Lightweight and Low-Parametric Network for Hardware Inference of Obstructive Sleep Apnea

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Obstructive sleep apnea (OSA) is a serious sleep disorder with significant health risks.
• Overnight polysomnography (PSG) is the diagnostic standard but is costly and labor-intensive.
• AI-embedded wearable devices offer a portable alternative, but face hardware limitations (memory, power).

Purpose of the Study:

• To evaluate depth-wise separable convolution (DSC) as a resource-efficient AI model for real-time sleep apnea detection.
• To compare DSC performance against spatial convolution (SC) using electrocardiogram (ECG) and oxygen saturation (SpO2) signals.

Main Methods:

• Implemented DSC and SC models for apnea detection using single-lead ECG and SpO2 data from PhysioNet.
• Developed individual signal models and fusion models for both convolution types.
• Acquired and analyzed ECG and SpO2 signals for model training and validation.

Main Results:

• Fusion models outperformed individual signal models for both DSC and SC.
• The DSC-based fusion model was significantly more energy-efficient (9.4x to 11.3x) than SC-based models.
• DSC-based fusion model achieved comparable accuracy (~95%), precision (~94%), and specificity (~94%) to SC models.

Conclusions:

• DSC presents a viable, energy-efficient alternative for AI-driven 1-D clinical signal analysis, particularly for wearable apnea detection.
• While SC models show slightly higher accuracy, DSC offers a practical trade-off for resource-constrained edge devices.
• This research highlights the potential of DSC in clinical applications beyond traditional computer vision tasks.