DeepJ: Graph Convolutional Transformers with Differentiable Pooling for Patient Trajectory Modeling.

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Graph learning for Electronic Health Records

Background:

• Electronic Health Record (EHR) data contains complex medical event interactions.
• Existing graph learning methods struggle with longitudinal data, failing to model cross-encounter temporal dependencies.
• Static graph approaches limit the analysis of patient journeys over time.

Purpose of the Study:

• To introduce Deep Patient Journey (DeepJ), a novel graph convolutional transformer model.
• To effectively capture intra-encounter and inter-encounter medical event interactions.
• To identify groups of temporally and functionally related medical events for patient outcome prediction.

Main Methods:

• Developed DeepJ, a graph convolutional transformer model.
• Incorporated differentiable graph pooling for enhanced interaction modeling.
• Applied DeepJ to structured EHR data for longitudinal analysis.

Main Results:

• DeepJ successfully captured intra- and inter-encounter medical event interactions.
• Identified key event clusters relevant to patient outcomes.
• Outperformed five state-of-the-art baseline models in prediction tasks.

Conclusions:

• DeepJ offers improved modeling of patient journeys using EHR data.
• The model enhances interpretability and demonstrates potential for patient risk stratification.
• DeepJ advances the application of graph learning in clinical informatics.