Reproduce “MiGHT-EHR: A Multi-task Graph Transformer for Heterogeneous Temporal Electronic Health Records” on OpenResearch

About this paper

Anirudh Rayas, Yuan Wang, Pavan Turaga

Researchers from Arizona State University and the University of South Carolina developed MiGHT-EHR, a multi-task graph transformer that learns comprehensive representations from heterogeneous temporal Electronic Health Records. The model integrates diverse clinical entities, longitudinal patient trajectories, and multiple prediction tasks, achieving substantial gains, particularly for rare clinical events such as mortality and readmission, on large-scale EHR datasets.

Key insights

• Learned representations effectively organize patient neighborhoods by clinical outcomes (e.g., mortality, readmission), rather than just diagnostic similarity.
• Clinically meaningful concept groups are recoverable as linear directions within the latent space, indicating the model's ability to capture structured clinical context.
• The multi-task learning framework provides robustness and generalization, with generally well-calibrated predictions across tasks.

Summary by alphaXiv.