AI in Medicine the Spectrum of Challenges from Managed Care to Molecular Medicine

Home , Molecular medicine, Mycin

AI Magazine Volume 20 Number 3 (1999) (© AAAI) Articles AI in Medicine The Spectrum of Challenges from Managed Care to Molecular Medicine

Russ B. Altman, M.D., Ph.D

AI has embraced medical applications from its inference algorithms required. There was (and inception, and some of the earliest work in suc- is) considerable debate about how complex cessful application of AI technology occurred in inference should be for expert performance, medical contexts. Medicine in the twenty-first cen- but it is clear that medicine is a field in which tury will be very different than medicine in the there is a lot of knowledge required for good late twentieth century. Fortunately, the technical performance. It is also clear that physicians challenges to AI that emerge are similar, and the prospects for success are high. constantly feel a sense of overload as they deal with the individual data associated with their patients as well as the content knowledge of hen I was asked to make this presen- medicine that they are trying to apply to the tation, the organizers specifically treatment of these patients. I can try to provide Wasked me to review a bit of the histo- a feel for the information-processing load on a ry of AI in medicine (AIM) and to provide an physician: A full-time general practitioner is update of sorts. I have therefore taken the lib- currently expected to longitudinally follow a erty of dividing the last 30 years of medical AI panel of 2000 to 2500 patients. Of course, the research into three eras: the era of diagnosis, severity of illness varies, but it is clear that the era of managed care, and the era of molec- physicians need systems (computer or other- ular medicine. A description of these eras wise) to track the data pertaining to these allows me to review for you some of the early patients and turn it into working hypotheses and current work in AIM and then tell you for diagnosis, treatment, and long-term prog- about some of the exciting opportunities now nosis. emerging. The other appeal to working in AI in medi- Why is AI in medicine even worth consider- cine is that the field is large, and so virtually all ing? In the late 1950s, medicine was already aspects of intelligent behavior can be studied drawing the attention of computer scientists in one part or another of medicine. You can principally because it contains so many stereo- study issues of image processing, automated typical reasoning tasks. At the same time, these management of database information, robotic tasks are fairly structured and so are amenable automation of laboratories, computer-assisted to automation. Every medical student learns diagnosis, multimedia for physician and that when one thinks about a disease, one patient education, virtual and telesurgery, and thinks in an orderly way about epidemiology, many other issues. For some, AI in medicine pathophysiology, diagnosis, treatment, and provides a kinder, gentler, “greener” applica- then prognosis. These are the bins into which tion area in which to apply their techniques. medical information is parsed. These sorts of structured reasoning methods made medicine Three Eras for AI in Medicine an attractive application area. In addition, medicine is clearly knowledge intensive, and The first era of AI in medicine was the “Era of so at places like Stanford (where knowledge Diagnosis.” The first aspect of medical reason- was power [Feigenbaum 1984]), it was very ing that caught the imagination of AI tempting to try to encode knowledge for the researchers was the process of collecting clini- purposes of reproducing expert performance at cal data and applying inference rules to make diagnosis and treatment. The working hypoth- the diagnosis of a disease. This is the common esis was that rich knowledge representations image of the doctor as sleuth, determining would be sufficient, with only relatively weak what disease is causing the patient’s symp-

toms. The second era of AI in medicine was bution from the NLM was the creation of an what I have called the “Era of Managed Care of online database of the published biomedical Chronic Disease.” This era has approached a literature, MEDLINE. Having gone through a set of problems quite distinct from those tack- number of transformations, the MEDLINE data- led in the preceding period, as I will discuss. base was recently made available to the general Finally, we are on the precipice of the “Era of public via the PUBMED resource on the World Molecular Medicine,” which is once again Wide Web (www.ncbi.nlm.nih.gov/PubMed/). going to raise issues that are different from For better or worse (I believe for the better), those occupying researchers during the first physicians and patients now have unprece- two. dented access to a literature that is growing exponentially. The challenges in indexing, The Era of Diagnosis searching, and parsing this literature represent In 1959, Ledley and Lusted (1959) published a a major challenge to AI investigators. paper in Science entitled “The Reasoning Foun- The 1970s brought the push for diagnostic dations of Medical Diagnosis.” This classic performance. De Dombal et al. in 1972 showed paper has the feature of many classic papers: It that you could make clinically accurate diag- puts forth a series of statements that are now noses using Bayesian inference (de Dombal et taken as almost self-evident. Ledley and Lusted al. 1972). Also in 1972, Kulikowski and the pointed out that medical reasoning was not team at Rutgers created the CASNET system in magic but instead contained well-recognized which they explored methods for using causal inference strategies: Boolean logic, symbolic models for somewhat deeper diagnostic tasks inference, and Bayesian probability. In partic- (Kulikowski and Weiss 1982). The models were ular, diagnostic reasoning could be formulated deeper because physiological models were now using all three of these techniques. Their paper being used to explain symptoms and describe mapped a research program for the next 15 diagnostic possibilities. Shortliffe, Buchanan, years, as investigators spun out the conse- and coworkers showed soon afterwards (with quences of applying these inference strategies the MYCIN system in Buchanan and Shortliffe to medical domains. [1984]) that production rules could be used to The research that followed was varied and make expert-level diagnosis of infectious dis- excellent, and I cannot properly review all the eases. Pauker and coworkers created the PIP sys- contributions but instead will pick some exem- tem (Presenting Illness Program) in which the plary efforts. For example, in 1965 Lawrence cognitive processes associated with short-term Weed introduced a computer system called and long-term memory were modeled in order PROMIS to support a problem-oriented medical to create programs that could consider multi- information methodology (Tufo et al. 1977). ple diagnoses but then focus on the few most Weed’s work was among the first to demon- likely solutions quickly (Szolovits and Pauker strate a truly electronic medical record. More- 1976). Figure 1 shows a ﬁgure from one of the over, this record was a highly structured, PIP papers, in which an associative memory strongly typed data structure (in many ways structure is modeled. As particular concepts are similar to our modern frame-based systems) activated and drawn into the river represent- that even today is rarely matched in its insis- ing active memory, they drag into the river tence on structured data input. Weed’s work with them associated ideas that then come to was limited by the absence of standard termi- the attention of the inference engine. nologies to use within his data structure, but A magnum opus during this period was the his belief in structured data is still a major goal INTERNIST knowledge base and inference pro- within the medical informatics community. gram published by Miller, Myers, and cowork- In the late 1960s the National Library of ers (Miller, Pople, and Myers 1982). INTERNIST Medicine (NLM) (www.nlm.nih.gov/) was had the goal of diagnosing any problem within established as one of the National Institutes of general internal medicine—basically any sys- Health (NIH). This was remarkable for many temic disease or disease of the organs between reasons, not least of which was that most insti- the neck and the pelvis. INTERNIST was based on tutes within the NIH are associated with an a very large knowledge base that was trans- organ or a disease (for example, The National ferred to a PC-based program called QMR, which Institute of Heart, Lung, and Blood or The now forms the basis for a commercial product National Cancer Institute). The NLM is still in (Miller, Masarie, and Myers 1986). The search of its organ or disease. Nevertheless, the INTERNIST/QMR knowledge base associated dis- extramural research program of the NLM has eases with ﬁndings using two numbers: a fre- been a principal source of research funds for AI quency of association and an evoking strength. in medicine. The principal intramural contri- There was then an algorithm created for col-

68 AI MAGAZINE Articles

Figure 1. A Representation of the Associative Memory Required for Medical Diagnosis from the PIP Work (Szolovits and Pauker 1976). A. A “fact” is searching for a connection to the network, so that the appropriate concepts can be pulled down into the short-term memory. B. A “fact” has found a matching concept and thus pulls the appropriate associated concepts into the short-term memory. lecting ﬁndings and computing the most likely Bayesian networks for diagnosis, in which the diagnoses. Since the introduction of this pro- conditional dependencies between variables gram, others have been introduced that are could be modeled in a somewhat natural man- based on similar ideas, including DXPLAIN (Bar- ner (Heckerman, Horvitz, and Nathwani nett et al. 1987) and ILIAD (Bouhaddou et al. 1992). They also were able to recast some of 1995). The performance of these programs has the assumptions behind the other (apparently been evaluated and compared by running nonprobabilistic) systems (MYCIN and INTERNIST) them on some challenging case reports (called to create a unified probabilistic “map” of the clinicopathological cases, or CPCs) such as those space of diagnostic algorithms (Dan and that appear each week in the New England Jour- Dudeck 1992; Middleton et al. 1991; Shwe et nal of Medicine (www.nejm.com) (Berner, Jack- al. 1991). So by the end of the 1980s, there was son, and Algina 1996; Wolfram 1995; Feldman a large and distinguished literature on medical and Barnett 1991). In most cases, the perfor- diagnosis. This literature has continued and mance of the programs is comparable to expert expanded to nonmedical areas such as the diagnostic performance (as judged by a blinded diagnosis of faults in electronic circuitry and review of diagnoses produced by both experts other engineering applications. and the programs, or unblinded evaluation of the performance using defined performance The Era of Managed criteria for success). The programs routinely Care of Chronic Disease outperform medical students and physicians in So what happened to the Era of Diagnosis? All training. of these systems were evaluated, and all of In the mid- to late 1980s, Heckerman and them seemed to perform near the level of coworkers showed that the preliminary work human experts. Well, there were a few prob- of De Dombal could be extended using lems. First, physicians did not embrace these

FALL 1999 69 Articles

technologies. Clinical data, unlike billing data, One of the ways to reduce the cost of health were not routinely available in a digital form, care is to move it out of expensive hospital so when you ran these programs there were rooms and into outpatient clinics. So instead these very awkward interfaces that asked you of intense episodes in the hospital, we have lots of questions in order to get the informa- these much more frequent less intense tion necessary to do the diagnosis. Clinicians episodes in the clinic where similar things are simply did not want to spend time entering being done but in a more fragmented manner. data that were already written into a note The fragmentation may cause confusion as we using natural language. The AI in medicine ask physicians to track the progress of 2500 community realized that they needed elec- patients with periodic interactions. tronic medical records as a prerequisite infra- One way to capture the look and feel of AI in structural element to allow the deployment of Medicine today is to look at the contents of a these technologies. Thus, issues of knowledge recent meeting. The AI in Medicine Europe representation, automatic data acquisition, (AIME) conference was held in Grenoble in federation of databases, and standard termi- 1997 (Shahar, Miksch, and Johnson 1997). An nologies became quite important. The second examination of the table of contents reveals problem for diagnostic programs was that three subjects, in particular, that reflect current physicians did not want help with diagnosis. concerns: (1) the representation and manipu- The AI in Diagnosis is fun, and physicians are trained to lation of protocols and guidelines, (2) natural do it well in medical school and only improve language and terminology, and (3) temporal medicine with years of practice. They did not want to reasoning and planning. Other areas of impor- community give up that fun to a computer. The most sig- tance include knowledge acquisition and nificant problem, however, was that diagnosis learning, image and signal processing, decision realized that is a actually very small part of what physicians support, and (our old friend) diagnostic rea- they needed do in the delivery of medicine. Most visits to a soning. physician are for an existing, previously diag- Protocols and guidelines have become an electronic nosed problem. The challenge to the physician important way to standardize care and reduce medical is to follow the problem and respond to its variance. Guidelines are created by panels of records as a evolution intelligently. Diagnosis is a relatively physicians who assess available data and rec- rare event, probably accounting for less than 5 ommend treatment strategies. For example, prerequisite percent of physician time. What physicians how should a newly discovered breast lump be infra- really need is help following chronic and slow- evaluated? The AI challenges follow directly: ly evolving disease in 2500 patients that are How do we develop robust and reusable repre- structural seen in brief episodes but require expert inter- sentations of process? How do we create adap- element to ventions. So we have the era of chronic care tive plans that respond to changes in available driving AI in medicine research. This problem information? How do we distinguish between allow the is compounded by an aging population with high-level plan recommendations and their deployment more chronic diseases. specific local implementation? How do we of these There is one other element of medicine that modify guidelines in response to data collected has changed the imperatives for AI research, during their execution? How do we model the technologies. and this is the emergence of new economic effects of guidelines on organizations? There is models for funding medicine (Selby 1997; Det- an increasing interest in the representation sky and Naglie 1990). The traditional model and simulation of organization systems in has been fee for service: A physician performs order to predict the effects of interventions in a service and gets paid an agreed-upon medical care. One recent development in this amount. If the physician performs lots of ser- area has been the development of a guideline vices, the physician makes more money. The interchange format (GLIF) (Ohno-Machado et new model of medical funding is based on a al. 1998). GLIF is a syntax for specifying clinical standard rate per patient that is paid to a protocols. It contains a language for represent- physician, regardless of the usage of services by ing actions, branch steps, and synchronization the patient. Now, the financial incentives are steps (among others) needed to specify a clini- reversed. If the physician provides a service, cal guideline (figure 2). then its cost in time and resources is taken out Natural language and standardized termi- of the pot of money that represents potential nologies remain a critical issue in medical profit. Now physicians still want to treat ill- computing. The medical goal is to create stan- ness, but there is now a huge incentive to dards for communication that move away deliver cost-effective, high-quality care. Sys- from hand-written natural language. Medicine tems for supporting these activities become is the only major industry still relying on the mandate. hand-written documentation. How do we

70 AI MAGAZINE Articles deﬁne formal semantics so that when we create these electronic medical records, we can populate them with clean data? What is the Guideline the_breast_mass_guideline underlying ontology for clinical medicine? { name = "Breast Mass Guideline"; authors = SEQUENCE 1 {"Max Borten, MD, JD";}; How do you map natural language into stan- eligibility_criteria = NULL; dard terminologies? How do we accommodate intention = "Evaluation of breast mass."; steps = local and global changes to these terminolo- SEQUENCE 40 gies? How do we integrate legacy databases { (Branch_Step 1); (Action_Step 101); with newer, semantically clean databases? (Action_Step 102); (Action_Step 103); How can we have machine-learning tech- (Synchronization_Step 1031); niques for extracting new medical knowledge (Conditional_Step 104); (Conditional_Step 105); from our semantically clean databases? It is Action_Step 102 important here to mention the Unified Med- { name = "Elicit Risk Factors for Breast Cancer in Personal History"; action = Action_Spec 102.1 ical Language System (UMLS), a project at the { name = "Elicit Personal History"; National Library of Medicine with the goal of description = "Elicit Personal History"; patient_data = SEQUENCE integrating a number of existing medical { Patient_Data 102.1 { name = "Personal risk factors for breast cancer"; vocabularies using a common semantic struc- type = "" ; ture (Bodenreider et al. 1998). The existing ter- possible_values = Sequence 0 {}; temporal_constraint = "valid thru all time"; minologies include those for specifying diag- didactics = SEQUENCE 0 {};}; noses, medical procedures, and bibliographic }; didactics = SEQUENCE 0 {}; indexing (Cote and Robboy 1980; Slee 1978). };

The UMLS is based on a semantic network and subguideline = NULL; has about 500,000 terms that have been classi- next_step = (Synchronization_Step 1031); didactics = SEQUENCE 0 {};} fied into about 150 semantic types with specified relationships. A fragment of its semantic network is shown in figure 3. Temporal reasoning and planning become critical in a setting where diseases are chronic, and interactions are episodic. The challenges Figure 2. The Guideline Interchange Format. are to integrate database and knowledge base This is a segment of a representation of the process of evaluating a breast mass. technology with temporal inferencing capabil- The language has a syntax for sequences of events, branches, and other struc- ities. How do we actually modify medical data- tured metainformation about the process of evaluation. bases so that we can do effective temporal inference with them? How can we recognize and abstract temporal trends in clinical data? Nonmonotonic reasoning becomes essential: As new data are collected, we retract old infer- The Era of Molecular Medicine ences and assert different ones. How do we cre- Although the management of chronic disease ate “smooth” models of patient state based on under conditions of capitated payment are episodic data collection? Finally, how can we likely to continue, I believe that there is an create plans for treatment over time? even more revolutionary set of changes com- My colleague at Stanford, Yuval Shahar, has ing to medicine. These changes will arise from done excellent work in the area of temporal the work being done in basic biology in deter- abstraction and has a system that is able to mining the complete DNA sequence of both automatically take a set of discrete data points the human organism as well as most major dis- and transform them into sensible intervals ease-causing organisms. There is an excellent that can, in turn, be grouped together into paper in the IAAI-98 proceedings by Rick Lath- even higher-level abstractions (Shahar, Miksch, and Johnson 1998) (as summarized in rop and coworkers (Lathrop et al. 1998) that is figure 4). an example of the opportunities in linking There are some other application areas with- molecular concepts with medical care and AI in medicine that deserve mention, including research. telemedicine, how to deliver medical care at a Some Biology First, it is appropriate to give distance using multimedia; intensive care med- some background about the genome sequenc- icine, with emphasis on reasoning with limited ing efforts. The entire development, structure, resources; and clinical trials, methods to auto- and function of an organism is specified by a matically recognize that a patient is eligible for sequence of four DNA letters: A, T, C, and G are a trial and to enroll them. the abbreviation of their chemical names

FALL 1999 71 Articles

Biologic Function

Physiologic Pathologic Function Function

Organism Organ Cell Molecular Disease Cell Experimental Function or Function Function or or Model Tissue Syndrome Molecular of Function Dysfunction Disease

Mental Genetic Mental Neoplastic Process Function or Process Behavioral Dysfunction

Figure 3. A Subset of the Semantic Net Created for the Uniﬁed Medical Language System (UMLS), in Which the Concept of Biological Function Is Specialized into Subsets. The semantic network is used to organize about 500,000 concepts in the UMLS.

BMT PAZ protocol

Expected CGVHD .

M[0] M[1] M[2] M[3] M[0] M[0] White Cell Platelet count counts ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ∆ ()• • • • • • ∆ • • () 150K • • • ∆ ∆ • • • • • 2000 100K • • • 1000

0 50 100 200 400 Time (days)

Figure 4. Temporal Abstraction in a Medical Domain. Raw data are plotted over time at the bottom. External events and the abstractions computed from the data are plotted as intervals above the data. BMT = a bone-marrow transplantation event; PAZ = a therapy protocol for treating chronic graft-versus-host disease (CGVHD), a complication of BMT; = event; • = platelet counts; ∆ = white cell counts; ∞ = context interval; = abstraction interval; M[n] = bone-marrow–toxicity grade n.

(energy). A human organism is specified by these 3 billion into subsegments for logistical three billion letters, arranged serially, that con- reasons, with an average length of 256 million stitute its genome. With 2 bits per DNA letter, it DNA letters. Genes are subsequences within the takes about 750 megabytes of data to specify a sequence of 3 billion that encode for particular human. There are 23 chromosomes that divide functions or structures that exist in your body.

72 AI MAGAZINE Articles

There are about 100,000 genes within a human culosis with these technologies (Behr et al. genome. More than 99.9 percent of the 1998; Blower, Small, and Hopewell 1996). Sec- genome is identical for all humans. And so all ond, clinical trials will have patients who are the diversity of human life is contained in the stratified by the differences and similarities in 0.1 percent that is different. One of the human their genes. We will be able to relate particular genes encodes a channel that allows a chloride clinical syndromes to particular treatments ion to pass from the outside of a cell to its based on a molecular understanding of the inside. This channel sometimes has a mutation basis of these syndromes. The diagnosis of that leads to the disease cystic fibrosis. An many diseases will become a simple lookup in understanding of how the DNA letters differ in the genome of the patient to see which variant patients with cystic fibrosis allows biologists to is present (Winters et al. 1998). We will be able begin to understand the mechanism of the dis- to focus treatments using this information, ease and ways to alter its course. The cystic once we have learned from the data the best fibrosis gene was isolated in a relatively expen- drugs to use against different disease/gene vari- sive, focused manner before the genome ations. Finally, we will have prognostic infor- sequencing project was under way. The logic mation beyond anything currently available behind a genome project is to isolate all genes because we will have access to the full genetic for an organism and catalog them using endowment of a patient and, when relevant, economies of scale. The primary genome pro- the infectious pathogens causing disease. In ject is the Human Genome Project, but there some cases, in fact, we may know decades are also genome projects for important model before a disease is evident that a patient is at organisms and organisms that cause human high risk for that disease. At this point, it is disease. The principal funding agencies for the important to mention the ethical, social, and genome project are the National Institutes of legal issues associated with the Human Health (via the National Human Genome Genome Project. A certain fraction of the Research Institute, www.nhgri.nih.gov/) and annual genome project budget is spent on the Department of Energy (www.er.doe.gov/ grants addressing these issues, including issues facepage/hug.htm). of privacy, ethical use of medical information, Associated with the genome sequencing patients rights to information, and the like projects are a number of other new technolo- (www.nhgri.nih.gov/About_NHGRI/Der/Elsi/). gies in biology that will allow data to be col- What’s the status of the genome sequencing lected on a large scale, never before possible. projects? Although this is not AI per se, it is Soon it will be possible to assess the complete useful to get a feeling for the amounts and set of genes that are active in a cell and com- types of data that are being generated. Consid- pare this set with the genes that are active in a er the GENBANK database of DNA sequences diseased version of the same cell (Marshall and (Benson et al. 1998). A recent release of that Hodgson 1998). Thus, we can find out which database contained 1.6 billion bases. (Remem- genes are active in a normal prostate cell as ber, there are 3 billion bases in a human). How- well as which genes are active in a prostate ever, this database contains DNA sequences cancer cell. The differences are the obvious from all organisms, and not just humans. Fig- places to look for new treatments of prostate ure 5 shows the growth in the size of this data- cancer. The differences may also provide new base since its inception in 1982. All these data ways to make a more sensitive and specific are available on the World Wide Web, and one diagnosis of prostate cancer in a patient. Final- of the remarkable aspects of the explosion of ly, the differences may be used to determine biological data is the ease in which it can be the likely prognosis of a particular prostate accessed, and so it becomes something of a cancer, based on its constellation of genes, and playground for information scientists who whether they are associated with an indolent need to test ideas and theories. Table 1 shows or aggressive type of cancer. the ranking of species in the DNA databases by Having defined all the genes in a biological the values of sequenced bases for each system, there are incredible opportunities for sequence. For example, we have roughly 700 information storage, retrieval, and analysis million bases of human genome sequence. The technologies. First, the epidemiology of dis- human genome is currently scheduled to be ease will now have a molecular basis. We will completed around 2003. Other organisms track the spread of infections by identifying include important laboratory test organisms the unique sequences of the offending bacteria (for example, mouse, rat, or fruit fly) or human and using this as a signature to follow its pathogens (for example, the HIV virus, malar- spread through the population. For example, ia, syphilis, or tuberculosis). One of the most investigators have tracked the spread of tuber- exciting challenges that arises as we learn the

FALL 1999 73 Articles

ear sequences of the DNA and of protein mol- Size of GENBANK DNA Database ecules (Durbin et al. 1998). Second, the technologies for defining 2500000000 ontologies, terminologies, and their logical relationships have been used to create formal 2000000000 theories for areas within biology (Schulze-Kre- mer 1998). Third, genetic algorithms and genetic pro- 1500000000 gramming have been used to create solutions that are in some cases superior to solutions cre- 1000000000 ated by hand (Koza 1994). Fourth, neural networks have been used, as 500000000 they have in many other fields, to achieve clas- sification performance that is often quite 0 impressive. The work in predicting aspects of DNA Bases in GENBANK three-dimensional structure from sequence information alone has received considerable Jun-88 Jun-90 Apr-94Oct-95Apr-97Oct-98 Dec-82May-86 Sep-92 attention (Rost and Sander 1994). Fifth, unsupervised cluster analysis of bio- GENBANK Release Date logical sequences and structures, including Bayesian approaches, has been successful in Figure 5. The Rapid Increase in Genetic Data Contained in a creating sensible categories within biological Major DNA Sequence Database. data sets (States, Harris, and Hunter 1993). Sixth, case-based reasoning has become very complete genetic background of bacteria is to important in areas that are still data poor. For develop comparative methods for understand- example, information about the three-dimen- ing how differences in genetic endowment cre- sional structure of biological molecules is still ate differences in preference for mode of infec- lagging behind the associated DNA sequence tion, type of infection, and virulence. Table 2 information. Thus, of the 100,000 proteins in shows some organisms whose genomes are a human, we only know the structure of about completely known. 700 of them. These examples represent valu- An international society (The International able “cases” that are constantly being used to Society for Computational Biology, extrapolate new information about the www.iscb.org/) has been formed to create a remaining 99,000 proteins (Jones 1997). community for researchers in biocomputing. Seventh, knowledge representation tech- In many ways, it is a spin-off from the AAAI niques have been used to represent the entire community. A conference entitled Intelligent set of metabolic capabilities of the bacteria Systems for Molecular Biology was first held in Escherichia coli. The resulting ECOCYC knowl- 1993 in association with AAAI (//ismb99.gmd. edge base has been used to infer metabolic de/). It subsequently was disconnected from capabilities and compare these capabilities AAAI for scheduling reasons but has been across organisms (Karp et al. 1998). remarkably successful. The proceeds from this Eighth, new knowledge representation and meeting provided the core funds to spawn the digital library techniques have been used to society. The name of the field (computational represent the complete literature of a subdisci- biology, bioinformatics, biocomputing, intelli- pline within molecular biology using ontolo- gent systems for molecular biology) has been gies for biological data, biological structure, the matter of some debate and often reflects and scientific publishing (Chen, Felciano, and the disciplinary training of the debate partici- Altman 1997). We have created a collaborative pants (Altman 1998). resource, RIBOWEB, that allows scientists to AI Contributions to Molecular Medi- interact with these data and compute with it cine I would like to briefly review some of over the web. the successes in the AI and molecular biology Ninth, intelligent agents are being designed community. In many cases, existing technolo- to assist biologists in understanding and min- gy has been transferred effectively, or new vari- ing the data that are accumulating from the ants have been created in response to the new high-throughput biological experiments. needs of the field. The National Center for Biotechnology Infor- First, Hidden Markov Models, developed mation (www.ncbi.nlm.nih.gov/) is the clear- originally for natural language processing, ing house for many sources of useful biological have become a powerful tool for analyzing lin- data. Their collection includes data about DNA

74 AI MAGAZINE Articles

Genes DNA Bases Species (Common Name)

1573906 1000128755 Homo sapiens (human) 403552 191558011 Mus musculus (mouse) 76540 142527757 Caenorhabditis elegans (soil nematode) 71079 78218600 Arabidopsis thaliana 56199 63799526 Drosophila melanogaster (fruit fly) 10581 28685645 Saccharomyces cerevisiae (baker’s yeast) 45849 28537572 Rattus norvegicus (rat) 4953 18023376 Escherichia coli 41866 17672014 Rattus sp. 32190 16498151 Fugu rubripes (puffer fish) 36345 16196521 Oryza sativa (rice) 9610 12068959 Schizosaccharomyces pombe 25383 11280798 Human immunodeficiency virus type 1 (HIV) 1094 9985595 Bacillus subtilis 4734 7009140 Plasmodium falciparum (malaria) 16688 6331052 Brugia malayi (filariasis) 5379 5922144 Gallus gallus (chicken) 685 5711838 Mycobacterium tuberculosis (tuberculosis) 5136 4648144 Bos taurus (cow) 10847 4413291 Toxoplasma gondii (toxoplasmosis)

Table 1. Number of Genes and Total Bases of DNA Sequenced for Various Organisms as of 10/98.

Aquifex aeolicus (bacteria that grows at 85° to 95° C!) Archaeoglobus fulgidus (bacteria that metabolizes sulfur, lives at high temperatures) Bacillus subtilis (ubiquitous soil bacteria) Borrelia burgdorferi (causes Lyme Disease) Chlamydia trachomatis (causes blindness in developing countries) Escherichia coli (can cause urinary tract infections, dysentery) Haemophilus influenzae (causes upper respiratory infections) Methanobacterium thermoautotrophicum (bacteria that produces methane, lives at 70° C) Helicobacter pylori (causes ulcers, maybe cancer) Methanococcus jannaschii (bacteria that produces methane) Mycobacterium tuberculosis (causes tuberculosis) Mycoplasma genitalium (smallest genome of known independent organisms) Mycoplasma pneumoniae (causes “walking pneumonia”) Pyrococus horikoshii (grows best at 98° C!) Saccharomyces cerevisiae (baker’s yeast) Treponema pallidum (causes syphillis)

Table 2. Some Completed, Fully Sequenced Genomes.

FALL 1999 75 Articles sequences, human genetic diseases, 7. We need systems to provide mission of Mycobacterium Tuberculosis. protein and nucleic acid structures, and document continuing educa- American Journal of Respiratory and Critical the biomedical literature, and other tion for physicians. Care Medicine 158(2): 465-469. important sources. The good news is Benson, D. A.; Boguski, M. S.; Lipman, D. J.; Evaluation Challenges that this information is easily avail- Ostell, J.; and Ouellette, B. F. 1998. GENBANK. Nucleic Acids Research 26(1): 1–7. able. The bad news is that it is not yet 8. We need demonstrations of the integrated in a manner that allows cost-effectiveness of advanced Berner, E. S.; Jackson, J. R.; and Algina, J. 1996. Relationships among Performance rapid discovery and inferencing. Thus, information technology. Scores of Four Diagnostic Decision Support a biologist with lots of time can even- 9. We need to create new medical Systems. Journal of the American Medical tually find the answer to a question knowledge with machine-learn- Informatics Association 3(3): 208–215. using these databases, but they have ing and/or data-mining tech- Blower, S. M.; Small, P. M.; and Hopewell, P. not yet been distilled to the point niques. Having established the C. 1996. Control Strategies for Tuberculosis where a physician in a clinical practice data infrastructure for clinical Epidemics: New Models for Old Problems. can use the data to guide clinical deci- data and biological data, there Science 273(5274): 497–500. sion making. The intelligent integra- will be unprecedented opportuni- Bodenreider, O.; Burgun, A.; Botti, G.; tion of biological information thus ties for gaining new knowledge. Fieschi, M.; LeBeux, P.; and Kohler, F. 1998. becomes one of the major bottlenecks Evaluation of the Unified Medical Lan- in progress for information processing 10. Finally, we need to ensure guage System as a Medical Knowledge (Markowitz and Ritter 1995). that there is equitable access to Source. Journal of the American Medical these technologies across patient Informatics Association 5(1): 76–87. Ten Challenges I want to end my and provider populations. presentation with the 10 grand chal- Bouhaddou, O.; Lambert, J. G.; and Mor- gan, G. E. 1995. Iliad and the Medical lenges to medicine I have previously House Call: Evaluating the Impact of Com- proposed (Altman 1997). These can be Conclusions mon Sense Knowledge on the Diagnostic divided into infrastructure, perfor- Accuracy of a Medical Expert System. In mance, and evaluation goals and are The Era of Diagnosis got things rolling Proceedings of the Nineteenth Annual summarized here: and created excitement as existing Symposium on Computer Applications in inferencing strategies were tested in Infrastructure Challenges Medical Care, 742–746. Salt Lake City, the real-world application domain of Utah: Applied Medical Informatics. 1. We need an electronic medical medicine. The current Era of Chronic Buchanan, B., and Shortliffe, E., eds. 1984. record based on semantically Disease and Managed Care has Rule-Based Expert Systems: The MYCIN Experi- clean knowledge representation changed the focus of our efforts. The ments of the Stanford Heuristics Programming techniques. coming Era of Molecular Medicine Project. Menlo Park, Calif.: Addison-Wesley. 2. We need automated capture of contains challenges that can keep Chen, R. O.; Felciano, R.; and Altman, R. B. 1997. RIBOWEB: Linking Structural Compu- clinical data, from the speech, information technologists busy for decades. tations to a Knowledge Base of Published natural language, or structured Experimental Data. Intelligent Systems for entry, in order to provide the data Acknowledgments Molecular Biology 5:84–87. required to move forward. RBA is supported by grants from NIH Cote, R. A., and Robboy, S. 1980. Progress 3. We need computable represen- in Medical Information Management. Sys- (LM-05652, LM-06422), NSF (DBI- tations of the literature. Both tematized Nomenclature of Medicine 9600637, agreement ACI-9619020), a clinical and basic biology data (SNOMED). Journal of the American Medical Faculty Scholar grant from IBM, and a should be structured and univer- Association 243(8): 756–762. gift from Sun Microsystems Inc. sally accessible for automated Dan, Q., and Dudeck, J. 1992. Certainty data analysis. Factor Theory and Its Implementation in a References Medical Expert System Shell. Medical Infor- Performance Challenges Altman, R. 1998. A Curriculum for Bioin- matics (London) 17(2): 87–103. formatics: The Time Is Ripe. Bioinformatics de Dombal, F. T.; Leaper, D. J.; Staniland, J. 4. We still need to do automated 14(8): 549–550. R.; McCann, A. P.; and Horrocks, J. C. 1972. diagnosis. Despite the passing of Altman, R. B. 1997. Informatics in the Care Computer-Aided Diagnosis of Acute its era, it is still worth under- of Patients: Ten Notable Challenges [see Abdominal Pain. British Medical Journal standing because there are times comments]. Western Journal of Medicine 2(5804): 9–13. it is useful. 166(2): 118–122. Department of Energy. 1996. Human 5. We need automated decision Barnett, G. O.; Cimino, J. J.; Hupp, J. A.; Genome Project. To Know Ourselves. Avail- support for providers who inter- and Hoffer, E. P. 1987. DXPLAIN. An Evolving able at www.ornl.gov/hgmis/tko/. Also act with patients episodically and Diagnostic Decision-Support System. Jour- available at www.ornl.gov/TechResources/ nal of the American Medical Association need help in making decisions Human_Genome/publicat/primer/intro.ht 258(1): 67–74. ml. about the treatment trajectory. Behr, M. A.; Hopewell, P. C.; Paz, E. A.; Detsky, A. S., and Naglie, I. G. 1990. A Clin- 6. We need systems for improving Kawamura, L. M.; Schecter, G. F.; and Small, ician’s Guide to Cost-Effectiveness Analysis access to information and expla- P. M. 1998. Predictive Value of Contact [see comments]. Annals of Internal Medicine nation for patients. Investigation for Identifying Recent Trans- 113(2): 147–154.

76 AI MAGAZINE Articles

Durbin, R.; Eddy, S.; Krogh, A.; and Mitchi- Miller, R.; Masarie, F.; and Myers, J. 1986. Medical Information Systems, 299–320. son, G. 1998. Biological Sequence Analysis. Quick Medical Reference (QMR) for Diag- Chicago: University of Illinois at Chicago. Cambridge, U.K.: Cambridge University nostic Assistance. MD Computing 3(5): Tufo, H. M.; Bouchard, R. E.; Rubin, A. S.; Press. 34–48. Twitchell, J. C.; VanBuren, H. C.; Weed, L. Feigenbaum, E. A. 1984. Knowledge Engi- Miller, R. A.; Pople, H. E.; and Myers, J. D. B.; and Rothwell, M. 1977. Problem-Orient- neering: The Applied Side of Artificial Intel- 1982. INTERNIST-1: An Experimental Com- ed Approach to Practice. I. Economic ligence. Annals of the New York Academy of puter-Based Diagnostic Consultant for Impact. Journal of the American Medical Sciences 246:91–107. General Internal Medicine. New England Association 238(5): 414–417. Feldman, M. J., and Barnett, G. O. 1991. An Journal of Medicine 307(8): 468–476. Winters, M. A.; Coolley, K. L.; Girard, Y. A.; Approach to Evaluating the Accuracy of Ohno-Machado, L.; Gennari, J. H.; Mur- Levee, D. J.; Hamdan, H.; Shafer, R. W.; DXPLAIN. Computer Methods and Programs in phy, S. N.; Jain, N. L.; Tu, S. W.; Oliver, D. Katzenstein, D. A.; and Merigan, T. C. 1998. Biomedicine 35(4): 261–266. E.; Pattison-Gordon, E.; Greenes, R. A.; A 6-Basepair Insert in the Reverse Transcrip- Heckerman, D. E.; Horvitz, E. J.; and Nath- Shortliffe, E. H.; and Barnett, G. O. 1998. tase Gene of Human Immunodeficiency wani, B. N. 1992. Toward Normative Expert The Guideline Interchange Format: A Mod- Virus Type 1 Confers Resistance to Multiple Systems: Part I. The PATHFINDER Project. el for Representing Guidelines. Journal of Nucleoside Inhibitors [in process citation]. Methods of Information in Medicine 31(2): the American Medical Informatics Association Journal of Clinical Investigation 102(10): 90–105. 5(4): 357–372. 1769–1775. Jones, D. T. 1997. Progress in Protein Struc- Rost, B., and Sander, C. 1994. Combining Wolfram, D. A. 1995. An Appraisal of ture Prediction. Current Opinion in Structur- Evolutionary Information and Neural Net- INTERNIST-1. Artificial Intelligence in Medicine al Biology 7(3): 377–387. works to Predict Protein Secondary Struc- 7(2): 93–116. Karp, P. D.; Riley, M.; Paley, S. M.; Pellegri- ture. Proteins 19(1): 55–72. ni-Toole, A.; and Krummenacker, M. 1998. Schulze-Kremer, S. 1998. Ontologies for Russ Altman is an assis- ECOCYC: Encyclopedia of Escherichia coli Molecular Biology. Paper presented at the tant professor of medi- Genes and Metabolism. Nucleic Acids Pacific Symposium on Biocomputing, 4–9 cine (medical informat- Research 26(1): 50–53. January, Maui, Hawaii. ics, and computer science by courtesy) at Koza, J. R. 1994. Evolution of a Computer Selby, J. V. 1997. Linking Automated Data- Stanford University. He Program for Classifying Protein Segments bases for Research in Managed Care Set- is interested in the appli- as Transmembrane Domains Using Genetic tings. Annals of Internal Medicine 127(8 Pt cation of advanced com- Programming. Intelligent Systems for Molec- 2): 719–724. putational techniques to ular Biology 2:244–252. Shahar, Y.; Miksch, S.; and Johnson, P. problems in biomedicine and molecular 1997. A Task-Specific Ontology for the Kulikowski, C., and Weiss, S. 1982. Repre- biology. He has an AB from Harvard Uni- Application and Critiquing of Time-Orient- sentation of Expert Knowledge for Consul- versity in biochemistry and molecular biol- ed Clinical Guidelines. In Proceedings of the tation: The CASNET and EXPERT Projects. In ogy, a Ph.D. from Stanford University in Sixth Conference on Artificial Intelligence in Artificial Intelligence in Medicine, ed. P. medical information sciences, and an MD Medicine, ed. E. Keravnou, 51–61. New York: Szolovits. Boulder, Colo.: Westview. from Stanford. He is a member of the Asso- Lathrop, R. H.; Steffen, N. R.; Raphael, M.; Springer-Verlag. ciation of Computing Machinery, the Deeds-Rubin, S.; Pazzani, M. J.; Cimoch, P. Shahar, Y.; Miksch, S.; and Johnson, P. American College of Physicians, the Amer- J.; See, D. M.; and Tilles, J. G. 1998. Knowl- 1998. The ASGAARD Project: A Task-Specific ican Association for Artificial Intelligence, edge-Based Avoidance of Drug-Resistant Framework for the Application and Cri- and the Institute for Electrical and Elec- HIV Mutants. In Proceedings of the Tenth tiquing of Time-Oriented Clinical Guide- tronics Engineers and is a founding mem- Annual Innovative Applications of Artifi- lines [in process citation]. Artificial Intelli- ber of the International Society for Compu- cial Intelligence Conference, 1071–1078. gence in Medicine 14(1–2): 29–51. tational Biology. He is a fellow of the Menlo Park, Calif.: American Association Shwe, M. A.; Middleton, B.; Heckerman, D. American College of Medical Informatics. for Artificial Intelligence. E.; Henrion, M.; Horvitz, E. J.; Lehmann, H. He is a past recipient of a U.S. Presidential Ledley, R. S., and Lusted, L. B. 1959. Rea- P.; and Cooper, G. F. 1991. Probabilistic Early Career Award for Scientists and Engi- soning Foundation of Medical Diagnosis: Diagnosis Using a Reformulation of the neers and a National Science Foundation Symbolic Logic, Probability, and Value The- INTERNIST-1/QMR Knowledge Base. I. The CAREER Award. His web page is ory Aid Our Understanding of How Physi- Probabilistic Model and Inference Algo- www.smi.stanford.edu/people/altman/. His cians Reason. Science 130(9): 9–12. rithms. Methods of Information in Medicine e-mail is [email protected]. Markowitz, V. M., and Ritter, O. 1995. 30(4): 241–255. Characterizing Heterogeneous Molecular Slee, V. N. 1978. The International Classifi- Biology Database Systems. Journal of Com- cation of Diseases: Ninth Revision (ICD-9) putational Biology 2(4): 547–556. [Editorial]. Annals of Internal Medicine 88(3): Marshall, A., and Hodgson, J. 1998. DNA 424–426. Chips: An Array of Possibilities. Nature States, D. J.; Harris, N. L.; and Hunter, L. Biotechnology 16(1): 27–31. 1993. Computationally Efficient Cluster Middleton, B.; Shwe, M. A.; Heckerman, D. Representation in Molecular Sequence E.; Henrion, M.; Horvitz, E. J.; Lehmann, H. Megaclassification. Intelligent Systems for P.; and Cooper, G. F. 1991. Probabilistic Molecular Biology 1:387–94. Diagnosis Using a Reformulation of the Szolovits, P., and Pauker, S. G. 1976. INTERNIST-1/QMR Knowledge Base. II. Evalua- Research on a Medical Consultation System tion of Diagnostic Performance. Methods of for Taking the Present Illness. In Proceed- Information in Medicine 30(4): 256–267. ings of the Third Illinois Conference on

FALL 1999 77