Deep Learning Models Generalization for Predicting 14-day Mortality in Traumatic Brain Injury Patients

One of the leading causes of morbidity and mortality in the world is Traumatic Brain Injury (TBI). Different outcomes are influenced by regional access and health infrastructure. In this study, using 17 predictor variables, we evaluate machine learning models performance and generalizability with two different datasets of Brazilian regions. The first region is Manaus, an isolated urban center with differentiated logistical challenges. The second, is São Paulo, an urban center. To the best of our knowledge, this study is the first one that evaluate predictive models in two distinct datasets in the same country. In the results obtained with 1-D convolutional neural network (CNN) models, the area under the ROC curve (AUC) in São Paulo and Manaus were 0.90 and 0.93, respectively. The model trained in São Paulo does not perform well in Manaus. The incorporation of context-specific features, such as time between trauma and admission, and pandemic-related variable significantly increased the model's accuracy in Manaus model, achieving a remarkable AUC of 0.98.Clinical Relevance- We highlighted the necessity of integrating local variables to improve TBI prediction in different healthcare environments.