MedDiffusion: Boosting Health Risk Prediction via Diffusion-based Data Augmentation.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Prediction

Background:

• Health risk prediction using Electronic Health Records (EHR) is crucial but often hindered by data insufficiency.
• Existing data augmentation methods struggle with task-unrelated designs, limiting their effectiveness.
• Novel approaches are needed to generate high-quality synthetic patient data for improved risk prediction.

Purpose of the Study:

• To introduce MedDiffusion, a novel end-to-end diffusion-based model for health risk prediction.
• To enhance risk prediction performance by generating synthetic patient data to enlarge the training sample space.
• To discern hidden relationships between patient visits for high-quality synthetic data generation.

Main Methods:

• Developed a diffusion-based model (MedDiffusion) for end-to-end health risk prediction.
• Utilized a step-wise attention mechanism to identify and retain vital information from patient visit sequences.
• Generated synthetic patient data during training to augment the dataset and improve model generalization.

Main Results:

• MedDiffusion significantly outperformed 14 baseline models on four real-world medical datasets.
• Achieved superior performance in Precision-Recall Area Under the Curve (PR-AUC), F1-score, and Cohen's Kappa.
• Ablation studies and comparisons with Generative Adversarial Network (GAN)-based models validated MedDiffusion's effectiveness and adaptability.

Conclusions:

• MedDiffusion offers a powerful and adaptable solution for health risk prediction, effectively addressing data insufficiency.
• The model's ability to discern patient visit relationships enhances synthetic data quality and interpretability.
• This diffusion-based approach represents a significant advancement in leveraging EHR data for proactive patient care.