OBJECTIVES: To empirically evaluate the predictive performance and risk factor identification of P-Course and compare it to logistic regression (LR) using patients screened for the presence of coronary heart disease (CHD). P-Course is a web-based naive Bayes classification (NBC) tool which special feature is its ability to utilize informative priors in model construction. METHODS: 597 CHD-suspected patients underwent coronary angiography. P-Course was compared to various forms of novel LR approaches (full LR, backward stepwise selection using Bayesian Information Criterion (BIC), and Bayesian Model Averaging (BMA) with medically supported interactions). Predictive performance was measured as the proportion of correctly classified patients. The dataset was randomly split into training and test sets. Performance was measured separately for four different sizes of training sets (150, 200, 300, and 450). For each size, the experiment was replicated 20 times to improve accuracy. RESULTS: P-Course outperformed LR approaches: the respective ranges for the average accuracies of P-Course, full LR, stepwise BIC, and BMA with same datasets were 0.78-0.81, 0.70-0.81, 0.70-0.80, and 0.78-0.80. In total splits modeled (N=3000-9000), P-Course predicted correctly on average 63 cases more than the best comparator, BMA. The analyses further illustrate that relevant prior information improves P-Course's accuracy, in particular when the training dataset is relatively small: average accuracies with informative and uninformative priors using training sets of sizes 25-50 were 0.70-0.71 and 0.66-0.68, respectively. In variable screening, P-Course yielded medically sensible choices of variables regarded as the most likely risk factors for CHD. CONCLUSION: In addition to previous work done with P-Course (e.g., Naïve Bayesian Fusion and decision rationality analysis), this analysis demonstrated the tool's additional value in comparison to the LR approaches. When the estimation of e.g. propensity scores or adverse events is of concern, NBC can offer additional predictive value compared to LR. In medicine, even one unnecessary faulty prediction is too much.