Machine learning prediction of the change in sleepiness following continuous positive airway pressure in obstructive sleep apnea: a multi-cohort analysis
Eric Staykov, Juha Töyräs, Dwayne L Mann, Timo Leppänen, Samu Kainulainen, Ali Azarbarzin, Scott A Sands, Philip I TerrillAbstract
Study Objectives
To predict the improvement in sleepiness following continuous positive airway pressure (CPAP) treatment for obstructive sleep apnea (OSA) using baseline characteristics.
Methods
Data from five polysomnography cohorts (HomePAP, BestAIR, ABC, STAGES, and MESA; total n=2332) were analyzed. Associations between polysomnographic metrics and Epworth Sleepiness Scale (ESS) were assessed using multivariable linear regression. Cross-validated LASSO regression was used to select predictor variables. Prediction of ESS improvement was evaluated in CPAP-treated participants (n=213) using linear regression models trained on 139 participants and tested in a holdout group (n=74).
Results
Cross-sectional analysis revealed significant associations between baseline ESS and ventilatory burden (0.56 points/SD; 95% CI 0.34-0.78), hypoxic burden (0.51; 0.28-0.74), the apnea-hypopnea index (AHI, 0.45; 0.23-0.68), time below 90% oxygen saturation (0.36; 0.15-0.57), and flow limitation severity (0.31; 0.10-0.53). Baseline ESS and baseline ventilatory burden were the most frequently selected predictors of ESS improvement. Using these two variables in a regression model, predicted and actual ESS improvement were correlated in the holdout group (adjusted R2=0.313). The model classified CPAP responders (≥2 points ESS improvement) more accurately than current clinical guidelines (78.1% [3.5% SD] versus 68.4% [3.7%]). Substituting the AHI for ventilatory burden slightly reduced accuracy (77.3% [3.4%]). Charts were developed to assist in predicting ESS improvement from baseline metrics.
Conclusions
Models incorporating baseline ESS and either ventilatory burden or AHI predict ESS improvement and may help identify patients with OSA likely to benefit from CPAP. The ventilatory burden model demonstrated modestly better performance, while the AHI-based model offers greater clinical applicability.