Interpretability

Methylation value had the highest node-attribute importance. CpG island features, next-base-pair attributes, and distance to transcription start site were important model contributors. In COPD, LIME indicated that lower SOFA scores, completed immunisations, and non-Medicare insurance supported classification as not readmitted. Graph-level interpretability found that the co-methylation graph had higher average importance than same-gene and same-chromosome graphs. Model transparency was supported through SHAP and LIME. The study combines Integrated Gradients with graph-branch occlusion for interpretability. Graph importance is estimated by masking each branch and normalizing the absolute prediction change. Mean absolute Integrated Gradients over feature dimensions gives node importance, and averaging over nodes gives feature importance. SHAP was used for global feature attribution and LIME was used for patient-level explanations. In HF, LIME associated Charlson Comorbidity Index above 8 and SOFA above 7 with increased local readmission risk. In T2DM, LIME associated SOFA above 6 and Charlson Comorbidity Index above 5 with readmission predictions.