Transferring and Adapting a Prognostic Model to Improve Care of Brazilian Traumatic Brain Injury Patients

Abstract

AbstractBackground: Traumatic brain injury (TBI) is a major cause of death and disability. About 10 million people annually are affected by TBI, with a prominent burden in low- and middle-income countries (LMICs). In Brazil, TBI is responsible for 125,500 admissions and 9700 hospital deaths annually. The poor prognosis could be caused by insufficient medical professionals and diagnostic machines. This study aims to find an optimum TBI prognostic model to serve as a diagnostic tool that can be adapted from prior work in Tanzania to Brazil. We aim to develop an effective TBI prognostic model that could be generalized in LMICs. Methods: The study was a secondary data analysis on clinical and sociodemographic variables of 3209 TBI patients at Kilimanjaro Christian Medical Center (KCMC) and 725 TBI patients at six Brazilian traumatic care hospitals. We trained and tested eight machine learning models using three strategies: 1) using Tanzanian dataset trained models to test Brazilian dataset, 2) using Tanzanian-Brazilian combined dataset for training and testing and 3) using Brazilian dataset for training and testing. We compared the performance of models using confusion matrix statistics: area under the ROC curve(AUC), sensitivity, specificity, positive predictive value, negative predictive value and accuracy. Findings: Models using Tanzanian-Brazilian combined dataset for training and testing outperformed models of other two strategies. The AUC of the models varied from 80.9% (K nearest neighbor) to 91.9% (Random Forest). The optimum model, Random Forest, had a strong predictive power of classification with sensitivity of 0.927, specificity of 0.756, positive predictive value of 0.960, negative predictive value of 0.620 and accuracy of 0.903. Interpretations: Our study shows the successful adaptation of TBI prognostic model from Tanzania to Brazil. Additionally, it indicates the possibility of generalizing a TBI prognostic model to LMICs. With larger multi-national data, we hope to develop an effective model that could accurately predict the potential outcome of TBI patients. The model could serve as a powerful auxiliary tool for diagnosis and help reduce mortality of TBI patients in LMICs. Source of Funding: The project is conducted with the funding from Duke Global Health Institute.