Abstract 18321: Identifying Predictors of Non-Home Discharge Dispositions Using Machine Learning Among Critically Ill Patients Undergoing Coronary Artery Bypass Grafting

Abstract

Introduction: Managing patients undergoing coronary artery bypass graft (CABG) procedures can be challenging. Research shows hospital discharges to locations other than home following CABG are associated with an increased risk of late mortality. We aimed to develop machine learning (ML) models to predict the risk of postoperative discharge dispositions other than home among these patients and assess their long-term postoperative prognosis. Methods: The Medical Information Mart for Intensive Care IV database was utilized to identify patients who underwent CABG procedures. The cases were divided into training (75%) and test (25%) sets. The AUROC, accuracy, and F1 scores were used to compare the performance of models. Several methods, such as penalized logistic regression, random forest, SVM, XGBoost, gradient boosting classifier, and lightGBM, were utilized, accounting for variables from demographics, labs, charting events, and comorbidity domains. Results: The study included 2228 eligible patients who underwent CABG procedures. The sample consisted of 23% females, 72% whites, and 43% were insured by Medicare. The median age of the sample was 69 years. Among these, 781 (35.05%) patients were discharged to locations other than home. Results show that the gradient boosting method had the highest accuracy (76.8%) with an AUROC of 80.8%. Age, BUN, male sex, PO2, hemoglobin, heart rate, mean arterial BP, chloride, platelet count, and creatinine levels were the top predictors for predicting non-home discharges. Results from the other models represented similar attributes as the top significant predictors. Conclusion: Predictors of discharge to locations other than home were identified with high accuracy, and this may aid in selective intervention to reduce the risk of death in this vulnerable patient population. ML can be used to investigate these patterns and possibly highlight particular patient subsets to enhance risk stratification and management.