PO-01-211 LONG-TIME MACHINE-LEARNING PREDICTION OF COMPLEX SPATIOTEMPORAL DYNAMICS DURING FIBRILLATION IN LIVE EXPLANTED PIG AND HUMAN HEARTS

Fibrillation is characterized by complex spatiotemporal electrical dynamics often driven the continuous creation and annihilation of reentrant waves. While it has been shown that this can be chaotic deterministic, hard to describe its evolution, which could help define for example best times defibrillate using lowest energy. To date impossible predict waves once initiated in space, not only experiments but also simulations most up-to-date ionic cell models. accurately forecast cardiac fibrillation time space several periods advance. Using an autoencoder echo state network (AE-ESN) approach introduced earlier, we present a novel integrated deep-learning architecture, called convolutional (Conv-ESN), (ESN) incorporated into (CAE) architecture. signals from optical-mapping and/or numerical simulations, CAE model constructed trained learn compressed representation samples bottleneck layer. The ESN, placed between encoder decoder, receives extracted features predicts next step, then passed decoder reconstruct dynamics. We applied our developed Conv-ESN both 2D simulated data 41-variable O’Hara-Virag-Varro-Rudy ventricular high-spatiotemporal-resolution experimental datasets obtained optical mapping live explanted pig human hearts during fibrillation, representing After training, technique successfully these over clinically relevant prediction horizon, about 600 ms short training 5s VF up 2 s long 25s. Figure shows AP APD sites (prediction black), MAE all pixels same time. show first possible generated multiple (both experiments) technique. Accurate predictions patterns made 8-10 dominant-frequency