Prediction of Clinically Significant Improvements During the Interdisciplinary Intensive Outpatient Program for Traumatic Brain Injury Using Machine Learning.

Purpose: The aim of this research was to assess the potential for machine learning to predict clinically significant patient improvement during a four-week interdisciplinary Intensive Outpatient Program (IOP) for traumatic brain injury (TBI) at the National Intrepid Center of Excellence (NICoE).

Methods: Assessment of brain injury characterization and outcomes were measured in 790 active duty service members at the NICoE, Walter Reed National Military Medical Center Bethesda Maryland. Demographic and self-reported measures of posttraumatic stress, depression, anxiety, post-concussion symptoms, and sleep were assessed upon admission. Total scores and symptom cluster scores for self-report measures were calculated. Clinically significant improvement from pre- to post NICoE IOP was operationally defined as clinically significant changes in posttraumatic stress and post-concussion symptoms. Two datasets were created: one including demographics and total scores on self-report measures and one including demographics, total scores, and symptom cluster scores for relevant self-report measures. Extreme gradient boosting (XGBoost) models were trained to predict group identification (clinically significant improvement vs. not significant improvement), where a binary logistic objective function is used to minimize the log loss between the predicted probabilities. Model performance and feature ranking were then evaluated on test datasets.

Results: The performance and feature importance of two models to predict group identification were evaluated, where the model including only demographics and total self-report measures performed with an AUC of 75% with the accuracy of 68%, compared to the model incorporating demographics and symptom cluster measures improved the AUC to 79% with 72% accuracy. The top five features contributing to the model with symptom clusters included the posttraumatic stress arousal, avoidance, and reexperiencing sub-scores, education, and postconcussive symptoms cognitive sub-score.

Conclusion: Utilization of the XGBoost models demonstrated acceptable discrimination for determining key factors associated with clinically significant improvement for SMs following participation in an interdisciplinary IOP using demographics and self-report measures. Severity of posttraumatic stress symptoms upon admission was the greatest predictors of clinically significant improvement in this model of care. Incorporating ML algorithms into clinical care is a precision medicine approach that may accurately predict treatment efficacy leading to improved healthcare resource allocation and patient outcomes.