A Decision-Tree Strategy for Combining Diagnostic Tests for Prediction

Abstract

In the context of medical screening, various diagnostic tests have been developed for determining whether a disease is presen in an individual. Similarly in the context of toxicological screening, a variety of short-term assays have been developed to predict whether a chemical would be carcinogenic if tested in a long-term bioassay. In both contexts, it is a challenge to combine the results of several predictive tests in a way that improves on the predictivity of the individual tests. Increases in positive predictivity can be accompanied by decreases in negative predictivity and vice versa. This article presents a decision-tree classification model for combining results from several independent short-term or diagnostic tests to quantify the likelihood of a true positive result (patient has disease, or chemical is carcinogenic). The decision-tree strategy determines the most advantageous sequence for conducting the predictive tests. The classification model is based on statistical confidence limits on the predictive probability of disease (carcinogenicity) rather than on the central estimate of the predictive probability. This model is applied to the assessment of the abilities of four short-term tests in the prediction of chemical carcinogenicity under the assumption of independence among the four tests, and is used to demonstrate a testing strategy for the application of three pancreatic cancer diagnostic tests.