Improving diagnostics and prognostics of implantable cardioverter defibrillator batteries with interpretable machine learning models

Medtronic Implantable Cardioverter Defibrillators (ICDs) and Cardiac Resynchronization Therapy Defibrillators (CRT-Ds) rely on high-energy density, lithium batteries, which are manufactured with a special lithium/carbon monofluoride (CFx)-silver F x )-silver vanadium oxide (SVO) hybrid cathode design. Consistently high battery performance is crucial for this application, since poor performance may result in ineffective patient treatment, whereas early replacement may involve surgery and increase in maintenance costs. To evaluate performance, batteries are tested, both at the time of production and post-production, through periodic sampling carried out over multiple years. This considerable amount of experimental data is exploited for the first time in this work to develop a data-driven, machine learning approach, relying on Generalized Additive Models (GAMs) to predict battery performance, based on production data. GAMs combine prediction accuracy, which enables evaluation of battery performance immediately after production, with model interpretability, which provides clues on how to further improve battery design and production. Model interpretation allows to identify key features from the battery production data that offer physical insights to support future battery development, and foster the development of physics-based model for hybrid cathode batteries. The proposed approach is validated on 21 different datasets, targeting several performance-related features, and delivers consistently high prediction accuracy on test data.