Methods

118 patients (67 F, 49 years, BMI 33kg/m²) suspected of SAS had an in-laboratory PSG together with Sleep Apnea Detector. From the pressure signal, Sleep Apnea Detector derives respiratory and cardiac signals and movements. From the microphone, snoring and snorting are detected. These features are used to detect sleep periods with a Random Forest classifier and apnea and hypopnea events with a Convolutional Neural Network. The Total Sleep Time (TST) and Apnea Hypopnea Index (AHI) deduced (TST sad, AHI sad) are compared with the PSG results scored according to AASM rules (TST psg, AHI psg). AHI and TST were compared using bias and Mean Absolute Error (MAE). Sensitivity, specificity, likelihood ratios (LR) and AUROC were calculated for AHI thresholds of 15 and 30/hr.

Results

The average (SD) TSTpsg was 367 (61) minutes. Sleep Apnea Detector overestimated TST by 25 minutes, 7.0% of the average duration in the sample. The precision is acceptable, with a MAE= 53 minutes. Average AHI psg was 32.5 (30.1) and AHI sad 32.8 (29.9). The bias was 0.3 (95% CI [-2.7, 3.3]), MAE= 10.3. The sensitivity (Se 15) and specificity (Sp 15) and their 95% confidence intervals were Se 15= 88.0%[79.0, 94.1] and Sp 15= 88.6%[73.3, 96.8]. Positive and negative LR were respectively LR+ 15= 7.70 and LR-15= 0.136. AUROC 15= 0.926. At the 30 threshold, Se 30= 86.0%[73.3, 94.2] and Sp 30= 91.2%[81.8, 96.7]. Positive and negative LR were LR+ 30= 9.75 and LR-30= 0.153. AUROC 30= 0.954.

Conclusion

Sleep Apnea Detector has excellent sensitivity and specificity, low bias and good precision. Thus it can be used as an unattended SAS screening device in patients likely to suffer from SAS.