Utilizing a Wireless Radar Framework in Combination With Deep Learning Approaches to Evaluate Obstructive Sleep Apnea Severity in Home-Setting Environments.

Abstract

Objective: Common examinations for diagnosing obstructive sleep apnea (OSA) are polysomnography (PSG) and home sleep apnea testing (HSAT). However, both PSG and HSAT require that sensors be attached to a subject, which may disturb their sleep and affect the results. Hence, in this study, we aimed to verify a wireless radar framework combined with deep learning techniques to screen for the risk of OSA in home-based environments.

Methods: This study prospectively collected home-based sleep parameters from 80 participants over 147 nights using both HSAT and a 24-GHz wireless radar framework. The proposed framework, using hybrid models (ie, deep neural decision trees), identified respiratory events by analyzing continuous-wave signals indicative of breathing patterns. Analyses were performed to examine correlations and agreement of the apnea-hypopnea index (AHI) with results obtained through HSAT and the radar-based respiratory disturbance index based on the time in bed from HSAT (bRDI_TIB). Additionally, Youden's index was used to establish cutoff thresholds for the bRDI_TIB, followed by multiclass classification and outcome comparisons.

Results: A strong correlation (ρ = 0.87) and high agreement (93.88% within the 95% confidence interval; 138/147) between the AHI and bRDI_TIB were identified. The moderate-to-severe OSA model achieved 83.67% accuracy (with a bRDI_TIB cutoff of 21.19 events/h), and the severe OSA model demonstrated 93.21% accuracy (with a bRDI_TIB cutoff of 28.14 events/h). The average accuracy of multiclass classification using these thresholds was 78.23%.

Conclusion: The proposed framework, with its cutoff thresholds, has the potential to be applied in home settings as a surrogate for HSAT, offering acceptable accuracy in screening for OSA without the interference of attached sensors. However, further optimization and verification of the radar-based total sleep time function are necessary for independent application.