RETRACTED ARTICLE: DynaGraph: interpretable dynamic graph learning for temporal electronic health records.

Area of Science:

• Machine Learning
• Computational Biology
• Health Informatics

Background:

• Electronic health records (EHRs) contain rich temporal data but are often analyzed with models that oversimplify their complexity.
• Existing machine learning models struggle to capture the dynamic relationships and temporal evolution within multivariate clinical time-series data.

Purpose of the Study:

• To introduce DynaGraph, a dynamic and interpretable graph learning framework for analyzing evolving spatio-temporal relationships in EHR data.
• To address challenges of class imbalance and temporal instability in clinical machine learning models.

Main Methods:

• DynaGraph constructs evolving spatio-temporal graphs from multivariate clinical time-series without predefined structures.
• It integrates sequential embeddings with contrastive graph augmentation and employs a pseudo-attention mechanism.
• A novel multi-loss objective combining focal, structural, and contrastive components is used for end-to-end training.

Main Results:

• DynaGraph consistently outperformed 14 state-of-the-art baselines across four large-scale EHR datasets.
• Achieved 6-8% relative improvements in area under the precision-recall curve (AUPRC) and 12-22% gains in sensitivity.
• Demonstrated time-specific interpretability, identifying temporally resolved risk factors and physiological relationships driving predictions.

Conclusions:

• DynaGraph offers a robust and interpretable framework for modeling complex temporal dynamics in EHR data.
• The method significantly improves predictive performance and provides clinically relevant insights into patient risk trajectories.
• This approach advances the application of machine learning in healthcare by effectively handling temporal instability and class imbalance.