CASNET: a glaucoma model before ophthalmologists

At IJCAI-77 in Cambridge (22-25 August 1977) S. M. Weiss, C. A. Kulikowski and A. Safir described CASNET, a program for the long-term management of glaucoma that reasons through a causal network of disease states. At the 1976 meeting of the American Academy of Ophthalmology and Otolaryngology 49 ophthalmologists rated its advice: 95% found it clinically acceptable, 77% expert or very competent.

The model describes glaucoma as a network of pathophysiological states joined by causal links and ties observations, diagnoses and treatments to them; the advice concerns long-term management. The program is written in FORTRAN and runs in 35K words of memory on a PDP-10. Since 1975 researchers at five medical centres, joined in the ONET network, had been testing it; it has over 200 diagnosis and therapy statements. On an early sample of 40 difficult cases the model, as it improved, came to diagnose all 40 correctly. The ONET members estimate, by personal communications the paper cites, that the program reaches reasonable and often sophisticated judgments in about 75% of difficult cases. The 1976 survey table: expert 25%, very competent 52%, acceptably competent 18%, inadequate 5%. Weiss and Safir were at Mount Sinai School of Medicine, Kulikowski at Rutgers University. The journal paper of 1978 (Artificial Intelligence 11, 145-172, with Amarel as a fourth author) was read in its reprint in a 1984 collection edited by Clancey and Shortliffe. It says that in November 1976 the program was a scientific exhibit at the Academy's annual meeting, that ophthalmologists were invited to present difficult cases to it, and that 77% of those who entered cases described it as performing at an expert or very competent level; it gives the current system more than 100 states, 400 tests, 75 classification tables and 200 diagnostic and treatment statements. The reprint has neither the survey table nor the 40 cases nor the 75%, which come from the IJCAI paper; the collection's editors say its papers are reproduced generally with only minor modifications, and the reprint was not compared with the journal pages, so the record does not say the journal lacks them. On the month the documents differ: the calendar of an audiology newsletter printed in January 1976 announced the meeting for 6-10 October in Las Vegas. What the record does not claim: the day of the 1976 survey, which no document read gives, or its month, on which they differ. The IJCAI figures come from the re-typeset text of the proceedings, not from a scan of the print.