Predicting Outcomes of Traumatic Brain Injury Using Machine Learning Models Among Patients at Kilimanjaro Christian Medical Centre, Tanzania: A Registry-Based Cohort Study.

Background: Traumatic brain injury (TBI) remains a major global health burden, disproportionately affecting low- and middle-income countries (LMICs) where access to neurocritical care is limited. Accurate and context-appropriate prognostic models are crucial to guide early clinical decision-making and optimize resource allocation in such settings. This study aims to develop and evaluate machine learning (ML) models for predicting TBI outcomes among adult patients using trauma registry data from Kilimanjaro Christian Medical Centre (KCMC), Tanzania. Methods: This retrospective cohort study utilized data from 4596 adult TBI patients recorded in the KCMC trauma registry between 2013 and 2024. The outcome was dichotomized Glasgow Outcome Scale (GOS): poor (1-3) versus good (4-5). Ten supervised ML algorithms, including Random Forest (RF), Decision Tree (DT), Logistic Regression, Support Vector Machine (SVM), and Artificial Neural Networks (ANN), were trained on 70% of the data after applying multiple imputation and synthetic minority Over-sampling Technique (SMOTE) to address missingness and class imbalance. Hyperparameter tuning was performed using 10-fold cross-validation. Model performance was assessed on a 30% test set using area under the ROC curve (AUC), accuracy, sensitivity, specificity, and predictive values. Results: Among the 4596 patients, 26.2% had poor outcomes. The RF and DT models achieved the highest AUCs of 0.83 and 0.82, respectively. RF also showed the highest accuracy (0.78) and strong positive predictive value (PPV = 0.87), while DT had the highest sensitivity for poor outcomes (84.5%). Predictors of poor outcomes included TBI severity, pupil non-reactivity, low oxygen saturation, lack of CT scan, alcohol use, and abnormal vital signs. Conclusion: ML models, particularly RF and DT, demonstrated strong predictive performance for TBI outcomes using routinely collected variables in a resource-limited LMIC setting. Their interpretability and reliance on admission-level data make them potential tools for real-time triage and risk stratification. Future research should focus on external validation and integration into clinical decision-support systems to support scaleup across similar settings.