Spatial radiomics-based interpretable multimodal machine learning model enhances outcomes prediction for minor...

Area of Science:

• Neuroimaging
• Machine Learning
• Stroke Medicine

Background:

• Accurate prediction of unfavorable outcomes is critical for managing minor stroke.
• Conventional radiomics models often overlook spatial lesion properties crucial for clinical insights.

Purpose of the Study:

• To develop and validate a novel spatial radiomics-interpretable model using machine learning to predict minor stroke outcomes.
• To quantitatively extract spatial features of lesions at various topological levels.

Main Methods:

• A cohort of 4,164 minor stroke patients was analyzed.
• Voxel-based and normative connection lesion analyses quantified spatial infarct features.
• A hybrid spatial radiomics model integrated these features with machine learning classifiers, interpreted using SHAP.

Main Results:

• The spatial radiomics model achieved superior prediction accuracy (AUC: 0.95/0.88/0.87) compared to conventional radiomics.
• Key predictors of unfavorable outcomes included lesion disconnection in corticospinal tracts, spinocerebellar tracts, and default mode regions.

Conclusions:

• The spatial radiomics model significantly enhances the prediction of unfavorable outcomes in minor stroke.
• This approach offers a novel, interpretable method within spatial-omics for improved stroke management.