Graph-spa: A Spatiotemporal Graph Neural Network based framework for ARDS prediction and interpretability

Objective: Traditional deep learning models for multivariate time-series data often fall short in capturing long-range temporal dependencies critical for early prediction of the onset of acute respiratory distress syndrome (ARDS). To address this gap, we introduce Graph-spa , a dynamic Spatiotemporal Graph Neural Network (STGNN) based framework that not only improves ARDS prediction by modeling evolving interactions among clinical variables but also enhances interpretability through model-agnostic feature attribution. Methods: Graph-spa at its core integrates temporal convolution layers with an STGNN model that dynamically updates the adjacency structure, capturing both local and non-local temporal dependencies across three datasets (HiRID, MIMIC-IV, and eICU). We benchmarked our model against four traditional deep learning models (GRU, LSTM, TCN, Transformer) and an STGNN baseline. To complement the prediction framework, we applied mask-based interpretability approaches to generate feature-time attribution scores. These scores guide a subsequent co-occurrence analysis that identifies clusters of sustained feature activations in the 12-h window preceding ARDS onset. Results: Our experiments demonstrate that Graph-spa consistently outperforms the baseline models in both internal and external validations. On the AUC F1–MCC metric, chosen for this imbalanced classification task, Graph-spa achieves 50.02% vs 45.61% on HiRID, 48.52% vs 46.88% on MIMIC-IV, and 46.64% vs 45.41% on eICU-CRD compared with the STGNN baseline. Graph-spa also outperforms recurrent, convolutional, and attention-based models evaluated under identical settings (Wilcoxon signed-rank; Holm-adjusted p-values < 0.05). The dynamic adjacency enhancement allows the model to capture complex, evolving feature interactions, as evidenced by more diversified connectivity patterns compared to the baseline. In addition, interpretability analysis reveals that sustained abnormalities in potassium levels, along with declining Glasgow Coma Scale scores, form a critical composite risk profile that may serve as an early indicator of ARDS. Conclusion: Graph-spa advances dynamic clinical event prediction and also offers significant promise for early detection of organ failure in acute care settings by illustrating an end-to-end approach covering spatiotemporal modeling, interpretability, and discovery of sub-clinical signatures. Because its core modules, dynamic spatiotemporal graph construction, mask-based attribution, and co-occurrence mining, are model-agnostic, the framework can easily be extrapolated to any dynamic classification or regression task in the ICU. The code is available at https://github.com/vsubbian/Graph-spa .