The history of computerized medical diagnosis is the history of intensive collaboration between physicians and mathematicians. As soon as electronic computers came into use in the 1950s and 1960s, the early diagnostic systems developed for different medical purposes. In the first papers published in 1959, Ledley and Lusted observed that physicians have an imperfect knowledge of how they solve diagnostic problems. They detailed the principles underlying work on Bayesian and decision-analytic diagnostic systems that has been carried out over subsequent decades. Also, they stated that both logic (as embodied in set theory and Boolean algebra) and probabilistic reasoning (as embodied in Bayes’ rules) are essential components of medical reasoning. Subsequently, logical systems based on ‘discriminating questions’ distinguishing among mutually exclusive alternatives developed by Bleich and his colleagues on Acid-Base and ELectrolytes program (ABEL). To this day, medical decision-support systems were applicable to narrow domains, especially those where it is fairly certain that only one disorder is present; when multiple independent disease processes usually interact in the one patient. Warner and colleagues in 1961 demonstrated that applicability of Bayes’ theorem to diagnostic problems is of more than theoretical interest. They developed one of the first medical application systems based on Bayes’ rule; however the first practical Bayesian system and one of the first medical diagnostic systems to be utilized at widespread clinical sites, was the system for the diagnosis of acute abdominal pain, developed by de Dombal.

Ability to implement programs on existing computers influenced the initial development of two dichotomous approaches to computer-based medical diagnostic systems: branching logic and probabilistic systems. In medical diagnosis, it is sometimes advantageous to reason categorically (causally or logically) and other times to reason probabilistically (actuarially). Yet two decades after the original contribution by Ledley and Lusted, Szolovits and Pauker provided a detailed discussion of deeper philosophical issues related to categorical and probabilistic reasoning. A third alternative to categorical and probabilistic reasoning combined features of both but retained a fundamental difference. That alternative was heuristic reasoning. The HEME program for diagnosis for hematologic disorders was one of the earliest systems to employ heuristics, and also one of the first systems to use, in effect, criterion tables for diagnosis of disease states. Programs that heuristically match terminology from stored descriptions of disease states to lexical descriptions of patient are conceptually similar to HEME. Gorry was an enlightened pioneer in the development of heuristic diagnostic systems employing symbolic reasoning. In a classical paper in 1968, he outlined the general principles underling expert system approaches to medical diagnosis that were subsequently developed in the 1970s and 1980s. Gorry proposed a formal definition of the diagnostic problem. He also pointed out the difference between the information value, the economic cost, and the morbidity or mortality risk of performing tests and described the ‘multiple-diagnosis’ problem faced by systems when patients have more than one disease and then suggested that the knowledge bases underlying diagnostic systems could be used to generate simulated cases to test the diagnostic systems. Gorry’s paper represented the intellectual ancestors of a diverse group of medical diagnostic systems, including the PIP (Present Illness Program) developed by Pauker et al., MEDITEL developed by Waxman et al., Internist-1 developed by Pople et al., QMR (Quick Medical Reference) developed by Miller et al., DXplain developed by Barnett and colleagues, ILIAD developed by Warner and colleagues and a large number of other systems. Shortliffe also introduced a clinical application of rule-based expert systems for diagnosis and therapy through his development of MYCIN in 1976. MYCIN used backward chaining through its rule base to collect information to identify the organism(s) causing bacteremia or meningitis in patients. The above systems are known as the first generation of clinical decision-support systems.

The conceptual basis for medical decision-support systems construction developed during the 1950s and 1960s, leading to exploratory and innovative implementations in the 1970s. Evolutions in medical decision-support systems during the 1980s and 1990s have been motivated by changes in hardware platforms and user interfaces, by philosophical changes in the perceived role of these systems, by new models for computerized diagnostic systems, and through expanded understanding of how to evaluate medical decision-support systems. One of the most important developments during the 1980s was the invention, evolution, and ubiquitous proliferation of the PCs. The PC made it possible for system developers to distribute clinical decision-support systems in a cost-effective manner for a large user community. The PC also encouraged development of new, sophisticated graphic user interfaces (GUI) for medical decision-support systems. The 1980s also heralded a new era of connectivity via local and national networks. With the advent of the computer came a change in philosophy regarding the development of medical diagnostic systems. The goal was to improve the performances of both the user and the machine over their native (unassisted) states. In addition, several innovative techniques were added in the 1980s and 1990s to previous models for computer-assisted medical diagnosis. The trend was to develop more formal models that add AI abilities to the successful but more arbitrary heuristic explorations of the 1970s and early 1980s; so systems based on fuzzy set theory and Bayesian Belief Networks were developed to overcome limitations of heuristic and simple Bayesian models. Reggia developed Set Covering Models as a formalization of medical diagnosis problem and on the other hand, neural networks represented an entirely new approach to medical diagnosis. Moreover, Adlassnig and others applied fuzzy set theory to diagnosis of medical conditions such as rheumatologic disorders and pancreatic diseases. Fuzzy set theory includes formal methods for addressing the incompleteness, inaccuracies, and in inconsistencies that are often found in medical data and medical knowledge. Also Bayesian belief networks, referred to as probabilistic causal networks or Bayesian networks, represented a mathematical formalism, consistent with the axioms of probability theory, that was developed to overcome the difficulties with data acquisition and reasoning associated with earlier, simple Bayesian approaches (especially the independence assumption). Meanwhile, Artificial Neural Networks and Case-Based Reasoning (CBR) methods promoted as medical diagnostic systems for focused diagnosis problems by a large number of research groups.

In conclusion it’s necessary to say when researchers in medical informatics encounter the term ‘medical decision-support system’, many think primarily of general-purpose, broad-spectrum consultation systems. However, a key distinction must be made in reviewing and analyzing medical diagnostic systems. There exist systems for general diagnosis (no matter how broad or narrow their application domains) and systems for diagnosis in specialized domains, such as interpretation of HIV genotype. The general notion of computerized diagnostic systems conveyed in the medical literature sometimes overlooks specialized, focused, yet highly successful, systems. Small, focused clinical decision-support systems are one of the most widely used forms of diagnostic programs and their use will grow as they are coupled with other automated medical devices.