Optimization Models for Healthcare Equipment Replacement Policies Using Reinforcement Learning

Abstract

Healthcare organizations must reconcile cost, reliability, and patient safety when replacing clinical equipment, a task made harder in Saudi Arabia by Vision 2030's rapid expansion of healthcare infrastructure and by the periodic patient-volume surges of Hajj and Umrah. Fixed-interval and threshold-based replacement rules assume stable operating conditions and cannot adapt to this variability.  This study proposes a two-stage decision-support framework that couples supervised failure prediction with a reward-driven reinforcement-learning (RL) replacement policy. A representative predictive-maintenance dataset (124,494 records from 1,169 devices) was pre-processed into 99 engineered features. Logistic Regression, Random Forest, and Gradient Boosting classifiers were trained to estimate failure probability; the best-performing classifier informed a risk-stratification layer and an RL policy weighing maintenance, replacement, and failure costs against equipment uptime. Random Forest achieved the strongest discriminative performance (ROC-AUC = 0.849), and device-level failure history was the most influential predictor of future failure. A moderate risk-based intervention threshold (probability ≥ 0.5) minimized total lifecycle cost (US$860 per cycle), a 38.6% reduction relative to a reactive baseline (US$1,400), while an overly aggressive, low-bar threshold (≥0.3) increased cost more than eleven-fold by over-triggering replacements. Under the optimal policy, 70.6% of interventions were timely replacements, 17.6% were preventive maintenance, and 11.8% ended in unplanned failure. Coupling predictive analytics with reinforcement learning shifts equipment-replacement decisions from reactive, calendar-driven practice toward adaptive, evidence-based policies, offering measurable cost savings and reliability gains for Healthcare Technology Management units pursuing Vision 2030's digital-health objectives.