Scenario-Based Optimization of RUL Models for Prescriptive Maintenance in Medical Imaging Equipment

Medical imaging systems rely on high-precision components that degrade over time due to thermal stress, mechanical wear, and electronic drift. Unanticipated failures can disrupt clinical workflows and incur significant operational costs. This study focuses on the optimization of machine learning (ML) based models for Remaining Useful Life (RUL) prediction, aimed at enhancing the effectiveness of prescriptive maintenance strategies in medical imaging devices. The developed framework adopts failure-specific predictive models (e.g. CT detector) that trigger targeted actions to prevent or resolve issues. Models are optimized to align with distinct service strategic goals under multiple operational scenarios, such as minimizing part replacement costs, ensuring service-level agreement (SLA) compliance, maximizing uptime, or improving maintenance efficiency by reducing service cost.
The approach leverages multivariate time-series data from historical system logs and service records. Feature engineering includes high-order statistical transformations to enhance signal quality. Feature selection is guided through statistical hypothesis test and domain expertise to isolate failure-relevant features. The labelled failurespecific datasets feed into a supervised ML pipeline to predict RUL classes for specific degradation modes. Model performance is evaluated using precision-recall metrics across multiple failure modes. We focus onto two scenarios: (1) balancing uptime and maintenance efficiency, where we balance confidence of the RUL predictions with coverage (e.g. recall > 0.75, precision > 0.75) to obtain both service productivity and asset owner satisfaction; (2) minimizing part replacement costs for high-cost components, where high-confidence RUL predictions (e.g. precision > 0.95 ) are essential to enable automated part ordering via integration with inventory management systems, ensuring timely availability and reducing unplanned downtime. These two scenarios integrated with specific, actionable recommendations for maintenance tasks demonstrate the feasibility of our approach implementation in healthcare prescriptive maintenance area.

Keywords: Predictive Maintenance, Prescriptive Maintenance, Machine Learning, Medical Imaging Devices, System Log, Remaining Useful Life, Degradation Mode.