Machine learning-based radiomics model for predicting hematoma enlargement in spontaneous intracerebral hemorrhage: A multicenter study

This study aimed to establish and validate three predictive models: a clinical model (CM), a radiomics model (RM), and a combined clinical–radiomics model (CRM) for forecasting hematoma progression in patients with spontaneous intracerebral hemorrhage (sICH) within 24 h of admission.

A total of 477 patients were included in the training cohort (TC), while 106 patients were assigned to the validation cohort (VC). For the RM, radiomics features were first extracted, and the Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to identify the most relevant predictors. Ten different machine-learning algorithms were evaluated to determine the optimal model. The CM was developed using univariate and multivariate logistic regression analyses to identify independent clinical risk factors for hematoma expansion. The CRM was then constructed by integrating both CM and RM features.

Nine radiomics features were ultimately selected, and the Extra Trees algorithm was applied to build the RM. Independent clinical predictors of hematoma expansion included diabetes, the presence of the whirlpool sign, CT attenuation value, and baseline hematoma volume. In the TC, the area under the receiver operating characteristic curve (AUC) values for the CRM, RM, and CM were 0.901, 0.892, and 0.821, respectively. In the VC, the corresponding AUC values were 0.797, 0.780, and 0.786.

The CRM demonstrated superior predictive performance compared to the individual CM and RM, offering a reliable tool for early identification of hematoma enlargement in patients with sICH.

This work is funded by the Science and Technology Bureau of Jiujiang (Grant No. S2024ZDYFN0059), Jiangxi Provincial Health Commission (Grant No. SKJP1220242116).