Influence of early temperature trajectories on clinical outcomes in traumatic brain injury: a multicenter validation study using machine learning.

objectivesTemperature management is a critical intervention to mitigate secondary injury in patients with traumatic brain injury (TBI). This study, based on the MIMIC database and externally validated using the eICU database, analyzed early 24-h temperature trajectories of TBI patients after ICU admission to investigate their association with clinical outcomes.

methodsLatent Class Mixed Model (LCMM) was employed to classify the 24-h temperature trajectories of TBI patients following ICU admission. Logistic regression models were constructed based on univariate selection, Boruta feature selection, and all variables to evaluate mortality risk across trajectory subtypes. Subgroup analyses were also performed. Furthermore, machine learning models were constructed using variables jointly selected by LASSO and Boruta, with multiple algorithms (Random Forest, XGBoost, LightGBM, logistic regression, SVM, and KNN) compared against traditional severity scores via DeLong's test.

resultsA total of 3249 TBI patients from the MIMIC database and 3246 patients from the eICU database were included, with temperature trajectories categorized into three distinct classes. According to the full-variable logistic model, patients in Class 1 and Class 3 exhibited significantly worse prognosis compared to Class 2 (p < 0.001). Sensitivity analyses yielded consistent results. Notably, the Random Forest model demonstrated superior predictive performance compared with conventional severity scores, such as SAPS II and OASIS.

conclusionsTemperature trajectories within the first 24 h of ICU admission are associated with clinical outcomes in TBI patients. Early identification of temperature trajectory subtypes facilitates timely recognition of high-risk patients with poor prognosis, enabling personalized temperature management strategies.