What question did this study set out to answer?

This study aims to validate an AI-driven autoscoring algorithm for accurately scoring central sleep apnea events.

May 10, 2026

0567 Enhancing Central Sleep Apnea Detection via AI-Driven Autoscoring by a Wearable Home Sleep Apnea Test

Key Points

This study aims to validate an AI-driven autoscoring algorithm for accurately scoring central sleep apnea events.
Multi-tiered validation using two datasets to assess accuracy of the autoscoring algorithm.
Primary analysis involved 45 patients comparing CAI scores from HSAT and polysomnography.
Secondary and tertiary analyses evaluated the algorithm against manually scored HSAT data from 45 and 129 sessions, respectively.
In primary analysis, the AA showed strong agreement with PSG for CAI (r = 0.95, 95% LOA = –2.8 to 2.9 events/h; p < 0.001).
Secondary analysis resulted in excellent concordance (r = 0.93, 95% LOA = –1.49 to 1.90; p = 0.0038).
In tertiary analysis, the AA reached r = 0.98 with 95% LOA of –3.53 to 3.45 for CAI, surpassing performance goals (p < 0.0001).

Abstract

Abstract Introduction Artificial intelligence (AI) is reshaping medical diagnostics by improving accuracy, consistency, and scalability across diverse clinical domains. The Wesper Lab Home Sleep Apnea Test (HSAT) Autoscoring Algorithm (AA) applies AI-based analysis to direct measurements of respiratory effort, airflow, and SpO₂. This study validated the accuracy of the AA for scoring central apnea events compared with manual human scoring (MHS). Methods We performed a multi-tiered validation using two datasets to assess the accuracy of the AA for determining the central apnea index (CAI). The primary analysis compared CAI from simultaneous HSAT and polysomnography (PSG) studies in 45 patients (27 males, 18 females; mean age 48.8 ± 14.7 yr) scored by blinded technologists. The secondary analysis evaluated the AA against 45 HSAT recordings scored by a blinded scorer. The tertiary analysis examined 129 HSAT sessions from 104 patients (81 males, 46 females, 2 undetermined; mean age 49.9 ± 13.8 yr; BMI 33.9 ± 7.6) manually rescored across 11 U.S. clinics. Agreement between AA and manual scoring for CAI was assessed using Pearson correlation and Bland–Altman limits of agreement (LOA). Results In the primary analysis comparing the AA to PSG studies, the AA demonstrated strong agreement for CAI (r = 0.95, 95% LOA = –2.8 to 2.9 events/h; p 0.001). In the secondary analysis of 45 HSAT studies manually scored by blinded technologists, the AA maintained excellent concordance (r = 0.93, 95% LOA = –1.49 to 1.90; p = 0.0038). In the tertiary analysis of 129 real-world HSATs manually rescored, the AA achieved r = 0.98 with Bland–Altman 95% LOA of –3.53 to 3.45 for CAI, exceeding the predefined performance goal (p 0.0001). Conclusion The AA demonstrated excellent accuracy for detecting central sleep apnea across all validation tiers. By integrating AI with direct respiratory effort and airflow signals, the system overcomes key limitations of traditional HSATs that rely on indirect or surrogate markers. Implementation of AI-based autoscoring may enhance diagnostic precision, reduce human scoring variability, and expand the scalability of home-based sleep testing for central sleep apnea. Support (if any)

Bookmark

Cite This Study

Rohrscheib et al. (Fri,) studied this question.

synapsesocial.com/papers/6a002147c8f74e3340f9c197 https://doi.org/https://doi.org/10.1093/sleep/zsag091.0566

Bookmark