Explainable machine learning reveals multifactorial drivers of early intracranial hematoma progression in traumatic brain injury: development of a SHAP-guided SVM nomogram

Background Early intracranial hematoma progression is a common and life-threatening complication of traumatic brain injury (TBI), associated with rapid neurological deterioration and poor outcomes. Accurate early identification of patients at risk remains challenging due to the multifactorial and nonlinear nature of underlying mechanisms. This study aimed to develop and validate an interpretable machine learning (ML) model for predicting early hematoma progression in TBI patients. Methods We retrospectively analyzed clinical data from 356 patients with TBI admitted to Qinghai University Affiliated Hospital. Patients were randomly divided into training (70%) and internal validation (30%) cohorts. A total of 25 demographic, radiological, and laboratory variables were evaluated. Predictive features were selected using least absolute shrinkage and selection operator (LASSO) regression and further confirmed by multivariable logistic regression. Five ML algorithms were constructed and compared. The optimal model was interpreted using Shapley additive explanations (SHAP), followed by the development of a nomogram. Performance evaluation and risk-stratification analyses based on both model-derived probability estimates and nomoscore stratification were performed to assess the clinical utility of the model. Results Early hematoma progression occurred in 49.7% (177/356) of patients. LASSO and logistic regression identified seven independent predictors: hematoma type, smoking history, age, D-dimer, monocyte-to-lymphocyte ratio (MLR), serum calcium, and multiple hematomas. Among the five algorithms, the support vector machine (SVM) achieved the best discrimination (training AUC = 0.937; validation AUC = 0.925), outperforming logistic regression, decision tree, XGBoost, and LightGBM. SHAP analysis confirmed the above variables as key contributors. The nomogram demonstrated strong predictive performance and interpretability. Rationality analyses showed that both model probability and nomoscore stratification exhibited stepwise increases in progression risk, validating the clinical robustness of the SVM-based model. Conclusion We developed and validated an interpretable SVM model that accurately predicts early hematoma progression in TBI patients. By integrating demographic, radiological, and laboratory features, this model provides a reliable tool for early risk stratification, guiding individualized management and timely intervention. Its strong performance across subgroups underscores its clinical applicability.