Multicenter Validation of an AI-Based Snoring Model vs. Wearables for Diagnosing Sleep Apnea.

Abstract

Background: Sleep apnea-hypopnea syndrome (SAHS) has high prevalence and cardiovascular/cerebrovascular comorbidities. Polysomnography (PSG) is the diagnostic gold standard but limited by complexity, cost, and accessibility; smart wearables are convenient but less accurate. This study validated an AI-based snoring-sound model and compared its diagnostic performance with a wearable, using PSG as reference.

Methods: Adults with suspected SAHS (n = 134) were included and underwent overnight PSG, with simultaneous snoring sound and wearable signal recording. The snoring algorithm used short-time Fourier transform (STFT) spectrograms and a multiscale encoder-attention-decoder neural network for classification. Model performance was assessed using correlation analysis, intraclass correlation coefficient (ICC), Bland-Altman plots, and receiver operating characteristic (ROC) analyses. Accuracy, sensitivity, specificity, and area under the curve (AUC) values were computed for both models.

Results: Disease severity was associated with variation in BMI (p < 0.001) and decreasing oxygen saturation (p < 0.001). The snoring model showed stronger correlation and higher agreement with PSG-derived AHI (r = 0.79, ICC = 0.753) than the wearable (r = 0.68). Although the snoring model yielded consistently higher AUCs, particularly in mild SAHS (AUC = 0.83 vs. 0.72, p = 0.07), the differences between methods were not statistically significant (all p > 0.05). Both models achieved excellent discriminative performance in severe SAHS (AUC ≥ 0.90).

Conclusion: The AI-driven snoring-sound model demonstrated comparable overall performance and numerically better performance in certain settings, particularly in identifying mild SAHS.