Integrating knowledge graph embeddings with clinical data: A case study on Acute Kidney Injury prediction

The integration of artificial intelligence (AI) and machine learning (ML) in healthcare has significantly advanced predictive analytics. Acute Kidney Injury (AKI), a condition marked by sudden kidney function loss, necessitates early intervention to improve patient outcomes. This study introduces a novel approach that leverages knowledge graph embeddings (KGE) to enhance the predictive accuracy of ML models for AKI detection. Knowledge graphs (KGs) model entities and their interrelations in a graph format, integrating heterogeneous data sources to provide a comprehensive view of complex biological systems. The AI contribution lies in the application of embedding techniques that transform these graphs into continuous vector spaces, improving the ability to capture semantic similarities and infer new relationships within the data. On the engineering side, we applied this AI-driven approach to healthcare by leveraging the Medical Information Mart for Intensive Care (MIMIC) III dataset to construct a KG, generate embeddings, and incorporate them into ML models for AKI prediction. This engineering application aims to demonstrate the utility of KGEs in clinical settings. Our approach involved extracting patient features, generating KGE, and training various models, such as those utilizing complex, translational, holographic, and multiplicative embeddings. The models were evaluated through ranking-based and binary classification protocols. The transparency of the AI models enhances their trustworthiness in clinical practice, and the findings underscore the need for continued collaboration with clinicians to refine these techniques and ensure successful deployment in healthcare.