PO-04-154 A NOVEL 12-LEAD ECG-BASED DEEP LEARNING ALGORITHM TO PREDICT CARDIOMYOPATHY IN PATIENTS WITH PVCS

Premature ventricular complexes (PVCs) are prevalent in 40-75% of patients on ambulatory monitoring. However, a significant patient minority, PVC-induced cardiomyopathy develops. Therapeutic decisions such as medical therapy or ablation based largely PVC burden (which correlates with reductions left ejection fraction (LVEF), symptoms, and whether not has already developed. To create assess deep learning algorithm to predict LVEF reduction PVCs the 12-lead ECG. Using records from 5 hospitals within health system, we retrospectively queried GE MUSE system for ECGs documented ≥18 years age. values were obtained corresponding electronic records. Patients an abnormal prior first ECG excluded. The primary outcome was diagnosis LVEF≤40% 6 months PVCs. We trained resnet50 neural network utilizing Group K fold splitting. dataset included 373,904 which 44,674 remained analysis (36,140 training 8,534 testing). An ROC curve created, explainability plots representative positive samples created using GradCAM (this methodology provides “heat map” those aspects identifies most relevant its prediction). median age cohort 70±18 (42% female 58% male). From among 44,754 (18,819 patients), successfully predicted newly reduced months. AUC model 0.81. framework highlighted QRS complex ST segment during sinus rhythm; interestingly, itself important feature prediction. Deep can new-onset While rhythm appear be predicting decrease LVEF, morphology does