What are Medical Expert Systems?
Expert systems are a branch of Artificial Intelligence or "AI" which also includes robotics, natural language processing, and other applications. Such systems can be defined as the collection of hardware, software, data, and embedded knowledge that demonstrate characteristics of intelligence. To avoid a convoluted discussion, I omit, at this point, the problem of defining "intelligence."
We are provided a futuristic glimpse into AI with science fiction and computer systems such as HAL, the commanding supercomputer in the film, "2001: A Space Odyssey." While HAL is fictional, the use of intelligent systems has been explored for more than 60 years. Beginning with embedding human expertise in computer programs for geological exploration, and ending today in control devices on consumer appliances, we see stunning applications of AI and expert systems permeating every aspect of our lives.
Expert systems have been used throughout the manufacturing industry, such as the design of vehicle components for cars, trucks, trains, and aircraft. Subway systems in Sendai, Japan, Washington, D.C.and major cities throughout the world use expert system software to control acceleration and deceleration of trains, read signals, and stop within inches of the right spot on the platform. Most motormen on these trains exist primarily to reassure passengers. "Fuzzy logic," an AI application programmed into chip controllers is used in refrigerators, washers, dryers, anti-lock brakes, sensors, security systems, video cameras, cell phones, and a wide variety of consumer electronics. In spite of the widespread use of expert systems and AI in our society, the application of this technology in medicine has fallen far behind.
Medicine's Intellectual Wrong Turn
The teaching and practice of medicine, in my opinion, has had it mostly backwards for over 100 years, when based on the concept: Given this disease, what are the most likely patient symptoms or findings? In fact the appropriate first question should be: Given this set of symptoms (findings), what are the most likely diagnoses (in order of likelihood)?*
By invoking possible diseases, then looking backwards at the patient, we end up Ruling Out diagnoses by a kind of scattershot approach, listing orderless possibilities, the famous "differential diagnosis," a kind of intellectual end run. This theology is as easy to find in hospital grand rounds as in the wildly popular TV shows, "House MD" and "Grey’s Anatomy." One of the inevitable results of teaching disease first and foremost is the frequent accumulation of useless data by performing mindless, useless often dangerous testing amounting to $200-$300 billion a year.
To Rule In a disease, you must start with the patient, not the test. Sir William Osler, (1849 – 1919) a Canadian physician, who has been described as the Father of Modern Medicine famously said, "Listen to the patient, he is telling you the diagnosis." If one eliminates routine, prenatal, and non-illness care, at least 84% or more of all patients visiting U.S. physicians did so because of a symptom or sign. Unfortunately for clinical science in this century there is a black hole of missing data relating to symptoms and signs. In other words, there is an epidemiology of disease, but not of patient complaints.
Expert Systems Can "Learn"
By the word "learn", I mean that computing machines can indeed improve their performance with "experience." In diagnosis this can be programmed by incorporating stored outcomes (perfect cases), and substituting these verified results automatically into their previously-stored code. In this way the more "correct" answers yielded by testing in the real world, the more the stored program improves its own performance.
There are many examples of disastrous failures of AI. Over twenty years ago a jumbo jet crashed outside Paris, killing 300 people, one (rare) example of air disasters linked to computer (and AI) failures. The reason for the crash: the pilot could not override a malfunctioning computer. Expert systems, like all technology, can be badly designed, but when this occurs it is not the fault of the technology, but the fault of ourselves. In the design of such systems, working closely with the users themselves in all stages of development. is the key to success. It is not enough to hand over the problem, as is often done, to the computer programmer. Moreover, this collaboration between designer and user must continue after development, because expert systems like database software are living, evolving organisms.
AI in Medicine and Health Care
Although expert decision-support systems in medical diagnosis and treatment have been around since the 1960’s, they have remained until recently an academic exercise waiting desperately for the medical and health professions to reach out and apply them. The pace has been excruciatingly slow, but gradually we are seeing successful applications. For example, some hospitals have studied the use of expert system analysis in estimating the optimal length of stay for a given patient. But other potential applications of AI for decision-making, such as optimizing testing, screening, cost variables, various trade-offs, etc., have hardly been studied. These are challenging problems requiring skilled analysis of large amounts of data, high level programming, testing, and finally implementation.
The medical profession and health care management have historically lagged behind industry in the use of expert systems and decision support. Yet theses applications could save countless lives, not to mention hundreds of billions of unnecessary expense. Computers are everywhere, but their main function in the 21st Century continues to be in the world of entertainment, games, finance, and the almost universal creation and management of databases. The potential of AI and expert systems has been largely overlooked possibly because it is much easier to construct and interrogate databases than to conceptualize decision support applications. This is going to change-but slowly, and only under the pressure of a health care system lurching into financial life support.
* Expressed in mathematical notation, the question is Not p(S|D) – the probability (p) of Symptoms (S) given (|) Disease (D) (including signs, findings, lab results) but Rather p(D|S) the probability of disease given symptoms. These probabilities are called conditional (or Bayesian) probabilities, after the pioneering 18th Century British clergyman/mathematician, Reverend Thomas Bayes. (1702-1761)
Martin F. Sturman, MD, FACP
Copyright 2009, Mathemedics, Inc.
EasyDiagnosis is an automated online service that analyses existing medical symptoms and predicts likely causes and conditions. Click here to find out more about this unique service and click here to try one of our modules for free.