From Basic to Extra Features: Hypergraph Transformer Pretrain-then-Finetuning for Balanced Clinical Predictions on...

Area of Science:

• Biomedical Informatics
• Machine Learning
• Data Science

Background:

• Electronic Health Records (EHRs) are vital for clinical research and practice, containing extensive patient data.
• Current deep learning models for EHRs often require extensive features, limiting their applicability to all patients.
• Integrating diverse patient data features into EHR modeling remains a challenge.

Purpose of the Study:

• To introduce HTP-Star, a novel framework for modeling EHR data using hypergraphs.
• To enable seamless integration of additional patient features into EHR analysis.
• To enhance the robustness of EHR models during the finetuning process.

Main Methods:

• Leveraging hypergraph structures for EHR data representation.
• Employing a pretrain-then-finetune framework for model development.
• Implementing Smoothness-inducing Regularization and Group-balanced Reweighting techniques for robustness.

Main Results:

• HTP-Star demonstrated superior performance compared to baseline models on two real-world EHR datasets.
• The model effectively balanced patients with both basic and supplementary features.
• Enhanced robustness was observed during the finetuning stage.

Conclusions:

• HTP-Star offers an effective hypergraph-based approach for modeling EHR data.
• The proposed techniques improve model robustness and feature integration capabilities.
• This framework advances the utility of EHR data in clinical research and practice.