Predictor of Mortality in Obstructive Sleep Apnea: Results of Explainable Deep Learning-Based Survival Analysis from a 15-Year Follow-Up.

INTRODUCTION

Obstructive sleep apnea (OSA) is a respiratory disorder with comorbidities of several nature, from cardiovascular to renal ones. OSA is typically treated through continuous positive airway pressure (CPAP), but more investigations are needed to confirm its benefits.

METHODS

In this work, survival analysis has been exploited and enhanced with eXplainable Artifical Intelligence (XAI) to investigate the impact of comorbidities and compare the model reliability. The dataset encompasses both clinical and polysomnography-based data for a total of 45 different features.

RESULTS

A total of 1,394 OSA patients followed for 15 years were enrolled. All the selected features have been studied by means of DL-models and time-dependent XAI techniques. The variables impacting on mortality are reported below in descending order with respect their importance: (1) age, years of CPAP, renal dysfunction, COPD, body mass index categories, sex, and anemia for the CoxTime; (2) age, apnea hypopnea index, renal dysfunction, SaO2 min, years of CPAP, COPD, and anemia for the LogHazard.

CONCLUSION

Advancing age, severity of OSA, comorbidity, including chronic kidney disease, COPD, and anemia, are crucial contributors to increased mortality.