Abstract
Abstract
Background
Spontaneous intracerebral hemorrhage (ICH) has extremely high rates of disability and mortality. However, traditional CT imaging assessment relies solely on hematoma volume and rough anatomical location, making it difficult to account for prognostic heterogeneity and lacking precise individualized predictive tools. In recent years, radiomics has been able to extract texture and heterogeneity features from CT images at high throughput, and combined with machine learning algorithms, it has provided new possibilities for prognostic assessment.
Methods
In this retrospective, multicenter study, we included 2680 consecutive patients with spontaneous intracerebral hemorrhage, comprising a training cohort (n = 1876), an internal validation cohort (n = 804), and an independent external validation cohort (n = 196). We used deep learning algorithms to automatically segment hematomas and then extracted 1690 radiomics features from admission NCCT images. After reproducibility assessment, univariate screening, and LASSO feature selection, 42 valid radiomics features were integrated into a single radiomics feature spectrum (Rad-score). The Rad-score was combined with demographic features, clinical variables, and conventional CT radiomics features to construct two complementary machine learning models, used for functional outcome prediction and comprehensive prognostic stratification, respectively. Model performance was evaluated using R², mean squared error (MSE), area under the receiver operating characteristic (AUC), calibration analysis, decision curve analysis, and independent external validation. Furthermore, the possibility of ambient temperature variation at symptom onset as a prognostic modifier was explored.
Results
Among the 1,690 extracted radiomic features, 42 reproducible features were selected to construct the Rad-score. Incorporation of the radiomics signature substantially improved predictive performance compared with models based solely on conventional clinical and imaging variables. For functional outcome prediction, the integrated model increased the coefficient of determination (R²) from 0.35 to 0.93 while reducing the mean squared error (MSE) from 387 to 61. For prognostic risk stratification, the model based only on clinical variables achieved an AUC of 0.59, whereas the fully integrated model achieved an AUC of 0.95 and maintained robust performance in the external validation cohort (AUC=0.86).
Conclusions
We developed and validated a machine learning–based clinical decision support system integrating conventional CT radiomics and clinical variables for prognostic prediction and risk stratification in patients with ICH. By capturing quantitative imaging information beyond conventional CT markers, this framework significantly improved prognostic discrimination and demonstrated good generalizability across independent cohorts. These findings support the integration of quantitative imaging biomarkers into clinical decision-making for spontaneous ICH. Furthermore, the observed association between ambient temperature variation and disease progression warrants further investigation in prospective studies.