Lightweight element-wise product enhanced neural network for efficient arrhythmia detection on embedded devices

Cardiovascular disease has emerged as a critical global health concern, where early and accurate arrhythmia detection is essential for preventing life-threatening complications and improving clinical outcomes. While deep learning approaches have shown promise in automated ECG-based arrhythmia classification, the development of computationally efficient models for resource-constrained embedded devices presents significant challenges. This study aims to address this gap by proposing a lightweight model optimized for real-time arrhythmia detection on embedded platforms.

We introduce an elementwise-product enhanced lightweight model (EPLM) featuring spindle-shaped architectures integrated with depthwise separable convolutions. This design reduces model parameters and computational costs while maintaining robust feature extraction capabilities. A novel elementwise product fusion mechanism is employed to combine dual-path features, enhancing high-dimensional feature representation. The model is rigorously evaluated using the standard MIT-BIH, SVDB, INCART, and PTB databases, with performance metrics including classification accuracy, precision, recall, F1-score, computational complexity, inference time and power consumption. Practical validation is conducted via deployment on a Raspberry Pi 5 and a smartphone to assess real-time applicability.

The proposed model achieves classification accuracies of 99.10% and 98.85% on the MIT-BIH and PTB databases, respectively. It requires only 25,981 parameters and 525,400 floating-point operations per second (FLOPs), enabling a single inference time of 2.29 ms. Deployment on Raspberry Pi 5 and Android 10 ×86 virtual machine confirm real-time operation, demonstrating suitability for wearable applications.

The EPLM framework delivers ultra-efficient arrhythmia detection with minimal computational overhead, enabling continuous cardiac monitoring on embedded devices. This advancement holds promise for early clinical intervention, improved patient outcomes, and scalable deployment in resource-limited settings. The model's efficiency and accuracy underscore its potential to transform real-time wearable healthcare technologies.