BackgroundEarly identification of delayed hematoma progression (DHP) in patients with frontal lobe contusion remains challenging in emergency settings. This study aimed to develop and externally validate an interpretable multimodal machine-learning model integrating routinely available clinical, laboratory, and CT imaging features to predict DHP.MethodsThis retrospective multicenter study included a development cohort of 799 patients and an external validation cohort of 443 patients. The development cohort was divided into a training set and an internal test set using stratified sampling. Feature selection was performed exclusively within the training set using seven complementary methods. Ten machine-learning algorithms were trained and compared using five-fold cross-validation. Model performance was assessed using AUROC, accuracy, sensitivity, specificity, precision, F1-score, calibration analysis, and decision-curve analysis. SHapley Additive exPlanations (SHAP) was used to interpret the final model.ResultsTen predictors were selected, including baseline contusion volume, hematoma density-related features, hematoma surface area-to-volume ratio, lymphocyte-to-monocyte ratio, admission Glasgow Coma Scale score, glucose-to-potassium ratio, time to baseline CT, and eosinophil count. The support vector machine (SVM) model showed the highest AUROC point estimate in the internal test set, with an AUROC of 0.801, and achieved an external validation AUROC of 0.724, indicating moderate external discrimination.ConclusionWe developed an interpretable multimodal model for early prediction of DHP in patients with frontal lobe contusion. The model may assist early risk stratification and clinical monitoring, but further prospective, multicenter, and geographically diverse validation is required before broad clinical implementation.