A NEURAL-NETWORK-BASED DIFFERENTIAL-DIAGNOSIS ASSESSMENT INSTRUMENT

Medical educators have been unable to produce convincing evidence of t he construct validity of written or simulation-based assessments of di fferential diagnosis (DDX) competencies. In 1987, a team of investigat ors at our institution introduced preliminary reports regarding the ps ychometric properties of an artificial intelligence-derived DDX assess ment instrument. These investigations produced evidence of the constru ct validity (experts' DDX performance > novices') of the measures deri ved from this instrument, a linear, fuzzy set-like expert system. In t his investigation, the authors used a non-linear, ''Back Propagation'' neural network as a DDX assessment instrument. An Acute Chest Pain kn owledge base was acquired from each of twenty-four board certified eme rgency medicine specialists and seventy-four junior and senior medical students. The neural network used these knowledge bases to simulate a nd assess each subject's individual DDX performance against twenty Acu te Chest Pain/Myocardial Infarction test cases. Student-t test reveale d that the DDX performance of experts was significantly superior to no vices (p < .001). This finding provides converging evidence of the val idity of DDX performance measures produced by both linear and nonlinea r, artificial intelligence-derived assessment instruments. These instr uments may prove to be a useful and powerful new assessment methodolog y.