Development and Validation of Outcome Prediction Model for Traumatic Brain Injury Using Classification and Regression Tree (CART) Technique

Traumatic brain injury is the leading cause of mortality, morbidity, disability, and socioeconomic losses all over the Globe. In Asia, many low and middle income countries including India are facing an increased risk for TBI... [ view full abstract ]

Traumatic brain injury is the leading cause of mortality, morbidity, disability, and socioeconomic losses all over the Globe. In Asia, many low and middle income countries including India are facing an increased risk for TBI due to a rapid surge in urbanization, motorization and economic liberalization. Our aim was to develop and validate a prognostic model, which was simple and easy to use, for outcome prediction in patients with traumatic brain injury. For this, we employed a classification and regression tree (CART) technique for predicting in-hospital mortality and unfavourable functional outcome at 6-months in patients moderate or severe traumatic brain injury patients of a level 1, tertiary care trauma database (1466 patients for model development and 316 patients for prospective validation) of Indian subcontinent using 14 prognostic indicators at the time of admission. Hypotension and motor score emerged as the strongest overall discriminating risk factor for both the outcomes. For in-hospital mortality and unfavourable outcome, optimal tree had 8 and 15 terminal nodes, respectively, and the test-sample relative cost was 0.392 and 0.390, respectively. For both the outcomes, the overall classification accuracy was 72.5% and 75.8%, respectively. In external validation dataset, area under the ROC curves (95% CI) to predict in-hospital mortality and unfavourable functional outcome at 6-months were 80% (75%-86%), and 87% (83%-91%), respectively. In both development and validation dataset, the performance of model was found good in terms of discrimination and calibration ability. Methodologically, CART is quite different from the more commonly used statistical methods like logistic regression, with the primary benefit of illustrating the important prognostic variables as related to outcome. This seems less expensive, less time consuming, or less specialized measurements. This technique may be useful in developing new therapeutic strategies and approaches for patients with moderate or severe brain injury.