|
|
Technological Infatuation and the Collapse of Clinical Medicine
20th Century medicine, despite enormous advances in diagnosis and treatment, remains largely the anecdotal science of the 19th Century. This is reflected in the continuing emphasis in medical teaching and practice on disease description, preoccupation with naming, and the explosion of diagnostic and therapeutic technologies which drive our $1 trillion yearly health care expenditures.
Physicians, moreover, have become progressively alienated from the clinical approach, preferring instead to "plug the patient into the laboratory" in a misguided and mindless search for a quick diagnosis. Medical practice has, in fact, become increasingly characterized by a dangerous reflexive behavior, driven not by experience or intellect, but rather by a perceived need to "do something" as fast as possible. This trend has accelerated in recent years because of cost containment pressures generated by utilization and peer review, monitoring and regulatory agencies, as well as fiscal intermediaries.
The True Nature of Illness
The practice of medicine requires a profound understanding not only of disease, but also of non-disease. For hundreds of years medical thinking has been driven by the nominalist approach in which we name something in order to substantiate its existence. By doing this we escape the need for dealing with the clinically inexplicable, what insightful clinicians have often described as "we see this." For centuries we physicians have regularly succumbed to a reductionism whereby we identify patient complaints as disease entities, when in fact, many so-called named "medical conditions," as in Thomas Szasz's formulation, do not exist, they are invented. For example, more than half the patients with chest or abdominal pain, (the first and second most common medical causes of emergency room presentation), and patients with back pain - the three categories accounting for the majority of hospitalizations in this country - do not have a precisely nameable diagnosis.
The present dogma which is taught in medical schools and forms the basis for medical reimbursement, whereby multiple diagnostic possibilities are invoked, is backward since it ignores the statistical nature of signs and symptoms occurrence and their relation to disease or non-disease. In terms of conditional probability, classical medical teaching and much of practice is mirrored in our obsession with possibilities (not probabilities), disease-oriented literature, and CD-ROM-style (textbook) databases: given a patient we invoke lists of unabridged possibilities, the differential diagnosis." Diagnosis and treatment, however, should be regarded as a hierarchical order of limited probabilities: In Bayesian probability terms, we are presented with a patient, given a specific clinical complaint or complaints. Thus, we have probability of disease (D), and a collection of signs and symptoms (S). In mathematical notation, our diagnostic question is always, p(D|S), probability of disease, given the symptom(S) and signs, NOT p(S|D), probability of clinical findings, given the disease.
The only way out of this flagrant intellectual blunder is by reorienting medical teaching back toward the clinical approach, and medical care and reimbursement toward disease definition derived from probabilistic analysis of signs and symptom. The official diagnostic codes (International Classification of Disease, ICD) first developed by the World Health Organization, now modified in this country as ICD-10-CM and used throughout most of the world are essential for the coder in the record room as well as the clerk at the insurance company. Yet an enormous number of routine presentations present a difficult exercise in equating clinical reality with reimbursement criteria. There are ways, however, of reconciling what many consider being the irreconcilable.
Medical Expert Systems and Deep Blue
Expert system software, often mistakenly called "Artificial Intelligence" or "AI" found its earliest application in the geological sciences. Today, such software is being widely applied in even the soft sciences, industry, and, in the form of smart or "fuzzy logic" controllers in a wide variety of consumer goods. Although attempts were made more than 40 years ago to develop and apply expert systems to the medical sciences, the field languished for decades. This came about through unrealistic expectations based on a fundamental misconception of their applications to clinical medicine. Early dedicated researchers, usually academically-oriented, believed, as memorialized in the time-honored Clinicopathological Case Conference, that the fundamental problem in medicine was the difficult diagnosis, usually rarely seen disorders. Ambitious, global programs such as Internist I evolving into later systems, attempted to embed enormous knowledge bases into scores of "clinical algorithms," i.e. deterministic flow charts. Obscure as well as common diagnoses were addressed, but little attempt was made to deal with symptom presentations such as headache, nonspecific chest, back, and abdominal pain, etc. This inability to deal with the probability of disease given a presentation was a direct outgrowth of the paucity of quantitative clinical knowledge and the plethora of laboratory data.
Physicians have always been fascinated with the exotic and the bizarre. Mundane, common misfortunes from which the overwhelming percentage of patients suffers have never enjoyed the intellectual fashion they so deserve, perhaps because the familiar has, in the phrase, always been seen as contemptible. Yet, what observing practicing clinician has not daily been reminded of these platitudes? Common conditions are overwhelmingly common; exotic conditions are overwhelmingly rare, and Many clinical presentations cannot be given an official diagnostic ICD code and therefore are not nameable or classifiable. Thus they persist in being undiagnosable ("unverifiable") by classic imaging or laboratory criteria.
With IBM, Deep Blue, and Gary Kasparov, the public has again been astonished and enthralled by expert systems, as if they did not exist for the past 4 decades. It should come as no surprise to the scientifically- minded that a computer chess program would ultimately be created to defeat the best human players(not even consistently, though the last match was won decisively by the computer).. Unfortunately, during the last encounter between Kasparov and the IBM computer, there were dark rumors of human interference in the "tournament." Now with a new World Chess Champion, other companies are working on the problem, and IBM has abandoned the project. That a digital computer would ultimately prove to be a better chess player than a man should occasion neither astonishment nor anxiety. After all, man did not merely design the program, more fundamentally, he invented the game of chess, something no computer is ever likely to do. If a computer could really "create," possibly the only true test of "intelligence," the act itself would be contradictory, a self-referential paradox: a program to create would itself have to be created by a human programmer.
Decision-making in medicine, of course, shares many similarities to certain types of game playing. In medicine, however, the problem is not nearly so enormous as in chess; compared to chess, in fact, programmable medical knowledge, in terms of knowledge- engineering, database size, and strategies is much more manageable. At the outset, we need to conceptualize the problem as a collection of smaller problems, and apply the strategy of "divide and conquer." To proceed effectively, we need the requisite epidemiological information, a generous set of patients, reliable clinical literature, and a collection of experienced clinicians. Even if we accept the fundamental uncertainties of all diagnosis, probability-based domain-specific expert medical systems could profoundly improve the practice of medicine. If such programs effectively dealt with the majority of routine clinical presentation, the savings to the health care system would be staggering.
MatheMEDics' Experience
Our experience in developing, designing, and testing medical expert system software has progressed enormously over the past 17 years. During this period the principals have had extensive experience in developing special probabilistic and other mathematical strategies as well as techniques of "knowledge- engineering," the daunting process of interrogating medical specialists and reducing that information, when possible, to numeric code. During this long period, we have progressed far on this expert medical system "learning curve." Moreover, we have had sufficient experience with clinical trials to appreciate their value as well as their pitfalls. We are now conducting retrospective and prospective validation studies with the collaboration of major medical centers.
Summary
Expert system conceptualization and techniques developed by MatheMEDics® are ideally suited for the creation of verifiable computer programs designed to optimize diagnostic and therapeutic strategies. These approaches have obvious implications, not only in various clinical , teaching, and managed care settings, but also for patients,.
In addition to probabilistic methods using a Bayesian formulation, we are now adapting a larger repertoire of mathematical techniques, and developing self-referential programs which can improve their own performance with the experience of stored clinical (perfect) cases. Building practical expert clinical systems is a process of formalization, systematization, and verification of domain-specific clinical knowledge and experience. The clinical conceptualization, organization, and programmatic presentation, perhaps more than the algorithmic construction remain key elements.
A critical requirement of any successful medical expert system is the ergonomic one, i.e. how the human user interfaces with the computer. Careful attention to this aspect of the software, it's "user-friendliness" qualities, spells the difference between success and failure, no matter how good other components of the software may appear. The programs must be compact, easy to use, and applicable PRIMARILY to the most common clinical presentations. In our opinion, our symptom and condition expert systems modules are an important initial step in fulfilling these requirements.
Martin F. Sturman, MD, FACP
|
| 
| 
| 
| 
| 