Abstract
In this paper, we present a memory-efficient ECG-based heartbeat classification for wearable devices enabled by multi-feature fusion and compressed bidirectional long short-term memory (Bi-LSTM). A multi-feature fusion technique based on time intervals and under-the-curve areas is proposed to extract the main characteristic points of the ECG waveform with high accuracy and robustness against noise and artifacts. A Bi-LSTM network is developed to process the input sequence in both forward and backward directions, resulting in higher accuracy and a 28% smaller network size compared with a conventional LSTM. Multiple neural networks with varying sizes, including tiny (84 k), small (150 k), medium (478 k), and large (1.25 M) models, are developed to achieve high accuracy across all classes. The overall accuracy is 96.4% for the large model and 94.6% for the tiny model, while the F1 score across all classes exceeds 89.1% and 85.1%, respectively. The proposed models, compressed using post-training quantization techniques, achieve state-of-the-art performance. The compressed large model with 8-bit integer quantization (INT8) achieves an accuracy of 88.4% with 1.3 MB of memory, while the compressed tiny model with dynamic range quantization achieves 94.6% accuracy with only 139 kB of memory.