Predictive Modeling with Temporal Graphical Representation on Electronic Health Records.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Medical Informatics
• Deep Learning for Predictive Modeling

Background:

• Electronic Health Records (EHR) are crucial for deep learning-based healthcare predictions.
• Existing methods struggle to integrate temporal and structural EHR information effectively.
• Sequential models capture time but miss structural data; graphical models capture structure but miss temporal dynamics.

Purpose of the Study:

• To develop a novel patient EHR representation that integrates both temporal and structural information.
• To introduce a temporal graph transformer (TRANS) for enhanced EHR analysis.
• To improve the accuracy of deep learning-based predictive models in healthcare.

Main Methods:

• Modeled patient EHR as a temporal heterogeneous graph with visit and medical event nodes.
• Developed TRANS, incorporating temporal edge features, positional encoding, and graph convolution.
• Integrated structured information propagation and time-aware nodes for health status changes.

Main Results:

• The proposed temporal heterogeneous graph effectively captures both temporal and structural EHR data.
• TRANS demonstrated superior performance in integrating diverse EHR information.
• Extensive experiments on three real-world datasets confirmed state-of-the-art results.

Conclusions:

• The novel temporal heterogeneous graph and TRANS model offer a significant advancement in EHR representation.
• This approach enhances deep learning model capabilities for healthcare predictions.
• TRANS provides a robust framework for leveraging complex EHR data.