Multimodal Graph Convolutional Networks for Patient Survival Analysis

Patient survival analysis is an important task in medical artificial intelligence, aiming to estimate the probability of survival or the risk of clinical events over time based on multimodal patient data. However, existing methods often rely on simple fusion strategies, such as direct concatenation of features, which limits their ability to capture meaningful cross-modal interactions. Many approaches rely on full fine-tuning of largelanguage models (LLMs), which is computationally expensive and prone to overfitting, especially in small-scale clinical datasets. In addition, most models fail to incorporate patient-level similarity, which hinders their generalization in complex clinical scenarios. To address these limitations, we propose MM-GSA, a Multi-Modal Graph Convolutional Network for Patient Survival Analysis. Our method integrates adapter-finetuned LLM with FTAttention to encode unstructured and structured data. Based on these multimodal embeddings, we construct a patient similarity graph and apply graph convolutional networks (GCNs) to capture relational dependencies and enhance representation learning. The model is trained in two stages: multimodal representation learning and graph-based survival modeling. Extensive experiments on two public clinical datasets, TCGA and MIMIC III, demonstrate that MM-GSA consistently outperforms existing baselines on standard survival analysis metrics. The results validate the effectiveness of combining efficient large language model fine-tuning, structured feature encoding, and graph-based patient modeling to improve both predictive accuracy and model robustness.