Fifty Years of AI in Medicine: The Reasoning Got Cheap, the Data Didn’t

In 1973, Charlie Safran, a logician from Tufts, went to MIT and started programming for a medical decision-making group. The group expected a computer-based doctor that could pass the Turing test by the year 2000. He doesn’t think the bet was crazy for its time. “I didn’t really understand what a doctor does,” he told us. “Diagnosis, in my opinion, is the least of what we do. It’s relieving pain. It’s supporting care.”

He went to medical school, then spent decades at Beth Israel building electronic health records, eventually as chief of the Division of Clinical Informatics at Beth Israel Deaconess. He is a past president and chair of AMIA. (Full disclosure: Steve has known him for decades, and the three of us recently co-wrote a paper.)

Rules that worked, and rules nobody could maintain

In 1986, a request to keep medication and problem lists for a satellite clinic grew into a full EHR. Clinicians typed far more than he expected: within a week, half a screen for a problem name wasn’t enough, and soon they wanted comments too. Once a controlled vocabulary pinned down which problems they were treating, rules of care followed, shown to clinicians at the moment they recorded a problem or ordered a test. By Charlie’s account, clinicians acted “five times faster and more than twice as complete” with a rule in front of them, including on things later shown not to be worth doing.

The cost came later. One health system he worked with had 700 interconnected rules at one hospital and 70 at another, and no way to carry a change between them. “Rules are a nightmare,” he said. Steve added that the guidelines behind those rules were “hopelessly poorly written” when he started, and still are. A word like “frail” means something to every clinician and nothing precise to a computer.

So, Leon asked, if rules come back as guardrails on large language models, do we inherit the maintenance problem? Charlie doubts the rules of care can be generated from scratch. He expects some mix of human-set goals and model customization, and scaling it, he said, “still seems daunting.”

What gets recorded

Almost no EHR records eye color, Charlie pointed out, because a second or two per encounter, across a billion encounters, is a lot of clinician time. What gets documented is what gets paid for and what fits in a clinician’s day. “It will always be biased,” he said, and that record is what models train on.

What persists

Leon’s summary at the close: the optimism of 1973 “was far from foolish. It was aimed at the wrong bottleneck.” Reasoning got cheapest, fastest, and the data is still the problem. Every decade has had its own version of AI, from Bayes in the seventies to expert-system rules in the eighties to large language models now. That is a reason to judge these systems by “what they do for patients and clinicians, not by what they’re called.”

The hard part was rarely proving a tool worked. Baby CareLink, which let parents see their premature babies in the NICU and talk with the care team from home, worked, even though Charlie’s team had to install DSL lines and computers in families’ homes to do it. The open questions were who pays, who gets it, and whether it reaches the truly sick. “The fundamental problem that we face in the United States,” Charlie said, “is how we pay for care.”

His own slogan points away from automating care: “High tech is high touch.” Use the technology to rebuild the connections families have lost, and to stop rushing to see a patient every ten minutes to meet salary goals.

Listen to the full conversation: https://practicalaiinhealthcare.com/episodes/#safran

You Might Also Enjoy

• S1E33 with Edward H. “Ted” Shortliffe — Another view of the same fifty years of clinical AI, from the creator of MYCIN, on what persists and what we’re forgetting.
• S1E27 with Charlie Harp — A framework for measuring the quality of clinical data, the problem Charlie Safran says never went away.
• S1E43 with Isaac “Zak” Kohane — Who controls healthcare data in the AI age, and why that decides what models can learn.