An Embedding Multitask Neural Network for Efficient Arrhythmia Detection

Abstract

This study addresses the critical need for prompt detection of life-threatening ventricular arrhythmias. We explore the application of neural networks within the constraints of Implantable Cardioverter Defibrillators to improve early arrhythmia detection. Our proposed neural network methodology leverages multitask learning, aiming to enhance detection efficiency by concurrently learning to identify ventricular arrhythmias and estimate RR intervals from intracardiac electrograms. Implemented on the NUCLEO-L432KC board, with limited memory and processing capacity, our approach achieved an Fβ score of 0.88, with a low inference latency of 59.96 ms. These results demonstrate the feasibility of integrating advanced neural network capabilities within Implantable Cardioverter Defibrillators (ICDs).