An interpretable machine learning model for predicting the risk of in-hospital death among ICU patients with traumatic brain injury

Background Traumatic brain injury (TBI) is one of the leading causes of death and disability; early and accurate prediction of a patient’s risk of in-hospital mortality is crucial for optimizing treatment and improving outcomes. Objective The aim of this study is to develop and validate a machine-learning-based model for predicting the risk of in-hospital mortality among ICU patients with traumatic brain injury, and to use interpretable methods to identify key predictive factors and assess their clinical utility. Methods Clinical data were retrospectively collected from patients with TBI admitted to the intensive care units (ICUs) of two Grade A, Level 3 hospitals between January 2022 and December 2025; After excluding cases of death due to causes other than TBI and those who were already brain-dead on admission, a total of 874 patients were ultimately included in the study. This was a 24-h milestone prediction study; the predictive model established is applicable only to patients with traumatic brain injury who remained alive 24 h after admission to the ICU. The primary outcome was in-hospital death due to traumatic brain injury; excluding deaths from causes other than traumatic brain injury and cases where brain death had already been confirmed on admission, a total of 381 cases experienced this outcome, representing an incidence rate of 43.59 per cent. Prior to analysis, the dataset was randomly divided into a training group ( n = 613) and a testing group ( n = 261). Lambda 1-se standard LASSO regression was applied solely to the training set to identify predictive features, resulting in the selection of 17 candidate variables. Seven machine learning algorithms were developed: logistic regression, random forest, XGBoost, artificial neural networks, decision trees, support vector machines and naive Bayes. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), recall, accuracy, F1 score, precision and negative predictive value, with 95 per cent confidence intervals calculated for all metrics; interpretability analysis and model pruning were performed using the SHAP method. Results LASSO regression identified 17 relevant key features; the random forest model demonstrated good predictive performance, with a training set cross-validation AUC of 0.954. The simplified RF model constructed after feature dimensionality reduction included 11 key predictive variables, with AUC on the test set was 0.924 (95% CI: 0.848–0.960). Conclusion The random forest-based predictive model demonstrated high predictive performance; the simplified model strikes a balance between performance and clinical interpretability and can effectively identify high-risk populations. Key features such as GCS and APACHE II can support early clinical risk stratification and personalized interventions, highlighting the potential of machine learning in the clinical management of TBI.