Medical Applications of Graph Convolutional Networks Using Electronic Health Records: A Scoping Review

Electronic health records (EHRs) contain complex, interconnected patient data that traditional analytical methods often fail to fully leverage. Graph convolutional networks (GCNs) offer a promising machine learning approach for modeling intricate relationships within tabular, multimodal clinical data to improve medical decision-making and patient outcomes. This scoping review presents the first comprehensive synthesis of current research applying GCNs to EHR data for medical applications. We conducted a systematic literature search across medical databases (MEDLINE Ultimate) and computational academic databases (Academic Search Ultimate, Business Source Complete, Supplemental Index, and Complementary Index), using combined search terms for GCNs and EHRs. Following the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) protocols, we screened 56 database-identified papers and five additional articles from citation review. After independent abstract screening and full-text review by multiple reviewers, 25 peer-reviewed English-language studies were included in the final analysis. Medical informatics and critical care were the most common application domains. Clinical tasks performed focused on diagnosis prediction, automated medical coding, mortality risk assessment, and medication recommendation. Hybrid architectures integrating GCNs with models such as Bidirectional Encoder Representations from Transformers (BERT) and recurrent neural networks were employed in 60% of studies. ICU EHR datasets were the most frequently used data source. Temporal trends revealed growing research interest, with eight publications (32% of the total corpus) appearing in early 2024. GCNs demonstrated measurable performance improvements in predictive accuracy, drug interaction identification, and risk stratification over other machine learning approaches. Enhanced clinical decision-making was supported through automated extraction of complex data relationships and improved handling of unstructured clinical text, with applications in critical care, oncology, hematology, and cardiovascular domains. Reported limitations included data-related challenges (68%), model-related issues (60%), and application-level barriers (28%), involving data heterogeneity, model generalizability, and clinical integration. GCNs effectively leverage complex relationships within EHR data to support clinical decision-making across diverse medical domains. The predominance of hybrid architectures and measurable performance improvements demonstrate the technology's potential to transform healthcare delivery by improving diagnostic accuracy, personalizing treatment recommendations, and facilitating early identification of high-risk patients requiring proactive intervention. A phased, clinically grounded implementation strategy, supported by strong data governance, multidisciplinary integration, and hybrid AI architectures, will be essential for realizing GCNs' full potential in real-world healthcare settings.