Towards Hardware Supported Domain Generalization in DNN-Based Edge Computing Devices for Health Monitoring

Area of Science:

• Biomedical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel in classification but face challenges in health monitoring due to robustness and resource constraints.
• Domain generalization (DG) is crucial for ECG analysis across different sensors and patients, especially on wearable devices.
• Integrating DG with ultra-low-power ECG accelerators for edge deployment is a significant technical hurdle.

Purpose of the Study:

• To provide a comprehensive overview of ECG accelerators and DG methods.
• To explore the synergy between DG and ECG accelerators for multi-domain health monitoring.
• To propose an efficient edge-based DG approach for DNNs in ECG analysis.

Main Methods:

• A novel approach using correction layers for deploying DG on edge devices.
• Limiting DNN fine-tuning to a single layer, keeping the core model unmodified.
• Algorithm-hardware co-optimization for emerging wearable systems.

Main Results:

• Reduced computational complexity (CC) for DG by over 2.5 times compared to full DNN fine-tuning.
• Minimal memory overhead achieved with the proposed correction layer method.
• Average F1 score increase of over 20% on generalized target domains.

Conclusions:

• The proposed correction layer approach enables robust DNN classification on edge devices for health monitoring.
• This method significantly reduces computational complexity and memory footprint for DG.
• Facilitates multi-domain ECG monitoring through algorithm-hardware co-optimized systems.