The computer and acute abdominal pain in Leeds
On 1 April 1972 the BMJ published a comparison that F. T. de Dombal's group at the University of Leeds ran from 1 January to 1 December 1971 on 304 patients with acute abdominal pain. The diagnosis to which the computer gave the highest probability matched the final one in 279 cases (91.8%); that of the most senior clinician to see the patient in 242 (79.6%).
Why it matters
Machine diagnosis was tested prospectively, in real time, at the bedside, on every patient of the unit who met the criteria, with the final diagnosis as the yardstick. The difference in error rates, 8.2% for the computer against 20.4% for the clinician in charge, is statistically significant. The authors themselves judged the system of practical value, but only in a small percentage of cases.
Criteria: abdominal pain as the presenting complaint, arising within a week, emergency admission through the receiving room, a patient able to give a history, and a diagnosis eventually made. Junior doctors scored 72.2% (house surgeon) and 77.0% (registrar); of all 858 clinicians' diagnoses 563 (65.6%) were substantiated. Among the 294 patients with diseases in the system's database the computer scored 94.9%; it classified 84 of 85 cases of appendicitis correctly; of the 137 patients it put in the non-specific abdominal pain group, 136 proved to belong there. Computer errors: 25 of 304; errors of the clinician in charge: 62; chi-squared 15.73, P = 0.0001. The system is described in the companion paper by Horrocks and colleagues in the same issue. The Bayesian analysis ran on an English Electric KDF9 in the university's computing laboratory, about 800 metres from the department of surgery, reached by a teletype in the department itself. Outside the KDF9's hours a desk-top Mathtronics 848 did the calculation, also with a variant of Bayes's theorem. The database on acute abdominal pain held 600 patients and took some three man-years to prepare. The system came into being at the beginning of 1971 and by the time of the paper had made 2,034 diagnoses in four areas. What the record does not claim: that the work rested on Ledley and Lusted's 1959 paper; the reference lists of both papers have only Lusted's 1968 book.