Automated selection of high-risk patients with acute pancreatitis

One hundred six early features were retrospectively collected on 44 patients with acute pancreatitis and stored in a database programmed on a microcomputer. Intraoperative/autopsy findings or clinical course categorized patients as having severe (n = 16) or mild (n = 28) pancreatitis. The frequency of 88 early available variables was analyzed. Twenty significant differences were found and used to program a Bayesian prediction of severity, assuming their independence. The computer's performance was assessed by retrospective testing of the patients, and a further 47 (14 severe, 33 mild) were studied prospectively, with a comparison carried out involving the unaided prediction of the clinicians. Finally, all 91 patients were retrospectively tested with the early predictive Ranson's signs and the results were compared with the computer's output. In the prospective evaluation, the computer's performance was 89.4% accuracy, 100% sensitivity, 84.8% specificity, while that of the physicians was 65.9%, 71.4%, 63.6%, respectively, like the performances of Ranson's criteria. Methodologic problems were: selecting patients to study, based on wide inclusion and exclusion in hindsight of the false-positives; dividing them into groups, based on both pathologic and clinical criteria; and the assumption of independence within the variables. The use of the procedure in other populations requires a specific analysis of the discriminant variables, but this study suggests a method to compare several samples of patients which allows prospective and controlled clinical trials.