CardioLike-Net, an edge-end arrhythmia classifier with quantization-aware-training, achieved over 96.4% inter-patient classification accuracy across 43 subjects with low hardware resource usage.
CardioLike-Net provides a highly accurate, resource-efficient edge-end arrhythmia classifier suitable for wearable ECG applications.
Wearable electrocardiogram monitors equipped with edge-end classifier is a transformative tool for managing cardiovascular diseases. Whereas, electrocardiogram pattern variations have posed difficulties in obtaining robust accuracy among different individuals. To address this issue, this paper proposes a novel neural network architecture, resembling the diagnosis procedure of cardiologists while utilizing the robustness of artificial intelligent models. Also, this work introduces a quantization-aware-training algorithm that enables weight and activation quantization within recurrent neural network layers, which has traditionally been difficult due to their temporal dependencies and complex internal structure. Evaluated on a widely used public electrocardiogram dataset, it shows over 96.4% inter-patient classification accuracy across 43 subjects, which is superior to other state-of-the-arts. The quantized model maintains accuracy with only a 0.1% loss while reducing model precision to 4-bit integer weights and 6-bit integer activations. When deployed on a field-programmable gate array, the proposed classifier consumes 24 mW at a maximum clock frequency of 40MHz with a latency of 420ms. Compared to the full-precision counterpart, the hardware resource usage is significantly reduced, including reductions of 47.2% in look up tables, 45.5% in flip-flops, 78.9% in block memories, and 93.0% in digital signal processors. The source code is available at https://github.com/xinziXu/CardioLike-Net.git . • Architecture: This work introduces a CardioLike-Net architecture, inspired by the diagnostic process of cardiologists, reducing feature redundancy and the adverse impact of electrocardiogram signal variations. • Quantization: A novel quantization-aware-training algorithm is proposed to quantize the recurrent neural network layers to low-bit while maintain high accuracy. • Implementation: An edge-end arrhythmia classifier with high inter-patient accuracy and low hardware resources is achieved.
Xu et al. (2025) studied Arrhythmia (n=43). CardioLike-Net with quantization-aware-training vs. State-of-the-art models and full-precision counterpart was evaluated on Inter-patient classification accuracy. CardioLike-Net, an edge-end arrhythmia classifier with quantization-aware-training, achieved over 96.4% inter-patient classification accuracy across 43 subjects with low hardware resource usage.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: