Construction of an Interpretable Model of the Risk of Post-Traumatic Brain Infarction Based on Machine Learning Algorithms: A Retrospective Study

Background: Post-traumatic cerebral infarction (PTCI) is a severe complication resulting from traumatic brain injury (TBI), which can lead to permanent neurological damage or death. The investigation of the factors associated with PTCI and the establishment of predictive models are crucial for clinical practice. Methods: We made a retrospective analysis of clinical data from 1484 TBI patients admitted to the Neurosurgery Department of a provincial hospital from January 2018 to December 2023. Predictive factors were identified using the Least Absolute Shrinkage and Selection Operator (LASSO) and multivariable logistic regression analysis. Several machine learning (ML) classification models were developed and compared. The interpretations of the ML models' predictions were provided by SHAP values. Results: Key predictors included age, bilateral brain contusions, platelet count, uric acid, glucose, traumatic subarachnoid hemorrhage, and surgical treatment. The logistic regression (LR) model outperformed other ML algorithms, demonstrating superior performance in the test set with an AUC of 0.821, accuracy of 0.845, Matthews correlation coefficient (MCC) of 0.264, area under the receiver operating characteristic curve (AUROC) of 0.711, precision of 0.56, and specificity of 0.971. It had stable performance in the ten-fold cross-validation. Conclusion: ML algorithms, integrating demographic and clinical factors, accurately predicted the risk of PTCI occurrence. Interpretations using the SHAP method offer guidance for personalized treatment of different patients, filling gaps between complex clinical data and actionable insights.