Abstract 14675: Developing Convolutional Neural Networks for Deep Learning of Ventricular Action Potentials to Predict Risk for Ventricular Arrhythmias

• Published in: Circulation (New York, N.Y.) N.Y.), 2019-11, Vol.140 (Suppl_1 Suppl 1), p.A14675-A14675
• Main Authors: Selvalingam, Anojan, Alhusseini, Mahmood, Rogers, Albert J, Krummen, David E, Abuzaid, Firas M, Zaman, Junaid A, Baykaner, Tina, Clopton, Paul L, Bailis, Peter, Zaharia, Matei, Narayan, Sanjiv M
• Format: Article
• Language: English
• ISSN: 0009-7322; 1524-4539
• DOI: 10.1161/circ.140.suppl_1.14675

BackgroundIt is unclear if machine learning can be trained to recognize electrophysiological signal features which represent pathological remodeling in patients with coronary disease (CAD) and left ventricular (LV) dysfunction at risk for sudden cardiac arrest (SCA).ObjectiveDevelop and optimize alternative convolutional neural network (CNN) models to learn the shape of ventricular monophasic action potentials (MAPs) and predict SCA risk.MethodsIn 26 patients with CAD and LV ejection fraction ≤ 40%, MAP voltage time series (fig. A) were recorded at 1 kHz in right (RV) and left (LV) ventricles at EP study. CNNs were iteratively developed in keras (Python) and trained to labels of ICD therapy (0/1) at 752+493 days. The network was trained on 3062 MAPs and validated in 518 independent MAPs.ResultsPatients had age 62.0±18.7 Y, LVEF 28.0±8.3%. Fig. A shows final deep neural network (DNN) architecture comprising 3 convolutional and 2 recurrent neural layers. Training was performed using k-cross validation (CV) with k = 7. Fig. B shows training accuracy of the DNN at >99%. In several independent validation sets, DNN predicted SCA with accuracies up to 78%.ConclusionsWe present the technical development of deep neural networks for ventricular action potentials, which predicted SCA at 2-3 year follow-up. Future studies should determine if these results can be extended by alternative learning models and by combining additional electrophysiological features.