Seventy years of AI in medicine, from expert systems to language models

In August 1955, four researchers, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, proposed a summer workshop at Dartmouth College. Their proposal rested on the conjecture that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." The workshop took place in 1956 and gave the field its name, artificial intelligence.

Medicine became one of its first testing grounds. A Tech Anatomy episode in French retraced that history in early 2025. This article follows the same arc, focusing on systems built to support clinical decisions; drug discovery and medical imaging have their own Tech Anatomy articles, on sixty years of computers in drug discovery and on AI in medical imaging.

The expert systems, 1960s to the 1980s

The first influential system was not clinical. DENDRAL, started at Stanford in the mid-1960s by Edward Feigenbaum and the geneticist Joshua Lederberg, later joined by the chemist Carl Djerassi, inferred the structure of organic molecules from mass spectrometry data. Its designers interviewed chemists to turn their know-how into rules, a process that became known as knowledge engineering. DENDRAL is often described as the first expert system.

Medicine followed. At the University of Pittsburgh, the computer scientist Harry Pople and the internist Jack Myers built INTERNIST-I during the 1970s to diagnose diseases across internal medicine. It worked from a large knowledge base linking diseases to findings, and ranked candidate diagnoses by scoring how well each explained the patient's findings. An evaluation published in the New England Journal of Medicine in 1982 found its performance on difficult cases broadly comparable to that of hospital clinicians, though below that of the case discussants who were experts. A consultation took a long time to run, and the system remained a research tool.

At Stanford, Edward Shortliffe developed MYCIN in the early to mid 1970s for his doctoral work. It asked a physician questions about a patient with a serious bacterial infection, applied a few hundred rules, and recommended an antibiotic, with the ability to explain why. In a blinded evaluation published in JAMA in 1979, infectious disease specialists rated its choices for ten meningitis cases acceptable 65 percent of the time, against 42.5 to 62.5 percent for the five Stanford faculty specialists it was compared with. MYCIN was never used in routine care. The reasons usually given were practical as much as cultural: data had to be typed in by hand, computers were not present at the bedside, and the question of responsibility for a machine's advice had no answer.

In the 1980s, the Laboratory of Computer Science at Massachusetts General Hospital developed DXplain, which suggests possible diagnoses from a list of findings. When national distribution began in 1987, its knowledge base covered about 2,000 diseases, according to the laboratory, a figure the episode gave as 500. Unusually for its generation, DXplain is still maintained today, with more than 2,600 diseases.

The winters

Expectations ran ahead of results. AI went through periods of reduced funding and interest, usually called "AI winters", in the mid 1970s and again from the late 1980s, after rule-based systems proved costly to build, hard to maintain and brittle outside the cases their designers had anticipated. The slowdown affected the whole field, not only medicine.

Data and computing power, 2000 to 2020

What changed in the 2000s was the supply of digital data and cheap computing power, which made machine learning practical: instead of writing rules, developers let models learn patterns from examples. Deep learning, a family of methods based on large neural networks, took off after 2012.

IBM's Watson became the public face of the period after winning the quiz show Jeopardy in 2011. IBM then built a health business around it, including a system to recommend cancer treatments. The episode describes Watson as an assistant offering diagnoses with unmatched speed. The record is more mixed: investigations by the health news site STAT in 2017 and 2018 reported that Watson for Oncology struggled to meet expectations and, according to internal documents, had produced unsafe treatment recommendations in testing. In January 2022, IBM sold the data and analytics assets of Watson Health to the investment firm Francisco Partners.

Deep learning delivered clearer results on narrow tasks. In 2016, a Google team reported in JAMA an algorithm that detected diabetic retinopathy in retinal photographs with high sensitivity and specificity. In April 2018, the FDA authorised IDx-DR, the first device allowed to provide a screening decision for diabetic retinopathy without a clinician interpreting the image.

The COVID-19 pandemic tested the field at scale. Hundreds of models were built to diagnose infection or predict severity. A living systematic review in the BMJ found almost all of them at high risk of bias, and a 2021 review in Nature Machine Intelligence of 61 studies of imaging models concluded that none of the models was of potential clinical use, mostly because of small, unrepresentative datasets and weak validation.

Generative models

Since late 2022, large language models have added a different capability: producing text, summarising records, drafting letters and answering questions in plain language. They are the first general-purpose systems in this history, where all earlier ones were built for a single task.

Competing readings

One reading sees steady progress. Each generation solved problems the previous one could not, from encoding expert rules to learning from images to handling language, and the limiting factors, computing power and data, have eased.

Another reading stresses a recurring pattern. MYCIN, Watson and many COVID models performed well in a test or a demonstration and then met the practical realities of clinical work: data entry, integration into workflows, validation on new populations, responsibility for errors. On this view, the history is less a march of capability than a repeated encounter with the same implementation problems.

What this history does not settle

Comparisons across eras are difficult. Early systems were evaluated on small sets of selected cases, often by their own developers, and today's systems on benchmarks that may not reflect practice either. The fact that a system never entered routine use does not mean it was wrong; MYCIN's advice was good, and it still went unused.

What to watch

The question this history raises for generative models is whether they will clear the barriers that stopped their predecessors: integration into clinical software, evidence of benefit for patients rather than accuracy on tests, and a clear allocation of responsibility when they are wrong. The episode closed by saying the revolution has only just begun. Whether the pattern described here repeats with this generation is still an open question.

Listen to the episode

La Folle Histoire de l'IA en Santé

Listen · 8 min