CTKGR: temporal knowledge graph enhanced prediction of in-hospital events for pneumonia patients using electronic health records.

OBJECTIVE: This study aims to develop an advanced clinical event prediction model leveraging the temporal characteristics embedded within electronic health record (EHR), with a specific focus on predicting clinical events during the hospitalization of pneumonia patients to enhance clinical decision-making. METHODS: We constructed a temporal knowledge graph (TKG) by systematically extracting clinical information from pneumonia patient EHR data across multiple time points. To effectively capture temporal dynamics and complex clinical interactions, we introduced the clinical TKG representation (CTKGR) model. This innovative model integrates graph convolutional networks (GCN), recurrent neural networks (RNN), and attention mechanisms to explicitly model clinical event progression. Additionally, a time length-aware module (TLAM) was developed to address challenges posed by irregular time intervals and varying frequencies of patient visits. RESULTS: The proposed CTKGR model exhibited superior predictive performance, achieving an AUC of 0.745 for symptom prediction and 0.798 for medication recommendation tasks, significantly outperforming traditional static and non-temporal methods. CONCLUSION: The CTKGR model effectively captures the temporal dynamics inherent in clinical data, demonstrating robust predictive capabilities for dynamically evolving clinical events. This study offers a novel and effective framework for AI-driven clinical decision support, with strong potential for application in pneumonia management.