GraphSurv: A Graph-Based Framework for Dynamic Survival Analysis With Longitudinal Data

Dynamic survival analysis in longitudinal biomedical studies updates individualized event-risk estimates as repeated clinical measurements accumulate. These data are often irregularly sampled, multivariate, and incomplete, requiring models to capture both temporal dependencies across visits and cross-variable dependencies within visits. To address these challenges, we propose GraphSurv, a graph-based framework that represents longitudinal histories through visit-level temporal states and variable-specific representations. Bidirectional message passing and explicit inter-visit interval encoding jointly capture cross-variable and temporal dependencies, and the learned representations are used for discrete-time hazard modeling and dynamic survival prediction. Experiments on real-world longitudinal survival datasets show competitive performance across discrimination and calibration metrics, with the most consistent improvements in time-dependent discrimination. Landmark evaluations further demonstrate adaptive risk updating as follow-up information accrues. SHAP-based interpretation of the predicted risk scores also indicates that the proposed framework can reveal clinically meaningful and time-varying feature contributions to risk prediction.