Automated Arrhythmia Classification System: Proof-of-Concept With Lightweight Model on an Ultra-Edge Device

Abstract

Arrhythmia can lead to severe complications and early detection of arrhythmia is crucial to prevent progression. Electrocardiograms are the most reliable measure for detecting arrhythmia. This study aims to implement a practical ultra-edge-computing system for automated arrhythmia classification, incorporating a lightweight deep neural network-based model and low-power wearable electrocardiogram sensing device. The lightweight model was designed based on the WavelNet architecture, which has been previously introduced for precise arrhythmia classification. Model compression methods including knowledge distillation, pruning, and quantization were employed to enhance arrhythmia classification performance while reducing computational complexity. The lightweight model was integrated into a wearable sensing device featuring a resource-constrained microcontroller unit. The efficiency of the lightweight model and edge-computing system was evaluated regarding arrhythmia classification performance and computational complexity. This study was conducted using a widely used publicly accessible database following a benchmark training and evaluation procedure. Compared to a standard convolutional neural network-based model which exhibited 81.5% overall accuracy, the proposed lightweight model achieved more precise arrhythmia classification with achieving 87.1% overall accuracy. The lightweight model also exhibited significantly reduced computational complexity, with 51.5% fewer parameters, 78.1% smaller model size, 40.3% fewer multiply-accumulate operations, and 79.6% reduced inference time. The completed edge-computing system featured sufficiently short inference time and low memory usage. The proposed lightweight model and ultra-edge-computing system demonstrated advanced performance in classifying arrhythmia, with a significantly reduced computational burden. This outcome ensures the practicality of the system to achieve on-device real-time arrhythmia classification in the real-world.

Keywords

• Edge computing
• knowledge distillation
• model compression
• on-device inference
• pruning