AI-Enhanced Smartwatch AHI Estimation and AI-Scored Polysomnography for Obstructive Sleep Apnea: Real-World Validation

Real-world validation of a smartwatch AHI algorithm against Level 1 polysomnography in Korean adults (53 of 90 with enough valid watch data). Watch eAHI agreed with expert-scored AHI at ICC 0.82, and detected moderate-to-severe OSA with 92.3% sensitivity and 92.6% specificity against AI-scored PSG, while underestimating AHI in mild OSA.

• Form factor: wrist

ObjectiveThis study validated the accuracy of an artificial‑intelligence (AI) smartwatch algorithm that directly estimates the apnea-hypopnea index (AHI) by comparing its performance with AI-scored Level 1 polysomnography (PSG) in Korean adults. The model was trained in South‑American cohorts, allowing inter‑ethnic validation.MethodsA total of 90 adults underwent simultaneous Level 1 PSG and smartwatch recording. Fifty‑three datasets with ≥ 3 hours of valid watch data were analyzed. AHI values were obtained as follows: expert‑scored PSG (pAHI), AI‑scored PSG (aiAHI), and smartwatch output (eAHI). Agreement was assessed with Spearman correlation, intraclass correlation coefficients, and receiver‑operating‑characteristic curves.ResultseAHI correlated strongly with aiAHI (ρ = 0.88, ICC = 0.87) and pAHI (ρ = 0.85, ICC = 0.82). For detecting moderate‑to‑severe OSA (aiAHI ≥ 15 events/h), the smartwatch algorithm yielded 92.3% sensitivity, 92.6% specificity, and 92.5% overall accuracy. Bland-Altman analysis revealed systematic underestimation of actual AHI by the smartwatch, particularly in mild OSA.ConclusionThis study demonstrates that the evaluated smartwatch-based AHI estimation algorithm shows high concordance with PSG-derived values, particularly for the detection and classification of moderate to severe OSA. However, it should be noted that this smartwatch algorithm tends to underestimate the AHI of OSA due to limitations in scoring unit and recording duration calculation. These findings support the clinical utility of wearable technology as a practical and scalable tool for early identification and longitudinal monitoring of OSA in real-world environments, while highlighting the need for further optimization to accurately detect mild cases.

Verbatim from Europe PMC (opens in a new tab) · Licence: cc by-nc

3 rows are credited to this paper. Ranks apply within each leaderboard only, as reported by the source.