Time-aware Embeddings of Clinical Data using a Knowledge Graph.

Area of Science:

• Biomedical informatics
• Machine learning in healthcare
• Clinical data representation

Background:

• Meaningful representations of clinical data are crucial for machine learning (ML) inference.
• Electronic health records (EHRs) contain rich patient information but require effective processing for ML applications.

Purpose of the Study:

• To propose a time-aware embedding approach for EHRs onto a biomedical knowledge graph.
• To create machine-readable patient representations that capture temporal dynamics and biological context.
• To evaluate the predictive performance of this approach for early disease detection.

Main Methods:

• Developed a time-aware embedding method integrating EHRs with a biomedical knowledge graph.
• Proposed the Temporal and Non-temporal Dynamics Embedded Model (TANDEM) ML pipeline.
• Applied TANDEM to the early detection of Parkinson's disease using the generated embeddings.

Main Results:

• The TANDEM pipeline achieved a classification Area Under the Curve (AUC) score of 0.85 on an unseen test dataset.
• The model successfully captured temporal dynamics and biological information from clinical data.
• Biological insights were derived from the knowledge graph to explain predictions.

Conclusions:

• Time-aware embeddings of clinical data provide meaningful representations for downstream ML tasks.
• This approach enhances predictive capabilities for clinical decision-making, as demonstrated in early Parkinson's disease detection.
• Integrating temporal dynamics and biomedical knowledge improves patient representation for ML.