Predicting Discharge Destination From Inpatient Rehabilitation Using Machine Learning

Predicting discharge destination for patients at inpatient rehabilitation facilities is important as it facilitates transitions of care and can improve healthcare resource utilization. This study aims to build on previous studies investigating discharges from inpatient rehabilitation by employing machine learning models to predict discharge disposition to home versus nonhome and explore related factors. Fifteen machine learning models were tested. A total of 4922 patient encounters from 4401 patients undergoing inpatient rehabilitation at a Midwestern academic center's inpatient rehabilitation facilities were included. Input variables included demographic and hospital encounter-specific data. The total dataset contained 3687 discharges to home, and 1235 discharges to nonhome destinations. A bagging classifier utilizing a decision tree base classifier utilizing random undersampling and without feature selection performed the best in terms of area under the receiver operating characteristic curve with a score of 0.722. Shapley value analysis suggested that length of stay, intravenous medication administration, urinary dysfunction, age, abnormalities white blood cell count or plasma sodium, and fatigue were the factors with the greatest impact on model output. Machine learning can help predict inpatient rehabilitation discharge disposition and identify factors associated with home or nonhome discharges.