Early prediction of incident delirium in traumatic brain injury: a multicenter validated and interpretable machine learning approach

Objective This study aims to develop and externally evaluate a machine learning (ML)-based predictive model for incident delirium in patients with traumatic brain injury (TBI). Methods Patients diagnosed with TBI from the MIMIC-IV and eICU-CRD databases were included. Predictors were selected using Boruta and LASSO regression. Five ML algorithms were developed and compared, with logistic recalibration applied to the external cohort. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). Shapley Additive Explanations (SHAP) was utilized to decode individual risk contributions. Subgroup and sensitivity analyses were conducted to define clinical boundaries and evaluate model robustness. Results A total of 915 TBI patients from the MIMIC-IV database and 317 from the eICU-CRD database were included. Random Forest (RF) model achieved balanced performance with an internal AUC of 0.819 and an external AUC of 0.706. The model exhibited favorable internal calibration, adequate external recalibration, and positive clinical net benefits (internal: 0.155, external: 0.080). Overall SHAP analysis identified invasive ventilation, Glasgow Coma Scale (GCS), extracranial injury, Acute Physiology Score III (APSIII), hemoglobin and mixed intra-/extra-axial injury as primary predictors. Crucially, stratified SHAP analysis identified invasive ventilation as the primary driver across all strata, with baseline GCS scores attaining their maximum predictive weight in the medium-risk tier. Subgroup analyses of the external cohort indicated robust generalization in younger patients (AUC = 0.780) and those with extracranial injuries (AUC = 0.762), with expected attenuation in subgroups with higher clinical severity (AUC: 0.578–0.589). Sensitivity analyses confirmed the model's stable performance against competing mortality and missing data (all DeLong test p > 0.05). Conclusions The RF model demonstrated acceptable discriminative capacity and clinical utility for early delirium prediction in patients with TBI. Supported by SHAP, it translated complex predictions into an actionable three-tiered framework, serving as a valuable adjunct for guiding early monitoring and neuroprotective strategies.