Multi-Feature Fusion and Compressed Bi-LSTM for Memory-Efficient Heartbeat Classification on Wearable Devices.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Wearable devices require efficient algorithms for real-time health monitoring.
• Electrocardiogram (ECG) analysis is crucial for diagnosing cardiac conditions.
• Existing methods often face challenges with memory constraints and noise in wearable applications.

Purpose of the Study:

• To develop a memory-efficient ECG-based heartbeat classification system for wearable devices.
• To enhance accuracy and robustness against noise and artifacts in ECG signals.
• To create a range of model sizes suitable for diverse wearable hardware.

Main Methods:

• Implemented a multi-feature fusion technique using time intervals and under-the-curve areas for ECG characteristic extraction.
• Utilized a compressed bidirectional long short-term memory (Bi-LSTM) network for sequence processing.
• Developed and evaluated multiple neural network sizes (tiny to large) and applied post-training quantization (INT8, dynamic range).

Main Results:

• Achieved high classification accuracy, with the large model reaching 96.4% and the tiny model 94.6%.
• The Bi-LSTM network resulted in a 28% smaller model size compared to conventional LSTM.
• Compressed models demonstrated state-of-the-art performance, with the tiny model using only 139kB memory at 94.6% accuracy.

Conclusions:

• The proposed multi-feature fusion and compressed Bi-LSTM approach offers a memory-efficient and accurate solution for ECG heartbeat classification on wearables.
• The developed models provide a scalable solution, balancing accuracy and computational resources for different wearable devices.
• Post-training quantization effectively reduces model size while maintaining high performance, enabling practical deployment on resource-constrained devices.