Abstract Introduction The WatchPAT (WP, ZOLL Itamar) is a widely used PAT-based home sleep apnea test. The company recently applied modern supervised machine learning techniques to develop a new algorithm to detect apneic events, incorporating clinical features and accumulated knowledge gleaned over 20 years of field experience. Over 900 synchronized WP and polysomnography (PSG) recordings were used to train the algorithm. A hybrid neural network (NN) was trained on a wide range of clinically significant features derived from WP channels. The NN combines convolution layers targeting signal morphology and bi-directional LSTM to identify timing connections between sympathetic activation, oximetry, movement and respiration. The outputs are second-by-second probabilities of disordered breathing events, which are then processed to identify discrete events and differentiate from non-respiratory sympathetic activations. The study’s goal was to examine the performance of the new algorithm compared to gold standard PSG for diagnosing sleep apnea. Methods We conducted a prospective multi-center validation study with simultaneous WP and PSG. Each PSGs was scored according to AASM guidelines by two independent scorers. We compared the frequency of disordered breathing events according to AASM 1A/1B rules (AHI3% and AHI4%, respectively) between WP and each scorer, individually, with Pearson correlation and Interclass correlation coefficient (ICC). Sleep apnea severity classification at apnea-hypopnea index (AHI) thresholds of 5/15/30 events/hour was compared with a linear-weighted Cohen’s kappa. Diagnostic performance for moderate-to-severe sleep apnea (AHI≥15) was assessed via sensitivity, specificity, ROC-AUC and Cohen’s kappa. Results The validation set comprised 231 subjects (141 Males, Age 52.4±17.8 years, BMI 32.0±8.6). WP-AHI3% had excellent correlation with PSG-AHI3% (r=0.94-0.95) and ICC (0.93-0.94). Substantial agreement across AHI severity levels (weighted Cohen’s kappa 0.76-0.81) was observed. Sensitivity for AHI≥15 was 0.86-0.92. Specificity was 0.83-0.92, ROC-AUC was 0.96 for both scorers. Cohen’s kappa indicated substantial agreement (0.73-0.76). For AHI4%, the results showed excellent agreement (k=0.84-0.89) for AHI≥15. The remaining outcomes were similar to those presented for AHI3%. Conclusion In a rigorous prospective study incorporating double-scoring, the new WatchPAT algorithm demonstrated high accuracy and agreement with PSG for assessing sleep apnea severity. The updated algorithm provides reliable, consistent, automated sleep apnea diagnosis. Support (if any) ZOLL-Itamar, Caesarea, Israel
Pham et al. (Fri,) studied this question.