Design and Implementation of an ARM-Based AI Module for Ectopic Beat Classification Using Custom and Structural Pruned Lightweight CNN

Abstract

This paper presents the development and testing of lightweight and power-efficient CNN models for ectopic beat classification, tailored for a compact-sized (25x45 mm) ARM-based (STM32H7) AI module. Methods: Two custom lightweight architectures (LMUEBCNet and SEmbedNet) were introduced, and their performances benchmarked against conventional models (AlexNet and VGG19). A structural pruning method, filter pruning via the Taylor score, was employed for pre-trained models' parameter optimizations. Further resizing the input image size to pixel-56/112 for the efficiency of the embedded system. Validation was conducted utilizing a combined ECG simulator and the PhysioNet MIT-BIH arrhythmia ECG dataset, adhering to the ANSI/AAMI EC57 standard. Results: SEmbedNet with parameter size of 0.04M and LMUEBCNet with 0.15M parameters reported accuracy of 99.9%/97.6%. In comparison, the pruned models of VGG19 and AlexNet, which had parameter counts of 0.10M and 0.07M respectively, achieved accuracy of 96.7%/94.4%. Post-structural pruning, VGG19 and AlexNet models saw reductions in parameters by thousands of times, with accuracy decreases of 2% to 5%. Notably, the proposed LMUEBCNet and SEmbedNet were vastly leaner by 930 and 320 times than VGG19 and AlexNet. Conclusion: The proposed ARM-based AI module integrated with custom lightweight CNN models offers superior accuracy-memory trade-offs with 0.04M parameters, which is less than 1% the size of conventional models. The AI module with compact size and a power consumption of only 0.4 watts achieves a classification rate of 4.2 segments per second and 99.9% accuracy. The AI module shows the potential to transform ECG monitoring devices into ECG analyzers.