Machine learning-based triage model for elderly traumatic brain injury patients in Chinese emergency department.

Background: Elderly traumatic brain injury (TBI) patients pose triage challenges in emergency departments due to complex physiology and comorbidities. Traditional methods often miss high-risk cases, necessitating accurate models to optimize ICU allocation. Methods: We developed an XGBoost-based triage model using data from 413 elderly TBI patients (aged ≥60 years) at Putian University Affiliated Hospital (2015-2024). Features included symptoms, CT hematoma density, contusion severity, age, anticoagulant use, and GCS scores. The target was ICU triage disposition within 48 h of ED presentation. Data were split (80:20) into training and testing sets. Class imbalance was handled within the training process using SMOTE, and model performance was assessed with 5-fold cross-validation using AUC, recall, precision, F1 score, and MCC. Because this was a prediction study rather than a causal inference analysis, all model outputs were interpreted as prediction of historical ICU triage disposition rather than proof of true ICU need. Complete-case analysis was used for model development after exclusion of records with missing key variables. SHAP analysis was used to improve interpretability. Results: The XGBoost model achieved an AUC of 0.93, recall of 0.9288, precision of 0.9274, F1 score of 0.9249, and MCC of 0.7241, outperforming the comparator models. Symptoms, CT hematoma density, and contusion severity were key predictors. Decision curve analysis suggested a higher theoretical net benefit than the "treat all" and "treat none" strategies across a clinically relevant threshold range; however, this finding should not be interpreted as a direct estimate of reduced unnecessary ICU admissions without prospective outcome-based validation. Conclusion: The XGBoost model provides an interpretable tool for predicting ICU triage disposition in elderly TBI patients and may support, rather than replace, emergency physician decision-making. Because the endpoint was a single-center process-of-care surrogate, further prospective and multicenter validation against patient-centered outcomes, including mortality, neurological deterioration, neurosurgical intervention, and functional status, is required.