Implementation of machine learning to improve implantable cardioverter-defibrillator detection algorithms

Abstract

Background: Sudden cardiac death is a leading cause of death in the US and globally. Implantable cardioverter-defibrillators (ICD) prevent this through electrical therapy, but inappropriate therapy for non-life threatening heart rhythms remains pervasive. Objective: The goal of this study was to improve upon current ICD discrimination algorithms by using supervised machine learning techniques on an annotated database of ICD electrograms (EGM) preceding therapy to discriminate between appropriate (App) or inappropriate (InApp) therapies. Methods: A total of 54 EGMs of therapy events adjudicated by cardiologists were digitized from 49 cases. The signals were analyzed within either a single long window, or four short overlapping windows preceding therapy. The discrimination between App and InApp therapies was done using EGMs recorded over specific windows by separately calculating RR-based and nonlinear dynamic (NLD) based metrics, and creating RR- and NLD-scores, respectively. Linear and quadratic discriminant analysis (LDA and QDA) were then used on the obtained RR- and NLD-scores to predict the App or InApp therapy. These results were then compared to the App and InApp designation by cardiologists. Error rates based on incorrect classifications were used to evaluate the performance of both techniques. Results: We demonstrated that the optimal windows for LDA and QDA can both greatly improve upon modern error rates, with our QDA error going as low as nearly 2% when using an optimal window, as compared to the errors up to 25% found in recent studies. Despite QDA having a lower overall error across all windows, the QDA error came mostly from more dangerous false negatives, which made up 100% of misclassified points in the longest temporal window, and in the temporal window closest to therapy, whereas the majority of LDA error came from less dangerous false positives, which made up 100% of misclassified points in the same two windows. Conclusions: This novel strategy shows promise for use in retrospective discrimination of inappropriate and appropriate ICD therapy events and could improve real time decision making algorithms in ICDs. Additional studies should be completed to assess its utility in real time decision making and case adjudicated.