Knowledge graph driven medicine recommendation system using graph neural networks on longitudinal medical records

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Existing medicine recommendation systems often overlook patient medical history.
• Longitudinal models incorporating Electronic Health Records (EHRs) show promise but can be improved.
• Integrating diverse data sources like ontologies and drug-drug interactions is crucial for accurate recommendations.

Purpose of the Study:

• To propose a novel Knowledge Graph-Driven Medicine Recommendation System (KGDNet).
• To enhance medication recommendation accuracy and safety by utilizing longitudinal EHR data, ontologies, and drug-drug interaction knowledge.
• To develop a system that considers a patient's complete medical journey for personalized recommendations.

Main Methods:

• Constructed patient-specific clinical and medicine Knowledge Graphs using longitudinal EHR data, ontologies, and Drug-Drug Interaction (DDI) knowledge.
• Employed Recurrent Neural Networks (RNNs) to model historical patient data.
• Utilized Graph Neural Networks (GNNs) for learning embeddings from Knowledge Graphs.
• Implemented a Transformer-based Attention mechanism for generating medication recommendations based on current clinical state, medication history, and joint medical records.

Main Results:

• KGDNet demonstrated superior performance over existing methods on the MIMIC-IV EHR dataset.
• Achieved improvements in precision, recall, F1 score, and Jaccard score.
• Effectively controlled for Drug-Drug Interactions.
• An ablation study confirmed the importance of individual components and inputs for optimal performance.

Conclusions:

• KGDNet offers a significant advancement in medicine recommendation systems by integrating longitudinal EHR data with knowledge graphs.
• The system effectively balances recommendation accuracy with medication safety, including DDI control.
• Case studies indicate KGDNet's strong potential for real-world clinical application.