Improving Diagnostics and Prognostics of Implantable Cardioverter Defibrillator Batteries with Interpretable Machine Learning Models

Abstract

Medtronic Implantable Cardioverter Defibrillators (ICDs) and Cardiac Resynchronization Therapy Defibrillators (CRT-Ds) rely on high-energy density, lithium batteries. Consistently high battery performance is crucial for this application. 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. We identify key features from the battery production data that offer physical insights to support future battery development. The proposed approach is validated on 21 different datasets, targeting several performance-related features, and outperforms a benchmark model from the literature specifically developed for this application.

Keywords: Batteries, Defibrillators, Machine learning, Generalized Additive Models, Diagnostics, Prognostics