Knowledge graph embedding and alignment of incomplete electronic health Records for critical care applications

Data-driven AI models in clinical care increasingly rely on integrating data from heterogeneous sources to improve predictive performance. Traditional Electronic Health Record (EHR) systems however pose significant challenges for data integration due to the use of diverse standards and inconsistent semantics. Personalized Health Knowledge Graphs (PHKG), which leverage biomedical ontologies to encode and interconnect clinical knowledge, have emerged as a promising solution for unifying heterogeneous health data. Yet the process of PHKG integration can result in incomplete and ambiguous representations. Our primary objective is to apply advanced embedding techniques to mitigate the incompleteness problem in PHKGs. We propose a framework that combines schema-based semantic integration, domain-specific ontology alignment, and patient context representations through embedding techniques to improve incomplete and ambiguous PHKG representations. We further demonstrate the framework’s capabilities for producing enriched representations in a use case concerning cardiovascular outcome prediction by training machine learning models which utilize the learned embeddings. We embed EHR from a critical care unit as structured PHKGs mapped with two different schemas for comparison across knowledge completion and alignment tasks following a modular design. Each module is individually optimized using three baseline embedding methods with enhanced loss functions. We evaluate each task for both accuracy and semantic consistency and show the contribution of each module to the overall performance. We find and report settings for each module in which the proposed framework outperforms baseline results. The learned representations are subsequently used to generate patient contexts for the task of Heart Failure diagnosis as a use case. Our experiments demonstrate that semantically enhanced PHKG embeddings achieve better precision and recall scores compared to baseline models. Our proposed method addresses the challenges in generating heterogenuous Personalized Health Knowledge Graphs (PHKG) through a modular framework that integrates schema mapping, ontology alignment, and contextual patient embeddings. Results on real-world patient records support the potential for improved performance in clinical decision making. Particularly in scenarios where sparse and fragmented health records prove problematic for data driven applications, our method can provide a robust approach for the disambiguation and completion of coded information. Our implementation is available at: https://github.com/AIDAVA-DEV/kge-framework.