Atrial Fibrillation Detection on Low-Power Wearables using Knowledge Distillation

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Embedded Systems

Background:

• AI adoption in low-power wearables is limited by computational and memory constraints.
• 24/7 monitoring via wearables is crucial for personalized healthcare but requires efficient, low-power systems.
• Cardiac arrhythmia detection demands sophisticated AI models that are often too large for wearables.

Purpose of the Study:

• To apply Knowledge Distillation to a convolutional-recurrent neural network for cardiac arrhythmia detection.
• To reduce the memory footprint and computational operations of the AI model.
• To enable the deployment of an accurate AI model on resource-constrained wearable devices.

Main Methods:

• Applied Knowledge Distillation to a pre-existing convolutional-recurrent neural network.
• Evaluated the memory footprint (parameters) and computational operations (MOp).
• Quantized the distilled network and tested its performance on a nRF52832 System-on-Chip.

Main Results:

• The distilled network achieved a 50% reduction in memory footprint (138K parameters) and operations (1.84 MOp).
• Quantization led to improved accuracy, with an F1 score increase from 0.779 to 0.828.
• The network successfully ran on a resource-constrained nRF52832 System-on-Chip.

Conclusions:

• Knowledge Distillation is effective for creating efficient AI models for cardiac arrhythmia detection.
• The optimized network is suitable for deployment on micro-controllers (e.g., ARM Cortex-M) in low-power wearables.
• This work facilitates continuous, real-time cardiac monitoring in wearable health devices.