Creation and validation of a mortality risk prediction model for ICU patients with traumatic brain injury: a...

Traumatic brain injury (TBI) stands as a major global cause of mortality and disability. Accurate prediction of in-hospital mortality is crucial for optimizing clinical management of TBI patients in the intensive care unit (ICU). However, existing prognostic models demonstrate significant limitations in short-term prediction and clinical immediacy. This study aims to develop and validate a practical prognostic model to address these research gaps and provide clinicians with a precise risk stratification tool.

This study integrated data from two publicly available databases, MIMIC-IV and eICU-CRD. The MIMIC-IV database was utilized for model development and internal validation, while the eICU-CRD database was employed for external validation. The research enrolled TBI patients admitted to the intensive care unit as study subjects. During the data preprocessing phase, continuous variables with missing rates below 30% were handled using multiple imputation, while missing values in categorical variables were retained as a separate category. Variables with missing rates exceeding 30% were excluded. Subsequently, eligible cases from the MIMIC-IV database were randomly divided into training and testing sets at a 7:3 ratio. Based on multiple regression analysis supplemented by LASSO regression screening, we developed a risk assessment model to identify independent predictors of short-term mortality in ICU-admitted TBI patients. Finally, the model's performance was systematically evaluated across three dimensions: discrimination, calibration, and clinical utility.

Strictly adhering to the inclusion and exclusion criteria, we ultimately enrolled 3604 TBI patients across the two databases. The final model incorporated seven independent predictors: APS-III score, age, use of mechanical ventilation, respiratory rate, prothrombin time, sodium level, and anion gap. In the training set, the 7-day mortality prediction model demonstrated excellent discriminative ability, with an AUC value of 0.829 (95% CI 0.803-0.855), a sensitivity of 78.6%, and a specificity of 72.6%. The model's performance further improved in the test set, achieving an AUC value of 0.871 (95% CI 0.822-0.921), with sensitivity and specificity increasing to 83.1% and 77.6%, respectively. During external validation, the model also exhibited robust predictive performance, yielding an AUC value of 0.757 (95% CI 0.711-0.803) for 7-day mortality, along with a sensitivity of 67.3% and a specificity of 74.8%, further confirming its generalizability. Following bootstrap internal validation, the predictive model demonstrated excellent performance. It exhibited strong discriminatory power (corrected AUC = 0.8309, 95% CI 0.8241-0.8356) and favorable overall predictive accuracy (Brier score = 0.1126, 95% CI 0.1117-0.1138). Calibration analysis confirmed model reliability: the calibration intercept approached zero (0.0068, 95% CI 0.0078-0.0219), indicating no systematic overestimation or underestimation, while the calibration slope approached unity (0.9622, 95% CI 0.8677-1.0668), demonstrating excellent alignment between predicted probability ranges and actual risk variations. These metrics collectively demonstrate the model's strong clinical applicability, and its predictions can be reliably used to guide clinical decision-making. The calibration curve also demonstrated high consistency, while decision analysis revealed significant clinical net benefit across different risk thresholds.

This study developed a predictive model that estimates short-term mortality for TBI patients in the ICU using seven routinely available clinical variables. The model demonstrated robust performance in external validation. Its design, which enables multi-timepoint assessment, may facilitate risk stratification and has the potential to support clinical decision-making, pending future prospective validation.