Predicting Future Disorders via Temporal Knowledge Graphs and Medical Ontologies

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Knowledge Representation

Background:

• Electronic Health Records (EHRs) contain valuable clinical information, but free-text data remains underutilized.
• Structuring EHR free-text using Named Entity Recognition and Linking (NERL) can unlock insights.
• Integrating structured EHR data with medical ontologies enhances medical history analysis.

Purpose of the Study:

• To propose MedTKG, a Temporal Knowledge Graph (TKG) framework for leveraging EHR data.
• To model patient medical histories dynamically using time-series snapshots.
• To predict future patient disorders by analyzing temporal and ontological information.

Main Methods:

• Developed MedTKG, a framework combining dynamic patient history snapshots and static medical ontologies.
• Modeled medical history as a series of time-stamped data points.
• Utilized a knowledge graph approach to predict missing disorder information in patient records (s, r, ?, t).

Main Results:

• MedTKG effectively predicts future disorders from clinical notes.
• The framework was evaluated using the MIMIC-III clinical database.
• Incorporating medical ontologies significantly improved the prediction performance of the TKG.

Conclusions:

• MedTKG offers a robust framework for extracting value from EHRs.
• Temporal knowledge graphs are effective for modeling dynamic patient health trajectories.
• The integration of medical ontologies enhances the accuracy of predictive models in healthcare.