Accurate and interpretable prediction of sepsis mortality in intensive care units (ICUs) is a critical challenge due to the complexity of multi-modal electronic health records (EHRs), which includes irregularly sampled physiological signals, static demographics, and intricate spatiotemporal dependencies. To address these challenges, we propose GraIS, a graph-based interpretable network. GraIS integrates dynamic and static graph learning to model organ system hierarchies and real-time physiological interactions. It employs attention-based fusion to align heterogeneous data streams, including irregular time series and static features. In addition, GraIS performs interpretable anomaly analysis to generate clinically actionable explanations through residual-based scores and graph-based feature importance. Evaluated on two large-scale EHR datasets, MIMIC-III and PhysioNet 2012, GraIS achieves state-of-the-art performance with ROC-AUC scores of 0.913 and 0.851, respectively, outperforming existing methods by 2.1–13.4% across metrics. Ablation studies confirm the contributions of each module, while case analyzes demonstrate its ability to highlight critical risk factors (e.g., lactate thresholds >4 mmol/L) and interactions. By bridging predictive accuracy with clinical interpretability, GraIS offers a reliable tool for ICU risk stratification and decision support.