Current healthcare AI risks repeating past mistakes in patient record organization.
getty
Every technologist who works in medicine has to admit a debt to someone who wasn’t a technologist. For me, that person was Warner Slack. Dr. Slack’s unwavering commitment to making patients key partners and valuable resources in the healthcare team deeply inspired me. That commitment started with a simple observation: the patient record wasn’t serving the patient at all. Healthcare AI risks repeating that same mistake.
Before his passing, he talked to me about that epiphany: “It occurred to me that the way we were organizing our records was terrible. No index, no table of contents—all formidable problems. I was so busy putting down numbers that would conform to regulations that had no correlation with what was wrong with my patient, that I had little time for anything else.”
Slack wasn’t alone in noticing this. In 1968, Lawrence Weed published a paper in the New England Journal of Medicine titled “Medical Records that Guide and Teach.” In that seminal article, he introduced the concept of the problem-oriented medical record (POMR)—a gamechanger for patients and clinicians alike and, well before anyone used the term artificial intelligence, a first attempt at a problem that healthcare AI is still trying to solve.
The Record Weed Wanted
Before the POMR, the medical record was source based. A clinician would open up the record and see all the laboratory results in one section, the radiology results in another section, and the patient history and physical exam somewhere else. Weed felt that this disjointed approach hampered the way clinicians thought about problems. His answer was the POMR, which organized a patient’s chart around their specific problems, rather than by department or data source.
Weed also devised what became known as SOAP notes: Subjective, Objective, Assessment, and Plan. Every entry in the chart was meant to trace a single line of reasoning—what the patient reported, what the exam and labs showed, what the clinician concluded from it, and what came next. All of these elements fit together around a problem list. The idea was that you could go to the medical record, look at a problem list, and understand what was going on with the patient.
The model that revolutionized patient records.
Frank Naeymi-Rad
An Enduring Format, a Lost Objective
Weed’s fix took hold. The first page of every patient chart featured a problem list, which physicians used to document and track patient issues in a systematic way. Decades later, it became federal law: every certified EHR is required to maintain one.
However, somewhere along the way, the reason for the problem list was forgotten. Historically, EHR use has focused on billing and reimbursement—and problem lists became something coders translated into billing codes rather than something that captured a clinician’s actual reasoning. These codes often failed to capture the clinical reality of the patient’s condition, precisely because they were designed with insurance and reimbursement in mind, not clinical understanding. Today, healthcare AI risks the same failure, when systems optimized for a downstream use case neglect the reasoning they were supposed to preserve.
Warner saw this coming. Years later, near the end of his life, he took my hand and said to me: “Frank, don’t forget the patients.” He knew what happened when a record built to serve a patient starts serving something else—be it regulation or reimbursement—instead. And he didn’t want patients to suffer the consequences.
What Healthcare AI Owes Patients
I’ve spent my career trying to heed Warner’s call. The most important thing in the exchange of patient information is to preserve the clinical intent. Getting that right starts with the words themselves. In 1988, I published a paper called “A Feature Dictionary Supporting a Multi-Domain Medical Knowledge Base”—my first attempt at capturing clinical language precisely enough for a machine to understand it the way a physician does.
It was, in its own way, the same question Weed had asked in 1968: not just what happened to a patient, but why. I just came at it from another angle, not as a clinician reading a chart but as a computer scientist trying to teach a machine to understand what that chart actually meant. Now, nearly sixty years later, the technology has evolved, but the mission remains the same: preserve the meaning, protect the purpose, and serve the patient.
Leave a comment