Utilizing a Wireless Radar Framework in Combination With Deep Learning Approaches to Evaluate Obstructive Sleep Apnea...

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) diagnosis commonly relies on polysomnography (PSG) and home sleep apnea testing (HSAT).
• Current diagnostic methods require sensor attachment, potentially disrupting sleep and impacting results.
• There is a need for non-invasive, home-based screening tools for OSA risk assessment.

Purpose of the Study:

• To evaluate a wireless radar framework integrated with deep learning for home-based OSA risk screening.
• To assess the feasibility of using radar technology as a non-contact alternative to traditional OSA diagnostic sensors.

Main Methods:

• Prospective collection of home sleep data from 80 participants over 147 nights, utilizing both HSAT and a 24-GHz wireless radar system.
• Development of hybrid deep neural decision tree models to analyze radar signals and identify respiratory events.
• Correlation and agreement analyses between the apnea-hypopnea index (AHI) from HSAT and the radar-derived respiratory disturbance index (bRDI_TIB), including establishing cutoff thresholds using Youden's index.

Main Results:

• A strong correlation (ρ = 0.87) and high agreement (93.88%) were found between the AHI and bRDI_TIB.
• The radar-based model achieved 83.67% accuracy for moderate-to-severe OSA (cutoff: 21.19 events/h) and 93.21% for severe OSA (cutoff: 28.14 events/h).
• The average accuracy for multiclass OSA classification using established thresholds was 78.23%.

Conclusions:

• The wireless radar framework demonstrates potential as a non-contact, accurate surrogate for HSAT in home-based OSA screening.
• Established cutoff thresholds for the radar-derived index offer acceptable accuracy for OSA risk assessment.
• Further research is needed to optimize and validate the radar-based total sleep time estimation for independent clinical application.