Learning Temporal Clinical Knowledge Graphs for Early Detectionof Multi-Organ Dysfunction in ICU Patients from Electronic HealthRecords using Machine Learning Techniques

Identification of potential clinical events at an early stage is still a challenging issue in intensive care settings owing to the complexity of dynamic interactions between multiple physiological systems. In this work, we introduce a Temporal Organ Interaction Graph Network (TOIGN) for predicting clinical outcomes based on longitudinal electronic health data. Our TOIGN model builds dynamic clinical graphs which depict inter-relationships between organ systems, laboratory measurements, vital signs and medication information. To find influential physiological interactions, our model employs graph attention techniques. Moreover, it utilizes temporal learning algorithms to learn about the illness process through multiple time steps. TOIGN concurrently performs predictions of acute kidney injury, respiratory failure, septic shock and MODS using multi-task learning paradigm. We evaluate our method on intensive care patients’ EHRs from the MIMIC-IV dataset against baseline machine learning, deep learning and graph-based methods. The outcomes showed significantly enhanced predictive power in terms of several metrics including 94.36% accuracy, 93.57% precision, 95.18% recall, 94.37% F1-score, 0.978 AUROC and 0.971 AUPRC. Moreover, the introduced method enabled earlier detection of the patient’s deteriorating state compared to previously developed approaches, which provided additional time for interventions. These results confirm that the application of the temporal model of organ interaction is a viable approach for enhancing early prediction and patient management in clinical environments.