Bridging WorldsOrder
← All articles

The Feature Dictionary: Defining Disease Through Clinical Terminology

By Frank Naeymi-Rad, PhD, MBA, FACMI · Nov 06, 2026

Trustworthy AI in medicine will depend on accurate clinical terminology.
Trustworthy AI in medicine will depend on accurate clinical terminology.

In the early 1980s, I joined a NASA-funded project that would shape the rest of my career and teach me why accurate clinical terminology is so critical. Dr. Max Harry Weil and his colleagues had developed the Medical Emergency Decision Assistance System (MEDAS) to help astronauts manage medical emergencies when communication with Earth was impossible. It covered 95 emergency diagnoses on a Nova minicomputer, and my assignment was to move it to an IBM portable computer for potential use aboard the space shuttle.

I saw an opportunity beyond space. If MEDAS could distinguish 95 emergencies, could it support physicians across every clinical specialty?

Teaching Computers How Physicians Define Disease

MEDAS used a Bayesian, multimembership model to weigh patient findings against competing diagnoses. Expanding it meant more than adding diseases. Each diagnosis had to be defined by the findings, observations, and relationships physicians use to recognize it, and those definitions had to come from physicians themselves.

I called this structured representation the Feature Dictionary. Working with physicians and medical students across cardiology, neurology, internal medicine, surgery, and other specialties, my team set out to capture expert clinical reasoning in a form a computer could interpret.

Physicians Use Different Clinical Terminology

As contributions arrived, a problem surfaced. Physicians often described the same finding in different words. One wrote “dyspnea,” another “shortness of breath,” another “SOB.” To a clinician, these meant the same thing. To a computer, they were three unrelated pieces of information.

The pattern sharpened as medical students moved into their specialties. Their knowledge increasingly reflected the language and conventions of their disciplines. They did not disagree about the medicine. They expressed it differently.

Without a way to reconcile those differences, knowledge from one specialty could not combine with another, and a multi-domain diagnostic system was impossible. We had set out to teach computers to recognize disease. We learned we first had to teach them the language physicians use to describe it.

Preserving Clinical Meaning

The answer was not to force physicians into a single vocabulary. Instead, we connected synonyms, abbreviations, and lexical variants to shared, explicitly defined clinical concepts, preserving the relationships diagnostic reasoning required. We described the approach in our paper, “A Feature Dictionary Supporting a Multi-Domain Medical Knowledge Base.”

That principle became the intellectual foundation of what is now called interface terminology: clinical language infrastructure that links the words physicians use at the point of care to standardized medical concepts. It later became the basis of the company I went on to co-found, whose terminology is now embedded in electronic health record systems across the country.

What began as a problem in medical artificial intelligence became a foundation for preserving clinical intent in digital healthcare.

Why It Matters in the Age of Generative AI

Four decades later, the field faces a remarkably similar challenge. Large language models can generate medical narratives, summarize records, and answer clinical questions. But fluent language does not guarantee preserved meaning. Two different expressions can represent the same concept, while two similar expressions can describe clinically different conditions.

When those distinctions are lost, the consequences reach beyond documentation into patient identification, population analytics, clinical research, and AI-assisted decisions. Clinical terminology is not only a documentation tool. It is part of the knowledge infrastructure trustworthy AI requires.

It also keeps the physician in the loop. When AI works from the clinician’s own concepts, the person responsible for the patient can trace, question, and correct its output.

The Lesson for the Next Generation

When I began expanding MEDAS, I wanted to build a system that understood disease across specialties. I learned that before computers could reason about disease, they needed a reliable way to interpret the language used to define it.

As a new generation builds AI for healthcare, that lesson matters more than ever. The future will depend not only on the intelligence of our models but on the accuracy of the clinical knowledge that grounds them.

Before AI can reason reliably about a patient, it must preserve what the clinician means.