One of the key uses of Bayesian networks in Human Reliability Assessment is to capture the probabilistic dependencies among the factors that influence human performance. Their ability to integrate uncertainty and contextual features makes them particularly suitable for safety-critical applications. In this study, we employ a data-driven Bayesian network approach to classify operator success in alarm management tasks using data from a formaldehyde plant simulator in which task complexity, alarm display configuration, and support level were experimentally controlled. Three classifiers, Naive Bayes, Tree Augmented Naive Bayes, and Pearl-Rebane augmented Naive Bayes, were evaluated under both constrained and unconstrained feature-selection approaches (mutual information filter versus greedy forward wrapper), incorporating both controlled variables and participant characteristics. Across 100 Monte Carlo cross-validation trials, the Pearl–Rebane model restricted to the three task-related features achieves a higher average AUC than both the Tree Augmented Naive Bayes model and the Naive Bayes model.