Multimodal Graph-Based Model for Discrete-Time Survival Prediction in Liver Cancer

Liver cancer is a leading cause of cancer mortality; hepatocellular carcinoma (HCC), its predominant form, requires accurate survival prediction to guide prognosis and treatment decisions. We propose a multimodal framework for discrete-time survival prediction using histopathological and clinical data from the TCGA-LIHC cohort. Patch-level features extracted from Whole-Slide Images using the UNI2-h foundation model were aggregated within each slide through a Graph Attention Network, while clinical variables were encoded with a sentence transformer. The resulting embeddings were fused through intermediate fusion strategies and processed by a multi-layer perceptron trained using a negative discrete-time log-likelihood loss to handle censored data. Across two clinically meaningful survival intervals, [0, 1) and [1, 5] years, multimodal fusion outperformed unimodal baselines, with concatenation achieving the highest overall performance (AUROC = 0.818). These findings highlight the complementary prognostic value of histology and clinical data for robust and interpretable survival modeling in HCC.