A Temporally Validated Machine Learning Framework for Predicting Anti-Tuberculosis Drug Stockouts: A Synthetic-Data Evaluation

Abstract

Continuous availability of anti-tuberculosis drugs is essential for maintaining tuberculosis treatment services, yet future stockouts arise from interacting demand, inventory, procurement, and supply conditions that are not fully represented by static stock thresholds. This study evaluates a temporally validated machine-learning framework for future anti-tuberculosis drug stockout prediction using synthetic longitudinal data as a methodological test bed prior to evaluation with operational hospital data. Drug-day predictors were derived from four domains: demand and consumption, inventory, procurement and supply, and historical stockouts. A 14-day horizon was used as the primary prediction target, with 7- and 30-day horizons considered for sensitivity analysis. Logistic Regression, Random Forest, XGBoost, and LightGBM were evaluated using temporally separated training, validation, and locked out-of-time test periods and compared with reorder-point and coverage-based inventory rules. Model assessment included recall, precision, specificity, F1-score, ROC-AUC, PR-AUC, Brier score, calibration, alert burden, temporal robustness, feature ablation, and predictive explanation. Logistic Regression was selected during validation but its recall decreased to 0.556 in the locked test period, whereas LightGBM descriptively achieved a recall of 0.889 and XGBoost a PR-AUC of 0.960. The results demonstrate the importance of locked temporal evaluation, multidomain predictors, threshold selection, and operationally oriented assessment. These findings validate the analytical pipeline rather than establish real-world effectiveness; external evaluation using SIMRS data remains required.