Enhancing Central Sleep Apnea Detection via AI-Driven Autoscoring

Download Full Article (PDF)

 

Antonio Artur Mours, MS., Amir Reuveny, Ph.D., Motti Perlman, Sonia Ancoli-Israel, Ph.D.,Ruchir P Patel, M.D., Chelsie Rohrscheib, Ph.D.

Study Objectives: To evaluate the accuracy and clinical reliability of the Wesper Autoscoring Algorithm (WAA) for automated detection of central sleep apnea (CSA), as measured by the Central Apnea Index (CAI), compared with manual human scoring by certified sleep technologists.

Methods: Deidentified Wesper Lab home sleep apnea test (HSAT) recordings were retrospectively analyzed from a curated dataset of adult patients evaluated across multiple U.S. sleep clinics. The WAA analyzed respiratory effort, airflow surrogates, oxygen saturation, pulse rate, sleep stage, and body position to automatically detect central respiratory events. Manual human scoring (MHS) served as the reference standard and was performed by certified technologists blinded to algorithm outputs. Agreement between WAA and MHS was assessed using Pearson correlation, Bland–Altman limits of agreement (LOA), and mean absolute error (MAE). A secondary blinded analysis was conducted to assess algorithm robustness and scorer independence.

Results: A total of 129 recordings from 104 unique patients were used for algorithm training and validation, with a hold-out test set of 36 recordings. The WAA demonstrated strong agreement with MHS for CAI (r = 0.98), with narrow Bland–Altman LOA (−3.53 to 3.45 events/hour) and low MAE (1.34 events/hour; Figure 1). In the blinded secondary dataset (n = 45), agreement remained high (r = 0.93; LOA −1.49 to 1.90 events/hour), confirming algorithm stability and reproducibility (Figure 2). Agreement for AHI was also maintained following CSA optimization.

Figure 1: Correlation of Wesper autoscoring algorithm (WAA) and manual human scoring (MHS) for central apnea index. A. Pearson correlation. B. Bland-Altman plot.

Figure 2: Correlation of Wesper autoscoring algorithm (WAA) and manual human scoring (MHS) for central apnea index. The analysis was completed on a clinical data set where scorers were blinded to the suggestions of the WAA. A. Pearson correlation. B. Bland-Altman plot.

Conclusions: The Wesper Autoscoring Algorithm demonstrates excellent agreement with manual human scoring for detection of central sleep apnea, supporting its accuracy and scalability for clinical use. Although the WAA employs machine-learning–based analysis, clinical interpretability is preserved through standardized, clinician-reviewable outputs that include raw physiologic signals and annotated events, consistent with other FDA-cleared HSAT technologies. These findings support AI-assisted autoscoring as a reliable adjunct to clinical workflows for CSA detection and management.