International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Graph Attention Networks for Modeling Multi-Sensor Relationships in Early Prediction of Critical Events in ICU Patients

Authors :

Amir Akhavan Saffar¹ Danial Eskandari Faruji² Javad Hamidzadeh³

1- Faculty of Computer Engineering and Information Technology of Sadjad University 2- Computer engineering department, Hakim Sabzevari University 3- Faculty of Computer Engineering and Information Technology of Sadjad University

Keywords :

Graph Attention Networks،Multi-Sensor Physiological Signals،ICU Event Prediction،Temporal Dependencies،Explainable AI in Critical Care

Abstract :

Timely prediction of critical events like sepsis, cardiac arrest, acute hypotension, respiratory failure, and atrial arrhythmia in Intensive Care Unit (ICU) patients is crucial for reducing mortality and enabling preemptive care. These conditions stem from complex physiological interactions, necessitating advanced predictive tools to identify early signs before irreversible decline. We introduce a novel Graph Attention Network (GAT) framework that excels by exploiting inter-sensor relationships, dynamically tracking patient state changes with a multi-head attention mechanism, and using similarity edges to enable knowledge sharing among patients with similar profiles. The framework builds patient-specific graphs with nodes for physiological signals—electrocardiogram (ECG), arterial blood pressure (ABP), oxygen saturation (SpO₂), body temperature, respiratory rate, and heart rate variability (HRV)—within 300-second windows (50% overlap, resampled to 1 Hz). Edges reflect temporal sequences, physiological correlations , and inter-patient similarities , capturing unique patterns like HRV drops in sepsis or ABP-respiratory synchronization in hypotension while addressing data sparsity. Tested on a subset of the HiRID dataset from University Hospital Zurich, Switzerland (900 admissions), the model applies Z-score normalization, imputes 5% simulated missing data with forward/backward-fill and linear interpolation, and tackles class imbalance using weighted cross-entropy. Experiments yield a multi-class ROC-AUC of 0.94, surpassing baselines by 14.6% (e.g., LSTM at 0.82), with an F1-score of 0.88, precision of 0.89, and recall of 0.86. It processes 900 patients in 7 minutes on an NVIDIA RTX3080 GPU, supporting real-time use. Attention weights highlight HRV and SpO₂ as sepsis indicators, bolstered by similarity edges, enhancing interpretability. Challenges include single-center data and preprocessing leakage risks; future efforts will explore additional edges, multi-center validation, and refined data handling to elevate ICU care standards.