A DEEP NEURAL NETWORK TO IMPROVE SIGNAL DETECTION IN CONTINUOUS LOOP RECORDER MONITORING AND ENHANCE RISK STRATIFICATION IN HYPERTROPHIC CARDIOMYOPATHY

BACKGROUND Implantable loop recorders (ILRs) are EGM monitors allowing continuous EGM monitoring for up to 3 years. This enables providers to capture arrhythmias that may occur infrequently and briefly despite carrying significant implications for a patient. In patients with hypertrophic cardiomyopathy (HCM), 1-3s runs of non-sustained ventricular tachycardia (NSVT) has been associated with an increased risk of sudden cardiac death (SCD). ILRs can be used to detect these arrhythmias and aid in risk stratifying HCM patients according to their SCD risk. While these devices are effective at recording such arrhythmias, their ability to automatically identify and label them can be improved. Here, we enrolled 33 patients with HCM to test a novel ILR with bluetooth enabled technology (Abbott ConfirmRx). Using EGM data captured from 3 years of follow up, we applied a deep neural network to improve the automated signal labelling. METHODS AND RESULTS We trained a recently published deep neural network (DNN) to identify cardiac rhythms in 4 categories: supraventricular tachycardia, ventricular tachycardia, sinus bradycardia and atrial fibrillation. We trained this model using a dataset consisting of raw EGM transmissions from ILRs in 33 patients with HCM. EGM interpretations by an RN and MD electrophysiologist were used as the gold standard. Transmissions were spliced into 20 second segments and annotated as one of the 4 aforementioned categories. 762 EGM annotated segments were provided in random order to the DNN to train a model to classify raw EGM readings. Next, we tested the model using 327 new EGM segments to evaluate its classification accuracy (figure 1). Our results (table 1) show that the DNN achieved an overall arrhythmia identification accuracy of 98.2%. The model's sensitivity in identifying supraventricular tachycardia, ventricular tachycardia and atrial fibrillation was 97.2%, 84.2% and 100% respectively. CONCLUSION Automated labelling of EGM signals recorded by ILRs can be improved with the use of novel DNNs. Our DNN-based model achieved an overall classification accuracy of 98.2%. These data suggest that DNNs can serve as one method of improving ILR data acquisition and analysis. This may increase the clinical utility of these tools at detecting brief and paroxysmal but yet concerning arrhythmias such as NSVTs in patients with HCM.