This paper reports a comparison between two modes of computer-aided diagnosis in a real-time prospective trial involving 472 patients with acute abdominal pain. In the first mode the computer-aided system analysed each of the 472 patients by referring to data previously collated from a large series of 600 real-life patients. In the second mode the system used as a basis for its analysis “estimates” of probability provided by a group of six clinicians. The accuracy and reliability of both modes were compared with the performance of unaided clinicians.
Using “real-life” data the computer system was significantly more effective than the unaided clinician. By contrast, when using the clinicians' own estimates the computer-aided system was often less effective than the unaided clinician—especially when diagnosing less common disorders. It seems, firstly, that future systems for computer-aided diagnosis should employ data from real-life and not clinicians' estimates, and, secondly, that clinicians themselves cannot analyse cases in a probabilistic fashion, since often they have little idea of what the “true” probabilities are.