The combination of physiology and machine learning for prediction of CPAP pressure and residual AHI in OSA.

Abstract

Study Objectives: Continuous positive airway pressure (CPAP) is the treatment of choice for obstructive sleep apnea (OSA); however some people have residual respiratory events or require significantly higher CPAP pressure while on therapy. Our objective was to develop predictive models for CPAP outcomes and assess whether the inclusion of physiological traits enhances prediction. Methods: We constructed predictive models from baseline information for subsequent residual apnea-hypopnea index (AHI) and optimal CPAP pressure. We compared models utilizing clinical variables with those incorporating both clinical and physiological factors. Furthermore, we assessed the performance of regression versus machine learning. All performances, including root mean square error (RSME), R-squared, accuracy, and area under the curve (AUC), were evaluated using a five-fold cross validation with ten repeats. Results: For predicting residual AHI, random forest models outperformed regression models, and models that incorporated both clinical and physiological variables also outperformed models using only clinical variables across all performance metrics. Random forest using both clinical features and physiological traits achieved the best performance. In both regression and random forest models, central apnea index is found to be the most important feature in predicting residual AHI. For predicting CPAP pressure, there was no additional predictive value of physiological traits or random forest modeling. Conclusions: Our findings demonstrated that the combined use of clinical and physiological variables yields the most robust predictive models for residual AHI, with random forest models performing best. These findings support the notion that prediction of OSA therapy outcomes may be improved by more flexible models using machine learning, potentially in combination with physiology-based models.