0311 Artificial Intelligence (AI) to Predict Arousals in Home Sleep Testing (HST) Without Electroencephalography (EEG)

Abstract Introduction The American Academy of Sleep Medicine recommends scoring hypopneas terminating in either oxygen desaturation or arousals. A limitation of HST, when EEGs are unavailable, is not scoring hypopneas that terminate in arousals. This lowers the average apnea-hypopnea index (AHI) in HST studies and disproportionately affects patients who predominantly have hypopneas that terminate in arousals. Methods We report on a deep neural network, Nox BodySleep 2.0 experimental prototype, that predicts arousals and sleep stages using non-EEG signals. The model uses abdomen and thorax respiratory inductance plethysmography (RIP) and activity signals. The model outputs arousal events; and Wake, rapid-eye-movement (REM), and non-REM sleep epochs. It was trained on ~3200 PSG sleep studies from the United States, Europe, and Asia, and validated using 2,407 PSGs from clinical sleep labs in the United States. The performance was measured using epoch-based sensitivity, specificity, and accuracy for scoring arousals. Furthermore, the clinical performance was validated by the sensitivity, specificity, and accuracy for AHI severity classification for the diagnostic cutoff thresholds of AHI ≥ 5 and AHI ≥ 15. Results The model sensitivity, specificity, and accuracy was 65%, 85%, 80% for arousal scoring. For AHI ≥ 5, the sensitivity, specificity, and accuracy were 95%, 88%, 94%, respectively. For AHI ≥ 15, the results were 86%, 97%, 92%. When the sleep studies were scored as HST without using the model, the corresponding results for AHI ≥ 5 were 75%, 95%, 77%, and for AHI ≥ 15, 60%, 99%, 81%. When the studies were scored as PSG, the results for AHI ≥ 5 were 96%, 95%, 96%, while for AHI ≥ 15, they were 89%, 98%, 96%. Conclusion The model is a promising method of providing conclusive results from HST sleep studies. The model-based AHI is closer to the AHI from a PSG than when using HST. The AI model was trained on a large and diverse dataset and validated on a clinical population from the United States. Predicting arousals in HST studies without EEG improves patients' access to conclusive sleep apnea testing, and may improve health equity and the operation of sleep clinics. Support (if any) None