A new model for risk assessment of delayed imaging deterioration in moderate traumatic brain injury
GU Yufeng
HOU Yueru
BIAN Renjie
XIAO Haiming
LI Qiang
LI Wan
ZHANG Qihang
FENG Dayun
GE Shunnan
Abstract：Objective To explore the characteristics and related factors of delayed deterioration in patients with moderate traumatic brain injury(mTBI),and to construct an assessment model for delayed imaging deterioration.Methods A multicenter retrospective cohort study design was adopted.A total of 363 mTBI patients admitted from January 2015 to December 2022 were enrolled,including 233 patients in the training set(single-center)and 130 patients in the validation set(multi-center).Automatic hemorrhage/edema segmentation and brain region segmentation of CT images were achieved based on the nnUNet deep learning framework,and CatBoost machine learning algorithm was combined to construct a prediction model with"progression of cerebral contusion volume＞80％within 72 h of admission"as the outcome event.The key predictors were analyzed by Shapley Additive exPlanations(SHAP),and SMOTE resampling and cross-validation were used to optimize the performance of the model.Results The AUC of the final model was 0.80[95％CI(0.76,0.84)],which was significantly better than that of the traditional scoring system.SHAP analysis showed that temporal lobe hematoma volume(SHAP value 6.31),long axis of hematoma(SHAP value 5.15),white blood cell count(SHAP value 4.68),fibrinogen(SHAP value 4.07),edema volume(SHAP value 3.85)and serum calcium(SHAP value 3.72)were the key predictors.The model verification showed that the accuracy was 80.65％,and the F1 score was 0.79,which had good clinical discrimination ability.Conclusion The CatBoost model constructed based on the admission case data of mTBI patients integrates multimodal data and systematically quantifies the synergistic predictive value of medical history data,cerebral contusion lesion morphological parameters and systemic inflammation/coagulation indicators,providing an intelligent decision-making tool for the hierarchical management of mTBI.
Keywords：brain injuriestraumaticdisease progressiontomographyX-ray computeddeep learningmachine learninghematomaclinical deteriorationalgorithms
Publication Date：2026-02-28
Online Publishing Date：2026-08-26(First online date of this platform, not the publication date of the document)
Pages：7( 224-230 )