SEmbedNet: Hardware-Friendly CNN for Ectopic Beat Classification on STM32-Based Edge Device

This study proposed a hardware-friendly-CNN-based hardware implementation system for screening electrocardiogram (ECG) ectopic beat with an STM32 ARM microcontroller-based embedded artificial intelligence (AI) edge device. In single heartbeat classification, continuous wavelet transformation based SEmbedNet and simplified AlexNet/GoogLeNet with different pixels of 56/112 of input size were compared to choose the best and most efficient combination to implement in the hardware. Five classes of the ectopic beat are followed by the ANSI/AAMI EC57 guideline in the MIT-BIH arrhythmia database, including non-ectopic beat (N), supraventricular ectopic beat (S), ventricular ectopic beat (V), fusion beat (F), and unknown beat(Q). Besides, this study performed the model through k-fold cross-validation and choose the best model for hardware implementation. The classification result showed that using a 5-layer CNN (SEmbedNet) with an input image of pixel 56 could get better performance than an 8-layer CNN (simplified AlexNet) with a total accuracy of 99.89%. Besides, the combination of SEmbedNet with an input image size of pixel 56 and STM32 can achieve the benefits of 1.3s and 1.1 W per heartbeat in the classification task, and it only takes about 4 seconds. Moreover, a multiple-STM32 cross-validation platform was built to reduce the validation time. It can process more than a hundred thousand heartbeats in 6.4 hours.