Computer Evaluation of Statistical Properties of Clinical Information in the Differential Diagnosis of Chest Pain

Abstract

A pilot study was undertaken to determine quantitatively from a large number of signs, symptoms and laboratory tests of patients with, the differential diagnosis of chest pain, which information items could serve as optimal descriptors and/or discriminators of disease. Data were obtained from 1238 patients. For each subject 429 questions of the »yes-no« type were answered and 69 numerical data collected. Incidence rates of signs and symptoms were considered as descriptors. Out of those exceeding an incidence rate of 25 percent, between 60 and 95 percent referred to medical history, depending upon the disease entity under study.Contingency table analysis and chi-square tests were used first to determine the discriminative power of various items. Historical data predominated again. Out of the total of 498 information items tested only 46 reached a chi-square level of 40 which was considered the minimum for efficient separation of diseases. Many items with high incidence rates contributed little or nothing to disease differentiation.To test the discriminative power of the identified signs, symptoms and laboratory data, discriminant function analysis was used. The number of items could be further reduced to less than 10. More than 95 percent of the 1000 patients with Coronary Artery Disease and Pneumonia could be classified correctly with this reduced set.Data reduction and identification of optimal descriptors and discriminators can be considered as one of the most important preliminary steps in computer analysis of clinical information.