LAUSR

Research Objectives To determine whether admission characteristics can accurately predict outcomes measured at discharge using machine learning methods for patients with TBI receiving inpatient rehabilitation facility (IRF) care. Secondarily, to determine which models are most accurate to aid future research Design The following models were compared: neural networks, LASSO/Ridge/Elastic Net regularization, Support Vector Machines, and Generalized Additive Models. Parameters were determined by 10-fold cross-validation. Models were evaluated on a subset of patients who were not used to develop the models using root mean-square errors (RMSE) and areas under the ROC curve (AUC). Setting Data were obtained from the Uniform Data System for Medical Rehabilitation which consists of data from 70% of IRFs in the US. Participants Adult TBI patients (N=175,358) who were admitted to a participating IRF and received a traumatic brain dysfunction Impairment Group code of 02.21 (traumatic, open injury) or 02.22 (traumatic, closed injury). Interventions Not Applicable. Main Outcome Measures Discharge Functional Independence Measure; clinically significant gain in FIM (≥ 22 FIM points) between admission and discharge; and mortality. Results For the FIM score, RMSE was reduced to 14-16, depending on the model, compared to an RMSE of 24.7 using the intercept only. For clinically significant increases in FIM, AUCs for all models were between 0.71-0.80. For mortality, AUCs were between 0.8-0.89. Elastic Net regularization resulted in the most accurate models. Conclusions Machine learning models can improve rehabilitation specialists’ ability to predict discharge outcomes for TBI patients receiving IRF care. Elastic Net methods were the most accurate. More complicated statistical methods (e.g. neural networks) may not be necessary. Author(s) Disclosures There are no conflicts to declare.