Construction and validation of a machine learning based prognostic prediction model for children with traumatic brain injury.

Objective: This study aimed to establish a prediction model for the short-term prognosis of children with traumatic brain injury (TBI) using machine learning algorithms.

Methods: The clinical data of children with TBI who were treated in the First Affiliated Hospital of Zhengzhou University were retrospectively analyzed. All children were divided into a modeling group and a validation group. In the laboratory indicators of the modeling group, the least absolute shrinkage and selection operator (LASSO) and multivariate Logistic regression analysis were used to screen out the independent influencing factors of poor prognosis in TBI, and a laboratory indicator model (LIM) was established. The risk scores of all patients were calculated. Then, the risk scores and other indicators were used to construct an extended prediction model through the extreme gradient boosting (XGBoost) algorithm. The discrimination, calibration, and clinical utility of the model were evaluated, and the extended model was explained using SHAP analysis. Finally, a subgroup analysis was performed using the risk scores to assess the robustness of the laboratory indicator model.

Results: Among the laboratory indicators, lactate dehydrogenase (LDH), N-terminal pro-B-type natriuretic peptide (NT-proBNP), hydrogen ion concentration index (pH), hemoglobin (Hb), serum albumin (Alb), and C-reactive protein to albumin ratio (CRP/Alb) were the independent influencing factors for the prognosis of children with brain injury. The extended model demonstrated excellent predictive performance in both the modeling and validation populations. SHAP analysis showed the contribution values of the Glasgow Coma Scale (GCS), the laboratory indicator model, the location of the head hematoma, the pupillary light reflex, and the injury severity score in the prediction of the overall patient prognosis. The subgroup analysis showed that there were differences in the risk scores of children with different GCS scores, pupillary light reflexes, and head hematoma locations, and there were also differences in the prognosis between the high-risk score group and the low-risk score group within them.

Conclusion: The extended model can accurately predict the prognosis of TBI patients and has strong clinical utility. The core model has good stratification ability and provides an effective risk stratification and personalized patient management tool for clinicians.