Development and validation of machine learning models for obstructive sleep apnea risk stratification in high-risk pregnant women.

backgroundObstructive sleep apnea (OSA) affects approximately 15% of pregnancies and is associated with adverse maternal and fetal outcomes. Although polysomnography (PSG) is the diagnostic gold standard, increasing clinical demand creates substantial bottlenecks in manual PSG scoring and specialist review. Conventional screening tools often show limited discriminative ability in high-risk referred populations. Therefore, optimized risk stratification models are needed to streamline clinical workflows, prioritize diagnostic resource allocation, and facilitate timely intervention for high-risk pregnant patients.

methodsThis retrospective observational cohort study recruited pregnant women with suspected OSA who underwent level 2 portable PSG. Six machine learning algorithms, including XGBoost, logistic regression, GBM, neural networks, KNN, and AdaBoost, were constructed based on integrated clinical and oximetry features. Feature importance screening and DeLong's test-based pairwise comparison were performed to determine the optimal feature combination for model construction. The primary outcome was any OSA defined by an apnea-hypopnea index (AHI) ≥ 5 events/h, while the secondary outcome was moderate-to-severe OSA (AHI ≥ 15 events/h). All models were optimized using 10-fold cross-validation and externally validated on an independent testing set. Model performance was comprehensively assessed via receiver operating characteristic curves, calibration curves, and decision curve analysis (DCA).

resultsAmong the 667 enrolled participants, 305 (45.7%) were diagnosed with OSA. The 3% oxygen desaturation index, lowest SpO2, body mass index, waist circumference, and abdominal circumference were identified as core predictive variables. For the primary screening of any OSA, logistic regression and neural networks achieved robust and comparable discriminative performance; the logistic regression model attained a testing-set AUC of 0.872 with a sensitivity of 78.0%. For moderate-to-severe OSA prediction, AdaBoost and GBM exhibited excellent predictive efficacy, with testing-set AUCs of 0.956 and 0.955, respectively. DCA confirmed that the established models yield favorable clinical net benefit across broad risk threshold ranges, enabling optimized clinical screening and priority referral strategies.

conclusionsThis machine learning-based risk stratification framework demonstrates promising diagnostic performance for identifying OSA in symptomatic pregnant women under clinical referral. Leveraging structured medical records incorporating sleep history and physical measurements, this tool serves as an auxiliary triage strategy to assist clinical decision-making. The proposed models may help identify high-risk patients for expedited PSG assessment, which has the potential to optimize diagnostic workflows and improve resource allocation in specialized obstetric sleep medicine services.