Abstract 5282: A Dynamic Knowledge Graph Registry Architecture Reveals Distinct Longitudinal Trajectories Across the Congenital Heart Disease Spectrum and Accurately Phenotypes Single Ventricle Physiology
Abstract
Introduction: Congenital heart disease (CHD) affects approximately 1% of live births and requires lifelong care. Conventional electronic health record (EHR) analyses rely on fragmented data and billing codes, limiting accurate phenotyping and longitudinal disease modeling. Dynamic knowledge graphs preserve temporal and relational clinical information, enabling characterization of disease progression across large patient populations. Hypothesis: We hypothesized that a dynamic knowledge graph integrating multimodal clinical data would improve computable phenotyping of single ventricle physiology (SVP) and enable longitudinal disease analysis beyond code-based approaches. Methods: We developed a dynamic knowledge graph registry comprising 25,000 CHD patients and more than one million encounters from 2013-2023 ( Figure 1 ). Diagnoses, procedures, laboratory values, and imaging metadata were integrated into an encounter-resolved graph containing 1.8 million nodes and 42.9 million edges. Graph representation learning generated patient embeddings capturing temporal and relational dependencies. A multimodal classifier incorporating EHR features, imaging reports, and graph embeddings was trained using five-fold cross-validation to identify SVP and benchmarked against billing code-based and expert rule-based methods. Survival analyses and patient similarity network analyses were performed across CHD severity groups Results: Survival differed significantly across CHD severity groups, with the lowest survival observed in complex CHD (log-rank p<0.001; Figure 2 ). Embedding-derived patient similarity networks identified significantly enriched SVP clusters (p<0.001), demonstrating the ability of dynamic multimodal representations to capture clinically meaningful patient relationships and disease patterns The multimodal SVP phenotyping framework achieved excellent performance (AUROC 0.994 ± 0.002; AUPRC 0.959 ± 0.010), with precision, recall, and F1-score of 0.906, 0.909, and 0.907, respectively, substantially outperforming billing code-based (0.424, 0.941, 0.585) and expert rule-based approaches (0.938, 0.853, 0.894; Figure 3 ). Conclusions: A dynamic knowledge graph registry integrating multimodal clinical data reveals differences in survival across the CHD spectrum and improves SVP phenotyping compared with conventional code-based approaches. Knowledge graph-based frameworks may support scalable longitudinal analysis of complex cardiovascular disease. Figure 1. Dynamic knowledge graph framework for CHD. (A) Data integration. (B) Registry. (C) Knowledge graph. (D) Embeddings. (E) Applications. Figure 2. Longitudinal outcomes by CHD severity. Figure 3. Knowledge graph-based SVP phenotyping. (A) Similarity network. (B) Similarity distribution. (C) Receiver operating characteristic curves. (D) Precision-Recall curves.