Comparing Diagnostic Decision Support Systems for Pneumonia

OBJECTIVE we set the threshold to achieve a sensitivity of 95% To compare a Bayesian network (BN) and an and then calculated the specificity, the positive artificial neural network (ANN) in diagnosing predictive value (PPV), and the negative predictive community-acquired pneumonia. value (NPV). Over the range of all thresholds, we also plotted sensitivity against 1 specificity to get BACKGROUND receiver operating characteristic (ROC) curves for the In the past decade, BNs and ANNs have been BN and ANN. We determined the area under the increasingly used as decision support methodologies receiver operating characteristic curve (AUC) as a in medicine. Their clinical applications include measure of overall accuracy for each network. We diagnosis, imaging, signal processing, analysis of tested for statistical difference between the AUCs laboratory data, and pharmacology. Despite this using the correlated area z statistic. broad range of applications, it is not always clear whether a BN or an ANN is more appropriate for a RESULTS particular clinical situation. In evaluating different At a sensitivity of 95%, the values for specificity, decision support methods, comparisons between BNs PPV, and NPV were: and ANNs or between BNs, ANNs, and other decision support systems are helpful. Such Specificity PPV NPV comparisons, however, are still scarce. The goal of BN 92.3% 15.1% 99.9% this study, therefore, was to compare a BN and an ANN 94.0s 18.6% 99.9% ANN. The networks were designed to diagnose community-acquired pneumonia (CAP). We The BN had an AUG of 0.9795 (95% CI: 0.9736, investigated the networks under the following 0.9843). The ANN had an AUG of 0.9855 (95% CI: aspects: the behavior of the networks at a sensitivity 0.9805, 0.9894). The difference between the AUCs of 95%, and the overall diagnostic accuracy of the was 0.9894).significant bet0e0044UC networks. was statistically significant (p=0.0044).