A telemetry-free method for predicting failures of medical equipment based on operational event data

Abstract

Predictive maintenance is one of the key applications of machine learning in management of equipment lifecycle. Telemetry-based methods provide high accuracy but require substantial infrastructure and equipment modification, limiting their applicability to medical devices without built-in sensors. Classical statistical methods of analysis and reliability assessment (Weibull, Cox) do not capture the dynamics of event sequences, while process-mining approaches are not adapted to sparse service records. None of the considered areas of predictive maintenance methods simultaneously meets three key requirements of independence from telemetry, accounting for event dynamics, and operating on standard service logs. Thus, the aim of this study is to develop and verify a method to predict medical equipment failures based on discrete event data of operation, which does not require telemetry and uses only standard service logs. Binary failure classification has been performed using logistic regression with class balancing and gradient boosting (XGBoost) taking into account interpretability requirements and class imbalance (14 %). Validation has been carried out using object-wise splitting and stratified 5-fold cross-validation; the primary metric is ROC AUC. The method has been verified on operational data from electroencephalographic equipment of Neurosoft LLC (Ivanovo). A method to predict medical equipment failures based on discrete event data without telemetry has been developed and verified. A method for constructing a feature space from discrete event data with aggregation over 30-, 90-, and 180 day horizons has been developed. It is applicable to standard service-request logs. It has been established that the XGBoost model achieved a ROC AUC of 0,82 at the 90 day horizon, compared to 0,60 for logistic regression. SHAP analysis has revealed that a reduction in the interval between repairs to below approximately 281 days is a significant indicator of increasing failure probability. The ROC AUC value of 0,82 indicates good quality of the model under class imbalance, and the standard deviation of 0,13 indicates sufficient stability of the result. The model enables ranking the equipment fleet by failure risk over a 90 day horizon, planning unscheduled maintenance, and optimizing spare parts inventory. Further research includes validation of the proposed method on expanded samples, incorporation of workload features, and adaptation of the model to additional to 30 and 180 day prediction horizons.