Skip to main content
  • Precision Prediction of Individualized Outcome in Traumatic Brain Injury Patients after Acute Rehabilitation

    Final Number:

    Matthew Z. Sun MD; Yiling Chen; Sheila Rosenberg; Emily Rosario; Jingyi Li; Daniel C. Lu MD PhD

    Study Design:

    Subject Category:

    Meeting: Congress of Neurological Surgeons 2018 Annual Meeting

    Introduction: No model currently exists that predicts individualized patient outcome after traumatic brain injury (TBI). For the first time, we developed modern statistical models and assessed their accuracy in predicting individualized patient-specific outcome in the Functional Independence Measure (FIM) scale after acute rehab using clinical data before rehabilitation.

    Methods: All patients (n=629) admitted to Casa Colina Acute Rehabilitation Unit (ARU) from 2010-2015 with a TBI diagnosis were included. 37 variables (Table 1) were used to predict 17 different (Table 2) FIM scales (each scored initially from 1-7, then mapped to 4 predicted scores: 1 [total assistance], 2-4[moderate assistance], 5[supervision only], 6-7[independence]) at time of discharge from ARU for each patient. Three statistical methods: random forest, support vector machine (SVM), and adaptive boosting (adaboost), were used to create 2 different models each, using either sequential binary classification, or one-vs.-one multinomial classification, resulting in 6 different models. For each model, we randomly sampled with replacement the entire dataset, selected the first half (n=314) as the training set and the rest (n=315) as testing set, repeated the process 1000 times, calculated 1000 bootstrap classification errors, and derived the mean and standard deviation for the classification errors.

    Results: The average FIM prediction accuracy ranged from 63-75% in our 6 models with an average standard deviation of 2-3% (Figure 1). The 2 models from Random Forest predicted with higher accuracy rates compared to Adaboost and SVM (p<0.05), averaging 74-75% +/- 3%. FIM score for Bowel Control was the most accurately predicted category for all models, with Random Forest predicting correctly more than 80% of the time (Figure 2).

    Conclusions: Modern statistical methods, particularly random forest, can predict patient-specific outcome after TBI to a reasonable degree, with our best models averaging over 70% accuracy for every feature for every patient. As datasets become larger in the future, we expect improved accuracy rates.

    Patient Care: Our research is the first step toward precision predicting by providing each patient with his own personalized outcome prediction rather than probabilistic outcome that are provided by current models. Our research will improve physician patient communication and provide a powerful tool to help guide management and improve overall patient satisfaction.

    Learning Objectives: - To evaluate extent to which modern statistical methods and machine learning algorithms can precisely predict individual clinical outcome using large datasets - To assess the feasibility of creating modern statistical models to ultimately provide each patient with their own specific functional outcome after traumatic brain injury.


We use cookies to improve the performance of our site, to analyze the traffic to our site, and to personalize your experience of the site. You can control cookies through your browser settings. Please find more information on the cookies used on our site. Privacy Policy