An Auditable Decision Support System for Pharmaceutical Supplier Resilience Using Leakage-Aware Machine Learning

Pharmaceutical supplier audits use different evidence related to service continuity, quality, traceability, and temperature control. This study uses seven operational indicators to examine whether three resilience-related profiles, assigned by an internal committee of five members, can be reproduced. It does not aim to predict independently observed disruption outcomes. The dataset covers 12 months of 2025 and contains 13,318 recorded evaluations of 441 anonymized suppliers from seven distribution hubs. In the ten-fold nested evaluation, suppliers were grouped and tree depth was selected within the inner folds. The decision tree obtained an accuracy of 0.9532, with a 95% t-based confidence interval of 0.9503–0.9562. Its macro-F1 was 0.9552 (0.9524–0.9579). For the October–December temporal test, suppliers had already appeared in the January–September training data. The accuracy was 0.9499. A further time-aware evaluation, with suppliers kept separate between training and testing, gave an accuracy of 0.9496 within the same pharmaceutical distribution network. This result does not represent external validation. In the test data permutation importance results, order fulfillment performance, the positively oriented Good Manufacturing Practice (GMP) alert score, and cold-chain logistics made the largest contributions. The committee classifications had limited agreement with clusters obtained using K-means, hierarchical clustering, and density-based clustering. The level of agreement also changed depending on the method selected. The trained decision tree has a depth of five, and its decision paths can be examined by human reviewers. These results show the agreement of the model with the committee’s recorded profiles. They do not show that actual shortages, disruptions, or recovery times can be independently predicted by the model.