HEPANEXUS: a multimodal graph attention network-based topological embedding integrating radiomic textures and clinical biomarkers layers in hepatocellular carcinoma prognosis

Hepatocellular Carcinoma (HCC) is one of the leading causes of mortality due to cancer globally. Transarterial Chemoembolization (TACE) is a standard intervention for intermediate-stage patients which hinders the outcome estimation due to significant inter-patient variability. Present models often process features independently, overlooking topological and relational correlations between medical images and clinical records. This research presents HepaNexus, a customized Multi-Task Graph Attention Network developed to establish an integrated framework for predicting recovery and therapeutic response. We employed an aggregate of 128 contrast-enhanced CT scans, extracting patient feature vectors and standardized clinical biomarkers modelled as nodes in a topological embedding. This architecture predicts Recovery Rate, binary TACE response, and mRECIST-based Degree of Response using a shared feature backbone and weighted multi-task loss simultaneously. HepaNexus achieved an MAE of 0.0909 for recovery prediction, corresponding to an 80% relative error reduction when compared with standard MLP baselines. In the classification tasks, the model achieved accuracies of 97.44% and 97.44% for TACE response and response degree, respectively, with a Macro F1-score of 0.8828. By transitioning from tabular instances to graph nodes, HepaNexus leverages inter-patient dependencies to encode nuanced clinical phenotypes, providing a more robust and prognostic approach to personalized prognostic assessment.