Integrating Complete Blood Count Parameters with Demographic Characteristics for Obstructive Sleep Apnea Prediction in Chinese Adults: A Machine Learning Approach.

Abstract

Purpose: To develop and validate machine learning models integrating complete blood count (CBC) parameters with demographic characteristics for obstructive sleep apnea (OSA) risk stratification in adults with suspected OSA referred to a tertiary sleep clinic in China.

Methods: This retrospective study analyzed 5,828 adults with suspected OSA referred to a tertiary sleep clinic in China (2018-2024) who underwent home sleep apnea testing (HSAT), with OSA defined as an apnea-hypopnea index (AHI) ≥ 5 events/h. The cohort was temporally partitioned into a training cohort (January 2018 - December 2022, n = 4,330) and a validation cohort (January 2023 - March 2024, n = 1,498) for independent temporal validation. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression with 10-fold cross-validation exclusively within the training cohort, identifying 9 predictors from 16 candidates: three demographic variables (gender, age, body mass index [BMI]) and six CBC parameters (hemoglobin, mean corpuscular hemoglobin [MCH], lymphocyte count, red cell distribution width-coefficient of variation [RDW-CV], mean platelet volume [MPV], and platelet count). Four machine learning algorithms (logistic regression, Naive Bayes, random forest, XGBoost) were developed using the selected features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration metrics (Brier score, calibration slope), bootstrap optimism correction (1,000 resamples), and decision curve analysis (DCA).

Results: LASSO regression identified nine robust predictors from 16 candidate variables. In the final logistic regression model, hemoglobin demonstrated the largest standardized coefficient (1.38), followed by BMI (0.61), age (0.60), gender (|β| =0.40, with male gender associated with higher OSA risk), and MCH (0.36). Logistic regression achieved the best validation performance (AUC, 0.901, 95% confidence interval (CI): 0.882-0.919; sensitivity, 84.6%; specificity, 81.7%), with good calibration (Brier score, 0.084, calibration slope, 1.208) and minimal optimism (bootstrap-corrected AUC, 0.907). Five-fold cross-validation within the training cohort confirmed model stability (mean AUC, 0.906 ± 0.011). All models demonstrated superior net clinical benefit compared with treat-all or treat-none strategies in DCA. Secondary analyses at AHI ≥ 15 and AHI ≥ 30 thresholds confirmed sustained discriminative ability (validation AUCs 0.820 and 0.809, respectively).

Conclusion: Machine learning models integrating CBC parameters with demographic characteristics demonstrated good discriminative ability for OSA risk stratification in Chinese adults referred to a tertiary sleep clinic. These readily available biomarkers may facilitate risk stratification in pre-HSAT triage settings.