Explainable Dual-Stream Graph Learning With Attention Fusion for Multi-Stage Hepatitis C Classification

Hepatitis C is a chronic liver disease that can progress to fibrosis, cirrhosis and liver failure, requiring accurate multi-stage classification. This study proposes an explainable dual-stream deep graph neural network for five-class classification using 615 patient records from the UCI Hepatitis C dataset. A leakage-free pipeline employed two-stage stratified sampling, normalization, training-only SMOTE and sequential backward elimination, producing 350 training, 81 validation and 184 independent test samples. The framework integrates a clinical branch for biomarker representation with a KNN-based graph branch using attention-enhanced residual graph learning, followed by adaptive attention and multi-scale feature fusion. The proposed model achieved 98.91% accuracy, 98.95% precision, 98.91% recall and 98.90% $F$ 1-score, outperforming XGBoost (95.62%) and GAT (95.65%). Ablation analysis demonstrated progressive improvements from branch integration, attention fusion and multi-scale fusion. Class-wise ROC-AUC ranged from 0.9854 to 1.0000. McNemar's exact test showed a statistically significant difference compared with GAT ( $p$ = 0.0312). SHAP analysis identified AST, PROT, CHE and ALB as major biomarkers with significant global and stage-specific effects. Overall, the framework provides accurate and interpretable Hepatitis C stage classification while jointly leveraging clinical and patient-relational information.