A transformer-based framework for temporal health event prediction with graph-enhanced representations.

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Computational Health

Background:

• Deep learning shows promise for predicting temporal health events.
• Existing methods struggle with comorbidity interactions and irregular patient data.

Purpose of the Study:

• To develop a deep learning model, GLT-Net, that addresses limitations in predicting temporal health events.
• To improve the prediction of future health events by better utilizing patient data.

Main Methods:

• GLT-Net combines Graph Learning and Transformer frameworks.
• It constructs patient association graphs and utilizes diagnosis code hierarchies.
• Graph neural networks and Transformer-Encoders capture comorbidity and temporal relationships.

Main Results:

• GLT-Net demonstrated superior performance on temporal health event prediction tasks.
• Experiments on real-world datasets confirmed its effectiveness against baseline models.
• A case study validated GLT-Net's predictive capabilities.

Conclusions:

• Understanding disease progression, comorbidity, and patient characteristics is key for event prediction.
• GLT-Net offers new insights into patient health status and disease trends.
• The model's architecture is versatile and adaptable to other data sources.