Predictive modeling of TBI outcomes in Rwanda: Generalizability of Tanzania developed prognostic models

Abstract

Background: Globally, many low-income settings lack diagnostic tools to handle prognosis of TBI patients. In such settings, development of generalizable predictive models which indicate likelihood of patient outcomes may help improve decision-making for physicians and health care providers. Methods: An analysis of a Rwanda TBI registry (n=682) was conducted to determine key predictors of TBI mortality. A previously developed prognostic model of a Tanzania TBI registry (n=3209) was subsequently implemented in the Rwanda TBI registry for external validation. 8 different machine learning models were implemented in the Rwanda dataset. Subsequently, 6 Tanzania models and a combined model aggregating the Tanzania model predictions were used to compare and predict Rwanda patient outcomes. Results: The predictive models developed in Rwanda had satisfactory predictive ability, with the best performing model, Ridge Regression having an AUC of 90% (CI: 89.3%-90.7%). The models developed in Tanzania and used to predict outcomes within the Rwanda dataset showed similar predictive ability, the AUC of the best performing Random Forest model, 91.3% (CI: 88.0%-94.6%) and the combined Tanzania machine learning model, AUC 91.9% (CI: 88.7%-95.1%). Conclusions: The results from the Tanzania and combined models indicate satisfactory predictive ability and generalizability. The ability of the models to hold similar predictive power in an external dataset, with the use of indicators collectable at triage suggests potential applicability in other low-resource settings.