 64, 9 of users captured in Twitter. In Proceedings of the (2007), 1853–1862. may be unaware of how some advertis- SIGKDD Workshop on Healthcare Informatics (HI- ing companies may be collecting and KDD) (Philadelphia). ACM, New York, 2012, 1–8. distributing their information. Even [4] Sadilek, A., Kautz, H., and Silenzio, V. Modeling Biography spread of disease from social interactions. Munmun De Choudhury is currently an assistant professor though a lot of the health inferences In Proceedings of the Sixth International AAAI at the School of Interactive Computing at Georgia Tech found in prior research are derived Conference on Weblogs and Social Media (Dublin, and a faculty associate with the Berkman Center for Ireland). AAAI, 2012, 322–329. Internet and Society at Harvard. His research interests are from implicit patterns in activity and [5] Paul, M., J. and Dredze, M. You are what you tweet: in computational social science, with a specific focus on content, the ability to derive any infor- Analyzing twitter for public health. In Proceedings of reasoning about our health behaviors and well-being from the Fifth International AAAI Conference on Weblogs social digital footprints. mation about a person’s health from a and Social Media (Barcelona, Spain). AAAI, 2011, public venue like Twitter may have se- 265–272. © 2014 ACM 1528-4972/14/12 $15.00

XRDS • WINTER 2014 • VOL.21 • NO.2 27 feature Here Comes the #Engagement: A serious health initiative made trendy Creating a user experience to communicate the seriousness of HIV prevention and awareness can be both educational while entertaining. This combination along with a sense of cultural influence helps to both attract and engage millennials.

By Fay Cobb Payton and KaMar Galloway DOI: 10.1145/2691362

ay 2006. Anticipating the usual visit from students seeking advice about final projects, discussing some stubborn bug in their SQL code, or poring over the cryptic results of data analytics software, a professor sat working in her office The day took a different turn, however, when a female student entered with a worried Mlook on her face. “Dr. Payton, my boyfriend tested positive for HIV. I do not want to become a statistic.” In 2011, out of the 49,272 total cases of HIV diagnosed in the United States, an estimated 47 percent were among African Americans. Of the 2,294 cases diagnosed among teenagers, 67 percent were African-American teens. More recent data shows a growing trend of HIV

infections among persons between 13 ning of a journey on the road to My- ently. The Internet is expected to play and 24 years old [1]. These statistics in- HealthImpactNetwork.org. a vital role in reducing long-standing dicate HIV remains a potent threat to But what does HIV have to do with health inequalities. Its ability to reach those who are young and college aged, creating an online experience that is as those suffering from stigmatized as well as the Black population. This engaging as it is informative? The an- medical conditions and those inter- warrants increased attention from the swer lies in understanding that differ- ested in learning more about such general public and policy makers. That ent socio-ethnic groups perceive and illnesses is improving every day [2]. fateful spring day marked the begin- react to online information differ- However research scholars caution

28 XRDS • WINTER 2014 • VOL.21 • NO.2 that in normalizing discourses such ties. This often represents how con- SOCIAL EDHEALTH-TAINMENT as the “digital divide,” ethnic minori- tent providers reproduce inequality We decided no design or content ties, especially African Americans, are and fail to accommodate the cultural should be devoid of fun. Rather, it presented as being deficient in com- perspectives of ethnic minorities [7]. should be the very vehicle that drives puter skills and their ability to utilize With these considerations in mind, participation. In our case, fun im- online resources [3–6]. Such deficit our team of research scholars and col- plied a culture of socio-technical “ed- philosophy models rely on the faulty lege students pondered over how to health-tainment” [9]. That is, a user assumption that information avail- better tune the content of health in- experience (UX) that simultaneously able online is tailored to appeal to all formation to make it resonate with educates and entertains. Further- psychological and cultural tastes. For college-aged millennials, specifically more, we sought to account for the instance, a study of African-American Black women. HIV is a serious illness user community’s social identity and women’s attitude toward the Internet and carries a strong social stigma. cultural nuances. Our goal was to cre- as a resource of health information Quite understandably, the subject is ate a user experience, not just an IT ar- revealed participants differentiated hardly ever associated with the con- tifact. In doing so, we hoped to create between Internet access and content cept of entertainment. Nevertheless, a fun working environment. We were creation, with the latter largely per- fun experiences are more attractive, dealing with the serious topic of HIV ceived as being dominated by white offer intrinsic rewards, impact en- awareness and prevention informa- culture. Menus, wording, and even gagement, and have a powerful influ- tion. But by rethinking creative ways navigation impacts how users view, ence on how people persist in using a to disseminate information, not only form, and experience online identi- system [8]. were we generating engaging experi-

XRDS • WINTER 2014 • VOL.21 • NO.2 29 feature

ences, but, more importantly, deliver- Figure 1. Twitter use by race and age. ing relevant matters. Young African Americans have high levels of Twitter use For millennials, social means fun; % of internet users in each age group who use Twitter engaging with friends and family over social networks has become an White Black essential part of the daily recreation All of the younger generation. The wide- 16 Internet spread use of social media has made Users 22 it convenient to understand the living, 28 working, and playing habits of this 18-29 40 demographic. College-aged African- American millennials have a strong 21 30-49 digital presence and are particularly 21 heavy users of Twitter. Forty percent of 10 African-American Internet users aged 50-64 9 18–29 say they use Twitter; a figure that is 12 percentage points higher than the Pew Research Center’s Internet Project July 18–September 30, 2013 tracking survey. comparable figure for white people, N=6010 adults ages 18+. For results based on internet users, n=3,617 for whites and n=532 for African Americans. only 28 percent of whom are Twitter users [10, 11]. Therefore, given the project’s focus Figure 2. The MyHealthImpactNetwork.org landing page using two screen on young African Americans, Twit- captures. ter was the best tool for disseminating health messages. This notion was confirmed by the qualitative data collected [12]. The study’s findings from 40 Black women indicated lack of trust, stigma ascribed to HIV, and misconceptions create communication barriers. There- fore the empowering ethos of “nothing about us, without us,” engendered by the culturally aware design of the UX, resonated particularly well with Black women [12]. The team ultimately adopted “for students, by students” as the slo- gan for our project to reflect this spirit of empowerment. Several additional findings emerged from the qualitative data. Even among stigmatized health conditions, web-based user experiences can interject fun while communicating serious messages to grab the target audience’s attention. We were able to establish reciprocal fun for the user by creating cultural relevant messages via social media [13]. For our designers, fun manifested itself in team interactions and meeting our intended audience in the social and physical spaces where they reside. As researchers it was important that we understood the socio-technical impacts of technology in all populations, and in particular, those under-represented and underserved.

BUILDING A COMMUNITY There was eager excitement as our

30 XRDS • WINTER 2014 • VOL.21 • NO.2 team worked on the final stages of the we brainstormed, the male attendee inclusive of ethnic and gender voices, project. Deciding the color theme for stated: “I enjoyed myself. Black wom- too often void or minimal in the tech- the landing page of MyHealthImpact- en were everywhere. It was good to nology space, to shape to both create, Network.org was a particularly hot see the positivity, but I admit I was a consumer and disseminate the health topic of debate. We wanted to catch little intimidated. There were not a lot messages. This is the essence of social the user’s attention at first glance and of brothers around, but the few I saw, edhealth-tainment. increase the comprehensibility of our we all seemed to feel the same way. We The National Science Foundation content. We also wanted to take advan- gave each other the head nod.” The fe- grant IIS-1144327 supports this research. tage of the fact that people have com- male attendee replied: “I never would mon associations with colors, positive have thought Ken [name changed] References and negative. would be intimidated by women. Was [1] Center for Disease Control and Prevention (2013). Statistics on HIV Surveillance. http://www.cdc.gov/ The initial design envisioned a it too many in one place for you?” He hiv/pdf/statistics_surveillance_adolescents.pdf neutral color theme that would ap- replied, “I got comfortable after I un- [2] Berger, M., Wagner, T.H. and Baker, L.C. Internet use peal to both men and women, that derstood the situation.” This was an and stigmatized illness. Social Science & Medicine latter group being our primary fo- important exchange in that it demon- 61, 8 (2005), 1821–1827. [3] Brock, A. A belief in humanity is a belief in colored cus. The first set of landing page was strated how empowerment is central men: Using culture to span the digital divide. predominantly red. But the feedback to design and how offline interactions Journal of Computer Mediated Communication 11, 1 (2007), 357–374. from the design team, which was can inform the online health messag- [4] Kvasny, L. The role of the habitus in shaping equally divided by gender, was mixed ing. discourses about the digital divide. Journal of at best: “Do we really want red which Computer Mediated Communication 10, 2 (2005). is the HIV/AIDS ribbon color?” Anoth- CONCLUSION [5] Kvasny, L. and Payton, F. C. African Americans and the igital divide. In Encyclopedia of Information er team member added, “Some say it In addition to Twitter, a blog and a Science and Technology, 2nd Edition. M. Khosrow- means that you have tested positive; YouTube channel were created when Pour (Ed.). Idea Group Publishing, Hershey, PA, 2008, 78–82. that is not the message we want to the site was launched. The blog gives [6] Selwyn, N. Reconsidering political and popular send or brand we are taking on? We individual team members and guest understandings of the digital divide. are focused on prevention and aware- writers an opportunity to voice their [7] Nakumara, L. Cybertypes: Race, Ethnicity, and ness.” The next choice was blue, but opinions on critical topics such as the Identity on the Internet. Routledge, New York, 2002. [8] Carroll, J.M. and Thomas, J.C. Fun. SIGCHI Bulletin 1, the debate continued. Although blue stigma associated with HIV; provide 3 (1988), 21–24. New Media and Society 6, 3 (2004), is a color that appeals to men and a male perspective on health issues; 341–362. women, we couldn’t agree: “We don’t discuss health-related messages in [9] Siek, K.A. What are our responsibilities when designing sociotechnical health interventions? want to lock females out because the hip-hop music, popular culture, and Interactions 18, 5 (2011), 20–23. world says blue is for guys.” The team the news; and prompt readers to take [10] Nielsen. African-American consumers: Still vital, eventually compromised by choos- action for social justice. In addition, still growing 2012 Report. The Nielsen Company. 2012. ing purple, a color halfway between users are able to interact with the [11] Pew Internet Research (2014), African American red and blue. “It works for females. research team, giving myHealthIm- and Technology Use. I can see this,” said one male team pactNetwork.org a personal feel. The [12] Payton, F.C., Kvasny, L. and Kiwanuka-Tondo, J. member. But would it work for men? YouTube channel enables viewers to Seeking and perceiving online HIV prevention information: Black female college students’ To which another male team member see the team in action. Music and art, perspectives. Internet Research 24, 4 (2014), explained: “We can get with that [the which are “hidden” talents for some 520-543. purple]. It gets our attention, and we team members, have been incorpo- [13] Payton, F.C. and Kiwanuka-Tondo, J. Contemplating public policy in AIDS/HIV online content, then remember that the initial project fo- rated in our social media channels where is the technology spirit? European Journal cus is NOT about us [men]. It is not and videos, helping further our social of Information Systems 18, (2009), 192-204. about us, but … the information will edhealth-tainment approach in circu- help us as well.” lating health messages. Biographies Fay Cobb Payton directs MyHealthImpactNetwork.org, a Once the design was finalized, our HIV is a hard topic to discuss. My- social network experience that focuses on health disparities next concern was content. A couple HealthImpactNetwork.org has shown and social media technology interventions. She is the author of Leveraging Intersectionality: Seeing and Not of young team members, male and that information about serious health Seeing, an anthology of her research on STEM education and female, were especially excited about conditions can be made more interest- experiences in both academe and corporate environments. Dr. Payton is an editor for Health Systems, an OR Society the opportunity to see and, in some ing to a young audience by interspers- journal, and is an associate professor of information cases meet, celebrity speakers, musi- ing it with the right degree of levity, systems at North Carolina State University. She received the 2013 National Coalition of Women in Information cians, and entrepreneurs at a local communicating via social media, and Technology (NCWIT) Undergraduate Mentoring Award. She is health event, which they covered for using a carefully designed user ex- the corresponding author and principal investigator for this article. Dr. Payton is an active Twitter user; @drfayonline. MyHealthImpactNetwork.org. The perience. The “coolness” of the con- KaMar Galloway is a graduate of the Computer Science students captured digital content tent results from the minimal use of Department at North Carolina State University and an active and gathered HIV prevention and medical jargon, plenty of hooks to pop sneaker collector. Combining his love for technology and footwear, he is interested in the birth of wearable computing health information, which they later culture and news events, quick and and the impact it will have on making healthy decisions. He is interested in online platforms that are informative, blogged about. During the post-event effective communication, and a user welcoming and can assist the African-American community. review, we discussed how to best use experience created “for students, by the content and event experience. As students.” This user experience is also © 2014 ACM 1528-4972/14/12 $15.00

XRDS • WINTER 2014 • VOL.21 • NO.2 31 feature Challenges in Personal Health Tracking: The data isn’t enough Increasingly, personal health data can be tracked and integrated from numerous streams quickly and easily, but our feedback lingers in the land of “show the user a graph and hope.” How can we help people make sense of personal health data?

By Matthew Kay DOI: 10.1145/2678024

ersonal informatics—and particularly personal health tracking—is booming. Apple and Samsung have both announced apps for aggregating personal health data, such as step counts or sleep quality, into one unified interface. This is the first step in achieving a long-standing vision in personal informatics. Giving people Pone place to see visualizations of all of their health data. A lot of the low-level technical challenges for this vision are being solved: We can sense myriad aspects of health—from step count, to sleep quality, to stress, to respiratory function, to heart rate variability. As users and developers, we can also access this data more easily than before. While it is sometimes cumbersome to get all of that data out of each provider’s silo—your Wi-Fi

bathroom scale may not have the same ’em a graph and hope” approach: Sure- to bewilderment, or worse, the draw- API or data format as your pedometer— ly people will notice the correlations ing of spurious conclusions. People— projects like S Health (http://content. between their different data streams even trained statisticians—are partic- samsung.com/us/contents/aboutn/ and draw appropriate conclusions if ularly good at seeing patterns in data sHealthIntro.do) and Apple HealthKit we just show them the data! Never- where there are none. (https://developer.apple.com/healthkit/) mind that one third of Americans have Some of the core challenges for are breaking down these barriers by low graphical literacy [1] and would be personal health informatics, then, lie encouraging developers to support unlikely to make valid inferences from in finding and presenting health data these APIs or be left behind. supposedly straightforward graphical to people in useful and meaningful What we are left with is the ques- data. Asking people to make statistical ways—it isn’t enough to bring the data tion of what to do with all that data. judgments from graphical data—such into one place. This leads us to ask It remains the case that most health as identifying correlations—without questions like: What data do we actu-

apps take some variant of the “show providing scaffolding too often leads ally need? How can we summarize it Image by Syda Productions

32 XRDS • WINTER 2014 • VOL.21 • NO.2 XRDS • WINTER 2014 • VOL.21 • NO.2 33 feature

in a way that aids understanding? I Figure 1. Lullaby deployed on a bedside table. Visible here are a touchscreen tablet explored some of these questions in mounted in a stand for easy access from the bed and the sensor box with pivoting the context of sleep in a project called sensor enclosures. The sensor suite itself is about the size of a bedside lamp. Lullaby, a research prototype that includes a suite of environmental sensors—sound, light, temperature, and motion—to help people assess the quality of their sleep environments (see Figure 1) [2]. In that project, we built a system to track how people’s bedroom conditions (such as light levels or air quality) impact their sleep quality, and explored different methods of presenting that data in meaningful ways to people.

IS YOUR BEDROOM MAKING YOUR SLEEP WORSE? Research has shown environmental factors can be a major cause of poor sleep quality and interrupted sleep [3]; for example, a room that is too warm, has improper lighting, is noisy, or has poor air quality can disrupt sleep. While some of these environmental factors are observable, others may be subtle or difficult to recognize. Thus, individuals who have poor sleep quality often have trouble evaluating the cause or severity of their sleep difficulties.

Figure 2. A screenshot from the Lullaby feedback app showing a single night of sleep data.

34 XRDS • WINTER 2014 • VOL.21 • NO.2 While clinical sleep centers can to helping people draw meaningful evaluate an individual’s sleep quality Designers should conclusions here. One of the goals of effectively, these evaluations do not think about how Lullaby was to help people identify occur in individuals’ actual homes. the things that are causing disruption Thus, a sleep lab cannot directly to design better to their sleep. Some disruptors might identify environmental factors in feedback devices be easy to see just by looking at the the home that contribute to reduced graphs in tandem with the audio-visu- sleep quality. At the same time, com- given their inherent al stream, such as a co-occurrence be- mercial personal informatics devices uncertainty and the tween awakenings and motion caused like the Fitbit can help identify when by cats entering or exiting the bed. a person has had poor sleep or when properties of the However, more subtle causes may be they awaken at home, but generally data being collected. difficult to determine just by reviewing don’t measure environmental condi- the data manually. For example, users tions—they report measures like the may want to know if they awaken more proportion of time in bed actually frequently while the temperature is spent asleep (sleep efficiency), and warmer over a period of several nights. when during the night sleep has been With enough captured data, Lullaby disturbed. These measures provide might use Lullaby and how they inter- could help identify such relationships some indication of sleep quality, but pret the data they see on the device (see by running statistical analyses on the cannot give concrete guidance for Figure 2). There are particular chal- data in order to produce higher-level sleep environment improvement. lenges associated with building appli- summaries, but expecting people Ubiquitous computing technology cations like this when the data is re- to find such correlations by looking that helps people track both their sleep corded while people are unconscious. through time-series graphs them- habits and environmental factors that The domain of sleep is one where selves is probably unrealistic. We need affect sleep quality could help people events of interest are not known by us- to consider another approach. identify why their sleep was interrupt- ers until well after their occurrence— ed, not just when. That was the goal of until the time at which the user goes THE POWER OF NATURAL Lullaby. Using a tablet device kept by looking for such events. As a result, LANGUAGE SUMMARIES the user’s bed, Lullaby displays this users must sift through data with little To investigate the potential for higher- environmental data together with data or no knowledge of what they seek or level summaries and inferences, we from an off-the-shelf sleep-tracking when it occurred, so helping them dis- followed up with some participants device, like a Fitbit. It aims to help cover salient data is very important. after the study ended to present mock- people better understand their sleep, To aid this discovery, we gave users ups of possible future feedback inter- to understand what goes on in their a wider context in which to view their faces. These mock-ups ranged from sleep environment while they are un- data by highlighting data that is out of scatterplots of potentially related fac- conscious, and to help them make im- recommended ranges from the sleep tors, to aggregate statistics of various provements in their sleep habits. literature (e.g., too hot or cold, too measures (for example, the amount Lullaby was motivated by discus- noisy, etc.). We also showed all collect- of time a given sensor spent out of its sions with sleep doctors: Some pa- ed data together, chunked by sleep pe- recommended range from the sleep tients report sleeping poorly at home, riod, and allowed people to play back literature), to one-sentence summaries but sleep just fine when they come in sound and infrared images of their of factors influencing sleep (for exam- to the sleep lab for an assessment. Of sleep alongside the data to provide a ple, “Over the past two weeks, higher course, if there are any problems being more concrete frame of reference. temperature has been associated with caused by factors in their bedroom at People found this unconscious worse sleep.”). home, these issues will not show up data compelling: Imagine watching Participants responded particularly in a sleep lab. This is the perfect op- yourself sleep—or for one participant, well to the single-sentence summa- portunity for personal health track- sleepwalk. Another person found she ries. In one participant’s case, Lullaby ing to help answer questions that were coughed regularly in her sleep by ob- had found a possible correlation be- difficult or impossible to answer in serving consistent spikes in audio tween higher temperature in the bed- the past. To help people find answers data. This illustrates the potential of a room and sleep disturbance (when the to these questions—“Is my bedroom system like Lullaby. Chronic coughing temperature went up, his sleep qual- temperature affecting my sleep?”—we is a symptom of sleep apnea, a condi- ity tended to do down). Presented as a need to understand how to put togeth- tion that is widely undiagnosed largely scatterplot with a trend line, he found er feedback from a device like Lullaby. because its immediate symptoms are this to be an interesting finding, but difficult for the sufferer to observe (one given a single-sentence natural lan- CAN PEOPLE MAKE SENSE of the very reasons that motivated us to guage summary, he became excited: OF SLEEP DATA? develop Lullaby in the first place). “That would be really cool.” We conducted a feasibility study of Lul- Compelling aspects of the data More generally, natural language laby to better understand how people aside, there are significant challenges has many potential advantages in

XRDS • WINTER 2014 • VOL.21 • NO.2 35 feature CACM_TACCESS_one-third_page_vertical:Layout 1 6/9/09 1:04 PM Page 1

sense-making. It is intuitive and easily tive understanding of weight data. For understood, and more explicitly iden- example, digital scale readouts con- tifies connections in data for people. vey an unrealistic level of precision, While a statistician might find a scat- negatively affecting user perception. terplot easy to interpret, most people Scales often present data at an appar- need more scaffolding than this. Spot- ent resolution of a tenth of a pound or ting correlations in time-series data less, when fluctuations of a few pounds ACM is similarly difficult, and the almost within a single day are typical of nor- purely graphical focus of many exist- mal individuals. This level of apparent ing personal informatics applications precision can cause people to focus on Transactions on is insufficient in the face of the reali- small, insignificant changes in their ties of statistical and graphical literacy weight. Experts in weight change (such Accessible levels in the general population. as dieticians or personal trainers) will The results from the Lullaby project try to mitigate the effect of these fluctu- Computing made me realize the need for a better ations, both by telling people to always understanding of personal health data weigh under similar conditions and literacy. I decided to simplify the prob- by educating people about why weight lem. Lullaby has many interacting data can fluctuate over the short term. In streams, making it difficult to tease interviews with such experts, one de- apart the many facets of health data scribed clients as getting “the horrors” literacy. I turned to another domain, when they feel like their weight moves perhaps the simplest, most ubiquitous in an undesirable direction even a example of a personal health sensor: small amount. People will fixate on the bathroom weight scale. and form an identity around a particular weight—a particular number—they A TEST CASE FOR DATA LITERACY want to be. This is only reinforced by The bathroom scale is arguably the an interface that communicates little most ubiquitous health sensor of all, else but a number. and it is effective, too: Studies have We conducted an online survey of shown that frequent weigh-ins help more than 800 scale users and quizzed maintain weight loss [4]. At the same them on their knowledge of weight time, many people have a love/hate re- fluctuation. We found respondents lationship with their scale, and dread with a less sophisticated understand- standing on it. Despite its central- ing of how weight fluctuates during the ◆ ◆ ◆ ◆ ◆ ity to global health and wellness, the day were less likely to trust their scales. bathroom scale interface has barely This is exacerbated by the fact that the This quarterly publication is a changed since it was first introduced scale interface makes no attempt to quarterly journal that publishes about 100 years ago: It still produces a inform users about how weight fluc- refereed articles addressing issues single value representing one’s weight tuates: The single data point reflects of computing as it impacts the at the moment of measurement. Digi- nothing about weight change over time tal displays have replaced the analog or the accuracy and precision inherent lives of people with disabilities. needle, coarse measurements of body The journal will be of particular fat have been added, and some scales interest to SIGACCESS members log data for offline review. However, and delegates to its aﬃliated the singular data point is still the main display and is often the only informa- conference (i.e., ASSETS), as well tion presented at the time of weigh-in. We can use more as other international accessibility Most scales answer just one question— conferences. “What do I weigh right now?”—which appropriate ◆ ◆ ◆ ◆ ◆ may not be the best framing for weight sensors, with more www.acm.org/taccess data. However, its ubiquity makes it a good test case for literacy: If we can’t intelligent feedback, www.acm.org/subscribe get the scale right, what hope do we to give better, have for more complex health data streams? We conducted a series of clearer answers to studies [5] to investigate how the scale underlying health interface impacts data understanding. We found several issues with cur- and wellness rent scales that work against an effec- concerns.

36 XRDS • WINTER 2014 • VOL.21 • NO.2 to that measurement. It is no wonder curacy of a device or what constitutes people fixate on single numbers—it is Users must sift typical fluctuation—will manifest in the only thing they are given. through data these domains. Better reflecting the In a study of consumer reviews of underlying model and educating us- scales, we found many people would with little or no ers about uncertainty will likely be rate scales poorly based on some as- knowledge of what crucial here as well. Both graphical sessment of their accuracy, but often and natural language presentations the expectations of accuracy people they seek or when it of data must be sensitive to people’s had were unrealistic. For example, occurred, so helping mental models of the data being com- some people had expectations of in- municated. As we continue to push ter-scale reliability that are not even them discover low-cost novel health sensing into the realistic in hospital settings: One per- salient data is very wild, it is important to consider how son complained their scale was off the accuracy of these systems affects by about a pound from their doctor’s important. users, and how much we can gain scale. However, even scales in hos- from more intelligent feedback. pitals can reasonably be expected to Finally, we have an opportunity differ by as much as 1.5 pounds when to revisit the underlying health ques- measuring the same person [6]. tions users want answered, what data What we are left with is a device that we can use to answer them, and how to gives data to users without setting rea- stead handling the resulting data and help interpret that data—rather than sonable expectations for the accuracy consequent user feedback in a more falling back to the nearest convenient of that data, without providing context considered way. Scales are already measure (e.g., weight, sleep quality, to the data (such as trends over time), ubiquitous, cheap, and fairly accurate; step count) and simply reporting in- and without attempting to provide the greater gains may be had by pushing dividual data points without context scaffolding of education about weight the state of the art in feedback. Design- or interpretation. Instead, we can use fluctuation over time. It is no wonder ers should think about how to design more appropriate sensors—with more that people often react negatively to better feedback devices given their in- intelligent feedback—to give better, the data they see, or have an aversion herent uncertainty and the properties clearer answers to a person’s underly- to stepping on the scale. Further, as of the data being collected. ing health and wellness concerns. scales are part of a larger class of in- There are many things we can do creasingly ubiquitous health feedback to address these issues. For example, References devices, which provide single-point, in- device feedback should avoid false [1] Galesic, M. and Garcia-Retamero, R. Graph literacy: A cross-cultural comparison. Medical Decision Making: stantaneous measurements—such as precision. Instead, give people a sense An International Journal of the Society for Medical body fat estimators, thermometers, pe- of how precise the data is and how Decision Making 31, 3 (2011), 444–57. dometers, and blood pressure cuffs— much fluctuation they can interpret [2] Kay, M., Choe, E.K., Shepherd, J., et al. Lullaby: A capture and access system for understanding the that we have not gotten the scale inter- as meaningful change versus noise. A sleep environment. Ubicomp ’12, (2012). face right bodes poorly for the others. scale that adopts a more sophisticated [3] Kryger, M.H., Roth, T., and Dement, W.C. Principles model of the underlying data could and practice of sleep medicine. W. B. Saunders Co., Philadelphia, 2000. SO WHAT SHOULD THESE also set user expectations more ex- [4] VanWormer, J.J., Linde, J.A., Harnack, L.J., Stovitz, INTERFACES LOOK LIKE? plicitly. For example, a scale might re- S.D., and Jeffery, R.W. Self-weighing frequency is Between Lullaby and the bathroom port estimates using natural language associated with weight gain prevention over two years among working adults. Int J Behav Med 19, 3 scale we see similar issues encoun- explanations: “We estimate your daily (2012), 351–358. tered across the space of personal weight to within 3 pounds since your [5] Kay, M., Morris, D., Schraefel, M., and Kientz, J.A. There’s no such thing as gaining a pound: health informatics: Making sense of weight typically fluctuates about that Reconsidering the bathroom scale user interface. data is hard, and proper support for it much during the day,” or may involve Ubicomp ’13, (2013), 401–410. is not there yet in many existing sys- graphical depictions of weight vari- [6] Goldberg, R. and Hebbard, G. How accurate are hospital scales? The Medical Journal of Australia tems. At the same time, we have made ability. Combining multiple data 194, 12 (2011), 665. good progress on technical issues. streams, health devices could go even [7] Bentley, F., Tollmar, K., Stephenson, P., et al. Health Sensing and data infrastructure have further: “When you sleep poorly your mashups: Presenting statistical patterns between wellbeing data and context in natural language to both improved dramatically, leading weight goes up.” Frank Bentley has promote behavior change. ACM Transactions on us to turn back to questions of what to done excellent work exploring this ap- Computer-Human Interaction 20, 5 (2013), 1–27. sense and what to do with the data now proach [7]. that we can get it. More generally, I hope future de- Biography I believe, for example, a better scale signs of ubiquitous health sensors— Matthew Kay is a Ph.D. student in computer science and engineering at the University of Washington studying can be designed without investing in thermometers, blood pressure cuffs, visualization and feedback in ubiquitous computing more expensive equipment, better cal- blood glucose monitors, etc.—adopt systems. He is particularly interested in improving the design of user interfaces backed by data with inherent ibration, or even clearer instructions a more sophisticated approach to uncertainty, such as those in personal informatics. for obtaining better data (e.g. to always feedback. I expect similar issues— use the scale on a hard surface), by in- such as gaps in knowledge of the ac- © 2014 ACM 1528-4972/14/12 $15.00

XRDS • WINTER 2014 • VOL.21 • NO.2 37 feature Did I Take My Meds Today? People tend to believe they are more aware of their own health behaviors than they really are. In this article, we present technologies that employ ubiquitous home sensing to support awareness of healthy habits.

By Matthew L. Lee DOI: 10.1145/2676574

eople are creatures of habit. In our everyday lives, we naturally fall into routines where we, for better or worse, mindlessly perform regular actions that make up our day. For example, consider your daily drive to work. When you reach your destination, you follow the routine to turn off the car, exit the car, close the door, Pand lock the car doors. However, sometimes the action of locking the door is so automatic that you might not explicitly remember whether you actually locked the doors. You might actually have forgotten to lock the doors, but believe the doors are locked because it is

normally part of your routine. Simi- and the context in which it occurred consumer in the form of commod- larly, actions people take to manage to prevent it from happening again. ity apps (e.g., Lift, Samsung sHealth, their health, such as taking medi- As individuals are encouraged to take FitBit) easily downloaded to’ mobile cations or eating a healthy diet, are the driver’s seat in their own health smart devices. They also come in the often performed as part of a regular journeys, they need tools that not only form of affordable gadgets found at any routine with actions that can easily reinforce regular healthy behaviors, shopping mall, which can be installed be overlooked. but also make them aware of when and in the home simply by plugging them To address a health behavior prob- how their behaviors can be improved. in (e.g., Nest, GlowCap, CubeSensors). lem like missed medication or a series The once fairy-tale vision of ubiq- Monitoring apps and devices demand of unhealthy food choices, people must uitous monitoring of our everyday ac- little of our attention and quietly accu-

first be made aware of the problem tions has now reached the common mulate a log of our actions over time. Arhelger Tobias by Image

38 XRDS • WINTER 2014 • VOL.21 • NO.2 XRDS • WINTER 2014 • VOL.21 • NO.2 39 feature

viduals). For telephone use, a sensor Figure 1. Sensor-augmented pillbox unobtrusively monitored medication taking placed on the telephone line logged over 10 months. the sequence and timing of digits di- closure switches custom PCB wireless radio aled, and software looked for patterns of misdialed calls (interposed digits, missing/added digits, etc.). A coffee- maker was augmented with various contact sensors to detect the sequence of adding water and coffee to the machine, placing the carafe in the machine, and turning it on. Each sensor device wirelessly sent its data in real time to a laptop placed in the home using a Zigbee wireless network. The laptop uploaded the data accelerometer microprocessor to a remote server, where the data was processed and visualized on a tablet- based ambient display [see Figure 2]. They allow us to reflect on our actions dwellSense, which logs people’s medi- Using a user-centered design process, to see if they are consistent with our cation taking using a sensor-enhanced the tablet display showed near real- health goals. These systems capture pillbox [see Figure 1] while also track- time feedback on how individuals car- a good deal of our actions in context, ing a number of other tasks important ried out their everyday tasks based on but user feedback often consists of a for independence, including dialing the information needs of individuals. simple summary of the data shown to the phone and preparing meals/coffee. For medication taking, previous field users when they choose to view it. Both Unlike medication prompting devices work found individuals wanted to use the content and the timing of the feed- such as GlowCaps, dwellSense did their own routines to take their meds back is important for supporting an not prompt or remind the user to take but sometimes did not recall whether awareness of (un)healthy habits and his/her medications at a certain time. they took their pills earlier in the day empowering individuals to improve Rather, it was designed to 1. support the [5]. To provide feedback, while also their habits. user’s autonomy by allowing the user to supporting individual routines, the Medication taking is a health action follow his/her own routine to self-initi- tablet visualization merely reflected almost universally considered impor- ate medication taking and 2. provide whether the individual had not yet tant by both those who prescribe med- feedback soon afterwards so that user taken their meds or had completed ications and those who take medica- can verify the behavior. Timely feed- taking their meds, rather than pro- tions. Taking medications is difficult back is an important part of self-regu- viding proactive reminders or alerts because, unlike other daily habits like lation [4], the process by which people to prompt medication taking. Users bathing or brushing teeth, it is a hab- self-monitor their behaviors, judge the defined the latest time of day by which it learned later in life. Further, poor quality of the behaviors based on their they planned to have taken their meds. medication adherence is not inher- own goals, and make adjustments to The feedback display showed whether ently self-reinforcing because missing improve their behaviors. meds were taken “on time” (complet- medications often does not have im- dwellSense is a suite of sensors em- ed before the user-defined time), “late” mediately perceivable bodily effects to bedded in the home that monitors and (completed after the user-defined the individual. Approximately 20–50 logs “instrumental activities of daily time), or “missing” (not yet complet- percent of patients do not take at least living,” including medication taking, ed after the user-defined time). The 80 percent of their medications [1]. C. telephone use (dialing and misdial- display also showed the time of pill Everett Koop, former Surgeon Gen- ing patterns), and making a pot of cof- taking, and which pillbox doors were eral of the U.S. once said, “Drugs don’t fee (as a proxy for meal preparation). opened, so users could check whether work in people who don’t take them.” These specific tasks were selected be- they had performed their medication Indeed, the primary concern of pri- cause they are performed frequently taking promptly and correctly. The mary care physicians is whether their and require sequenced procedures tablet visualization updated at least patients take their medications, which indicative of an individual’s cognitive once every 30 minutes, providing near can lead to improved health outcomes and functional ability. For medication real-time, glanceable feedback to us- and reduced costs of care [2]. taking, a pillbox was augmented with ers about their behavior for the cur- In our research, we aimed to de- contact sensors to record when each rent day before resetting every night termine whether feedback about how pillbox door was opened and closed. while the individual was sleeping. people take their medications would An accelerometer in the pillbox also The dwellSense system was de- help them take their medications more logged when individuals handled the ployed for 10 months to monitor the consistently and correctly [3]. We de- pillbox and inverted it to pour out their medication taking of 12 older adults signed a home sensing system, called pills (a common action for some indi- who were living alone and taking

40 XRDS • WINTER 2014 • VOL.21 • NO.2 multiple medications for multiple other words, we hypothesized people group increased from 94.5 to 98.4 per- chronic conditions. Each participant need to know when they are making cent, while the control group remained believed taking medications was im- mistakes so they know how to fix the relatively unchanged (from 93.7 to 92.1 portant; they also claimed they were problem. The feedback must also be percent), which was not statistically taking their meds consistently with- given promptly so the reason for the significant (F[1,67]=2.33,p=0.131). out any mistakes. sub-optimal performance can still be Participants in the study actually be- In the first two months of the de- remembered or inferred. The results gan the study with a high baseline ad- ployment, no feedback was given to of the 10-month study corroborate this herence rate averaging greater than the participants about their monitored hypothesis. Participants in the feed- 90 percent, as measured by dwell- behaviors. This baseline period pro- back group significantly improved in Sense. We observed a ceiling effect vided time for participants to grow two ways after the feedback display in the feedback group, increasing accustomed to being monitored and was introduced: 1. promptness (taking close to the 100 percent adherence behave as they normally would. Be- medications before their pre-defined rate. In fact, all six participants in the ginning in the third month, half the late time, going from 75.1 to 90.8 per- feedback group had at least one two- participants (“the feedback group”) cent, F[1,67]=11.40 , p=0.001); and 2. month streak of 100 percent adherence received the real-time feedback tablet correctness (opening the pillbox door with the help of the feedback display, a display, and the other half (“the con- that matches the current day of the trend not observed before the feedback trol group”) received no feedback. The week, going from 94.2 to 99.2 percent, display was introduced. feedback group placed the display in F[1,67]=4.18, p=0.0448), [see Figure 3]. Participants reported the display their living or dining room so that the In the same period, the control group, helped them self-regulate their medi- display was easily viewed throughout not surprisingly, remained unchanged. cation taking by increasing their the day. At the end of the fourth month, Furthermore, the feedback group also awareness, identifying errors, and participants in the control group were significantly decreased in the variance confirming their memory. Participant shown a graph of their medication in the time of day they took their meds P05 said the display increased her taking data from the preceding six to (15.5 to 5.0 hours), after the display was awareness because, “It always tells eight weeks to ensure both groups un- introduced while the control group ac- you what time you get your medication derstood the type of data that was be- tually increased, which likely indicates then in the meanwhile sometimes it ing monitored. After the eighth month, individuals in the control group were tells you when you have missed your the tablet display was removed from less conscientious about taking their medication. That way, it’s good.” P10 the homes of the feedback group. In medications despite being monitored. remarked the display helped her iden- the remaining two months, the dwell- For the simple metric of the adherence tify when she made mistakes: “Before, Sense system continued to monitor rate (how often medications were tak- if I made a mistake and forgot it, I the participants’ medication-taking en at all), participants in the feedback might not notice until the end of the behaviors. The lengthy 10-month study was intentional; it was designed to give Figure 2. A tablet-based display shows feedback about how well individuals carry time for participants to become ac- out their medication taking, phone use, and coffee making. It shows information customed to being monitored by sen- about what time meds were taken; whether it was on time, late, or missing; sors and to integrate the feedback into what pillbox doors were opened; and a glanceable visual rating (right) for the their everyday routines. Shorter-term promptness and correctness of medication taking. studies of behavior change technology often show results that point to an immediate novelty effect or performance effect, leaving open the question of the longer-term results of the intervention. By the end of the study, more than 5,785 episodes of medication taking were observed across all participants. Moreover, the length of the study also allowed for the well-regarded A-B-A experimental paradigm for the feedback group, in which the effects of introducing an intervention (the B phase) and removing the intervention (a return to the A phase) can be studied. Medication taking is a self-regulatory process, which means individuals require timely feedback about whether they are taking medications according to their personal standards. In

XRDS • WINTER 2014 • VOL.21 • NO.2 41 feature

The results of this study demon- Figure 3. The feedback group increased in the promptness (taking meds before the strate providing feedback can be one user-defined late time) of their medication taking after the feedback display was effective means for supporting behav- introduced into their homes, while the control group did not change significantly. ior change. This study also suggests simply providing behavioral feedback introduced is not enough, rather, the timing and feedback display control group frequency of the feedback is important feedback group for triggering and sustaining behavior 100% change. Confronting individuals with a long-term record of their medication taking behaviors convinced individu- 90% als in the control group they needed to improve. But without subsequent 80% reinforcement and feedback about their actual performance they could not troubleshoot their routines and 70% were only able to sustain improved Average Promptness Average medication taking for a few weeks. In 60% contrast, although providing real-time feedback had only a gradual, subtle effect individuals’ awareness of their 50% medication taking performance, it of- 1 2 3 4 5 6 7 fered the timely feedback necessary Month for self-regulating their daily medication taking. The frequent feedback allowed individuals to identify when week or something. [With the display] ery day, to make sure I was on time or they have forgotten or made a mis- you notice it, you see it right away.” make sure I was right or make sure I take, which gave them an opportunity P01 mentioned the display gave her a did the thing I was supposed to.” P06 to either correct the problem or avoid way to check whether she had taken also developed this same habit. In fact, making the same mistake in the next her evening pills yet: “I look up and after the display was removed from pill-taking episode. Real-time feed- I see ‘Oh! I didn’t take my evening their homes, both P02 and P06 report- back and longer-term feedback appear pills,’ so it’s a nice reminder, it’s not a ed they had grown so accustomed to to have different effects on aware- nasty kind of thing, you know.” Even glancing at the display that they still ness and follow-through. Combining though participants reported the turned their head to look in the direc- these types of feedback likely will have display helped them be more aware tion of the tablet after taking their pills complementary effects that support and identify errors, it did not make or during TV commercial breaks. behavior change through increasing participants realize when they were The improvements in medication awareness and supporting the actual performing their medication taking taking for the feedback group did not change in behavior, making this com- sub-optimally. In contrast, when the persist after the display was removed. bination an interesting topic of fu- control group reviewed a graph of the Measures of adherence, promptness, ture investigation that can influence 6–8 weeks of their medication-taking and correctness decreased to a level the manifold health apps that rely on behaviors, most found it shocking that that was slightly higher (but not statis- tracking and reflection. they missed so many medications and tically significantly higher) than their Another important lesson from vowed to make a change. However, left baseline levels before they received the this study is technological solutions without an easy means of feedback for feedback display. Variance in the time need to respect individuals’ autonomy. self-regulation, the members of the of day also increased to a level slightly Most individuals in the beginning of control group were only able to main- lower (but not significantly lower) than the study reported they performed tain a temporary and slight increase baseline. Participants relied on the their medication taking without any in medication taking performance for feedback to assist them in self-regu- problems and did not feel the need a few weeks before reverting back to lating their medication taking. Once for technology to support this seem- their baseline performance. the feedback display was removed, ingly simple task. Instead of introduc- Individuals need not only an aware- they did not have an easy way to verify ing the sensors and tablet display as ness of errors but also confirmation whether they had taken their medica- an assistive technology that reminded that they perform behaviors correctly. tions already and to identify when they (or even worse, nagged) individuals to Feedback group participant P02 re- made a mistake. Ongoing frequent take their medication according to a ported he regularly used the display feedback seems to be instrumental rigid schedule, the technology was in- to verify he took his pills correctly: “I for maintaining the improvements in troduced as a safety net that could be find myself [checking the display] ev- medication taking behavior. ignored or used in whatever way the

42 XRDS • WINTER 2014 • VOL.21 • NO.2 individual deemed useful to them. The applied to other behaviors important display required neither any button Ongoing frequent for maintaining independence. Moni- presses nor any other explicit inter- feedback seems toring other cognitive-related tasks actions; instead, it simply offered an in the home such as dialing the tele- easy-to-read, up-to-date information to be instrumental phone, multi-step meal preparation, display, driven by the natural interac- for maintaining or operating new electronic devices tions individuals already did (e.g., open like a TV can also provide feedback and close the pillbox). Users could con- the improvements to individuals and their care network sult it to verify their medication taking in medication-taking about health information, such as with the system and use their own self- changes in vision or cognitive abili- regulatory processes to adjust their behavior. ties. Smart glucometers can moni- daily routines to take their medica- tor how often people with diabetes tions more promptly and correctly. measure their blood sugar, support Older adults, just like younger their awareness of how often they adults, often do not believe they re- take measurements, and also provide quire help to do simple things even some assistance in interpreting the though their physical and cognitive values and identifying the causes for abilities may be declining. Ask any the individual’s autonomy is critical deviations. Cars, equipped with GPS, older adult and they will likely tell you because self-initiated behaviors driven distance sensors, and other drive- that they feel young: They report hav- by the user’s intentions are generally by-wire sensors can detect when an ing just a few more aches and pains more effective and more sustainable individual’s reaction time might be and just a little more forgetfulness, than behaviors driven by an external slowing and provide feedback for in- but not enough to affect their day-to- agent. Individuals should be allowed dividuals to drive more safely. Future day functioning. To design systems to set reasonable goals for themselves research should look into addressing that respect this “I still feel young” be- (e.g., “I want to take my medications on these challenges while considering lief, the technology needs to support time at least five times a week.”) and be the task-supportive effects of real- the user’s autonomy by empowering supported with feedback technology time feedback while also considering users to adapt their routines in the to make themselves aware of the prog- the impacts of long-term feedback. way they see fit rather than imposing ress toward their goals and to trouble- Moreover, it should apply technolo- a rigid regimen that makes users feel shoot their behaviors. gies that support individual auton- like they have a disability. Technology Routines help us carry out regular omy to self-initiate and perform the for older adults should be designed to tasks with minimal demands on our behaviors important for healthy, inde- be convenient and helpful rather than attention and effort. With this auto- pendent living. rigid and stigmatizing. maticity comes the occasional risk A good example of technology that of substituting the performance of a References preserves the user’s autonomy is the task with the mere thought of carry- [1] Kripalani, S. Interventions to enhance medication adherence in chronic medical conditions. Archives of spell checker found in word processing out the routine, as is often the case Internal Medicine 167, 6 (2007), 540. ing programs. Spell checkers only with medication taking. Sensing and [2] McDonnell, P.J. and Jacobs, M.R. Hospital intervene when the user has spelled feedback technologies in the home admissions resulting from preventable adverse drug reactions. The Annals of Pharmacotherapy 36, 9 something incorrectly, and when they can unobtrusively monitor the actions (2002), 1331–1336. do intervene, they give the option to of individuals and provide individu- [3] Lee, Matthew L. and Dey, Anind K. Real-time feedback for improving medication taking. In the user to accept or reject the spell- als with timely feedback about their Proceedings of the SIGCHI Conference on Human ing suggestion. Almost everyone acci- behaviors to support the mundane Factors in Computing Systems (CHI ‘14) (Toronto, dentally mistypes or misspells words, routines important for maintaining April 26–May 1). ACM, New York, 2014, 2259–2268. [4] Bandura, A. Social cognitive theory of self- especially when typing quickly. Spell their health and independence. These regulation. Organizational Behavior and Human checkers merely make typing and technologies must provide the right Decision Processes 50, 2 (1991), 248–287. spelling correctly more convenient amount of information and deliver [5] Lee, Matthew L. and Dey, Anind K. Reflecting on pills and phone use: supporting awareness of functional rather than making users feel badly it at the right time without requiring abilities for older adults. In Proceedings of the SIGCHI for being bad spellers or typists. Word too much attention from the user or Conference on Human Factors in Computing Systems (CHI ‘11) (Vancouver, May 7–12). ACM, New York, processors could enforce a rule that imposing a rigid routine that com- 2011, 2095–2104. imposes only correctly spelled words, promises the individual’s autonomy. encroaching upon the autonomy of dwellSense is an example of an unob- Biography the user, but they do not. Instead, spell trusive sensing system that provides Matthew L. Lee is a researcher at Philips Research North America. He is interested in bringing about the future checkers step in only when the user has glanceable feedback to help individu- of mobile and ubiquitous computing by tracking and made a mistake, and they conveniently als self-regulate, adapt, and improve making sense of the simple actions that people do in their everyday lives to reveal rich patterns about people’s give the user a chance to correct it by their medication taking routines behaviors and help people achieve their goals. exercising his/her judgment about as demonstrated in a longitudinal, © 2014 Copyright held by Owner(s)/Author(s). how to spell the word. Similarly with 10-month study. The same sensing Publication rights licensed to ACM. health behavior change, supporting and feedback technique can also be 1528-4972/14/12 $15.00

XRDS • WINTER 2014 • VOL.21 • NO.2 43 feature Seeing Is Believing Why visualization will play a critical role in bringing big data decision making to a hospital bed near you.

By Megan Monroe DOI: 10.1145/2676576

t’s 2 a.m. in the Neonatal Intensive Care Unit (NICU) of a downtown hospital when an alarm pierces the night. A bleary-eyed resident rushes in to investigate, weaving his way to a computer that monitors the vital signs of each infant. The screen is adamantly flashing an instruction to administer antibiotics to one of the babies. This is a daunting Irequest. For those of you who don’t speak baby medicine, administering antibiotics to an infant is not the same as popping a couple extra aspirin on a Sunday morning. You are

altering the bacterial landscape of an antibiotics as it pleases, and the resi- threshold. This can be problematic for immune system that is still developing, dent isn’t a resident at all because he’s doctors, who are trained in biology, which can result in all sorts of “bad” finishing up a Ph.D. in bioinformatics. chemistry, psychology, but not neces- down the road. In this case, however, That distant future is hardly some sarily in statistics. Probabilities are not the computer has determined it is worth singularity-dependent pipe dream. their lingua franca, especially at 2 a.m. the risk. But why? The vitals appear to Current pattern detection algorithms The point is this: In order to arrive be normal. There is no visible cause for can isolate trends that no human could at a distant future in which computers concern. So the question is, should the hope to unearth by hand. If a spiking are entrusted with critical, potentially resident follow this instruction blindly, fever in the NICU is a sure sign of dis- life-or-death decisions, we need to cre- without explanation? tress, but the symptom doesn’t present ate a near future in which humans and This scenario, by the way, is taking itself early enough for an effective inter- computers can arrive at data-driven place in the near future. Today, there vention, these algorithms can scan the decisions together. If these machine- would be no alarm until the infant data for invisible patterns that precede learning algorithms really want to showed an obvious sign of distress, the fever and suggest a course of action. prove their worth, they need to be able such as spiking a fever. In this case, the Before we break out the champagne, to learn directly from humans, not just resident would immediately adminis- however, there is one key thing to re- from the data trail we leave behind. ter antibiotics, but it would probably be alize: These algorithms don’t actually This means we need to find a way for too late to make a difference. It is also produce answers, they produce prob- humans, across a wide range of back- worth considering the distant future, abilities. To a computer, a “decision” grounds and expertise, to effectively in which the computer administers the is simply a probability paired with a communicate with their data-crunch-

44 XRDS • WINTER 2014 • VOL.21 • NO.2 ing counterparts. And while there is medical domain, to highlight three rea- sonal jabs at it, and Target has made much debate as to how this communi- sons why visualization will play a criti- untold millions using it’s predictive cation will ultimately take shape, there cal role in bridging the gap to a more marketing strategies. These isolated is general agreement that a large com- automated future in health analytics. missteps are a natural consequence of ponent of it will involve visualization. probabilities, casualties of a low and My graduate work at the University of REASON NO. 1: THERE IS SUCH pre-calculated error rate. And yet still, Maryland’s Human-Computer Interac- A THING AS A STUPID ANSWER an average human could flag them as tion Lab was spent on the development “What is Toronto?” This answer, incorrect in a single glance. of EventFlow [1], a visualization tool that chirped by IBM’s supercomputer Wat- The blunders of artificial intelli- allows users to assess event patterns son in response to a televised “Jeop- gence, however, become a lot less comi- in the context of a large population of ardy” question about U.S. cities, was cal when they involve actual casualties. patients (see Figure 1). This is done by seriously stupid. The word stupid also What if the only reason the vital-sign- grouping together patients with the comes to mind when you read about monitoring computer in the NICU is same event pattern, and shifting the Target, an American discount retailer, recommending antibiotics is because most common patterns to the top of the and their algorithmic decision to send the infant in question is in bed No. 4, display. It allows researchers to quickly pregnancy coupons to a teenager’s and the last three infants who spiked a answer population-level questions such home before she had even told her fever also happened to reside in bed No. as “Where were patients typically trans- parents that she was expecting. A com- 4? Perhaps bed No. 4 is indeed cursed, ferred after the emergency room?” and puter will never skip an instruction or but it doesn’t take a certified doctor to “How often are patients re-admitted calculate a percentage wrong, but when realize this logic seems flawed; it takes to the ICU?” Throughout the design they do err, the result can be obviously a human. But in order for this common and development of the software, I col- and laughably stupid. sense smell test to take place, the com- laborated with nearly a dozen medical It is worth pointing out both of the puter must provide some form of expla- research teams on long-term, multi-di- aforementioned stupidities were mi- nation for the decisions it’s outputting. mensional data analytics projects. This nor blemishes on the face of indisput- Even if these oddities are extremely article draws on my experiences with able victories. Watson had dominated infrequent, no one is going to under- these EventFlow case studies, as well as “Jeopardy” so absolutely that one of the write an error rate that involves cursed broader lessons learned outside of the contestants had resorted to taking per- cradles and dead babies.

XRDS • WINTER 2014 • VOL.21 • NO.2 45 feature

The good news here is even a basic and ultimately misunderstood copula what prompted the very existence visualization tool can typically reveal function was used to price hundreds of EventFlow. Medical researchers answers that are rooted in stupidity. of billions of dollars worth of mortgage were having trouble analyzing com- For example, one of our EventFlow case risk, a gross overextension of the func- mon event sequences within a large studies involved a dataset of the events tion’s original intention. Wired’s 2009 dataset of patients. They had to that precede and follow a surgery. The article on the ensuing financial crash scroll through page after page of pa- researchers had previously been using concluded the following: “One reason tient records; although this allowed command-based query tools to explore was that the outputs [of the copula them to easily understand the indi- the dataset, which had been produc- function] came from ‘black box’ com- vidual attributes of any one record, ing results that, again, seemed flawed. puter models and were hard to subject it made it impossible to glean in- Most notably, the calculation for the to a commonsense smell test. Another sights about the dataset as a whole. average duration of a surgery seemed was that the quants, who should have Our goal with EventFlow was to pro- slightly askew. When the dataset was been more aware of the copula’s weak- vide these researchers with a single loaded into EventFlow, however, and nesses, weren’t the ones making the screen on which they could see every each patient record was displayed in big asset-allocation decisions. Their event sequence in their dataset. To sequence, it became immediately clear managers, who made the actual calls, do this, details about the time laps- the calculation was being thrown off by lacked the math skills to understand es between events were not included surgeries that spanned midnight. The what the models were doing or how in the primary display. This allowed date of each patient’s surgery had been they worked” [2]. the visualization to group patients recorded in a separate column from the If you swap out some of the nouns by event sequence alone, creating individual event timestamps, so events in that previous quote, it can start to a much more compact display that that took place after midnight were ap- sound eerily similar to our curious in- better supported a dataset-level un- pearing at the start of the day. While cident of the NICU resident. When a derstanding of event sequences. these surgeries were rare, the data solution works, the immediate temp- anomaly was extreme enough to affect tation is to deploy it at scale. Comput- REASON NO. 3: THE BLIND SPOT aggregate calculations. The simple ad- ers will always be up for this challenge. We’ve established computer logic is dition of a date to every timestamp im- To them, one data point is the same as imperfect and visualization can help re- mediately corrected the problem. one million data points, regardless of veal these imperfections, but we haven’t This example is not meant to high- whether the logic scales as gracefully touched on the fact that these visualiza- light an overt stupidity on the part of as the dataset. Ironically, it is the hu- tions can be interactive. Not only can the humans or the computer, but rath- mans, with our comparatively lim- they help humans understand complex er how the disconnect between human ited memory and processing power, algorithmic computations across large logic and computer logic can result in who must decide how well the logic datasets, they also provide an interface stupid. The fact that a surgical team will scale across hundreds of billions through which humans can contribute cannot pause halfway through a proce- of computations. This is a steep chal- new information and feedback to the dure, travel back in time, and continue lenge, particularly when uncertainty machine. This contribution will be crit- the surgery at the beginning of the is involved (uncertainty is always in- ical because, even when computer logic day is so painfully obvious to humans volved). And the price of getting it is functioning exactly as intended, it is that it is easy to overlook this explicit wrong can be globally severe. always functioning within a bounded clarification. But from the computer’s Visualization can address the corpus of information. perspective, based on the data and the difficulties of scale by abstracting One of the most interesting Event- instructions that it was given, that is ex- non-critical information out of the Flow case studies came from the trau- actly what happened. Why not? display. In fact, this capability is ma bay at a children’s hospital. Nurses Visual representations can serve as and physicians in the trauma bay are a potent common language between required to adhere to the Advanced these two logics. They offer our best Trauma Life Support (ATLS) protocol, hope at explaining the inner workings We need to find a a series of five steps that must be com- of computational processes to profes- pleted in a specific order during every sionals outside of the technical do- way for humans, trauma resuscitation. The research- main. Not only can visualization high- across a wide range ers with whom we collaborated were light overt logical inconsistencies, but interested in identifying common as we’ll see next, they can also address of backgrounds deviations from the protocol in hopes more subtle errors that arise from the and expertise, of providing feedback to the trauma natural proliferation of new technology. team as well as making systemic ad- Reason No. 2: Infinity and Beyond Pre- to effectively justments to better support adher- dicting the behavior of a process at communicate with ence. In one of the most meticulous scale can be extremely difficult. The data collections I observed, videos of financial industry learned this lesson their data-crunching more than 200 traumas were reviewed the hard way in 2008, when a complex counterparts. by hand, and each step of the ATLS

46 XRDS • WINTER 2014 • VOL.21 • NO.2 protocol was recorded at the time it was executed. Figure 1. EventFlow consists of three panels: The control panel and legend (left), When this data was visualized the aggregated view (center), and the individual view (right). in EventFlow, two unexpected findings became immediately clear. The first was adherence to the ATLS protocol was far lower than expected, suggesting this protocol does not have as critical an impact on patient outcome as initially assumed. The second finding was instead of one or two common deviations from the protocol, the dataset contained 29 unique deviations, making it impossible to codify the results into a single action plan for improvement. More importantly, these findings shifted the foremost research question from “how can we better adhere to this protocol?” to “which components of this protocol are actually essential?” The Datasets like this one, which tracks patient transfers and outcomes, are typically researchers wanted to know whether displayed as a vertical list of records, as seen in the individual view. Scrolling is certain deviations were more predic- required to see the entire dataset, making it difficult to make holistic assessments. In the aggregated view, common event patterns rise to the top of the display (1), and we tive of negative outcomes than oth- can immediately see anomalous errors (2) and outliers (3). ers. The problem was there was no way to explore this new direction. Pa- tient outcomes had not been recorded ticular infant had an allergic reaction IN SUMMARY because the data collection had been to the lotion a nurse was wearing ear- One thing is certain: The answer to our designed under the assumption that lier that day. If those symptoms are original question is no, our resident, or the correlation between outcome and appearing on the timeline, but have any other medical professional, should ATLS adherence was a given. already been resolved, the resident not blindly execute an instruction Compiling a perfect dataset is like might ask the computer to factor out without explanation or clarification. trying to catch a stream of water in your those particular events and recom- Regardless of how much a computer hands. Some data is inherently inaccu- pute its findings. These changes may contributes to a medical decision, it is rate, some data is not available for lo- reduce the risk of fever to only 35 per- the human who will ultimately be held gistical or legal reasons, and some data cent, causing the system to rescind its accountable for the decision-making simply isn’t included because it didn’t initial recommendation. process. If computers hope to make a seem necessary at the point of design. If the medical domain is looking contribution at all, they must find a way Systems must be built to support a data for inspiration on this front, it can to integrate their input into the human- landscape that is always in flux, and a take cues from the self-driving car thought process. And if these machines large component of that will be seam- industry. This industry has already possess true intelligence, they will do lessly combining analytics from prede- embraced the fact that real solutions this by leveraging one of the most pow- termined data sources with in situ hu- can come from imperfect data by in- erful and mysterious components of man feedback. corporating human feedback. For ex- our human brains: the visual cortex. Let’s return to our example of the ample, if every car on the road were resident in the NICU. What if, instead networked, collisions could be avoid- References of being dictated a blind instruction, ed through car-to-car communication [1] EventFlow: Visual Analysis of Temporal Event Sequences and Advanced Strategies for Healthcare our resident arrives at the computer and driving routes could be optimized Discovery. Human-Computer Interaction Lab (HCIL), to find a timeline of events and at- at a massive scale. Humans could hap- University of Maryland. tributes that have been mined from pily sleep through their morning com- [2] Salmon, F. Recipe for disaster: The formula that killed Wall Street. Wired February 23, 2009. the infant’s vital signs? The computer mute. Ten years down the road (pun has found a particular pattern that intended), this might be realistic, but has preceded a fever in more than 95 right now this data is simply not avail- Biography Megan Monroe completed her Ph.D. in computer science percent of other NICU cases. This in- able. As a result, self-driving cars, in at the University of Maryland, and is currently working for formation has been perfectly visual- their initial incarnations, will require IBM Research in Cambridge, MA. Her research interests include data analytics, visualization, and college football. ized, making it extremely easy for the some degree of engagement from resident to understand how the com- their human drivers. The cumulative puter arrived at its decision. However, task of driving will be a collaboration maybe the resident knows this par- between man and machine. © 2014 ACM 1528-4972/14/12 $15.00

XRDS • WINTER 2014 • VOL.21 • NO.2 47 feature Wearable Technologies: One step closer to gait rehabilitation in Parkinson’s patients Wearable computing has the potential to fundamentally alter healthcare by enabling long-term patient monitoring and rehabilitation outside of the lab.

By Sinziana Mazilu and Gerhard Tröster DOI: 10.1145/2676578

earable technologies have begun to take root in our everyday life. On-body sensors are used to quantify how often and how well we perform our favorite sport, while our smartphones have become our personal assistants— monitoring how active we are, our daily routines, the places that we visit, and Wthe activities we do. Wearable systems are already being developed for bipolar disorder, cerebral palsy, chronic pain, and Parkinson’s disease. The latter ranks among the most common neurological disorders, with an estimated 7 to 10 million people worldwide living with Parkinson’s disease.

The symptoms of Parkinson’s ate symptoms through medication or in the same way as we tend to ignore disease include tremors, slow move- deep-brain stimulation with a brain the rhythm of a song that we have lis- ments, rigidity of the limbs, shuffling pacemaker. Freezing of gait is often tened to too often. gait, and, in advanced stages, the freez- treatment-resistant. However recent ing of gait. Freezing of gait is a sudden clinical studies have shown people THE WEARABLE ASSISTANT inability to move, people with Parkin- with Parkinson’s disease respond to In the Wearable Computing Lab at son’s describe it as the feeling as if “the rhythmic external stimuli [2], such as ETH Zürich we have been researching legs are glued to the ground” [1]. Of metronome beats, which might help to the use of wearable sensors and smart- all the symptoms of Parkinson’s, gait alleviate gait freeze. There is a simple phones to provide rhythmic auditory freezing is the most feared, being the explanation: The human gait is rhyth- stimuli to people with Parkinson’s dis- main cause of falls and mortality in mic, and humans respond to rhyth- ease, helping them to alleviate or even people with the disease. mic tones in everyday life, for example overcome freeze episodes. Although widespread, the causes when listening to music while walking Rhythmic stimulation combined of Parkinson’s disease remain un- or jogging. However, the impact of a with gait-training exercises is a prom- known and there is currently no cure. continuously rhythmic pattern on the ising practice toward gait rehabilita- Existing treatments can only allevi- Parkinsonian gait wears off with time, tion in Parkinson’s disease [3]. A wear-

48 XRDS • WINTER 2014 • VOL.21 • NO.2 able system will allow people to follow tAssist, a wearable gait assistant, which to a smartphone. On the smartphone these treatments in their homes, with- supports rehabilitation exercises in Par- an Android application analyzes the out any clinical supervision. Gait-train- kinson’s disease [4]. The development of sensing data and detects in real-time ing experts and clinicians are replaced GaitAssist threw us into the adventur- whether the subject is experiencing a with the wearable system, which can ous intersection of wearable systems, gait freeze. If a freezing of gait is de- also act as a walking assistant during machine learning, and human-comput- tected, a rhythmic metronome ticking daily-life activities. er interaction for healthcare. sound is started, which is synchro- There are two main research ques- GaitAssist uses on-body sensors nized with the rhythm of the user’s tions that underpin the development and a smartphone to detect gait freez- gait. The rhythmic sound continues of such a system: What kind of sensors ing in real-time. Upon a gait freeze or a until the user resumes their normal and what methods do we need to best similar gait pattern, a rhythmic audito- walking pattern. A detailed descrip- capture the gait-anomalies characteris- ry cueing is given to the user, support- tion of GaitAssist and its components tics in Parkinson’s disease? Where will ing the subject in regaining the motor is presented in Figure 1. the wearable components be placed on function and to resume walking. the patient’s body and how will future GaitAssist consists of up to two in- SENSOR TYPES users accept wearables? ertial sensors mounted on each ankle, AND ON-BODY POSITION

Image by Ocskay Bence Our answer to these questions is Gai- which send real-time data readings Following previous research cues, we

XRDS • WINTER 2014 • VOL.21 • NO.2 49 feature

decided to use wearable inertial mea- Figure 1: Eighteen people with Parkinson’s disease and five healthy subjects were surement units (IMUs) to capture the asked to perform walking protocols in a laboratory setting while wearing nine on- characteristics of gait in order to detect body inertial measurement units (IMUs) and a smartphone. the motor freezing episodes. On-body IMUs containing an accelerometer, a gyroscope, and a magnetometer are useful in capturing human motion. They can be used to detect the activity the user is performing [5] —e.g. sitting, walking, cycling, or running— and even offer precise-enough informa- Back sensor tion to determine the correctness of a movement, allowing (for instance) running skills to be assessed. We asked five healthy participants and 18 people with Parkinson’s disease and freezing of gait to each perform a single three hour walking session in a laboratory setting, over a period of three Hand sensor weeks. The walking tasks were specifically designed by clinicians to induce gait freeze in Parkinson’s disease subjects and included turns, u-turns, and Thigh sensor walking in narrow, crowded corridors or with varying cognitive load. During the walking protocol, the participants wore a system composed from nine in- Smartphone ertial measurement units and a smartphone. The placement of the wearable sensors is shown in Figure 2. Ankle sensor Each subject wore the system con- Foot sensor tinuously for a period of two to three hours, and sensor data was collected as they completed the tasks in the walking protocol. Subjects then completed Two of the IMUs were placed on the feet, two on the ankles, two on the thighs, two on questionnaires and participated in the arms, and one on the lower back of the subject. The smartphone was placed in a open-ended discussions with clini- trouser pocket. cians, regarding the components that the wearable system should have, and Figure 2: The GaitAssist system consists of up to two wearable sensors. where those components should be placed on the body, in terms of comfort and wearability. This co-design process concluded participants perceived (b) foot and ankle positions to be the most comfortable for wearing on-body sen- (a) sors. The smartphone in the pocket was not noted at all, being already accepted as an everyday wearable tool. Independent of the co-design study, we analyzed the sensor data collected (c) during the walking sessions to check which body position gave the best information to distinguish the motor block- (a) Attached on the ankle of the users, sensors are attached using specially designed ing episodes from the other walking Velcro straps. Data sample from the wearable IMUs is sent in real time to a Samsung events. Clinicians labeled all the freez- S3 Galaxy phone (b) that acts as a wearable computer. Sensing data is analyzed in ing of gait episodes from the videos tak- order to detect the gait-freezing episodes. Upon motor block, a rhythmic sound is provided for a limited period of time. The subject can choose to use a single earbud en during the data collection, and these (c) to hear the rhythmic biofeedback given by the system. labels were then synchronized with the sensor data. Applying the same detec-

50 XRDS • WINTER 2014 • VOL.21 • NO.2 Figure 3: A data sequence from an ankle IMU’s accelerometer, gyroscope, and magnetometer from a subject with Parkinson’s disease Data contains a gait-freezing episode and diverse walking events, such as getting up, turning, or stopping. We can easily observe raw sensor data during the gait-freeze episode is similar to other gait events, such as sitting, starting to walk, or slowly turning.

3−Axis Accelerometer Raw Data

40 freezing of gait

–20 sitting walking turning 360 degrees turning 360 degrees walking turn walking sitting 0 5 10 15 20 25 30 35

3−Axis Gyroscope Raw Data

4 2 0 –2 –4 0 5 10 15 20 25 30 35

3−Axis Magnetometer Raw Data

400 200 0 –200 –400 0 5 10 15 20 25 30 35 tion algorithm to IMU data from five dif- the IMU data to characterize the motor models and decision trees in particular ferent body positions showed the ankle block events and to develop methods to are robust in terms of both detection was the most informative position to detect such episodes in real-time. performances and computation time extract the properties of freezing events. Fast Fourier Transformation (FFT)- for human activity recognition and gait- based features from acceleration are freeze detection. The feature computa- HOW TO DETECT FREEZING OF GAIT used to characterize and quantify the tion time and the classifier decision time FROM WEARABLE SENSOR DATA gait. We followed this direction and are very important in detecting risky Figure 3 shows a data sequence from an compute four gait-freeze specific FFT gait anomalies such as freezing of gait. ankle IMU’s accelerometer, gyroscope, features in a sliding window manner A freeze episode needs to be detected as and magnetometer from a subject with from the raw acceleration data. Figure soon as it starts, to avoid the onset of a Parkinson’s disease. This sequence con- 4 contains details about FFT-based total motor block or a fall, and to help in tains a gait-freeze episode and diverse features and an example of how they resuming walking as early as possible. walking events such as getting up, turn- change during normal walking and ing, and stopping. We can easily observe freezing of gait. We extended our fea- CLOSING THE LOOP: A PRELIMINARY that raw data during the gait-freezing ture extraction to statistical features, EVALUATION OF GAITASSIST episode is similar to the data collected e.g., mean, standard deviation, min, The best evaluation a healthcare wear- during other gait events, such as sitting, and max from accelerometer, gyro- able system can have is to pass ac- starting to walk, or slowly turning. scope, and magnetometer data. But the ceptance tests with its users. To test In our problem it is just as important most informative features for our data GaitAssist, we asked five additional to avoid false gait-freeze detections, as are the FFT-based ones extracted from subjects with Parkinson’s disease and it is to detect the freezing episodes, e.g., acceleration data. freezing of gait to set up and wear Gai- to incorrectly classify a “stand up” event The gait-freeze detection algorithms tAssist during a gait-training protocol as a gait-freezing episode, only because consist of decision trees, trained on the in a hospital setting. The subjects had their IMU data patterns share similari- acceleration data from the ankle posi- not participated in the previous design ties. Thus a very important part of our tion, collected from the 18 Parkinson’s and development process of the sys- work was to find the best features in disease subjects. Machine learning tem, and had not used a wearable sys-

XRDS • WINTER 2014 • VOL.21 • NO.2 51 feature

tem in other circumstances before the sist started rhythmic auditory cueing comfort of the system. The usability evaluation protocol started. at moments without gait freeze. This part assessed the participants’ experi- Subjects were asked to perform a typically happened during unusual ences with the rhythmic audio feed- gait-training and rehabilitation proto- and anomalous gait motions that re- back during freeze episodes. The feasi- col assisted by the wearable system, as sembled freezing, such as turning bility part inquired about the possible shown in Figure 5. The protocol includ- with very small steps, sudden stops, use of the GaitAssist system at home, ed sessions of walking designed to pro- or during sit-to-stand and stand-to-sit and the participants’ motivation to use voke gait freeze events, such as u-turns, movements. The latter triggers were the system in daily life. In the final part 360-degrees turns, or sit-to-stand tasks, due to the lack of these types of move- of the questionnaire, we addressed the but also sessions of natural walking in ments in the training data. usage of the system from a technical crowded and narrow corridors, stairs, The total time required to compute point of view, and the user-friendliness and elevators of the hospital. Each ses- the FFT features from the sensing of the software and sensor setup. Ad- sion lasted around 30 minutes and sub- data,and for the gait-freeze detection ditionally, we performed informal dis- jects were asked to repeat the protocol algorithm to make a decision, is at most cussions between a clinician and the on three different days. The protocol was six milliseconds, measured in real-time user to gather more detailed feedback. video recorded and synchronized with on the Samsung Galaxy S3 mini phone. Participants reported they were the GaitAssist sensing data, and as in the Regarding resource consumption, in a satisfied with the wearability and per- previous experiment clinicians detected realistic usage setting GaitAssist con- formance of GaitAssist. They enjoyed 102 gait-freezing events from the videos. sumes less than 1 percent of the battery using the system and felt it was reli- The first test: How well does the power per hour. However, the IMU bat- able and accurate. The sensor attach- system function? GaitAssist success- teries last for at most four hours, which ments did not trouble the users during fully detected 99 gait-freezing epi- is sufficient to use the system for the walking, although they expressed the sodes and started auditory cueing in gait-training exercises but limits the du- desire to wear the sensors under their response, typically with latency small- ration for which the system can be used trousers. Subjects reported they often er than 0.5 seconds after the start of as an assistive device. forgot they were wearing the sensors a gait-freeze event. The three missed The second test: Is the system ac- during the trial. gait-freezing events were shorter than ceptable to users? At the end of each They perceived the audio feedback 0.5 seconds, and therefore difficult to testing day we asked the five subjects played through earphones as easy to detect; 57 false alarms occurred in to- to complete a feedback questionnaire listen to, even in noisy environments tal during the study, meaning GaitAs- regarding the usability, feasibility, and They also reported the auditory feedback was always triggered at the right Figure 4: Twenty seconds of walking and turns, including a gait-freezing episode. time when a long freezing of gait epi-

40 sode occurred, and with low response ] 2 freezing of gait Acceleration magnitudes [raw data] latency, i.e., 0.5 seconds, which corre- 30 sponds to the latency we measured. 20

10 All participants felt they benefited Acceleration [m/s from the support of GaitAssist during 0 0 2 4 6 8 10 12 14 16 18 20 freezing episodes. Figure 6 shows their

Power freezing of gait band [3−8] Hz Power locomotion band [0−3] Hz Total power [0−8] Hz perception on whether the rhythmic 10 auditory feedback provided by GaitAs- sist could reduce the number of freez- 5 Power ing events and whether it could reduce the duration of those episodes. They re- 0 0 2 4 6 8 10 12 14 16 18 20 ported the freeze duration was shorter 10 compared to their normal experience, Freze index feature but using the system would probably 5 not decrease the number of such epi-

Freeze index Freeze sodes. Although rhythmic stimulation

0 0 2 4 6 8 10 12 14 16 18 20 was provided in some false-positive Time [s] cases, i.e. without a motor block in We compute the raw-acceleration magnitude and compute 4 FFT-based features progress, subjects reported this might in a sliding-window manner: Power on the freezing of gait band [3–8] Hz, power on the locomotion band [0.5–3] Hz, the total power in the [0–8] Hz band, and the freeze in fact have prevented them from expe- index, which is the ratio between the power on the “freezing of gait band” and the riencing a gait freeze, as they felt they power on the “locomotion band.” Power-on freeze band and freeze index features were having difficulties walking in increase prior to or during the gait-freeze compared with the walking periods, while power on the locomotion band (PL) values decrease before or during freeze. While some of these instances. the locomotion-power feature is similar in cases of turning, standing, sudden stops, and freezing of gait, the total power feature helps in distinguishing between turns INSTEAD OF CONCLUSIONS and freezing of gait, for example. The combination of the four features can give an accurate description of the freezing-of-gait episodes. At the end of the day, nothing is more rewarding for researchers than seeing

52 XRDS • WINTER 2014 • VOL.21 • NO.2 and raw information about sensing Figure 5: The GaitAssist system during in-the-lab validation with five Parkinson’s data and gait-freeze detection. Clini- disease subjects. cians are able to remotely monitor the training and progress of the patients, and change the treatment and exercises setting accordingly, without requiring the user to visit the hospital or the clinician to visit the user’s home. The detection of gait freeze is a difficult yet very interesting problem to solve, and one, which is still open-ended. But detection implies gait-freeze events still occur. A different approach would be to be able to predict periods of gait difficulty, and to provide a rhythmic cueing or alerts, in order to help the From left: A subject wearing the wearable sensors during the protocol, a subject user to completely avoid these episodes. performing a figure-eight-like walking task, and another subject preparing to perform a protocol session, which includes sit-to-stand and turning tasks. Meanwhile, we hope wearable technology, in general, will enrich the the positive impact of their developed sistant for its users and even to replace healthcare system, and GaitAssist will system on its target users. GaitAssist the clinician’s help outside the lab, help in building our knowledge of Par- was well received and accepted by its opening the opportunity for people kinson’s disease rehabilitation. users, who stated they “don’t feel alone, with Parkinson’s disease to perform and the system is like someone who gait-training protocols in their homes. References supports them,” they “could wear it all Up to now such treatments were given [1] Nutt, J.G., Bloem, B.R., Giladi, N., Hallett, M., Horak, F.B., and Nieuwboer, A. Freezing of gait: Moving day long, as it helps with decreasing the in clinical settings or under the super- forward on a mysterious clinical phenomenon. gait freeze severity,” and it “is very use- vision of a physiotherapist—a burden Lancet Neurology 10, 8 (2011), 734–744. ful and improves the gait.” for the healthcare system in terms of [2] Donovan, S. Lim, C., Diaz, N. et al. Laserlight cues for gait freezing in Parkinson’s disease: An open-label Following these positive results and costs and personnel. study. Parkinsonism & Related Disorders 17 (2011), users’ feedback, clinical researchers In this article we’ve focused mainly 240–245. are running an in-home trial with more on one function of the wearable sys- [3] Espay, A., Baram, Y., Dwivedi, A. et al. At-home training with closed-loop augmented-reality cueing than 20 subjects with Parkinson’s dis- tem—the real-time detection of gait- device for improving gait in patients with Parkinson’s disease. Journal of Rehabilitation Research & ease and freezing of gait to study wheth- freezing episodes and the start of Development 47 (2010), 573–581. er GaitAssist can help improve their rhythmic auditory stimulation upon [4] Mazilu, S., Blanke, U., Hardegger, M., Tröster, G., gait and alleviate gait freeze in the long- these events—but the wearable system Gazit, E., and Hausdorff, J.M. GaitAssist: A daily-life support and training system for parkinson’s disease term. Users are performing specially also supports gait-training exercises patients with freezing of gait. In Proceedings of the designed gait-training exercises without designed by physiotherapists. It has 32nd annual ACM Conference on Human Factors in Computing Systems. ACM Press, New York, 2014, any clinical supervision, assisted only by a simple to use interface allowing the 2531-2540. the wearable system. The hope is after a user to choose the type of training and [5] Bulling, A., Blanke, U., and Schiele, B. A tutorial on few weeks of such training, users will exercises to be performed. human activity recognition using body-worn inertial sensors. ACM Computing Surveys 46 (2014), 1–33. have improved overall gait quality. The system is connected to a tele- GaitAssist is designed to be an as- medicine server and sends statistics Biographies Sinziana Mazilu received her Dipl.-Ing. and M.Sc. degrees in computer science and information technology from Figure 6: The five participants’ individual scores to the following statements: Politehnica University of Bucharest in 2009 and 2011, (a) “GaitAssist helps in reducing the freezing of gait number of episodes,” and respectively. She joined the Wearable Computing Laboratory at ETH Zürich as research assistant in 2011. Her research (b) “GaitAssist helps in reducing the freezing of gait durations.” interest lies at the intersection of wearable computing for • Reduce the gait-freeze number • Reduce the gait-freeze duration healthcare, human computer interaction, applied machine learning, and context recognition from on-body sensors. 5 Gerhard Tröster received his Dipl.-Ing. degree in electrical 4 engineering from Darmstadt and Karlsruhe in 1979 and his Dr.-Ing. degree from the Technical University 3 (Darmstadt, Germany) in 1984. He was involved in the research on design methods of analog/digital systems in 2 CMOS and BiCMOS technology for eight years at Telefunken Semiconductors. Since 1993, he has been a full professor Likert score Likert 1 of electronics at ETH Zürich, heading the Electronics Laboratory. At ETH, he established the multichip module 0 Subject1 Subject2 Subject3 Subject4 Subject5 (MCM) electronic packaging group. In 2000, he founded the Wearable Computing Laboratory, where he is involved in Subjects were asked to give answers in a Likert score format, where 1 means they interdisciplinary research combining IT, signal processing, electronic platforms, wireless sensor networks, smart strongly disagree with the statement and 5 means they strongly agree with the textiles, and human-computer interaction. statement. All of them appreciated that GaitAssist supports them in decreasing the freezing of gait duration, but not the number of episodes. © 2014 ACM 1528-4972/14/12 $15.00

XRDS • WINTER 2014 • VOL.21 • NO.2 53 profile

PROFILE DEPARTMENT EDITOR, ADRIAN SCOICĂ Trevor van Mierlo The Story of Building a Startup in Health Informatics DOI: 10.1145/2685368

Anyone who’s been of video games or doing some minor informatics. In his own words, the late a member of tech programming on VIC-20s or Commodore ‘90s were “a really exciting period for culture for a while 64s, so technology wasn’t my forte,” he the dot-com boom, because people were can acknowledge the explained. He pointed out that as head throwing money at anything you could startup phenomenon of a company with its own server farms, digitize.” He was playing a technical, a has been responsible developers, designers, and programmers, business development, and a product for infusing even more technology into he now needs to have an understanding of management role in the company, which our lives. However, despite the fact that all the languages and be able to work with allowed him to earn experience with we’ve managed to move almost every all the technologies involved. both the business side of pharma and facet of our mundane existence online— “My first degree was a double major the research side of behavioral health. It from working to bonding—healthcare has in English and history, and I really loved was also at Mediconsult that he first got stubbornly resisted digitization attempts school, and I was going to pursue my a taste of the power of data, which was over the years, making health-startups doctorate when, quite frankly, one of my appealing to both the business people, today seem like an uncharted land of advisors said to me that I was probably for whom it was providing return on uncertainty and untapped potential. making a bad choice to do a Ph.D. in their investment, and to the behavioral To find out what it takes to seize the theory and criticism, because it was going psychology researchers, for whom it was opportunity and build a successful to cost me a lot of money, and it would providing new avenues for study. business in the risky environment of be very difficult for me to find a job,” he Thus, when Mediconsult later went healthcare startups, I went to the source. admitted. Since he was the type of person out of business in the same way many Trevor van Mierlo lives in Canada. who really loved data and really loved dot-com companies did, he recognized He is the founder and CEO, as well as academics, he was encouraged to go out the opportunity to go out into the acting CSO, of Evolution Health, which into the world and find a job in something field and develop programs with fairly specializes in building digital solutions that was new and unique—something that specific return on investment criteria to improve patients’ adherence to he loved— because school would always for the different stakeholders he had treatment and medication. The company come back to him later. previously worked with. In 2000, he now has offices in both Toronto and founded his own consulting company Silicon Valley, and it does business LIFE AFTER GRADUATION called Evolution Health. with partners from all over the world. After graduation, van Mierlo took the As his work at Evolution Health was According to van Mierlo, the backstory advice of his mentors and went on to taking off, van Mierlo was excited to see is a long, inspiring journey of constant work for a health informatics startup his occupation merge into the academic learning and self-improvement. company called Mediconsult.com, of side of things when he began going which he was the ninth employee during to international medical conferences PREPARING FOR A CAREER the dot-com bubble of the 1990s. and presenting the outcome data. It IN HUMANITIES “The company ended up creating became apparent to him that he needed “I was born in a small town in Northern 360 employees within a two-year time a master’s degree in health in order to Ontario, where I grew up in—for lack of a period, so it was baptism by fire for be accepted as an insider within the better term—a family of proletarians; I me,” he recalled, adding the stock for international research community, so was actually the first person in my family the publicly held company went from he decided to go back to university. to go to university,” van Mierlo recalled. about 49 cents to between $23 and He earned a master’s of science and And to make his story truly surprising, $24 in those two years. They company community health, with a concentration even though he is now the CEO of a had managed to raise (and spend) as on addictions and mental health. healthcare tech company, his career much as $80 million in its last round “I think it’s a little bit of a story in started out with an undergraduate of financing. It was at Mediconsult itself how when you’re setting out in your degree in humanities, not a STEM field. that van Mierlo saw potential in the career, sometimes it’s good to follow an “Actually, as a teenager, I just spent healthcare industry, and began to lay the academic path that’s laid out for you. most of my time playing hockey instead foundations for his future work in health Getting a master’s of science, as I did,

54 XRDS • WINTER 2014 • VOL.21 • NO.2 was not something that would have the demographic and psychographic naturally occurred to me after my first characteristic of “superusers” in health degree,” he confessed. support communities . As things snowballed with Evolution Health, and the business started MANAGING A HEALTHCARE to grow, van Mierlo again realized STARTUP while talking to potential investors While talking about the process of and clients that he was still lacking founding and growing a healthcare something—this time, it was business startup business, van Mierlo explained speak. He was quick to take his own the key challenge he had to face was advice, and enrolled in a program at the overcoming the systemic problem of University of Toronto called the Omnium digital health, which is marrying three Global Executive MBA, which was a joint disciplines that often don’t get along Evolution Health had made the mistake program between the Rotman School with each other: business, healthcare, of being stuck with legacy in the past. He of Management and the University and technology. explained agile development is crucial of St. Gallen in Switzerland. He was “On one hand, you have technology, for the success of a health informatics therefore able to go to class at different which moves very fast, on the other hand startup, which is not always true for universities in countries such as Brazil, you have healthcare, which moves very large software companies, China, India, as well as Europe and slowly, and then you have the business North America. This opportunity allowed and financing side of things, which is A CLOSING WORD him to learn about doing business really focusing on maximizing profit. While van Mierlo thinks e-health is going internationally, in different cultures, Getting these three types of people in a to become as ubiquitous in the future while at the same time getting the room is very interesting, because their as social media is today, he does warn financial, accounting, and supply chain underlying philosophies are not the that the path has not yet been cleared knowledge that he needed in order to same, and they often come from very despite investments of hundreds of advance Evolution Health. different perspectives.” millions of dollars in digital health over He didn’t stop there, however. “By He also warned being an entrepreneur the past couple of years. now, we’ve presented at almost 100 requires having a self-motivated “The only thing I’d really say,” he academic conferences; we have at outlook on life, and should be fully concluded, “is to keep an open mind. least 20 peer-reviewed publications, prepared to make more mistakes than When I was in school and I had my double so we’ve done a lot; and we work in you have successes—which is in itself major in English and history, I would have diverse areas, from smoking cessation, a very powerful life experience. To be never believed you if you had told me to depression, panic disorder, colitis, successful, he explained, one needs that I would end up having four master’s and arthritis, and more recently even to realize with an environment like degrees, that I would have a number of schizophrenia. The ability to develop a healthcare and digital health, everything publications under my belt, and that I methodology and be able to apply it to is progressing so fast if you don’t remain would also be running a company that’s different disease conditions and also light on your feet and you become too doing business in a number of countries. different cohorts of patients is really invested in a certain type of technology, However, the strength that I do have is fantastic, because we just keep learning it can cost you everything. being able to recognize opportunities and developing new things every day,” “Being able to adapt technology is and find holes in both academia and van Mierlo enthusiastically explained. extremely important, because what business. So I think that if you’re able to He is currently pursuing a doctorate worked in 2010 no longer works in 2015, go into both business and school with degree with a focus on patient profiling and, from a business perspective, it an open mind, the places you can go are and the forecasting of health behaviors can become very expensive to invest quite amazing.” through predictive algorithms, through in things which are outdated quickly,” which he hopes to share his findings on he added. According to van Mierlo, Copyright held by Owner(s)/Author(s). Photo Credit TK

XRDS • WINTER 2014 • VOL.21 • NO.2 55 end

A world map listing the latest known OpenMRS implementations and their intention of use.

LABZ greatly toward the formal establish- ment of the GHI group, and helped The Regenstrief Global formulate many of the group’s key aims, values, and areas of focus. The GHI group strives to: Health Informatics Group ˲˲ Be a world leader in pragmatic health information solutions. Indianapolis, Indiana ˲˲ Promote a global community where the promise, outcome, and Editor’s Note: In this issue Suranga pion and architect of many of Regen- real-world value of health information Nath Kasthurirathne talks about his strief’s global health initiatives. The technologies are recognized. research and fellowship experience in GHI team focuses specifically on ad- ˲˲ Ensure the environments we supporting and contributing to a number dressing issues pertaining to global serve are empowered to implement of Open Source Health Informatics projects, health information technology, and and maintain this technology on their specifically the Open Health Information is widely known for its efforts to im- own. Exchange (OpenHIE) project—Somdip Dey prove healthcare across some of the ˲˲ Directly contribute to the most underserved populations on strengthening of large health care s a student, my association earth. Regenstrief, in collaboration systems. with the Regenstrief Global with Partners In Health (Boston) and Currently, the main areas of focus Health Informatics (GHI) the Medical Research Council (South of the GHI group include: the Academ- group began in 2011, when Africa), helped found the OpenMRS ic Model Providing Access To Health- AI was an intern for one of their open project in 2004. OpenMRS is an open care (AMPATH), Open Medical Record source projects. Regenstrief GHI falls source effort to build a framework System (OpenMRS) and Open Health under the domain of the Regenstrief for electronic medical record (EMR) Information Exchange (OpenHIE). Institute’s Center for Biomedical In- systems in resource constrained en- My internship was part of the formatics .The group is led by Dr. Paul vironments. The rapid expansion and Google Summer of Code program. My Biondich, who has been a key cham- acceptance of this project contributed work on OpenMRS boosted my inter-

56 XRDS • WINTER 2014 • VOL.21 • NO.2 Real-time data collected on mentions of symptoms on Facebook and Twitter allows sites like Sickweather.com to plot a “disease heatmap” of the world.

est in health informatics, and encouraged me to continue contributing to BACK the project as a volunteer. After continuing to work on the project under various roles, I went on to become a Radiography mentor for other interns who were new to OpenMRS. Modern medical imaging encompasses a diverse set of techniques In August 2013, I became a fel- used for scientific study, and non-invasive diagnosis of many medical low of the Regenstrief GHI group. As conditions. Many of you are probably familiar with the names of part of my fellowship, I was able to several such techniques, and most of you have likely been subject to continue working under the capable at least one imaging technique at some point in you life. Perhaps the guidance of Drs. Paul Biondich, Burke most familiar category of techniques though is radiography, which Mamlin, and Shaun Grannis. And as refers primarily to methods using X-rays. part of my commitment to OpenMRS, The X-ray was discovered in 1895 by Professor Willhelm Röntgen. I am heavily involved in mentoring During his experimentation with cathode rays, he observed his the next generation of budding health equipment was producing some other invisible rays that were capable informaticians. I also assist in the de- of penetrating books on his desk. After studying the phenomenon in velopment of OpenHIE, focusing on greater detail, he produced the first X-ray photograph, radiographing the evaluation and measurement of his wife’s hand. The potential medical application was clear and clinical care delivered by the Rwanda quickly put to use. Soon after this discovery, in 1896 John-Hall Health Information Exchange—the Edwards became the first to use radiography in clinic and surgical first ever implementation of Open- practice, while Thomas Edison’s fluoroscope became the common HIE. As a graduate student, I feel priv- imaging device in use. ileged to see this project grow, and X-rays were not well understood in the very early days, however. have learned tremendously by watch- The image quality was lower and the dangers of radiation were not ing, and contributing toward its de- immediately acknowledged. Since then radiography has come a long sign and development. way in the last century, with advances in X-ray emitters and digital I also play a significant role in ex- detection technology that allow for faster imaging at lower doses of ploring the future of healthcare in- radiation. More recent developments in high-resolution radiography, teroperability standards and their and computed tomography with fast 3-D reconstruction of scans, has application to both OpenMRS and pushed the popularity of the technology even further. It continues to OpenHIE. I recently led an effort to expand into more applications, medical and otherwise. design and build support for Fast —Finn Kuusisto Healthcare Interoperability Resourc- es (FHIR)—a widely accepted interoperability standard used to exchange healthcare information between different hospital systems and clinics. In addition, I am also involved with other efforts to improve the efficiency and reliability of public health registry reporting. We believe the automated diagnosis and reporting of health conditions can help address this limitation, and are working on Early Radiography Modern Radiography production worthy approaches to ad- Timeperiod 1890s Present dress these issues. Image Dimensionality 2-D 2-D, 3-D X-ray attenuation, phase shift, Radiation measurement X-ray attenuation Biography backscatter Suranga Nath Kasthurirathne is a second year Ph.D. Detector Type Photo plate or film Digital detector student in health informatics at the Indiana University - Medical, industrial, security, Purdue University (IUPUI), Indianapolis. He is also a fellow Uses Medical microscopy Early Radiography photographRöntgen. construction Wilhelm Bone by photograph Zygorfi by (https://commons.wikimedia.org/wiki/File:Bonereconstruction.jpg) of the Regenstrief Global Health Informatics (GHI) group.

XRDS • WINTER 2014 • VOL.21 • NO.2 57 HELLO WORLD The Anatomy of a Human Disease Network BY MARINKA ZITNIK

ools of network analysis have recently been applied to many Figure 1: The human disease network. Every node corresponds to a distinct complex systems, to both disorder and is colored based on predicted network community to which it simplify and highlight their belongs. The size of each node is proportional to the number of genes im- underlyingT structure and the relation- plicated in the corresponding disease. We labeled diseases associated with ships that they represent. The results many genes. obtained from network-based approaches provide not only insight into interactions between online users, but also new clues about how to improve our understanding of biological systems. Network medicine in particular, a network-based approach to studying human disease, has proven effective in studying interdependence between molecular components in cells, and in identifying disease modules and biological pathways [1]. The data we consider is usually in the form of a set of records, each of which describes one object in the system. These objects may be physical entities such as planets, stars, or people, for example customers; or may be abstract entities such as diseases. Given a social or a biological system, many kinds of data can be collected. For example, a star might be described by the strength of its spectral emissions at particular wavelengths; a customer may be profiled using his or her demographic data; and a disease might be described with a set of genes whose mutations are implicated in this disease. Data about these systems are often represented in the form of networks. Network abstractions often turn out to be practical, but remain highly complex and are hence an active area of research.[2, 3, 4, 5]. In this column, we explore the human disease network [2] and demonstrate how network-based tools can help us understand relations between diseases at a higher level of organismal organization without considering any prior biomedical knowledge.

58 XRDS • WINTER 2014 • VOL.21 • NO.2 ON CONSTRUCTING THE NETWORK Listing 1: A Python script to construct the human disease network. The list To construct the human disease of diseases, genes, and associations between them was obtained from the network we follow the influential work Online Mendelian Inheritance in Man and from Goh et al. [2]. of Goh et al. [2]. Nodes in the network represent diseases and two diseases from igraph import Graph are connected to each other if they # reading weighted edge list from a file have at least one gene in common f = open(‘disease.net.w.txt’) whose mutations are associated f.readline() with both diseases. Disease data and g = Graph.Read_Ncol(f, weights=True, names=True) their associations with genes are g = g.as_undirected(mode=’collapse’) obtained from the Online Mendelian f.close() Inheritance in Man (OMIM), which # setting disease names as node attributes is a comprehensive and regularly f = open(‘supplementary_tableS1.txt’) updated online resource (http://www. f.readline() omim.org). It contains information f.readline() on all known Mendelian diseases and did2name = dict([line.strip().replace(‘”’, ‘’).split(‘\t’)[:2] for line in f]) tens of thousands of genes. Readily f.close() prepared human disease network g.vs[‘disorder’] = [did2name[did] for did in g.vs[‘name’]] data are available for download from supplementary material [2]. The reader may also access the dataset by visiting the diseasome website Listing 2: A Python script to calculate various structural properties of the (http://diseasome.eu). Once one has human disease network. downloaded the network dataset, the human disease network can be print ‘Nodes: %d’ % g.vcount() constructed as shown in Listing 1. We print ‘Edges: %d’ % g.ecount() also use Igraph (http://igraph.org), cmp = g.components(mode=’strong’) a network analysis package suitable print ‘Strong conn. comp: %d’ % len(cmp.sizes()) for explorative analysis of small and print ‘Giant component: %d’ % cmp.giant().vcount() medium-sized networks. Alternatively, print ‘Closed triangles: %d’ % g.as_directed().triad_census().t300 one may want to check SNAP (http:// print ‘Diameter: %d’ % cmp.giant().diameter() snap.stanford.edu), a scalable graph d = sorted([(n.degree(), n[‘disorder’]) for n in g.vs], reverse=True)[:10] mining library that can handle massive print ‘Max deg.:\n %s’ % ‘\n ‘.join(‘%d (%s)’ % (deg, dis) for deg, dis in d[:5]) networks. We visualize our network print ‘Clustering: %3.3f’ % g.transitivity_avglocal_undirected() using Vis.js (http://visjs.org), which is a dynamic, browser-based visualization library allowing manipulation and interaction with the data. Listing 3: A Python script to find network communities with structure based on random walks, label propagation method and the Infomap method. STRUCTURAL PROPERTIES OF THE DISEASE NETWORK We first examine some properties # detecting communities in the HDN based on random walks of the human disease network as rndwlk = cmp.giant().community_walktrap(steps=3).as_clustering() computed by the script in Listing 2. mmbr_rndwlk = rndwlk.membership Of the many diseases in the OMIM print ‘Modularity (random walks): %3.3f’ % rndwlk.modularity compendium (listed at the time of the # detecting communities in the HDN with label propagation-based method Goh et al. paper), 867 diseases have at lblprp = cmp.giant().community_label_propagation() least one connection to other diseases mmbr_lblprp = lblprp.membership and there are 1,527 edges altogether. print ‘Modularity (label propagation): %3.3f’ % lblprp.modularity If every disease would be independent # detecting communities in the HDN using the Infomap method of others in terms of mutated disease infomap = cmp.giant().community_infomap() genes, then the network would mmbr_infomap = infomap.membership fall apart into many single nodes print ‘Modularity (Infomap): %3.3f’ % infomap.modularity corresponding to individual diseases print ‘Codelength (Infomap): %3.3f bits/step’ % infomap.codelength and small disconnected components of few closely related diseases. However,

XRDS • WINTER 2014 • VOL.21 • NO.2 59 the largest connected component average clustering coefficient of 0.61 node labels in a way that each contains 516 diseases, suggesting (http://snap.stanford.edu/data). node adopts the label that most of there are many connections between its neighbors currently have. The diseases that are related at both NETWORK COMMUNITY assumption is that by the end of this a lower and higher level of cellular DETECTION iterative process densely connected organization. Whereas most diseases The discovery of community structure groups of nodes reach a consensus are connected to only few other nodes, is a challenge of great interest, and on a unique label, hence defining there is a small number of highly methods for community detection a community (Listing 3). Another connected disease nodes known as have attracted considerable attention method we apply to the disease hubs, such as colorectal cancer (its across many disciplines [2, 3, 4, 5]. As network is Infomap [3]. This is a rather node degree is 50), breast cancer (it we have just seen, and as concluded by different technique that attempts has 30 links) and gastric cancer (it has many studies [1,2], there is a strong to model information flow between 27 links). indication that human disease network diseases. It uses probability flow Next, we are interested to see if the contains communities of closely of random walks on a network as a network exhibits community structure. interconnected diseases. We shall surrogate of information flow and Communities or modules are densely consider here three approaches to formulates a coding or compression connected subsets of nodes with only search for communities (Listing 3) in problem to efficiently describe random sparse connections between them. our network. walks. Infomap uses computational Many studies [2, 3, 4, 5] discovered Our first approach to organize search to find a partition into communities that corresponded to the disease network is based on network communities that minimizes functional or behavioral units within random walks [5]. The algorithm is the expected description length networks, such as social groups in called Walktrap [5] and builds upon of a random walk. The particular social networks or protein modules the intuition that random walks on partitioning of the disease network in biochemical networks. This implies a network tend to get “trapped” picked up by the Infomap yields that we may be able to gain insights into densely connected parts that description length of 5.7 bits per step into systems, whose operation is less correspond to communities. The (see codelength in Listing 3). well understood, as is in the case of Walktrap algorithm runs short random While these approaches are human disease network, by detecting walks (see steps parameter in Listing relatively efficient, they are not and examining their communities. A 3) to estimate similarities between appropriate in all situations. simple way to estimate the overall nodes and between communities, thus Recent algorithms aim to identify presence of community structure in defining a distance. Distance scores hierarchically nested—as well as a network is to calculate triad census are then used to iteratively merge the overlapping—communities, where an (Listing 2) and compare it to that nodes into communities and to obtain individual can be placed in more than obtained from an appropriately defined a hierarchical community structure. one community or no community at and randomly generated network. Triad An induced hierarchy of network all. There is also a great emphasis on census classifies every triplet of nodes partitions is then scored against efficiency of community detection to one of four possible types of triads modularity, a quality function widely tools as large networks are becoming and for each type reports the number used in many community detection increasingly common in many areas. of triads present in the network. A approaches, to select a partition that triplet consists of three nodes that captures well the community structure References are connected by either zero, one, of the data. Figure 1 shows partition of [1] Barabási, A. L., Gulbahce, N., & Loscalzo, J. Network medicine: a network-based approach to human two, or three links, the latter is known the disease network as was detected disease. Nature Reviews Genetics, 12, 1 (2011), as a closed triad. The human disease by Walktrap. Disease nodes are colored 56–68. network has 1,517 closed triads based on communities to which they [2] Goh, K. I., Cusick, M. E., Valle, D., Childs, B., Vidal, M., & Barabási, A. L. The human disease network. labeled as triplet type “300” [2]. We belong and nodes associated with most Proceedings of the National Academy of Sciences, also calculate clustering coefficient disease genes are labeled. One can 104, 21) (2007), 8685–8690. of the network, which provides an see Walktrap was able to successfully [3] Rosvall, M., & Bergstrom, C. T. Maps of random walks on complex networks reveal community estimate of probability that two group diseases and automatically structure. Proceedings of the National Academy of neighbors of a node are connected. recognize a number of disease classes Sciences, 105, 4 (2008), 1118–1123. [4] Raghavan, U. N., Albert, R., & Kumara, S. Near linear More precisely, clustering coefficient related to cancer, hematological, time algorithm to detect community structures is a ratio of the number of closed muscular, and ophthamological in large-scale networks. Physical Review E, 76, 3 triads to the number of connected diseases, among others. (2007), 036106. [5] Pons, P., & Latapy, M. Computing communities in triplets in the network. The human Our second method finds large networks using random walks. Journal of disease network has a high average communities with a label propagation Graph Algorithms and Applications, 10, 2 (2006), 191–218. clustering coefficient, which is 0.81. algorithm [4]. The algorithm initializes For comparison, a recent dataset of every node with a unique label and friends lists from Facebook has an then in an iterative manner reassigns Copyright held by Owner(s)/Author(s).

60 XRDS • WINTER 2014 • VOL.21 • NO.2 POINTERS ACRONYMS

technology to make access to HEALTH 2.0 healthcare information timely and easy. AMIA American Medical Informatics Association: A non-profit organization, Health informatics are a big part http://www.forbes.com/sites/ founded in 1989 as a result of of the wearable tech boom, but stevenbertoni/2014/05/14/disruptive- the merger of other NGOs. AMIA are increasingly pervasive in the healthcare-start-up-oscar-raises-80- supports the development and tech landscape. According to million-valuation-nears-1-billion/. application of health and biomedical informatics for the improvement of Wikipedia, “[Health informatics] research, teaching, patient care, and deals with the resources, devices, administration in healthcare. and methods required to optimize IN MEMORIAM the acquisition, storage, retrieval, Maybe the person who knows best CDSS Clinical Decision Support and use of information in health and about Health 2.0 is someone who System: It is an expert system that biomedicine.” This is particularly founded and ran the country’s largest links health data with appropriate important as machine learning and provider of managed, medical care. health observations, which in turn data analysis algorithms are able to Morris F. Collen, who co-founded supports doctors, clinicians, and patients with medical decision-making. produce remarkable insights. The Kaiser Permanente in the 1940s, was a future is here, with watches and body champion of computerized medicine., CDW Clinical Data Warehouse (also monitors hoping to predict a heart He designed the first system for Clinical Data Repository): It is a real-time attack before it happens. Here are a automating “multiphasic health database that integrates information few resources to get you started. checkups.” Dr. Collen passed away on from multiple sources to present a —Ashok Rao September 27, 2014 and is a hero to unified, clear medical history of (usually) many in this field. a single patient. Typical sources of data for the CDW include medical test http://www.nytimes.com/2014/10/05/us/ reports, hospital admission, progress HEALTH TECH STARTUPS morris-collen-computerized-medicine- notes, discharge details, etc. pioneer-dies-at-100.html?_r=0 HITECH Act Health Information Predilytics Technology for Economics and Clinical Part of dealing with information is Health Act, USA: Enacted as a part of the HEALTH INFORMATICS ABROAD knowing how to use it well. Sometimes American Recovery and Reinvestment that’s difficult given the scattered Act of 2009, this act aims to promote nature of various feeds, without any Asia Pacific Association the “adoption and meaningful use” of common thread. Predilytics aims health information technology. It also for Medical Informatics addresses various concerns regarding to use data to “answer healthcare’s While improvements in technology privacy and security. most important questions.” Applying are going a long way to systematized preditictive analytics to healthcare healthcare in the U.S., a shocking MEDLINE Medical Literature Analysis may prove promising not only for dearth of basic, computerized and Retrieval System Online (also Predilytics’s bottomline, but for the processes abroad is a critical MEDLARS Online): It is a freely available, healthcare industry as a whole. challenge for many poorer countries. online database, maintained by the U.S. http://www.predilytics.com Rapid advances in mobile National Library of Medicine. It contains life sciences and biomedicine related technology are particularly helpful data in the form of journal citations and in poorer countries where remote “Disruptive Healthcare Start-up abstracts from academic and scientific villages, without access to modern Oscar Raises $80 Million, literature around the world. technology, can obtain and transmit Valuation Nears $1 Billion” health information (including TBI Translational Bioinformatics: By Steven Bertoni critical prescription data) quickly and One of the sub-fields of health Maybe HealthCare.gov is the clearest efficiently, saving lives in the process. informatics, it is a combination of example of the lack of quality APAMI was formed in 1993 as clinical informatics, biostatistics, and technology in health insurance. In molecular bioinformatics. It supports the Asia and Pacific regional branch fact, obtaining information from the storage, analysis, and integration of IMIA (International Medical pretty much any insurer is painfully of massive amounts of biomedical and Informatics Association). While the difficult, bureaucratic, and the genomic data using data mining—and APAMI 2014 Conference has past, be other data-based technologies—to transaction feels like it’s from the sure to check back next year. formulate disease diagnosis strategies, 1950s. Oscar aims to change this by http://www.apami.org health outcome predictions, treatment applying simple, mobile-integrated, therapies, etc.

XRDS • WINTER 2014 • VOL.21 • NO.2 61 end

EVENTS

CONFERENCES SIGAI Career Network and Conference Seventh International Conference on Radisson Hotel & Suites Austin Bioinformatics and Computational Downtown Biology International Conference on Austin, TX Waikiki Beach Marriott Resort & Spa Distributed Computing and January 26, 2015 Honolulu, HI Networking http://sigai.acm.org/cnc March 9-11, 2015 Birla Institute of Technology & http://www.cs.umb.edu/bicob Science, Pilani - K K Birla Goa Campus Goa, India Eighth ACM International Conference January 4-7, 2015 on Web Search and Data Mining Sixth International Conference http://www.icdcn.org Crowne Plaza Shanghai Fudan of Augmented Human Shanghai, China Marina Bay Sands Expo and February 2-6. 2015 Convention Centre Pacific Symposium on Biocomputing http://www.wsdm-conference.org/2015 Singapore, Singapore Fairmont Orchid March 9-11, 2015 Kohala Coast, HI http://asg.sutd.edu.sg/ah2015/home January 4-8, 2015 International Conference http://psb.stanford.edu on Information Systems Security and Privacy Architectural Support ESEO Grande Ecole d’ingénieurs for Programming Languages Keystone Symposia: Precision Genome Angers, France and Operating Systems Engineering and Synthetic Biology February 9-11, 2015 Istanbul, Turkey Big Sky Resort http://www.icissp.org March 14-18, 2015 Big Sky, MT http://asplos15.bilkent.edu.tr January 11-16, 2015 http://www.keystonesymposia.org Sixteenth Workshop on Mobile Computing Systems and Applications 14th International Conference Santa Fe, NM on Modularity International Joint Conference February 12-13, 2015 Colorado State University on Biomedical Engineering Systems http://www.hotmobile.org/2015 Fort Collins, CO and Technologies March 16-19, 2015 Sana Lisboa Hotel http://www.aosd.net/2015 Lisbon, Portugal Third ACM/SIGDA International January 12-15, 2015 Symposium on Field-Programmable http://www.biostec.org Gate Arrays Monterey Conference Center CONTESTS & EVENTS Monterey, CA International Conference on Tangible, February 22-24, 2015 Facebook Hacker Cup Embedded, and Embodied Interaction http://www.eecs.ucf.edu/isfpga Stanford University The Facebook Hacker Cup is a Stanford, CA worldwide computer programming January 15-19, 2015 Second International Conference on competition that takes place each http://www.tei-conf.org/15 Perception and Machine Intelligence year. The contest features algorithmic Saha Institute of Nuclear Physics problems and several rounds of Kolkata, India challenges in order to take home the International Conference on February 26-27, 2015 coveted Hacker Cup Trophy as well as High-Performance and Embedded http://www.permin.in/permin15 cash prizes. The qualification period Architectures and Compilers begins in early January 2015 culminating Forum Centre at Amsterdam RAI with the final round at Facebook’s 46th ACM Technical Symposium on Amsterdam, Holland headquarters in Menlo Park, CA. Computer Science Education January 19-21, 2015 https://www.facebook.com/hackercup/ http://www.hipeac.net/2015/amsterdam Kansas City Convention Center Kansas City, MO March 4-7, 2015 Launch48 Weekend Accelerators http://sigcse2015.sigcse.org Throughout November 2014 to February 2015, Oxygen Startups will be featuring Launch48 Weekends

62 XRDS • WINTER 2014 • VOL.21 • NO.2 FEATURED EVENT

Accelerators. In this program, a GPA of 3.0 or better participants’ pitch, build, and launch Benefits: $2,000 - $8,000 a startup within 48 hours. Members Explanation: In line with its mission to will work with mentors and peers advance the study of cost-engineering to develop business and customer and cost-management through the plans. Other Oxygen Startups “total cost management” process, programs include accelerators to AACE International awards hard- help existing, but early, startups find working students in related programs. guidance and funding. http://oxygenstartups.com/ Spencer Foundation Small Research Grant CS Games Website: http://www.spencer.org/ CS Games is an annual computer content.cfm/budgets-50000-or-less science competition that will take Deadline: Multiple Deadlines place next year at Sherbrooke throughout 2015; February 5th, June University in Canada from March 13 2nd, and August 20th. to 16. The competition is open to any Eligibility: Grants are awarded for university and each university can research related to any of Spencer’s Third International Conference on register teams of six to 10 members. areas of inquiry: education and Bioinformatics and Computational Within the competition, there are social opportunity; organizational Biology (ICBCB 2015) events focusing on various aspects of learning; purposes and values of Hong Kong, China computer science such as artificial education; and teaching, learning, March 12-13, 2015 intelligence, algorithms, gaming, and instructional resources. scavenger hunts, and more. Trophies Benefits: Project budgets up to $50,000 Bioinformatics uses principles and awards will be presented at the Explanation: Established by Lyle M. of computer science and closing ceremony. Spencer, the Spencer Foundation aims mathematics to interpret and http://csgames.org/ to investigate ways education can be study data in biological research. improved around the world. This has allowed researchers and jQuery UK 2015 practitioners to make tremendous On March 6, 2015, a conference will be Department of Energy Computational strides, especially in the fields taking place in the UK for the jQuery Science Graduate Fellowship of genetics and genomics. As framework. jQuery is a very popular Website: https://www.krellinst.org/csgf/ research problems continue to frontend JavaScript framework used how-apply become harder and more complex, for interactive websites, animations, Deadline: January 2015 developments in computing and graphical interfaces. The event Eligibility: U.S .citizens and will be critical to the future of features several industry speakers permanent residents planning study biological research. from well-known companies such as toward a Ph.D at an accredited U.S. In March, the South Asia Google, Twitter, and GitHub. university, and have not yet started Institute of Science and http://jqueryuk.com/2015/ their second year of doctoral studies. Engineering (SAISE) and WIT Benefits: $36,000 stipend, tuition, and Press will host ICBCB 2015 in fees for up to four years. An allowance Hong Kong. The conference welcomes the latest research GRANTS AND SCHOLARSHIPS for a computer workstation is added. Explanation: The Computations in bioinformatics and bridges Science Graduate Fellowship is meant the gap between academia and AACE International Scholarship to encourage interdisciplinary work industry. Topics of interest Website: http://www.aacei.org/awards/ and collaboration in the field of will include sequence analysis, scholarships/overview.shtml computational science. The fellowship biological data analysis, and Deadline: Feburary 7, 2015 includes a practicum at a Department genetics. Attendees of the Eligibility: Full-time students in a of Energy laboratory. A conference is conference will be able to step field related to cost-engineering or held each summer for recipients. into Hong Kong, which offers cost-management, such as electrical a bustling environment and engineering or computer science, with a beautiful skyline. Learn more at http://www.icbcb. org/. ——Rohit Goyal Photo by Leung Cho Pan

XRDS • WINTER 2014 • VOL.21 • NO.2 63 end

BEMUSEMENT From A to B and Back Again

Puzzles: Thomas the Truant Thomas has missed an excessive number of days of school, so he must meet with Principal Davis. Mr. Davis asks him, “Why on Earth have you missed so many days?” Thomas replies: “There just isn’t enough time for school. I need 8 hours of sleep a day, which adds up to about 122 days a year. Weekends off is 104 days a year. Summer vacation is about 60 days. If I spend about an hour on each meal, that’s 3 hours a day or 45 days a year. I need at least 2 hours of exercise and relaxation time each day to stay physically and mentally fit, adding another 30 days. Add all of that up and you get about 361 days. That only leaves 4 days for school.” The principal knows Thomas is full of it, but can’t figure out why. Why is Thomas wrong? Source: http://goodriddlesnow.com/riddles/ view/743

Post-Bachelor’s Disorder Find the solution at: http://xrds.acm. org/bemusement/2014.cfm

SUBMIT A PUZZLE Can you do better? Bemusements would like your puzzles and mathematical games (but not Sudoku). Contact [email protected] to submit yours! PhD Comics ©JorgePhD Cham

64 XRDS • WINTER 2014 • VOL.21 • NO.2 acm STUDENT MEMBERSHIP APPLICATION Join ACM online: www.acm.org/joinacm CODE: CRSRDS Name Please print clearly INSTRUCTIONS

Address Carefully complete this application and return with payment by mail or fax to ACM. You must City State/Province Postal code/Zip be a full-time student to qualify for student rates.

Country E-mail address CONTACT ACM

Area code & Daytime phone Mobile phone Member number, if applicable phone: 800-342-6626 M E M B E R S H I P B E N E F I T S A N D O P T I O N S (US & Canada) • Free software and courseware through the ACM • ACM e-news digest TechNews (thrice weekly) +1-212-626-0500 Academic Initiative • ACM online newsletter MemberNet (monthly) (Global) ® • Free e-mentoring services from MentorNet • Student Quick Takes, ACM student e-newsletter (quarterly) hours: 8:30am–4:30pm US Eastern Time • Electronic subscriptions to Communications of the ACM • Free "acm.org" email forwarding address plus filtering and XRDS: Crossroads magazines through Postini fax: +1-212-944-1318 • Online courses, online books and videos • Option to subscribe to the full ACM Digital Library email: [email protected] • ACM's CareerNews (twice monthly) • Discounts on ACM publications and conferences, mail: Association for Computing valuable products and services, and more PLEASE CHOOSE ONE: Machinery, Inc. General Post Office J Student Membership: $19 (USD) P.O. Box 30777 J Student Membership PLUS Digital Library: $42 (USD) New York, NY 10087-0777 J Student Membership PLUS Print CACM Magazine: $42 (USD) J Student Membership w/Digital Library PLUS Print CACM Magazine: $62 (USD) For immediate processing, FAX this application to +1-212-944-1318. P U B L I C AT I O N S Please check Check the appropriate box and calculate Issues amount due on reverse. per year Code Member Rate Air Rate* PAYMENT INFORMATION • ACM Inroads 4 178 $16 Ë $58 Ë • Communications of the ACM 12 101 $25 Ë $58 Ë Payment must accompany application • Computers in Entertainment (online only) 4 247 $43 Ë N/A $ Computing Reviews 12 104 $55 Ë $39 Ë Member dues ($19, $42, or $62) • Computing Surveys 4 103 $36 Ë $32 Ë To have Communications of the ACM Evolutionary Computation (MIT Press) 4 177 $32 Ë $30 Ë sent to you via Expedited Air Service, • interactions, new visions of human-computer interaction 6 123 $22 Ë $35 Ë (included in SIGCHI membership) add $58 here (for residents outside of • Int’l Journal of Network Management (online only) (Wiley) 6 136 $92 Ë $30 Ë North America only). $ Int’l Journal on Very Large Databases 4 148 $83 Ë $30 Ë • Journal of Educational Resources in Computing (see TOCE) N/A N/A N/A N/A Publications $ • Journal of Experimental Algorithmics (online only) 12 129 $31 Ë N/A • Journal of Personal and Ubiquitous Computing 6 144 $65 Ë $30 Ë Total amount due $ • Journal of the ACM 6 102 $55 Ë $58 Ë • Journal on Computing and Cultural Heritage 4 173 $49 Ë $23 Ë Check or money order (make payable to ACM, Ë Ë • Journal on Data and Information Quality 4 171 $49 $25 Inc. in U.S. dollars or equivalent in foreign currency) • Journal on Emerging Technologies in Computing Systems 4 154 $42 Ë $23 Ë • Linux Journal (SSC) 12 137 $31 Ë $33 Ë • Mobile Networks and Applications 6 130 $72 Ë $29 Ë JVisa/Mastercard J American Express • Wireless Networks 4 125 $72 Ë $29 Ë • XRDS (included with membership) 4 XRoads $39 Ë N/A Transactions on: Card number Exp. date • Accessible Computing 4 174 $49 Ë $24 Ë • Algorithms 4 151 $52 Ë $23 Ë • Applied Perception 4 145 $43 Ë $23 Ë Signature Ë Ë • Architecture & Code Optimization 4 146 $43 $23 Member dues, subscriptions, and optional contributions • Asian Language Information Processing 4 138 $39 Ë $23 Ë Ë Ë are tax deductible under certain circumstances. Please • Autonomous and Adaptive Systems 4 158 $41 $23 consult with your tax advisor. • Computational Biology and Bioinformatics 4 149 $20 Ë $49 Ë • Computer-Human Interaction 4 119 $43 Ë $25 Ë • Computational Logic 4 135 $44 Ë $25 Ë EDUCATION • Computation Theory 8 176 $49 Ë $32 Ë • Computer Systems 4 114 $47 Ë $25 Ë • Computing Education (formerly JERIC) 277 $25 Ë N/A • Database Systems 4 109 $46 Ë $25 Ë Name of School • Design Automation of Electronic Systems 4 128 $43 Ë $25 Ë • Economics and Computation 4 192 $49 Ë $23 Ë Please check one: Ë High School (Pre-college, Secondary • Embedded Computing Systems 4 142 $44 Ë $23 Ë School) College: Ë Freshman/1st yr. Ë Sophomore/2nd yr. • Graphics 4 112 $51 Ë $25 Ë Ë Junior/3rd yr. Ë Senior/4th yr. Graduate Student: Ë • Information and System Security 4 134 $44 Ë $23 Ë Masters Program Ë Doctorate Program Ë Postdoctoral • Information Systems 4 113 $47 Ë $25 Ë • Intelligent Systems and Technology 4 179 $46 Ë $72 Ë Program Ë Non-Traditional Student • Interactive Intelligent Systems 4 191 $49 Ë $76 Ë • Internet Technology 4 140 $42 Ë $23 Ë • Knowledge Discovery From Data 4 170 $50 Ë $23 Ë Major Expected mo./yr. of grad. • Management Information Systems 4 190 $47 Ë $22 Ë Ë Ë • Mathematical Software 4 108 $47 $25 Age Range: Ë 17 & under Ë 18-21 Ë 22-25 Ë 26-30 • Modeling and Computer Simulation 4 116 $51 Ë $25 Ë • Multimedia Computing, Communications, and Applications 4 156 $42 Ë $23 Ë Ë 31-35 Ë 36-40 Ë 41-45 Ë 46-50 Ë 51-55 Ë 56-59 Ë 60+ • Networking 6 118 $29 Ë $52 Ë • Programming Languages & Systems 6 110 $59 Ë $32 Ë Do you belong to an ACM Student Chapter? Ë Yes Ë No • Reconfigurable Technology & Systems 4 172 $49 Ë $24 Ë • Sensor Networks 4 155 $42 Ë $23 Ë I attest that the information given is correct and that I will • Software Engineering and Methodology 4 115 $43 Ë $25 Ë abide by the ACM Code of Ethics. I understand that my • Speech and Language Processing (online only) 4 253 $33 Ë N/A membership is non transferable. • Storage 4 157 $42 Ë $23 Ë • Web 4 159 $41 Ë $23 Ë Marked • are available in the ACM Digital Library * Check here to have publications delivered via Expedited Air Service. Signature For residents outside North America only. PUBLICATION SUBTOTAL: CAREERS at the NATIONAL SECURITY AGENCY

EXTRAORDINARYWORK Inside our walls, you will find the most extraordinary people doing Computer/Electrical Engineering the most extraordinary work. Yet it’s not just finite field theory, Computer Science quantum computing or RF engineering. It’s not just discrete Cybersecurity mathematics or graph analytics. Information Assurance Mathematics Foreign Language It’s all of these and more – rolled up into an organization that leads Intelligence Analysis the world in signals intelligence and information assurance. Cryptanalysis Inside our walls you will find extraordinary people, doing Signals Analysis extraordinary work, for an extraordinary cause: Business Management Finance & Accounting The safety and security of the United States of America. Paid Internships, Scholarships & Co-op

U.S. citizenship is required. NSA is an Equal Opportunity Employer. All applicants for employment are considered without APPLY TODAY regard to race, color, religion, sex, national origin, age, marital status, disability, sexual orientation, or status as a parent.

Search NSA to Download WHERE INTELLIGENCE GOES TO WORK®

14CNS-06_7.875x10.5(live7x10).indd 1 10/7/14 12:53 PM