Prediction of hematoma expansion in deep intracerebral hemorrhage using baseline non-contrast CT and clinical variables: a multicenter multimodal study

Hematoma expansion (HE) is a major determinant of neurological deterioration and poor outcome after intracerebral hemorrhage (ICH), yet early risk stratification remains challenging. We developed and validated a multicenter multimodal model for predicting HE in patients with deep ICH using baseline non-contrast computed tomography (NCCT) and clinical variables. In this retrospective study, 539 patients from three hospitals were included, including 229 in the training cohort, 99 in the validation cohort, and 211 in the independent external test cohort; 98 patients (18.2%) experienced HE. Baseline NCCT images were converted into 2.5D inputs using the slice with the largest hematoma area and adjacent slices. Deep imaging features were extracted and fused with clinical variables, followed by feature selection and classifier construction. The Transformer-hybrid model achieved the AUC of 0.839 in the external test cohorts. In the external test cohort, the model achieved high specificity at the default threshold and showed a higher AUC than all six baseline NCCT radiological signs, whose highest AUC was 0.569. This 2.5D multimodal framework may serve as a candidate adjunctive approach for early HE risk stratification in deep ICH and requires prospective validation before clinical use.