Enhanced Atrial Fibrillation Prediction in ESUS Patients with Hypergraph-based Pre-training.

Area of Science:

• Biomedical Informatics
• Cardiology
• Neurology

Background:

• Atrial fibrillation (AF) is a significant complication post-embolic stroke of undetermined source (ESUS), increasing stroke recurrence and mortality risk.
• Current AF detection methods lack accuracy, scalability, and cost-effectiveness.
• Machine learning (ML) shows potential but is limited by small ESUS patient cohorts and complex medical data.

Purpose of the Study:

• To develop advanced ML strategies for improved AF prediction in ESUS patients.
• To address limitations of existing prediction tools using novel pre-training techniques.

Main Methods:

• Introduced supervised and unsupervised hypergraph-based pre-training for patient embedding models.
• Pre-trained models on a large stroke cohort (7,780 patients) to capture complex interactions.
• Transferred embeddings to a smaller ESUS cohort (510 patients) for efficient, lightweight model prediction.

Main Results:

• Hypergraph pre-training approaches significantly outperformed traditional models trained on raw data.
• Achieved improved accuracy and robustness in AF prediction.
• Demonstrated effective dimensionality reduction while preserving key clinical information.

Conclusions:

• The proposed hypergraph-based pre-training framework offers a scalable and efficient solution for AF risk prediction in ESUS.
• This method enhances early identification of AF, potentially reducing recurrent stroke and improving patient outcomes.