Interpretable machine learning for in-hospital mortality prediction in ICU patients with traumatic brain injury

Background TBI is associated with high ICU mortality, yet traditional prognostic scores often lack accuracy due to linear assumptions. This study aimed to develop an interpretable machine learning model to predict in-hospital mortality in TBI patients, combining high predictive performance with clinical transparency. Methods This retrospective analysis utilized TBI clinical records (2008–2019) retrieved from the MIMIC-IV database. We collected comprehensive baseline data including demographics, comorbidities, vital signs, laboratory parameters, disease severity scores, and therapeutic interventions. To identify the most robust predictors, we employed a rigorous intersectional feature selection strategy combining Univariate Logistic Regression ( p 0.05), LASSO regression, and the Boruta algorithm. Seven supervised ML algorithms (Logistic Regression, Decision Tree, Random Forest, XGBoost, LightGBM, SVM, and ANN) were developed and compared. Predictive performance was benchmarked using AUROC, Brier score, and DCA. Additionally, SHAP were implemented to enhance the interpretability of the final model’s decision-making process. Results Among the 2,691 included TBI patients, 1,716 (63.8%) were male, and the median age was 66 (IQR 48–81) years. The primary outcome, in-hospital mortality, occurred in 316 patients (11.7%). A final predictive set of 12 variables was identified: demographic and clinical metrics (age, GCS, temperature), lab results (anion gap, glucose, PT, RDW, WBC, urea nitrogen), and key clinical interventions or conditions (acidosis, mannitol, and sedatives). Among the seven algorithms, the XGBoost model achieved the best performance with the highest AUROC of 0.873 (95% CI: 0.842–0.906) and superior calibration (Brier score = 0.078). SHAP analysis identified GCS, age, anion gap, and glucose as the top mortality drivers, and revealed critical non-linear relationships, such as a U-shaped association between body temperature and mortality risk. Conclusion We successfully developed and validated an interpretable XGBoost-based model that demonstrates robust discriminative capacity, excellent calibration, and high sensitivity for predicting in-hospital mortality in ICU patients with TBI. By integrating SHAP analysis, the model provides intuitive explanations for both population-level risk drivers and individual predictions, effectively bridging the gap between predictive accuracy and clinical interpretability to facilitate personalized decision-making in the ICU.