Predicting in-hospital mortality of older adult trauma patients in Northern Tanzania: using machine learning to improve triage algorithms and resource utilization in low resource contexts

Background: Older adults (>60 years) account for up to 25% of all trauma patients worldwide and up to 14% in Tanzania. Half of the older adult trauma patients are under-triaged at the Emergency department (ED), which is associated with a four-fold higher risk of in-hospital mortality compared to younger adults. To guide the optimization of scarce resources, this study aimed to develop a machine-learning predictive algorithm for triaging older adult trauma patients at the ED. Methodology: Adult trauma registry data from 2018-2024 from the Kilimanjaro Christian Medical Centre (KCMC) was analyzed using Python programming language to predict in-hospital mortality. Six models were developed and tested with a supervised stratified 5-fold cross-validation approach: the LASSO Logistic Regression, the Support Vector Machines (SVC), the Artificial Neural Network (ANN), the Bayesian Algorithm (Naive Bayes), the Random Forest and the Gradient Boosting Decision Trees. Model performance was assessed using the accuracy, sensitivity, specificity, precision, negative predictive value, the Area Under the Receiver Operating Curve (ROC-AUC), the precision-recall curve, and the calibration plot. Of the top two best-performing models, model selection was based on the balance between model performance, sparsity, interpretability, and transparency. Results: The registry had 623 older adult patients, of which 606 were included in the analysis; the rest were missing age or outcome data. The patient’s mean age was 72 years (SD: ±9), more than half 335 (55%) were males and the in-hospital mortality rate was 9.2%. The top two best-performing models were the Support Vector Machines SVC (ROC-AUC: 0.78, 95%CI:0.75-0.81) and the LASSO Logistic Regression (ROC-AUC: 0.74, 95%CI: 0.68-0.79). We selected the LASSO Logistic Regression as the best model for triaging older adult trauma patients at the ED. From the LASSO Logistic Regression: the patients’ level of consciousness (GCS), age, pulse oxygen saturation, heart rate, and systolic blood pressure were the most important predictors driving mortality predictions among geriatric trauma patients.Conclusion: The LASSO Logistic Regression was the best model for potential translation to triage older adult trauma patients at the Emergency Department (ED). The model shows age and vital signs on admission as strong predictors of mortality among geriatric trauma patients. The next steps will involve further re-training and re-calibration to ensure better contextual performance in translation to clinical care.