Using Machine Learning to Predict Discharge Destination from Inpatient Rehabilitation

Research Objectives: To evaluate machine learning approaches to predict discharge destination after Inpatient Rehabilitation.

Design: Cross-sectional study.

Setting: Hospitalized care.

Participants: Adult inpatient rehabilitation patients (N=4404) totaling 4919 encounters from seven years at a large academic medical center in the midwestern United States.

Interventions: Not applicable.

Main Outcome Measures: Three approaches were tested: linear discriminant analysis (LDA), regularized linear models with stochastic gradient descent (SGD) learning, and logistic regression (LR) models. The goal was to predict home versus non-home patient discharges. Input variables included demographic data, diagnosis, insurance payor, medical history, lab values, length of stay, and administration of high-risk drugs (antipsychotics, anticoagulants, antibiotics, opioids, antiplatelets, and hypoglycemics). Performance and variables important to predicting discharge destination were evaluated for each approach.

Results: Encounters were randomly split into a training (N=3443) and a test set (N=1476). Using class balancing and 10-fold cross validation with hyperparameter optimization, LDA achieved area under the curve (AUC) of the receiver operating characteristic curve of 0.662, with macro-averaged accuracy=0.628, SGD achieved AUC=0.678 (accuracy=0.623), and LR achieved AUC=0.660 (accuracy=0.621). Shapley values indicated greater importance of intravenous medication use, antibiotics, presence of urinary and swallowing dysfunction, managed care payors, and fatigue on predicted discharge destination.

Conclusions: Approaches incorporating machine learning methods may play a role in helping to predict inpatient rehabilitation discharge and identifying factors that increase risk of non-home discharges. These findings may help facilitate more expeditious transfers to care and preparations for discharge home. Subsequent models could incorporate additional information such as therapy notes and functional scoring which may improve predictive ability.

Author(s) Disclosures: Authors declare they have no conflicts of interest to disclose.