Beginnings of Artificial Intelligence in Medicine (AIM): Computational Artifice Assisting Scientific Inquiry and Clinical Art – with Reflections on Present AIM Challenges

Casimir A. Kulikowski’s 2019 article asks what early medical AI tried to model and why successful research prototypes did not automatically become useful clinical tools. It contrasts discovering biomedical knowledge with advising a clinician about an individual patient. Its method is historical overview and participant commentary, drawing on earlier publications and the author’s own experience, rather than a new clinical trial or a systematic comparison of treatment outcomes.^[1]

What the early systems represented

The article distinguishes several approaches to putting expertise into a program. DENDRAL used specialist knowledge to infer molecular structures from mass spectra: its immediate problem was scientific discovery, rather than choosing a patient’s treatment. MYCIN represented advice about antimicrobial therapy in modular rules, with confidence factors expressing uncertainty. Its rule-based framework was subsequently generalized as EMYCIN.^[1]

CASNET, which Kulikowski helped develop, combined a model of disease mechanisms with empirical associations among findings, diagnoses, prognoses, and treatments. Glaucoma was its application: a causal account of how disease progresses could support explanations of patient-management advice. By contrast, the Pittsburgh work that became INTERNIST attempted the much broader domain of internal medicine, organizing extensive disease knowledge and diagnostic findings. The commentary describes its initial testing on published clinical case descriptions and its later evolution into the QMR reference system.^[1]

These examples explain a research contribution beyond simply computerizing records: different representations made different kinds of inference and explanation possible. They also required expertise to be elicited, encoded, and maintained, rather than acquired automatically from a single dataset.^[1]

Demonstrations and clinical use

Kulikowski reports a successful public demonstration of CASNET at the Academy of Ophthalmology in 1976. That account and the use of case descriptions for INTERNIST establish the kinds of testing being discussed; this retrospective supplies no common denominator or comparative patient-outcome estimate across the systems. A demonstration of consultation advice is a different evidential claim from showing that routine care improved.^[1]

His broader assessment is that many early consultation systems became more useful for explanation and medical education than for clinical practice. He identifies clinicians’ limited time to interact with computers and the effort of keeping knowledge bases current as obstacles. These are the participant author’s historical conclusions, rather than a controlled causal estimate of adoption barriers.^[1]

Explanation and responsibility

Writing during the 2019 deep-learning boom, Kulikowski proposes combining the explanatory resources of earlier rule-based and causal systems with newer learning architectures. This is a research direction, rather than a demonstrated solution. He also distinguishes the responsibilities of biomedical researchers, practitioners treating particular patients, and healthcare enterprises pursuing economic goals. Predictions alone cannot settle how those responsibilities should be shared.^[1]

The article appeared online on April 25, 2019, in the Yearbook of Medical Informatics, volume 28, pages 249–256; its issue date is August 2019. Its personal perspective and assessments belong to that setting. Read alongside Some Expert Systems Need Common Sense, it makes a concrete distinction: representing specialist knowledge, fitting clinical work, and assigning responsibility are separate requirements for useful assistance.^[1]