Predictive value of NCCT-based radiomics of hematoma and edema combined with clinical factors for 90-day poor prognosis in patients with spontaneous intracerebral hemorrhages

CHEN Xiaoyu; CHEN Sheng; LIN Pan; LI Chongyun; CAI Yong

doi:10.20258/j.cnki.1006-9011.2026.01.003

Abstract：Objective To investigate whether non-contrast CT(NCCT)radiomic signatures of hematoma and perihematomal edema,integrated with multidimensional clinical variables,can stratify 90-day poor outcome risk in spontaneous intracerebral hemorrhages(sICH).Methods Clinical and imaging data of 429 consecutive sICH patients were retrospectively collected.Pa-tients were dichotomized by 90-day modified Rankin Scale(mRS):good outcome 0-3(n=167)and poor outcome 4-6(n=262).After splitting into training(n=326)and validation(n=103)sets,radiomic features were extracted from manually segmented he-matoma and edema regions on NCCT.Gradient-boosting machines were trained separately for each region following mRMR and LASSO feature selection.Multivariable Logistic regression identified independent clinical predictors of poor outcome.Radiomic scores(RS)for hematoma and edema were computed and fused with clinical factors to build a combined model.Discrimination was assessed with ROC-AUC and clinical utility with decision-curve analysis(DCA).Results Eight hematoma and nine edema radiomic features remained after selection.Seven independent predictors of poor 90-day outcome were identified:age,ad-mission glucose,hematoma volume,edema volume,GCS,hematoma-RS,and edema-RS.The hematoma-only model achieved AUCs of 0.824(training)and 0.769(validation)and the edema-only model reached 0.844 and 0.756,respectively,with no sig-nificant differences.The combined radiomic-clinical model significantly outperformed either single-region model,with AUC of 0.933(training)and 0.851(validation),and provided the highest net benefit on DCA.Conclusion Quantitative NCCT ra-diomic features of both hematoma and perihematomal edema,when combined with readily available clinical variables,enable ac-curate early stratification of poor-outcome risk in sICH.