Event-Level Identification of Sleep Apnea Using FMCW Radar.

Area of Science:

• Biomedical Engineering
• Medical Devices
• Artificial Intelligence in Healthcare

Background:

• Sleep apnea diagnosis is hindered by costly and inconvenient methods like Polysomnography (PSG).
• Current wearable devices lack the quality control and patient compliance needed for effective screening and monitoring.
• There is a critical need for accessible, accurate diagnostic tools for widespread sleep apnea assessment.

Purpose of the Study:

• To develop a contactless, low-cost, and accurate method for event-level sleep apnea detection using FMCW radar.
• To enhance detection accuracy by integrating SpO2 prediction as an auxiliary task within a multitask-learning framework.
• To evaluate the system's performance against established methods for sleep apnea diagnosis and severity grading.

Main Methods:

• Utilized frequency-modulated continuous-wave (FMCW) radar technology for non-contact signal acquisition.
• Developed a novel deep-learning model based on U-Net architecture with self-attention and SE modules for event-level segmentation.
• Implemented a multitask-learning framework incorporating SpO2 prediction to improve feature extraction and clinical relevance.

Main Results:

• Achieved high accuracy in event-level sleep apnea segmentation with an F1-score of 0.8019.
• Demonstrated superior performance compared to baseline methods (Base U-Net and CNN-MHA).
• Successfully graded Obstructive Sleep Apnea (OSA) severity with 91.43% accuracy.

Conclusions:

• The FMCW radar-based system presents a promising, non-contact, low-cost, and accurate solution for OSA assessment.
• This technology has the potential to revolutionize sleep apnea diagnosis, enabling large-scale screening and continuous home monitoring.
• The findings suggest improved patient outcomes and public health impact through accessible sleep apnea diagnostics.