ICD-BAS: Detecting Ventricular Arrhythmia using Binary Architecture Search for Implantable Cardioverter Defibrillators

Qing Lu, Zhenge Jia, Jingtong Hu, Yiyu Shi

Abstract: Ventricular arrhythmia (VA) detection plays an important role in preventing sudden cardiac death, and is one of the most crucial components in current implantable cardioverter defibrillators (ICDs). Recently, convolutional neural networks (CNNs) have been widely acknowledged to outperform traditional methods for arrhythmias detection. However, most modern CNN-based VA detectors have relied on complex network design and expensive computation, which are not affordable to implantable devices. Meanwhile, some networks are successfully tailored to meet high resource constraint but their designs are by heuristics and specific to the individual project. Therefore the existing methods are considered inexplicable, hardly generalizable, and publicly inapplicable. In this paper, we propose to use the common techniques namely quantization and architecture search for designing hardware efficient VA detectors. Specifically, we demonstrate binary neural network (BNN) is outstanding to perform low-power VA detection, and propose an extremely efficient architecture search framework, named ICD-BAS, for accurate yet lightweight BNN design. Experiment shows the proposed binarization method can lower the power consumption of multiple networks by up to 70% with less than 0.3% accuracy loss in classifying intracardiac electrogram (IEGM) segments for life-threatening VA. By jointly optimizing accuracy and computation complexity, ICD-BAS achieves 97.9% accuracy in detecting VA at an average power of 4.33 mW.