The proposed adaptive attention fusion spatiotemporal graph neural network achieved 96.3% anomaly detection accuracy and a 38.5-minute early warning capability, outperforming existing methods.
Does an intelligent vital sign monitoring system based on adaptive attention fusion ST-GNN improve anomaly detection accuracy and early warning capability in ICU patients compared to traditional and existing deep learning methods?
A novel spatiotemporal graph neural network integrated with reinforcement learning significantly improves the accuracy and early warning time for detecting vital sign anomalies in ICU patients.
Absolute Event Rate: 96.3% vs 91.7%
This study proposes a vital signs monitoring framework that addresses the limitations of traditional threshold-based alarms and existing deep-learning models in capturing multimodal physiological interactions and spatiotemporal dynamics.The method integrates an adaptive attention fusion mechanism that dynamically adjusts the importance of heterogeneous physiological parameters, a spatiotemporal graph neural network that jointly models inter-parameter correlations and temporal evolution using multi-scale windows, and a reinforcement learning module that enables active, strategy-driven early warning and clinical decision support.Evaluated on the MIMIC-III and eICU datasets, the proposed system achieves 96.3% anomaly detection accuracy, 38.5-minute early warning capability, and a 0.912 F1-score, outperforming existing methods.Ablation studies confirm the contributions of adaptive fusion, spatiotemporal graph modelling and policy optimisation.
Cheng et al. (2026) studied Intensive Care Unit (ICU) patients requiring vital sign monitoring (n=253,859). Adaptive attention fusion spatiotemporal graph neural network (ST-GNN) with reinforcement learning vs. Existing methods (MEWS, LSTM, MTAN, TCN, GAT) was evaluated on Anomaly detection accuracy on the MIMIC-III dataset. The proposed adaptive attention fusion spatiotemporal graph neural network achieved 96.3% anomaly detection accuracy and a 38.5-minute early warning capability, outperforming existing methods.