Acil serviste kullanılan tıbbi sarf malzemelerinin makine öğrenmesi yöntemleriyle tüketim tahmini: Bir özel hastane örneği

Abstract

Objective: This study examines whether monthly consumption of high-volume medical consumables in an emergency department can be forecast using machine learning. Method: Six algorithms were compared using 480 monthly observations for 10 items from a private hospital in Ankara (2022–2025). Hyperparameters were tuned through chronological, month-grouped cross-validation. The best-performing model was selected solely from cross-validation results, while the held-out test set was used only for independent validation. Emergency-department visit volume was included in lagged form. Findings: Random Forest was selected as the best-performing model (cross-validation R²=0.792; SD=0.008) and achieved an independent test R² of 0.887. Ablation analysis showed no measurable incremental contribution of visit volume to Random Forest, although positive contributions were observed in linear models. Conclusion: The relationship between visit volume and consumption was limited at the product level. The model may support aggregate budgeting and human-supervised decision-making, but it is not sufficient as a standalone tool for automated product-level ordering.