Modern financial systems face escalating challenges in systemic risk prediction due to increasingly complex interconnections and heterogeneous contagion pathways, especially under the pressure of digital transformation. Existing approaches often fail to capture the nonlinear, multi-relational nature of risk propagation across dynamic market networks. This paper introduces a Spatio-Temporal Decoupled Heterogeneous Graph Network (STDHGN) that innovatively addresses these limitations. The framework combines spatio-temporal propagation for modeling dynamic market interactions with a structure-decoupled graph learning network that disentangles heterogeneous risk transmission patterns. By integrating hierarchical graph refinement and cross-temporal fusion, STDHGN effectively traces multi-layered contagion pathways while preserving temporal market dynamics. Extensive experiments on datasets from both the U.S. and China’s markets show that our model consistently outperforms state-of-the-art baselines in identifying high-risk financial entities, particularly during periods of elevated volatility. A real-world case study further demonstrates the practical value of our approach in anticipating and mitigating systemic financial risks. The proposed approach offers a robust analytical tool for monitoring systemic vulnerabilities in evolving financial ecosystems.