ICU-TGNN: A Hybrid Multitask Transformer and Graph Neural Network Model for Predicting Clinical Outcomes of Patients in the ICU

Predicting clinical outcomes for patients in the intensive care units (ICUs) is crucial for physicians to assess clinical risk and provide timely and appropriate interventions. Current models often fail to predict multiple clinical outcomes simultaneously, and patient similarity has not been fully utilized to improve prediction accuracy. Additionally, electronic health record (EHR) data in the ICUs often contain observations recorded at irregular time intervals, which existing prediction methods do not effectively model. To address these challenges, we propose the ICU-TGNN model. This model combines time attention-based Transformer and graph neural network (GNN) architectures to leverage temporal and relational patient data. The Transformer component analyzes time series EHR data to capture dynamic patient states over time, while the GNN component exploits the relational structure among patients based on comorbidities to enhance model generalizability. We evaluated the ICU-TGNN model on the eICU-CRD dataset, demonstrating its superior capability in predicting general clinical outcomes such as ICU mortality and length of ICU stay. Our findings highlight ICU-TGNN's capability to provide accurate outcome predictions by effectively handling the complexity of ICU data, thereby holding great potential to optimize patient management and improve clinical outcomes.