Construction of an interpretable prediction model for poor functional outcome in conservatively managed basal ganglia hemorrhage based on CT radiomics and multiple machine learning algorithms.
Authors
Affiliations (2)
Affiliations (2)
- Department of Neurosurgery, The First Affiliated Hospital of Shaoyang University, Shaoyang, Hunan, China.
- Department of Neurosurgery, The Traditional Chinese Medicine Hospital of Longhui County, Shaoyang, Hunan, China.
Abstract
Among patients with conservatively managed basal ganglia intracerebral hemorrhage (BG-ICH), a 3-month mRS ≥ 4 signals a poor functional outcome. Routine non-contrast CT reveals intralesional density heterogeneity, yet conventional clinical indicators largely overlook this information, weakening long-term prognostic accuracy. We retrospectively enrolled 254 conservatively treated BG-ICH patients and randomly split them into training and internal validation sets (7:3). Clinical variables were compared between outcome groups, and radiomic features were extracted from admission plain CT. Univariate filtering plus LASSO regression isolated robust predictors. Ten machine learning classifiers, including a SuperLearner ensemble, were trained and compared head-to-head. The best random forest (RF) model was assessed via ROC, calibration curves, decision curve analysis, and the KS test. SHAP analyses unpacked feature contributions and nonlinear links to continuous mRS scores. After dimension reduction, 14 predictive variables-mostly wavelet radiomic features-were retained. RF delivered the highest AUC (0.809), rising to a bootstrap-corrected 0.867 (95% CI, 0.797-0.929), together with good calibration and net clinical benefit. Admission GCS score and key wavelet markers dominated prognostic importance. This interpretable CT radiomics-based random forest model can stably predict poor functional outcomes at 3 months in patients with basal ganglia intracerebral hemorrhage receiving conservative treatment. Following adequate external cohort validation, this model may assist in individualized risk stratification and serve as a reference for subsequent clinical evaluations. At present, the model cannot be directly implemented in clinical practice, and external validation using independent multicenter cohorts is still required.