Resource Constrained CVD Classification Using Single Lead ECG On Wearable and Implantable Devices

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Wearable Technology

Background:

• Electrocardiogram (ECG) is crucial for cardiovascular disease (CVD) detection.
• Wearable devices offer accessible ECG monitoring, necessitating efficient algorithms.
• Resource-constrained platforms require optimized deep learning models.

Purpose of the Study:

• To develop a compressed deep learning model for CVD detection from ECG signals.
• To optimize a sophisticated model for micro-controller platforms in wearable devices.
• To minimize performance loss during model compression.

Main Methods:

• Knowledge distillation was used to compress a baseline deep neural network (teacher model) into a TinyML model (student model).
• Piecewise linear approximation was employed for model compression.
• The model was evaluated for compression factor, computational load reduction, and performance metrics.

Main Results:

• Achieved a ~156x compression factor, fitting within 100KB memory for wearable deployment.
• Reduced computational load by an estimated ~5782 times compared to state-of-the-art ResNet models.
• Demonstrated negligible performance loss (less than 1% in accuracy, sensitivity, precision, and F1-score).

Conclusions:

• The proposed ECG TinyML model is highly efficient for CVD detection on resource-constrained wearable devices.
• The small model size (62.3 KB) is suitable for deployment on micro-controllers and potentially implantable devices.
• Enables advanced smart healthcare ecosystems through on-device ECG analysis.