Predicting Outcome of Patients With Cerebral Hemorrhage Using a Computed Tomography-Based Interpretable Radiomics...

Area of Science:

• Neurology
• Radiology
• Medical Informatics

Background:

• Cerebral hemorrhage prognosis prediction is crucial for patient management.
• Existing models may lack generalizability and interpretability.

Purpose of the Study:

• Develop and validate a multimodal radiomics model for predicting cerebral hemorrhage prognosis.
• Enhance model interpretability and generalizability.

Main Methods:

• Retrospective study with 237 cerebral hemorrhage patients across 3 centers.
• Extracted 1762 radiomics features from CT scans; incorporated clinical and macroscopic imaging features.
• Developed radiomics and radiomics-clinical models using random forest; validated with SHAPley Additive exPlanations (SHAP) for interpretability.

Main Results:

• The radiomics-clinical model achieved an AUC of 0.88, outperforming the radiomics model (AUC 0.85).
• SHAP analysis identified significant contributions of fusion features (rad score, clinical rad score) to model predictions.
• The model demonstrated high sensitivity, specificity, and calibration.

Conclusions:

• The multimodal radiomics-clinical model significantly improves prognostic prediction for cerebral hemorrhage.
• The model offers enhanced predictive performance and a basis for decision-making in risk prognosis.
• Interpretability through SHAP analysis aids in understanding model predictions.