Association for the A A Advancement of and B u l l e t i n P& P

From the Editor President’s Column

Welcome to another lively discus- In the latter part of 2019 I started seeing colleagues using digital signatures sion, this one led by Dr. Phoebe Frie- on administrative paperwork and investigated setting up a digital signature for sen, who is tackling the issue of digital myself, but it seemed like it would take more work than it was worth. By late psychiatry. This is a very broad area March 2020 we were in lockdown in response to COVID-19 and I found myself that leads in many directions. Phoebe at home. As the associate dean in my college I had to sign anywhere from three focuses in on the ability of digital tech- to seven grade change forms a day nearly every day of the week. I printed them nology to generate huge data sets that out, signed them, and emailed a scanned copy to the registrar’s office. By early allow prediction and diagnosis, all with May I decided I had to implement the digital signature option and it took me no awareness on the part of the subjects maybe 10 minutes to set one up. I do not see myself signing administrative forms of the research. She questions both the by hand ever again if I have a choice. Nor I pull a credit card out of my wal- utility and the of such research. let if a store takes Apple Pay or write a check to deposit money into my mother’s As you will see from the commentaries, bank account rather making a direct transfer using Zelle. her analysis has stimulated a lot of dis- Technology forces itself into our lives and changes our worlds. We some- cussion. times adopt it out of openness and curiosity, or because those we work with have As is now our routine, this issue of adopted it, or we may be shamed into it, or in the case of COVID-19, forced into the Bulletin will be accompanied by a it by the situation. In this issue of the AAPP Bulletin, Phoebe Friesen and several target article and call for commentaries of our colleagues raise important considerations about yet another potential tech- for the next Bulletin issue. Our author nological development – informatics-assisted diagnosis and is Dr. Louis Charland, and his piece is under the name of digital psychiatry. titled Consent and Capacity in the Age Rather than summarizing the many important concerns and reservations of the Opioid Epidemic: The Drug about digital psychiatry articulated in this issue, I let me briefly ponder a world Dealer’s Point of View. that introduces new technology at a dizzying pace. Meanwhile, my own commentary If we could bring someone from the late 19th century into our current world, on Phoebe Friesen’s article. what might they think of it and of us? Quite likely, the farther back through the centuries we reach, the harder it would be for our time traveling ancestor to adapt to and become part of society. Assuming our society continues to progress along the same trajectory it has been on since the Renaissance, the same would be true Data and People of us if we could be transported to late 22nd or 23rd century. More proximately, we should think about adapting to the society that will Phoebe Friesen has eloquently exist 25 years into the future. Many of us will end up there one day no matter introduced us to a language game than what. includes data sets, machine learning, I remember how 25 year ago one of the upper level administrators at my algorithms, digitalization, AI, and natu- university took pride in not having an email account. During that same time, I ral language processing. And she has had colleagues who insisted on computing an analysis of variance by hand rather used this cluster to examine a new type than using SPSS or SAS. I thought they were ridiculous and never wanted to be of research that rates high on diagnosis ‘that person” and still don’t, but neither do I tweet or use TikToc. Last year I and outcome prediction, but with the found myself trying to explain to my incredulous younger brother why I was still scientific and ethical challenges posed buying CDs rather than paying ten dollars a month for a streaming music service. by such research. In this brief commen- It is stunning, but the amount of time it goes from an early adopter to an tary I wish to focus on one aspect, clus- (Continued on page 12) ter, data, and to do this I will include another feature of the new research protocol, metrics. Working with data the study and manipulation of huge data sets. And the manipulation involves always involves measurement, and measurements. That is what happens with the kind of research Friesen is describ- measurement in research always in- ing. When a suicide prediction model can predict with 70% to 85% accuracy who volves data. is at risk for suicide, the model is measuring the accuracy of the prediction. And Data are pieces or units of infor- finally, as Friesen points out, all this manipulation and measurement can be done mation, and such pieces may exist in by machines and algorithms. many different formats. Whatever form Friesen contends that the research findings in prediction do not lead in an ob- they start as, the that they can be vious way to patient care. It’s not hard, however, to invoke examples of data meas transformed into digital units allows for urements where they do accomplish something meaningful and important. A (continued on page 12)

Volume 27, Number 1 2020

Digital Psychiatry: Promises behaviors will change significantly many cases, it does not get us any clos- after birth, signalling a risk of post- er to understanding causation, which is and Perils partum depression [2]. In another what is likely to support the develop- example, using social media data ment of novel interventions [5]. Argua- Phoebe Friesen, Ph.D. from 547 users who had either post- bly, the gap between prediction and [email protected] ed publicly about a past suicide causation is even more significant in attempt or donated their data to digital psychiatry than the one we saw Hype is not new to the field of OurDataHelps.org, Coppersmith et in previous waves of psychiatric re- psychiatry. Following waves of enthu- al. built a suicide prediction model search enthusiasm involving brains and siasm related to neuroscience and ge- that demonstrated a 70% to 85% biomarkers; since the body is no longer netic research, psychiatry is now look- true positive rate (depending on involved, putative targets for interven- ing towards digital technologies to how false positives were weighted) tion seem even more distant. Unfortu- help relieve suffering. Machine learn- [3]. Similarly, Corcoran et al. re- nately, in the digital realm, it’s hard to ing techniques, combined with enor- ported the use of linguistic analysis see how predictions based on speech mous data sets, have opened up a new to predict the onset and severity of patterns, social media use, and texting frontier, in which algorithms can pre- in a sample of high-risk behaviors are going to translate into dict and identify the diagnoses of pa- youth. Using an automated machine novel interventions. tients with tremendous accuracy and learning speech classifier, the au- Similarly, gaps related to resources efficiency. It may seem obvious that thors were able to identify which loom large in predictive analytics relat- this development is worth celebrating. features of speech best predict the ed to . After a recent pa- However, while there is no shortage of onset of psychosis and use the clas- per reported the development of a mod- predictive success in this domain, it is sifier to predict the onset of psycho- el that could predict loneliness based not always clear whether and how sis with 83% accuracy and distin- on social media data, the study’s lead these technologies are contributing to guish between patients and healthy author, Sharath Chandra Guntuku, was patient wellbeing. Furthermore, these controls will 72% accuracy [4]. quoted as saying that this identification techniques now allow for psychiatric While such predictions are un- system, combined with early interven- research to take place outside of exist- doubtedly impressive, more than tions, could have “long-lasting effects ing structures of research ethics gov- mere predictive power is often on public health” [6, 7]. The distance ernance. Predictions and diagnoses promised in these manuscripts. In between identifying lonely Facebook can now be made about unsuspecting their abstract, Corcoran et al. claim users and offering effective interven- users who are not aware that their data that “automated linguistic analysis tions to those users is a significant one is being used for research purposes, can be a powerful tool for diagnosis though. The ‘epidemic of loneliness’ and yet few protections are in place. and treatment across neuropsychia- has been under discussion for decades AI-driven techniques, including try” and that these findings can help now and it’s far from clear that we machine learning, natural language to “identify linguistic targets for have the tools or resources to address it processing, and predictive analytics, remediation and preventive inter- [8]. Pinpointing which Facebook users are changing the of health re- vention” [4]. It is unclear precisely are most lonely may help us to under- search. These technologies, combined what is being suggested by the au- stand who, and when, people are most with enormous and widely available thors here in terms of preventative lonely, and perhaps where to direct our data sets, now allow for medical re- intervention though. The features scarce resources, but it is unlikely to search to take place in new settings (in picked out by the classifier as pre- solve the problem. online forums, on mobile phones) us- dictive of the onset of psychosis Much of this research takes place ing novel data sets (Twitter posts, included the use of possessive pro- online and uses publicly available data Google searches) and to make predic- nouns as well as decreases and vari- sets, creating challenges for the sys- tions well in advance of medical ance in semantic coherence. Are tems of research ethics governance that events. This new frontier has made its these the “linguistic targets” the were established in the mid-twentieth way into all domains of medicine, but authors refer to? It seems very un- century, when health research looked psychiatry, in particular, is diving in likely that by teaching patients to very different. Importantly, these new with both feet. A recent systematic correct these language patterns, by technologies allow for the creation of review of research that utilized ma- using more possessive pronouns ‘emergent medical data’ from non- chine learning techniques to analyze (e.g., her, his, mine, our) for exam- medical data sets without the awareness online personal health data by Yin et ple, that they would be less likely to of those whose data is being used [9]. al. found that mental health was the develop psychosis. Unfortunately, While the users on Twitter that De most common investigational target in nothing more is said within the Choudhury et al. collected data from their sample, appearing in 39 of 103 manuscript about how the research had publicly announced their new papers [1]. might contribute to the develop- motherhood, they were never told that In many of these papers, the pre- ment of such preventative interven- their data was being collected for re- dictive results are impressive, if not tions. search purposes and that they were incredible. For example, De This example points to a much being assessed in relation to their risk Choudhury et al. analysed Twitter larger issue that arises frequently of post-partum depression. Because the posts of 376 new mothers and devel- within digital psychiatry: the gap data the researchers were interested in, oped a classifier that can predict, with between making a prediction and that which represented the social en- 71% accuracy, which mothers’ online identifying causal pathways. Pre- gagement, emotions, social networks, diction is not causation, and in and linguistic styles of the new moth-

2 Volume 27, Number 1 2020 ers, were publicly available online, 2018. 10: p. 1178222618792860- tion, prediction is not treatment, and there was no need to ask for consent. 1178222618792860. data mining is not ethically innocuous. In their paper, De Choudhury et To this list of important perils, I al. acknowledge that “people may be 4.Corcoran, C.M., et al., Prediction would like to suggest adding another: uncomfortable with others performing of psychosis across protocols and machine learning is not a comprehen- and sharing these predictions”, while risk cohorts using automated lan- sive theory of mind. By this I mean also pointing out that the authors guage analysis. World Psychiatry, that, while we are tempted to move “chose Twitter because it is public and 2018. 17(1): p. 67-75. from considering machine learning as a provides a longitudinal record of the description of how we learn some events, thoughts, and emotions experi- 5. Kleinberg, J., et al., Prediction things, to its being a description of enced in daily life” [2]. The creation policy problems. American Eco- how we learn anything, and, therefore, of medical data about unsuspecting nomic Review, 2015. 105(5): p. of what knowing is, or even (a mathe- and nonconsenting users is especially 491-95. matical description) of what we know, worrisome in relation to psychiatric these moves we should not make. research, where the medical data being 6.Guntuku, S.C., et al., Studying It seems natural to apply associa- created is stigmatized and can intro- expressions of loneliness in indi- tionist theories, however vaguely for- duce significant risks. Despite the in- viduals using twitter: an observa- mulated, including machine learning, creases in risk that arise when psychi- tional study. BMJ Open, 2019. 9 very widely. For example, in an intro- atric data is created about unsuspect- (11): p. e030355. duction to machine learning Alpaydin ing users, De Choudhury and col- reminds us in completely general terms leagues were not required to submit 7. Tweets from Twitter users could that, “When we learn the best strategy their research protocol for ethics re- predict loneliness, in Penn To- in a certain situation that is view. As the research was conducted day. 2019. stored in our brain, and when the situa- by employees of Microsoft, federal tion arises again – when we recognize regulations requiring ethics review, as 8. Killeen, C., Loneliness: an (‘cognize’ means to know) the situa- well as legal protections for health epidemic in modern society. tion – we recall the suitable strategy data, do not apply [2, 10]. Even if the Journal of Advanced Nursing, and act accordingly.” (1) The scope of same project had been pursued within 1998. 28(4): p. 762-770. scientific applications for machine an academic setting, however, it learning and other associationist ap- would have likely been exempt from 9. Marks, M., Tech companies dan- proaches has been breathtaking – lan- ethics review, because such ‘publicly gerous practice: Using AI to infer guage acquisition and reading, vision available’ data does not qualify as hidden health data, in STAT and facial recognition, learning and ‘human subjects’ research [11]. News. 2019. problem-solving, memory and motor While the predictive promise of action (2) – suggesting that it may be computational psychiatry is real, it is 10. Health Insurance Portability universally appropriate. In psychiatry, worth keeping in mind that there is no and Accountability Act (HIPAA), as well, it has recently been proposed golden road from prediction to treat- E.B.S.A. U.S. Dept. of Labor, that all effective for all ment, and much of this research is Editor. 2004: Washington, D.C. forms of psychopathology (that is, taking place outside of current protec- therapy which is ‘transdiagnostic’) can tions. So, along with the enthusiasm, 11. Metcalf, J. and K. Crawford, be explained in terms of cognitive- we might want to keep on hand a little Where are Human Subjects in behavioral processes – which are es- pile of salt. Big Data Research? The Emerg- sentially associationist. (3) And Witt- ing Ethics Divide. The Emerg- genstein appears to be addressing this References ing Ethics Divide (May 14, temptation to generalize an associa- 2016). Big Data and Society, tionist model of mind when he opens 1. Yin, Z., L.M. Sulieman, and B.A. Spring, 2016. Malin, A systematic literature re- Philosophical Investigations with St. Augustine’s version of associationist view of machine learning in online *** personal health data. Journal of the language learning and his own charac- American Medical Informatics As- terization of it: “These words, it seems sociation, 2019. 26(6): p. 561-576. to me, give us a particular picture of Another Peril: Machine the essence of human language.” (4) 2. De Choudhury, M., S. Counts, and Learning as a Comprehen- But as appealing as the idea of a E. Horvitz. Predicting postpartum sive Theory comprehensive machine learning or changes in emotion and behavior associationist model of mind may be, via social media. in Proceedings of there are causes for concern, if not Paul Lieberman, M.D. alarm, because such a model would the SIGCHI conference on human [email protected] factors in computing systems. 2013. pass by or leave out much of what we ordinarily think about when we think Dr. Friesen is right to warn us about our mental lives. For example, if 3. Coppersmith, G., et al., Natural of the perils of over-hyping ma- Language Processing of Social Me- thoughts were neural connections, then chine learning in psychiatry. We thinking would be a brain process and dia as Screening for Suicide Risk. are apt to forget that, as she re- Biomedical informatics insights, the content of thoughts in mind, their minds us, correlation is not causa- presentation if you like in conscious-

3 Volume 27, Number 1 2020

ness, would have no causally explan- examples of what Wittgenstein fa- including machine learning, must be atory role (5), for example, providing mously referred to as our ‘forms of situated within a much broader uni- reasons for or for action. Not life’ which all share the properties verse of thought, talk and practice. only would machine learning be uni- of being holistic (terms and propo- These divergent views about ma- versal, we would be nothing more sitions having meaning only within chine learning and other associationist than learning machines. The model webs or networks of beliefs and approaches – one seeing them as fun- could also lead to or skep- practices), normative (practices, damental and comprehensive, the oth- ticism. If thoughts were neural con- including using language, requiring er as specialized and partial – are apt nections and thinking were a brain standards which we accept as their to have divergent effects on the peo- process, why should I believe that correct applications and implica- ple, including , who hold my thoughts are true – since they are tions), and externalist (contents of them – as any models or pictures do. only associations in my mind, or thoughts being identified and con- Like psychoanalysts and psychophar- weighted connections in my brain? stituted by objects, events, proper- macologists, cognitive-behavioral psy- And why should you believe some- ties and behaviors which are public- chiatrists are also susceptible to self- thing I tell you is true because I be- ly observable). imposed limits on the questions they lieve it (6), since you also know that From these considerations, it ask and the answers they suggest. For my belief is nothing but neural con- seems we should conclude that ma- example, modifying weightings, nections? chine learning or other association- whether in fact or in the imagination But we don’t have to accept ist models of mind are a specialized of a cognitive-behavioral therapist these unwelcome conclusions be- subset of a much more general and treating someone who feels anxious or cause machine learning is not the diverse set of formulations, and that in a learning machine which models an comprehensive, universally applica- it is this larger, more diverse uni- anxious patient simply passes by a ble theory of mind that some might verse of our multiple, shared forms number of questions which can be imagine. of life which is in fact the more asked. Is there something about this Machine learning or other asso- comprehensive model of mind person (what his are like ciationist approaches can be mean- which we searched for in machine or how he thinks) in this situation (of ingful and empirically testable only learning or other associationist uncertainty or exposure) which brings within larger networks of beliefs and models, which turned out to be only on or contributes to anxiety? Why practices. For example, in order to specialized subsets.1 does his being here now cause a symp- determine what a machine has If this is correct, then a psychi- tom (an inside that some- ‘learned,’ or to test and confirm that atrist should accept that while ma- thing is wrong) as opposed, for exam- an associationist model is true, one chine learning or other association- ple, to a behavior, and might that must correlate inputs and outputs and ist approaches may have very useful ‘inner-ness’ be telling us something also have knowledge of the structure applications in her field – as ma- important about what anxiety, or any and functioning (program) of the chine learning may have for high- symptom is (a retreat or adaptation, for intervening machine or brain – all of lighting vulnerable groups and po- example)? Does it matter, or is it just which must be identified and meas- tential treatments and as CBT in its coincidence, that anxiety phenomeno- ured which, in turn, occurs within a various forms has for ameliorating logically resembles fear (and anger)? much broader set of processes and distress – these approaches are not And why is it something that has those practices: reliable recognition, label- comprehensive and other useful resemblances and not others ing and correlation with other obser- means of understanding, explaining (depression-connected to-grief, for vations and with numbers, by multi- and treating psychiatric conditions example) that this person experiences ple observers, using specified proto- may exist (although whether they here and now? Why do we feel that cols, calculations made and repeata- do exist is an empirical matter). anxiety normatively, usually, has a ble using procedures appropriate to As obvious as this conclusion psychological explanation (seems to the task. These further rely upon the may be, it is often lost sight of in allow for ‘a certain kind of ‘why ques- capacities of people, suitably trained, psychiatry which, as Dr. Friesen tion”’ (7)) when feeling pain or cold to reach agreement which may de- notes, is prone to hype and, we do not? Why is anxiety embarrassing pend, in turn, on their experience, might say, a weakness for sectarian- or stigmatizing? We don’t for the most selection, education, motivation, ism. The predictive successes of part know if these questions have any temperament, intellectual endow- machine learning, the treatment cash out in clinical psychiatry but psy- ment, physical health, economic se- successes of cognitive-behavioral chiatrists may reasonably wonder if curity, and so on. and a number of they might. Ignoring them altogether Some of these beliefs and prac- discoveries in neural (for in enthrallment to over-hyped machine tices are highly specialized and re- example, of processes which inhibit learning and other associationist theo- quire extensive formal education (for or potentiate neural activity at the ries is a real peril for us. example being able to program or cellular or network level) are col- use statistical methods), while others lectively reinforcing of the notion Note are also parts of and necessary for that a fundamentally associationist ordinary, everyday life (such as be- model of mind may be a compre- 1. At this point, one might think that, ing able to see and remember relia- hensive one. Dr. Friesen’s warning if a machine learning or associationist bly, in agreement with the relevantly is that, rather than being compre- model is true, it would apply to all similar judgments of others). All are hensive, associationist approaches, our thoughts, talk and practices, so

4 Volume 27, Number 1 2020

the argument we’re making seems Digital Selves and Bodied ness (Csillag et al. 2016, 540). And it is flawed. But it should be recalled, first, in Psychiatry: within FEP that excitement for digital technologies has grown in a dramatic that we can’t learn or know that the More Perils? application is universal without al- fashion, related to the hopes of approxi- mating cure vis a vis early intervention. ready knowing and participating in Suze Berkhout, MD, PhD, Digital platforms have been lauded our ordinary forms of life and, sec- FRCPC here for their role in self-management ond, that we can’t even formulate or [email protected] as well as their ability to generate bio- mean the thought (that all learning is metric data and digitized “phenotypes” machine-like or associationist) with- It was with great enthusiasm that may accomplish a much sought- out a prior acceptance of our ordinary that I read Dr. Friesen’s sharp and after goal of understanding heterogene- ity in clinical presentations. Much of forms of life, since the terms in which insightful piece on the promises and this is based on a collection of research the thought is formulated have mean- perils of digital psychiatry. The no- tion that a digitized psychiatry will practices termed “ecological momen- ings only within such webs of belief modernize the field to respond to tary analysis” (EMA) or “experience and practice. challenges in diagnosis, prognosis, sampling method” (ESM), which can and intervention holds sway across be likened to one another in light of numerous domains of psychiatric their use of digital platforms medicine. As Friesen explains, much (smartphone apps predominantly) to References of this is directed toward the promis- track behaviour and phenomenological es of both prediction and preven- experiences (Ben-Zeev et al. 2014; Bell 1. Alpaydin E. Machine Learning. tion—whether in relation to perinatal et al. 2018; Firth and Torous 2015). Cambridge, MA, The MIT Press, depression, suicidal behaviour, or If we turn to considering the lan- 2016, p. 18. psychosis. It is the latter field—and guage and metaphors used to depict the area of first episode psychosis digital care tools in FEP, something (FEP) in particular—that I want to interesting is revealed. There is a strik- 2.Posner, ML (editor). Foundations of ing overlap between how digital tools Cognitive Science. Cambridge, MA, take a moment to reflect on, as a way to extend the cogent argument are conceptualized, and the imagery The MIT Press, 1989. that Dr. Friesen puts forward. that shapes FEP and the early interven- Friesen aptly points to a number tion itself. Within early in- 3.Caspi, A and Moffitt, TE. All for of underlying issues within the tervention in psychosis, hope is in- one and one for all: mental disorders claims of digital psychiatry: (1) the dexed through the temporality of inter- in one dimension. American Journal gap between prediction and causa- ventions. The framing of biological of Psychiatry 175(9):831-844, 2018. tion (presumably an important one to treatments in FEP, for instance, is one address if intervention is the end of a medical “cure” (or near-cure)— 4.Wittgenstein, L. Philosophical Inves- goal); (2) the distance between con- intervention in FEP is frequently re- ferred to in editorials, scientific articles, tigations, Third Edition, New York, ceptualizing a broad public issue and how one best directs limited re- and infographics as a new era, a bridge The Macmillan Company, 1958, sources to address it; (3) the ethical to the future; early intervention as it para.1. challenges in using “emergent medi- relates to FEP additionally links hope cal data,” i.e. clinical information to novelty, knowledge, and innovation 5.Kim.J. Supervenience and Mind. generated from large, non-medical (McGorry et al. 2015; Saraceno 2007). Cambridge, Cambridge University data sets for which consent is not Similarly, digital care tools are a “new Press, 1993. required. Most (if not all) of these hope” (Torous et al. 2019); innovation concerns likewise apply to digital and novelty are predominant themes. 6.Moran, T. Getting told and being technologies in FEP. That said, the The temporality of the early interven- scope of ethical concerns regarding tion paradigm as well as of digital care believed. Philosopher’s Imprint 5:1- tools attaches hope and optimism to the 29, 2005. digital technologies in psychiatry can be extended even further in the rational deployment of scientific . field of FEP. In both instances, we hear of a futurity 7.Anscombe, GEM. Intention. Pro- FEP refers generally to an early that casts mental disability as an obsta- ceedings of the Aristotelian Society point in time in the diagnosis and cle to the arc of progress (Berkhout 57:321-332, 1957. management of psychotic illnesses 2018). This enlightenment-laden pro- stemming from a range of potential gress narrative is part of what Allison *** causes (Breitborde et al. 2009) and Kafer calls the curative imaginary: an as a clinical organizing , FEP understanding of disability that sees is structured by what is termed the medical intervention as unquestionable; “early intervention” paradigm. This expected and assumed, any other way paradigm holds that the ability to of living is unimaginable (Kafer 2013, prevent or reduce morbidity (in this 27-28). And, to Friesen’s point about case, from psychosis) is best accom- the attention of digital psychiatry as plished through the provision of being focused on prediction as well as treatment early in the course of ill- prevention, prognostication is likewise part of a futurity that demarcates disa-

5 Volume 27, Number 1 2020 bled bodyminds as unquestionably in and case illustration.” Internet Inter- Prediction Without Explana- need of biomedical intervention ventions. 14: 18-25. tion in Digital Psychiatry (Clare 2016). Digital technologies in FEP are meant to predict and control Ben-Zeev, Dror, et al. 2014. Robyn Bluhm, Ph.D. uncertainty so as to return individu- “Feasibility, acceptability, and pre- [email protected] als to (or approximate, as closely as liminary efficacy of a smartphone possible) a pre-morbid sate, presum- intervention for schizophrenia.” Phoebe Friesen’s paper raises a ably with the help of antipsychotic Schizophrenia Bulletin. 40(6): 1244- number of important questions about medications that are also more easily 1253. digital psychiatry research. I will focus managed through digital care tools— Berkhout, S. 2018. “Paradigm here on the issue of the gap between micro-political technologies of Shift? Purity, Progress, and the Ori- production and identifying causal path- health at the individual and molecu- gins of First Episode Psycho- ways or developing treatments. Friesen lar level (Berkhout 2018). To say sis.” Medical Humanities. 44: 172- acknowledges that a number of studies that such technologies give “voice” 180. have reported impressive success in to lived experience as some propo- Breitborde, N., Srihari, V.H., and predicting, for example, post-partum nents of digital care tools in FEP do Woods, S. 2009. “Review of the Op- depression or the onset of psychosis. (It (see Torous et al. 2019) misses a erational Definition for First-Episode is worth noting, though, that a 50% larger issue of health politics. When Psychosis/EIP.” Early Intervention in success rate is expected by chance, and biodata is seen as somehow speaking Psychiatry. 3(4): 259-265. that the true test of a predictive algo- for itself while serving as a proxy for Clare, E. 2017. Brilliant Impfe- rithm comes when it is used on datasets the complexity of experiential tion: Grappling with Cure. Durham: other than the one used to generate it.) knowledge, it needs to be acknowl- Duke University Press. Yet even when specific features of the edged that this lived experience Csillag, C., Nordentoft, M., Mi- data can be shown to drive successful serves as knowledge only when zuno, M. et al. Early Intervention production, the predictive features can- translated through expert-designed Services in Psychosis: From Evi- not be situated with a clear causal path- tools (Swartz 2018). dence to Wide Implementation. Early way. As Friesen points out, knowing What might we say if we hold Intervention in Psychiatry 2016; 10 that possessive pronoun use is linked to crip and disability-informed insights (6): 540-546. the onset of psychosis does not suggest regarding futurities alongside the Firth, J. and Torous, J. 2015. a means of intervening to prevent the notion of co-production? As a con- “Smartphone apps for schizophrenia: episode. Thus, these statistical relation- cept stemming from the field of sci- A systematic review.” JMIR mhealth ships fail to explain how a predictive ence and technology studies, co- and uHealth. 3(4): e102. features is related to a clinical outcome, production refers to the ways in Kafer, A. 2013. Feminist, Queer, or to suggest a potential avenue for which evolving scientific , Crip. Bloomington: Indiana Universi- developing a treatment. This does not, technological artifacts, and associat- ty Press. though, stop researchers from promis- ed beliefs may emerge hand-in-hand Jongsma, K., Bredenoord, A.L, ing more than predictive power. I sug- with representations, discourses, and and Lucivero, F. 2018. “Digital Med- gest that their optimism is unwarranted, identities (Jongsam et al. 2018). icine: An Opportunity to Revisit the not merely because the field of AI- Within FEP, EMA/ESM apps built Role of Bioethicists.” American Jour- driven psychiatry is fairly new, but into smartphones and tablets can be nal of Bioethics.18(9): 69-70. because the relationships it discovers said to be coeval with beliefs about McGorry, P. Early Intervention are almost never going to be the right what constitutes psychotic phenome- in Psychosis: Obvious, Effective, kind for explanation. na and with psychiatric service users Overdue. Journal of Nervous and The philosopher Heather Douglas themselves. To paraphrase Latour, Mental Disease 2015; 203(5): 310- has written about the relationship be- digital care tools have never been 318. tween prediction and explanation in modern: such tools do not simply Saraceno, B. New Knowledge science, arguing that the purpose of carve up the world at its joints but and New Hope to People with explanations is (or should b) to inform materialize particular kinds of sub- Emerging Mental Disorders. Early better predictions. Her claim takes on jects. I see this very much akin to the Intervention in Psychiatry 2007; 1(1): an additional urgency in medical re- concerns Dr. Friesen raises in their 3-4. search, where prediction is necessary excellent piece and look forward to Swartz, A. 2018. “Ethical Chal- not only for accurate prognosis, but more discussion on this topic from lenges of Digital Medicine for Seri- also to ground good treatment deci- within philosophy of psychiatry and ous Mental Illness.” American Jour- sions. Some of the studies Friesen dis- psychiatric ethics. nal of Bioethics.18(9): 65-67. cusses do show promise for prognosis; Torous, J., Woodyatt, J., examples are the prediction of post- References Keshavan, M., and Tully, L. 2019. partum depression or of an episode of “A New Hope for Early Psychosis psychosis. But the kind of predictive/ Bell, Imogen H., et al. 2018. Care: The Evolving Landscape of explanation relationships Douglas dis- “Smartphone-based ecological mo- Digital Care Tools.” The British cusses can’t be achieved by digital psy- mentary assessment and intervention Journal of Psychiatry. 214: 269-272. chiatry. Detailed causal explanations in a blended coping-focused therapy of phenomena that can lead to new for distressing voices: Development *** scientific insights are notoriously im-

6 Volume 27, Number 1 2020 possible in the kinds of AI-driven tech- mercial messages are fed directly to health information as formulated in niques used in digital psychiatry, where him, appearing as holographic images the HIPAA regulations We might the algorithm driving prediction cannot fading in and out to him throughout consider the consequences of this. If be mapped onto known causal varia- the store. The results of his unique currently innocuous, seemingly trans- bles. Nor can what we know about big-data digital phenotyping gets its parent, non-clinical information about causal processes be used to inform pre- ultimate capitalist expression: selling us turns out to be parts of clinically- dictions in any meaningful sense: this stuff to him, of course! significant diagnostic or prognostic requires the development of hypotheses In her Bulletin essay Phoebe Friesen sets of information, this poses a sub- that can be experimentally tested, or at raises more than a few very interest- stantive epistemic and practical chal- least a model that works “forward” ing questions, for which I can only lenge in protecting health privacy. from potentially causal variables to response to a few, and briefly. Dr. Should we expand the kinds of infor- outcome data. By contrast, in digital Friesen wonders about the utility of mation about us to include, well, eve- psychiatry, prediction works AI profiling, or digital phenotyping rything counting as PHI? Besides “backwards” from existing data to iden- (Insel, 2017) in terms of treating peo- being a practical challenge for en- tify patterns within the data, with no ple with mental disorders. My point forcement (of which already there is guarantee that those patterns have any through mentioning Minority Report little of, if HIPAA convictions are deeper causal, or more clinically useful, is to mention one implication: As- telling), I can't imagine the mental meaning. This means that, ultimately, suming that the Western world main- health industry will tolerate con- the prospect for these techniques to tains some substantive commitment straints upon the kinds of precise mar- contribute to the development of new to free- or minimally-constrained keting information available that also therapies is low. market capitalism, we can count on happens to be the same information digital phenotyping to sell that is 'clinical' information. Further, References (presumably more effectively) prod- the potential for my digital phenotype ucts to people with mental disorders to identify me is real, if a currently- Douglas, HE (2009) Reintroducing pre- These products will most likely be unrealized ideal sought by commer- diction to explanation Philosophy of medications or somatic therapies of cial interests illustrated in Minority Science 76(4):444-463. the future. It's possible that services, Report. A looming difficulty already such as psychotherapies, will also be identified in the PHI front is that one's *** 'sold' or promoted, though if the past DNA profile is already capable of and present are any indication, psy- identifying me; meaning my personal identifiers are carried around with me, Digital healthcare can expand chotherapy and other mental health services are obligations framed by and through my skin cells, shedded the mental health industrial clinician's conscientious interests and everywhere (shades of another sci-fi complex not very appealing to lobbyists in film, Gattaca). So what we seem to Washington, nor to politicians, whose be facing is the loss of the feasibility John Z Sadler M.D. campaign coffers are much more like- of health privacy. As two millenia of ly to be swollen by pharmaceutical stigma has demonstrated, this doesn't [email protected] seem to be a very attractive develop- and industry money (Sadler 2013). Will psychosocial treatments be fur- ment for people with mental illnesses In his prescient 2002 science- and disabilities. fiction film, Minority Report, Steven ther marginalized by digital psychia- try? The answers are not clear. We Spielberg's digital future portrays a References world where crime can be prevented already see computerized and online psychotherapies, now aimed at ena- before it happens, through pre-emptive Insel, T. R. (2017). Digital phe- bling these services for patients in intervention based upon technology. notyping: technology for a new sci- underserved areas. What he doesn't get quite right, is he ence of behavior. Jama, 318(13), A second, more explictly philo- depends upon a triad of gifted humans, 1215-1216. 'precogs', who can foresee with com- sophical question implied by Dr. Frie- sen's essay, concerns the concept of Sadler, J. Z. (2013). Considering plete accuracy criminal actions. The the economy of DSM alternatives. data for their input is supplied by net- 'health information', and what counts as protected, and protect-able health In Making the DSM-5 (pp. 21-38) worked big data which feeds the pre- edited by J. Paris and J Phillips. information. While my electrocardio- cogs constantly. What the ensuing 18 Springer, New York, NY. gram, my MMPI, and my serum sodi- years of real-world digital development um are 'personal health infor- has shown is that the precogs are not *** needed, because AI potentially can do mation' (PHI), the information col- lected and even synthesized by digital criminal, and other forms of surveil- lance, without superpowered . phenotyping appears to not be pro- Only supercomputers are needed. tected health information of the 18 What Spielberg gets exactly right kinds mentioned in the HIPAA regu- in the film is the following world- lations, cited by Friesen. While my building component of a digital future. preference for Levi's jeans, arugula, In a particularly revealing scene, the and Archie Shepp's jazz may predict star Tom Cruise's character walks some health vulnerability or outcome through a department store where com- for me in the digital health universe, this kind of information is far from 7 Volume 27, Number 1 2020

and genomics to machine learning, review, Shatte et al. show that most ( Don’t) Believe the Hype the hype appears to be similar. On a machine-learning applications are be- ing developed in mental health for the more general level, the current hype Paolo Corsico, Søren Holm on machine learning extends far be- detection and diagnosis of mental paoo.corsico@postgrad. yond psychiatry to a variety of medi- health conditions [6]. Digital manchester.ac.uk cal domains. One thing must be not- phenotyping—the collection and analy- ed. The hype is evident not only in sis of cognitive and behavioural data What is the problem with digital medicine but also in ethics. It is evi- via digital technologies—holds promise psychiatry? In her article, Phoebe dent in the growing field of Artificial to transform mental health care because Friesen effectively highlights three Intelligence (AI) and digital ethics, to it uses widespread low-cost technolo- key challenges posed by the develop- which we are contributing right now. gies such as smartphones. As Martinez- ment of digital psychiatric applica- Does ethics also suffer from the hype Martin et al. claim, “Because digital tions. One of these challenges is prac- problem? Are we co-responsible for phenotyping uses a ubiquitous technol- tical whilst the other two are ethical. the hype on digital ethics, or are we ogy and is inexpensive to deploy, it First, there is the practical issue of justified in investigating the ethical will likely transform the diagnosis and research governance: if digital psychi- implications of digital innovation? treatment of mental illness globally by atric research takes place outside of Udo Schuklenk has interestingly sug- enabling passive, continuous, quantita- clinical contexts, how can we regu- gested that we should reflect on the tive, and ecological measurement- late—and who should oversee—such ‘ethics of AI ethics’, and to avoid that based care” [7]. Digital psychiatry research so that participants are safe- only the voices of those who are will- promises to fill the two gaps identified guarded according to established re- ing to investigate the moral implica- above. Hence, the hype. search ethics principles? Second, there tions of AI whilst not being overly If our hypothesis is true, then only by is the ethical issue of beneficence: will critical of the digital agenda are heard addressing issues (1) and (2) in the first digital applications be beneficial to [2]. place may we properly understand the psychiatric patients? Third, and close- So, why the hype on digital psy- hype on digital psychiatry and evaluate ly linked to the second, is the issue of chiatry? We propose a hypothesis. the potential role of digital innovation the gap between prediction and inter- The hype on digital psychiatry may in improving people’s mental health. vention in digital psychiatry: what do be the result of not only technological The important point seems to be, do we we do with increased predictive capa- innovation but also of how new tech- have good reasons to believe the hype? bilities if we cannot provide (access nologies interact with longstanding Do we have good reasons to think that to) timely intervention? This is not issues in a given clinical speciality. digital innovation will really help us to only a clinical problem but also an Not only technological development address those two issues? If we do not ethical one. It raises the moral ques- but also the cultural background have good reasons to believe the hype tion of how we ought to act upon our where this unfolds may be held ac- then we should probably be wary of the increased predictive capabilities in countable for the hype. Two major claimed disruptive potential of digital mental health care. These three chal- issues seem to characterise psychia- technology in psychiatry. Conversely, lenges, Friesen argues, make digital try: if we have good reasons to believe the psychiatry currently problematic. hype then it may be worth to sketch out We do not directly address these 1. Diagnostic uncertainty, mean- the practical and ethical issues that three challenges here. Rather, we be- arise in digital psychiatry, and to strive lieve that Friesen poses a more funda- ing that (i) a lot of patients re- to regulate emerging practices. Figur- mental question that embeds and puts ceive many different diagnoses ing out whether we have, and what the others into context. We address over their lifetime, and that (ii) these good reasons may be can stem this question here: Why the hype? diagnostic categories have been from a thorough discussion amongst How cautious should we be in believ- heavily debated and contested the clinical community, ethico-legal ing that digital psychiatry will bring and are subject to constant revi- scholars, and patients. After all, the about the revolution it promises to hype problem of digital psychiatry is sion [3, 4]. bring about? situated at the interface between We might call this question the (promising) research results and appro- ‘hype problem’. As Friesen argues, 2. The structural lack of access to priate clinical translation. In describing the hype problem posed by digital effective mental health care [5]. their results, researchers are often psychiatry has much in common with This refers to both (i) access to prompted to consider—and sometimes the hype problem posed in the last mental health services, and (ii) overstate—the potential clinical impact decades by genomics and neurosci- the difficulty in developing novel of their findings. ence, which often promised to bring and effective interventions for Assessing the hype of digital psy- chiatry and the ethical issues thereof about a revolution in the way we un- severe mental illness. derstand, diagnose, and treat mental hence implies that we do two things. illness [1]. The extent to which we Firstly, we should carefully evaluate believe that genomics and Our hypothesis is that these two the real potential of digital applications neuroscience are (or may be) in a issues may create the necessary to ameliorate diagnostic and treatment position to hold on to that promise ‘room’ for the hype on digital psychi- practices in psychiatry. We should in- defines the boundaries amongst atry. Digital applications seem to be vestigate how digital applications different aetiological theories of currently used to address these two might effectively reduce diagnostic mental illness. Yet, from neuroscience longstanding issues. In their scoping uncertainty and improve access to

8 Volume 27, Number 1 2020

(quality) mental health care. By doing Analog Follies in the Age of as it is even if successful digital mark- so, we might be able to start distin- Digital Psychiatry ers are discovered. There is a mistak- guishing what is only hype from what en belief in psychiatry that a biologi- may indeed benefit patients. Second- cal or digital marker of some sort will Awais Aftab, M.D. somehow establish the disorder are ly—not chronologically but theoreti- [email protected] cally—we should ensure the ethical more real, as less abstract. A digital marker may therefore lead to a false conduct of digital psychiatric research Phoebe Friesen does a splendid within and outside of clinical settings, sense of validity of the and job in outlining how the of may further contribute to the already and ensure the appropriate (ethical) computational psychiatry falls short clinical translation of research find- rampant reification of DSM diagno- of the hyped promise and how much ses. It will be essential for clinicians ings in mental health care. In doing of this research is happening without so, we may end up contributing to the to understand that digital markers will oversight from the traditional appa- be designed through machine learning hype problem in digital ethics by inci- ratus for protection of human sub- dentally highlighting its promise, but to detect the presence of a construct. jects. I find myself in agreement with I am reminded of something that is what bioethicists do; they in- her, and I will use this opportunity to vestigate the ethical implications of Douglas Adams wrote: elaborate on some of my own con- technological innovation in medicine. cerns about the rise of computational Indeed, they might help to provide This is rather as if you imagine methods in psychiatry. a puddle waking up one morning that little pile of salt that is needed I will do so in the setting of a along with technological enthusiasm. and thinking, 'This is an interest- partly hypothetical scenario in which ing world I find myself in — an methods of digital psychiatry have References interesting hole I find myself in made it possible to analyze publicly — fits me rather neatly, doesn't 1.Insel, T.R., Translating scientific available online data (as well as any it? In fact it fits me staggeringly opportunity into public health impact: personal data that users may be will- well, must have been made to a strategic plan for research on mental ing to share from their social media have me in it!' This is such a illness. Archives of General Psychia- profiles) and make some psychiatric powerful idea that as the sun rises try, 2009. 66(2): p. 128-33. diagnoses with a high degree of accu- in the sky and the air heats up and racy. as, gradually, the puddle gets 2.Schuklenk, U., On the ethics of AI smaller and smaller, frantically ethics. Bioethics, 2020. 34(2): p. 146- 1) Concerns about Diagnostic Reifi- hanging on to the notion that eve- 147. cation rything's going to be alright, be- cause this world was meant to 3. Read, J. and J. Dillon, Models of Dr. Friesen makes an excellent have him in it, was built to have madness. Psychological, social and point that methods of digital psychia- him in it; so the moment he disap- biological approaches to psychosis. try are not likely to be causally in- pears catches him rather by sur- 2nd ed. 2013, London & New York: formative. It’ll be valuable for us to prise.” (2) Routledge. look at this point in the context of Derek Bolton’s discussions of bi- Like the puddle, we should not be 4. Guloksuz, S. and J. van Os, The omarkers: caught by surprise. Our digital mark- slow death of the concept of schizo- ers may fit psychiatric diagnoses stag- phrenia and the painful birth of the …if and when a biomarker geringly well, but that is not evidence psychosis spectrum. Psychological were found, the science of mech- that we are carving nature at its joints. Medicine, 2018. 48(2): p. 229-244. anisms, causes, treat- ment, preven- tion may well stay as it is. It de- 2) Concerns about Clinical Validity 5. NHS England and Department of pends on the extent and nature of the and Ethical Use Health, Achieving Better Access to causal of the biomarker. At one of Mental Health Services by 2020. the sp ectrum, a biomarker may be Until now, for the most part, di- 2014. just an (other) sign of the illness, agnosis of a psychiatric disorder has in ternal (inside the skin) as op- relied on patients or families seeking 6. Shatte, A.B.R., D.M. Hutchinson, posed to external, but as yet hard- help, thereby creating a certain thresh- and S.J. Teague, Machine learning in ly worth distinguishing from the old of “clinical significance”. This is mental health: a scoping review of external signs and symptoms of important because it means that the methods and applications. Psycholog- the illness, from the point of view diagnosis is mostly, in some sense, ical Medicine .2019 ,49(9): p. 1426- of the etiological model, which, invited. Either the patients or the pa- 1448. we may suppose, stays as highly tient's social system is experiencing complex and multifactorial as distress/impairment/harm and is in 7.Martinez-Martin, N., T.R. Insel, P. before. (1) Dagum, H.T. Greely, and M.K. Cho, need for help. That is why psychiatry and psychology are healing profes- Data mining for health: staking out Digital markers of diagnostic the ethical territory of digital pheno- sions. entities will in all likelihood be sur- The ability to make a diagnosis typing. npj Digital Medicine, 2018. face phenomena and epiphenomena. 1.68, based on publicly available online The science of mechanisms, causes, data such as social media use divorces treatment, prevention may well stay *** 9 Volume 27, Number 1 2020

diagnosis from that invitation, from perhaps psychiatric diagnoses will which large data sets can erase individ- that threshold of clinical signifi- become ubiquitous, a benign and ual narratives. John Sadler alerts us to cance. We can legitimately wonder common aspect of our ? the potential for digital psychiatry, par- what meaning a diagnosis holds if it Facebook apps will inform us of our ticularly in the context of capitalism, to is not tied to clinical significance. psychiatric morbidity profiles with lead to the further marginalization of This is an objection that has been as much nonchalance as they tell us psychosocial treatments. Suze Berk- raised with regards to epidemiologi- which wizarding house at Hogwarts hout raises concerns about how digital cal surveys (which evaluate presence we belong to? tools can serve to reinforce the assump- of descriptive symptoms and typical- Digital psychiatry is unlikely to tion that mental disability is an obstacle ly produce inflated estimates of revolutionize understanding of etiol- for which a cure is the solution to work prevalence of psychiatric disorders). ogy and treatment, but it may very towards. Furthermore, Paul Lieberman The same objection will also apply well drastically alter society’s rela- examines how the growth of digital to digital psychiatry. tionship with psychiatric diagnoses. psychiatry may lead us to adopt a nar- When digital psychiatry divorc- Like the ancient philosopher Thales, row theory of the mind, causing us to es diagnosis from the ‘invitation’ by we may grow so absorbed in con- view ourselves as “nothing more than consumers/clients who are in need templating the digital heavens that learning machines” and to neglect com- of help, it creates the tools with we may stumble into the well at out plex and interesting research questions which anyone can be diagnosed feet. that do not present themselves in such a without their consent or their paradigm. Especially worrisome is the knowledge. The diagnosis is deper- References picture of “political psychiatric war- sonalized and decontextualized. fare” painted by Awais Aftab. In this Consider the implications of 1. Bolton D. Classification and (oh-so-) possible future, digital diag- this using the highly controversial causal mechanisms: a deflationary nostic tools are widely available for use example of the Goldwater rule: with approach to the classification prob- on unconsenting users; journalists, who regards to public individuals, Ameri- lem. In Philosophical Issues in Psy- have no obligation to follow the Gold- can Psychiatric Association main- chiatry II: Nosology, K. S. Kendler water rule, can use these tools to feed tains that “it is unethical for a psy- & J. Parnas (Eds). 2012. Oxford: “a hungry public” who desire daily chiatrist to offer a professional opin- Oxford University Press, pp. 10. psychiatric assessments following pres- ion unless he or she has conducted idential tweets! an examination and has been granted 2. Adams D. The Salmon of While it may not balance out these proper authorization for such a state- Doubt: Hitchhiking the Galaxy One considerable worries, I enjoyed the tiny ment.” (3) This rule is often defend- Last Time. Pan Macmillan. 2002. bit of optimism I found within the com- ed partly on an epistemological and mentaries, voiced by Jim Phillips. Nod- partly on an ethical basis. The epis- 3.American Psychiatric Associ- ding to our current pandemic context, temological argument states that any ation. Goldwater Rule’s Origins he aptly points out the many ways in diagnosis made in the absence of a Based on Long-Ago Controversy. which epidemiological data can be of personal examination is highly falli- https://www.psychiatry.org/ use, and how we might also find such ble. The epistemological argument is newsroom/goldwater-rule (accessed uses within digital psychiatry. He also less convincing, since the profes- 5/12/2020) reminds us that digital predictive tech- sional lives of psychiatrists and psy- nologies won’t always have the last chologists are full of instances *** say. Just as psychiatrists (and patients, I where we offer professional opin- would add) can interpret the conclu- ions in the absence of personal ex- sions of meta-analyses as relevant to an amination (when supervising train- individual’s care or not, the outputs of ees, doing chart reviews, doing psy- digital diagnostic algorithms will ideal- chological autopsies, etc.) The ethi- ly be filtered through additional routes cal argument states that even if it is Response to Commentaries whereby false positives and false nega- possible to make a diagnosis without tives can be caught, and individual ex- personal examination, a Phoebe Friesen periences and values can be taken into shouldn’t do it because such an account. opinion is likely to be exploited for Thank you for a number of fasci- The particular worries I raised in political purposes. nating and creative responses to my my commentary, related to jumping If all you need to make a diag- discussion of digital psychiatry, each from predictions to the promise of in- nosis is a computer algorithm, how- of which has supplied me with an terventions, regulatory gaps that re- ever, digital psychiatry may possibly ample serving of food for thought. search takes place within, and the remove the psychiatrist from the While I sought to map out a few growing hype surrounding digital psy- equation entirely. Any newspaper emerging issues with digital psychia- chiatry, were also illuminated and ex- may use such an algorithm to exam- try within my commentary, I will say panded upon within several commen- ine the publicly available tweets of a that I have a lot more to worry about taries. Robyn Bluhm is entirely on president and will have all the pre- now. The responses point to a num- point in her explanation of why the gap dictive power of digital psychiatry to ber of possible harms, implications, between predictions and intervention is offer a diagnosis to a hungry public! and misuses, of digital psychiatric so significant in digital psychiatry. Pre- Will this be the beginning of technologies. Both Jim Phillips and cisely because of the nature of these political psychiatric warfare? Or Paul Lieberman note the way in

10 Volume 27, Number 1 2020 technologies and the way they seek to ticipatory approaches, claiming to and so will merely affirm our existing exploit any predictive features that incorporate the voices of those with systems of psychiatric classification, might be present, the statistical rela- lived experience into their tools and rather than generating alternatives; this tionships discovered “are almost never treatments. In doing so, she offers means it’s unlikely that the contested going to be the right kind for explana- another kind of warning about the categories will become any less con- tion”. As Bluhm puts it, drawing on the narratives produced by those working tested within the digital revolution. In work of Heather Douglas, a good scien- in this digital space. Beyond concerns terms of accessing care, predictive tific explanation generates predictions about overpromising interventions as technologies certainly promise to iden- by developing models and hypothesis a result of predictive power, such tify more individuals at risk or in crisis. that can then be tested empirically. initiatives may also make unfounded But, as noted above by Bluhm and my- However, in machine learning, promises related to representation. As self, such predictions, which narrow in “prediction works ‘backwards’ from Berkhout points out, however, these on finger strokes, the use of personal existing data to identify patterns within “tools do not simply carve up the pronouns, and prosodic features of the data, with no guarantee that those world at its joints but materialize par- speech, are unlikely to generate effec- patterns have any deeper causal, or ticular kinds of subjects”. I see this as tive treatments anytime soon. As is more clinically useful, meaning”. I an especially important warning, giv- often the case, it seems, the proof will couldn’t agree more. Although I fear en the increasing pressure to democ- be in the pudding. The more important that the enthusiasm surrounding ma- ratize psychiatric research, and the question in this case, however, may be chine learning may be inspiring others concerns related to tokenism and co- who gets to decide whether the pudding to move backwards rather than for- optation that often arise within such constitutes proof. As John Sadler points wards [1]. initiatives [2, 3] out, in this realm, as in all parts of psy- John Sadler further advances my Paolo Corsico and Søren Holm chiatry, market forces will be shaping worries regarding how frequently digi- also narrow in on the issue of hype, determinations of efficacy. As a result, tal psychiatry evades research oversight asking why such hype exists and how along with Patyusha Kallari, I suggest and health privacy protections. Noting we ought to response to it. They in- we focus on how not only on how good that it is not merely health data that can sightfully identify two features of or fair uses of AI tend to be, but how be exploited today, but any large data psychiatry that render it vulnerable to they shift, or fail to shift, distributions sets that can be translated into health such digital hype: diagnostic uncer- of power [4]. data, he notes that “what we seem to be tainty and the lack of access to effec- That said, if Corsico and Holm are facing is the loss of the feasibility of tive treatments. They suggest further right, by merely writing this commen- health privacy”. Given this predica- that we should “be wary of the tary, I’ve already contributed to the ment, he asks whether we might ex- claimed disruptive potential of digital hype. So I might not sleep very well pand health privacy laws to “include, technology in psychiatry” if we do tonight! well, everything”? While I share Sad- not have good reasons to believe the References ler’s concerns, I’m not ready to give up hype. If we do have good reasons, hope on demarcating what ought to fall however, they suggest that “it may be 1. Yarkoni, T. and J. Westfall, Choos- within the boundaries of regulatory worth to sketch out the practical and ing prediction over explanation in psy- oversight just yet. While it seems likely ethical issues that arise in digital psy- chology: Lessons from machine learn- that ‘medical data’ may no longer serve chiatry, and to strive to regulate ing. Perspectives on Psychological Sci- the function of providing such a bound- emerging practices”. While I agree ence, 2017. 12(6): p. 1100-1122. ary, it may be that we need to expand with these directives, I think we protection to include not only types of should embrace them both, regardless 2. Kalathil, J., Beyond Tokenism: Par- data, but also models and technologies of whether we have good reasons to ticipation of mental health service users that can create such sensitive data; believe the hype or not. I would sug- from racialised groups in mainstream through this route, perhaps we can still gest we should always be wary of user involvement initiatives. Agenda, salvage some form of health data priva- hype, particularly in a field like psy- 2010. 34: p. 16-18. cy. chiatry, which has such a long history Other responses engaged with the of hype followed by disappointment. 3. Eriksson, E., Four features of coop- topic of hype which surrounds digital Ethical and practical analyses, as well tation: User involvement as sanctioned psychiatry. Suze Berkhout notes the as regulatory reform, are perhaps resistance. Nordisk välfärdsforskning| resemblance and overlap between dis- even more necessary if unfounded Nordic Welfare Research, 2018. 3(1): cussions of digital psychiatric tools and hype is being generated, particularly p. 7-17. early intervention , particu- as it corresponds with mountains of larly those which focus on first episode research dollars being spend on such 4. Kalluri, P., Don’t ask if artificial psychosis (FEP). In a fascinating char- digital developments. intelligence is good or fair, ask how it acterization of the language and meta- In terms of Corsico and Holm’s shifts power. Nature, 2020. phors used within these two realms, she discussion of digital psychiatry filling observes the way in which they both in the gaps of diagnostic uncertainty *** portray themselves as associated with and access to care in psychiatry, I hope, novelty, and innovation, and as can’t say I’m feeling very optimistic frontiers which offer “a bridge to the on this front. As Awais Aftab points future” that can help us enter “a new out, digital tools in psychiatry are era”. She also observes how these initi- “designed through machine learning atives often align themselves with par- to detect the presence of a construct”,

11 Volume 27, Number 1 2020

(President, continued from page 1) (Editor, continued from page 1) rather to underline the fact that, along with symptoms, the findings of EBP emerging dinosaur might be around 25 ready example is the current corona- and RCT form the data of psychiatry, years, or less. virus crisis. We depend heavily on the and that as data they pose the challeng- We should not just passively ac- gathering of data and its measurement es to data described above. cept that recent technology is the new to assess where we are in relation to the To illustrate these challenges I will best practice (it is not always better), virus. Think of incidence of illness, focus on psychiatric treatment and the nor try to deny it once it arrives. It incidence of death, positive and nega- use of RCTs and meta-analyses in the would be better to actively make it ours tive test results, choice of the best medication. In an and, in the process, both remake our- effects of social distance, benefits RCT each subject in each group is a selves and mold the technology to our of masks, declining or rising rates of data point, in all important respects the purposes and goals. hospitalization, etc. – all involving the same as other subjects in its group. We Each of us will always be who we measurement of data. are not interest in their differences. are - children of a particular time and On the other hand, in addition to Inasmuch as the meta-analysis is an place - but part of who we are should the limitations suggested by Friesen, I analysis of RCTs, it includes huge include asking who we want to be. We wish to point out one big limitation of numbers of data-point/subjects. Fur- should be grounded enough to not fol- data analyses – that they leave out indi- ther, the decision as to the best answer low every fad or buy the hype about the vidual narratives. To appreciate this, is made semi-automatically and algo- next big product, but we should also consider how the PBS i datum, we rithmically. The challenge here is the appreciate and benefit from what hu- weep. Data are what they sound like same one raised by Friesen with her man and creativity produces. – impersonal, numerical, anonymous, examples: does the meta-analysis lead For example, I find this array of all- shallow. to better psychiatric care? In treatment access max plus pay for television be- How might these thoughts apply to the subject is not a datum; it’s the per- wildering. As someone who was in psychiatry? Let’s begin with the notion son in the office with me. And the elementary school in 1970s, I almost of a symptom. Is a symptom a piece of treater is not an algorithm. Although in miss those days of only three television data – a datum? If so, symptoms are the hierarchy of EBP, the opinion of networks plus PBS, not but quite. What among the t the gravity of the diagnosis the treater is at the bottom of the hierar- people are calling ‘quality television’ is in each case. As we can see, in this chy, way below the meta-analysis, the better. Likewise, in the past month, diagnostic process data and metrics are fact remains that it is this treater who several artists I follow have released keeping close company. decides whether the meta-analytic rec- new albums and I have listened to all of The above represents the diagnos- ommendation fits the needs of this pa- them via my streaming music service. tic manual approach to diagnosis. tient. No piece of data, and no meta- It is better in many ways. Scales are also used in the diagnostic analysis will answer that question. Now I have to decide what to do process. The Beck, Hamilton, and PhQ- with all these CDs. To be honest, many 9 scales all represent ways of counting JP of them are unopened because they symptoms, with thresholds to meet for came with free digital downloads into the diagnosis. For psychiatric treat- *** iTunes. It seems ridiculous to have ment, on the other hand, we rely on the been buying them given what the new claims of Evidence Based Psychiatry technology offers. But I will keep (EBP) and the randomized controlled them. Who knows? I should have held trials (RCT) that provide the on to all my old vinyl records instead ‘evidence’. EBP and RCT have been of giving them away in 1995 – today’s soundly criticized, but that is not the teenagers might pay good money for point I want to develop here. I want the ‘outmoded’ analog technology. They say it’s better.

***

12 Volume 27, Number 1 2020

The Association for the Advancement of Philosophy and Psychiatry was estab- Bulletin Editor Edwin L. Hersch, M.A., M.D., F.R.C.P. James Phillips, M.D. lished in 1989 to promote cross- Ginger Hoffman, Ph.D., Ph.D. disciplinary research in the philosophi- 88 Noble Avenue Brent M. Kious, M.D., Ph.D. Milford, CT 06460 cal aspects of psychiatry, and to support Robert S. Kruger, Ph.D. educational initiatives and graduate Phone (203) 877-0566 Paul Lieberman, M.D. Fax (203) 877-1404 training programs. Douglas Porter, M.D. E-mail [email protected] Nancy Nyquist Potter, Ph.D. OFFICERS Kathryn Tabb, Ph.D. President G. Scott Waterman, M.D., M.A. Philosophy, Psychiatry, & Psychology Peter Zachar,Ph.D. J Melvin Woody, Ph.D. Jon Tsou, Ph.D. K.W.M. Fulford, D.Phil., MRCPsych. Vice-president Founding Editor Christian Perring, Ph.D.

Secretary EMERITUS MEMBERS John Z. Sadler, M.D. James Phillips, M.D. Editor-in-Chief Treasurer Jerome L. Kroll, M.D. John Z. Sadler, M.D. Jennifer H. Radden, D. Phil. AAPP Web Site Louis Sass, Ph.D. www3.utsouthwestern.edu/aapp Executive Secretary Serife Tekin, Ph.D. Immediate Past President INTERNATIONAL ADVISORY Claire Pouncey, M.D., Ph.D. BOARD

K.W.M. Fulford, FRCP, FRCPsych EXECUTIVE COUNCIL Gerrit Glass, M.D., Ph.D. Emilio L. Mordini, M.D.

Giovanni Stanghellini, M.D> Awais Aftab, M.D. Robyn Bluhm, Ph.D. Jeffrey D. Bedrick, M.D. Louis C. Charland, Ph.D. Phoebe Friesen, Ph.D.

ASSOCIATION FOR THE ADVANCEMENT OF PHILOSOPHY & PSYCHIATRY (AAPP) MEMBERSHIP APPLICATION

Membership in AAPP is open to all individuals interested in the subject of philosophy and psychiatry by election through the Membership Committee. The Association welcomes Student Members (enrollees in degree-granting programs in colleg- es and universities and physicians enrolled in approved psychiatric training programs and post-graduates in post-doctoral programs). In order to join AAPP please detach this form and mail to: Ms. Alta Anthony, Journal Subscriptions/ Memberships, The Johns Hopkins University Press, P.O. Box 19966, Baltimore, Maryland 21211.

Annual Dues: $115 Members; $42 Student Members (this includes a year’s subscription to Philosophy, Psychiatry, &Psychology (PPP). Make checks payable to The Johns Hopkins University Press.

Name Qualifications (clinical and/or philosophical)/Speciality/Interests

Address Telephone FAX

Amount Enclosed: Check:______VISA:______Exp.Date:______

MasterCard:______Exp.Date:______13