(1136) A Machine Learning Algorithm to Predict Mortality in ICU Patients With Obstructive Sleep Apnea

Introduction: Obstructive sleep apnea is a prevalent comorbid disease in patients admitted to the intensive care unit (ICU). While obstructive sleep apnea (OSA) has previously been found to be an independent risk factor for all-cause mortality, recent retrospective studies have noted an association between a diagnosis of OSA and decreased hospital mortality. Further studies are required to elucidate the factors contributing to mortality in patients with OSA.

Methods: Data extraction was performed using structured query language to create a data set using the Mart for Intensive Care IV database (MIMIC-IV) consisting of all patients admitted to the intensive care unit with a diagnosis of obstructive sleep apnea. This study included demographic characteristics, comorbidities, vital sign and ventilator information, and common laboratory results prior to and during ICU admission. A prediction model for in-hospital mortality was constructed using eXtreme Gradient Boosting (XGBoost). SHapley Additive exPlanations (SHAP) analysis was performed to quantify feature contribution.

Results: Of 43,549 total ICU admissions, a population of 7,644 patients with a diagnosis of OSA were split into a training and test set. Patients with OSA had a mortality rate of 2.8% compared to 5.1% in the general ICU population. They otherwise had similar Acute Physiology Score (APS III 42.5 vs 42.0), Systemic Inflammatory Response Syndrome (SIRS 2.5 vs 2.5), and invasive ventilation days (1.91 vs 1.93) respectively. In the training set, The XGBoost model performance was measured using a receiver operating characteristic (ROC) curve, with an area under the curve (AUC) of 0.82. The strongest predictors of in-hospital mortality in SHAP analysis were red cell distribution width (RDW), age, blood urea nitrogen (BUN), and number of days requiring mechanical ventilation.

Conclusions: Preliminary results from this predictive model have shown the ability to predict in-hospital mortality in critically ill patients with OSA. Further studies to understand the importance of the variables highlighted by SHAP analysis and external validation with a second test set will increase reliability for use in the clinical setting.