Leveraging explainable machine learning models to predict moderate to severe obstructive sleep apnea in heart failure with preserved ejection fraction patients: A comorbidity perspective.

BACKGROUND

Obstructive sleep apnea (OSA) is commonly observed in patients with heart failure with preserved ejection fraction (HFpEF), exacerbating the severity of the disease. Identifying the composition of disease clusters and understanding the associated risk factors for comorbidities is essential for early identification and management of HFpEF patients with concurrent conditions.

METHODS

This observational multicenter study enrolled 1140 patients diagnosed with HFpEF during hospitalization from January 2017 to December 2019, who were randomly assigned to the training (n = 684) and internal verification (n = 456) cohorts in a 6:4 ratio. An additional 717 participants were enrolled in the external verification cohort from January 2022 to August 2025. The predictive performance of K-nearest neighbors, light gradient boosting machine, random forest (RF), support vector machine, extreme gradient boosting, and logistic regression models were evaluated using a comprehensive set of performance metrics.

RESULTS

Among the six models, the RF model best predicted moderate to severe OSA in patients with HFpEF, achieving an area under the receiver operating characteristic curve of 0.974 (95 % confidence interval [CI]: 0.962-0.986) in the internal verification cohort and 0.910 (95 % CI: 0.889-0.930) in the external verification cohort. Additionally, decision curve analysis indicated that the RF model provided the highest net clinical benefits across different threshold levels. SHAP analysis revealed the significant contributions of individual variables and their associations with moderate to severe OSA in patients with HFpEF.

CONCLUSION

RF model-driven prediction methods allowed for accurately diagnosing moderate to severe OSA in patients with HFpEF.