Faculty of Public Health - Andalas University - OCS, 13th IEA SEA Meeting and ICPH - SDev

Font Size: 
Statistical models for prediction of outcomes after traumatic brain injury in India
Vineet Kumar Kamal

Last modified: 2018-08-07

Abstract


Introduction Traumatic brain injury (TBI) is a significant public health problem in all regions of the globe. Statistical modelling is essential for prognostication, hypothesis generation and stratification of patients in research studies in case of TBI. Depending upon need, settings, utility, performance and context, one can highlight and chose any prognostic (statistical) model for predictions of outcome in future patients with TBI. Our aim was to develop and compare statistical models for prediction of outcomes after traumatic brain injury based on admission characteristics.

Methods Comparison of prognostic models were done in both development and external validation data set in terms of overall performances (Brier score), discrimination ability (Area under ROC), and calibration ability (Hosmer-lemeshow test’s p-value) using logistic regression (LR), classification and regression tree (CART), and artificial neural network (ANN), Random forest (RF) and by hybridizing two popular classification tools, CART and LR method to predict in-hospital mortality and unfavourable functional outcome at six months post admission using demographic, clinical, secondary insult, and CT variables of patients with moderate or severe traumatic brain injury. For this, we utilized trauma database (n=1782) of India’s largest tertiary care level one trauma centre.

Results For LR, CART, ANN, RF, and the hybrid CART-LOGIT models in external validation data set, area under the ROC curves (95% CI) were 0.86 (0.81, 0.90), 0.80 (0.75, 0.86), 0.86 (0.82, 0.91), 0.77 (0.72, 0.81) and 0.86 (0.82, 0.90), respectively to predict in-hospital mortality; and 0.91 (0.87, 0.94), 0.87 (0.83, 0.91), 0.91 (0.87, 0.94), 0.82 (0.77, 0.86) and 0.90 (0.86, 0.93), respectively to predict unfavourable functional outcome at six months.

Conclusions LR, ANN, and the hybrid CART-LOGIT models seem to be having similar performance, but both these methods outperformed CART, and RF to predict both outcomes. The stand-alone CART model outperformed the random forest approaches. The hybrid CART-LOGIT model is excellent in covering high order interaction as well as local and global structure of data.