Hybrid machine learning for predictive spare parts replacement in medical equipment

Abstract

Stability of medical equipment is critical for uninterrupted healthcare services, yet hospital computerized maintenance management systems (CMMS) records are often incomplete, heterogeneous, and partially labeled for spare parts usage. In the analyzed CMMS dataset, spare-part-related work orders accounted for approximately 65.1% of maintenance cases, indicating that spare-part readiness is a major contributor to maintenance response time and potential equipment downtime. This study proposes a semi‑supervised hybrid random forest-neural network (RF-NN) model to predict whether a work order will require spare‑parts replacement and to support proactive inventory planning. Retrospective CMMS work‑order records from six hospitals (N = 295) were pre‑processed using leakage‑safe pipelines (categorical one‑hot encoding, text feature extraction from fault descriptions, and standardization of numeric fields). Semi‑supervised self‑training was used to exploit unlabeled or weakly labeled records through iterative pseudo‑labelling. Models were evaluated using stratified 10‑fold cross‑validation and reported as mean ± SD. The hybrid RF-NN achieved Accuracy = 0.986 ± 0.018, precision = 0.980 ± 0.026, recall = 1.000 ± 0.000, and F1‑score = 0.990 ± 0.013, outperforming RF, NN, and Markov baselines. These results demonstrate the practical advantage of combining semi‑supervised learning with hybrid modelling to reduce missed spare‑parts cases, minimize downtime risk, and strengthen CMMS‑ready procurement decision support.