Enhancing atrial fibrillation detection in PPG analysis with sparse labels through contrastive learning

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Photoplethysmography (PPG) is a promising wearable technology for detecting atrial fibrillation (AF).
• Current deep learning methods for AF detection require extensive labeled data.
• Self-supervised contrastive learning offers a potential solution to data scarcity.

Purpose of the Study:

• To explore the efficacy of self-supervised contrastive learning for AF detection using PPG signals.
• To investigate optimal data augmentation strategies within contrastive learning frameworks for PPG data.

Main Methods:

• Utilized 1,209 hours of unlabeled PPG data for self-supervised pretraining with SimCLR and BYOL frameworks.
• Investigated seven data augmentation operations, including single-sided and double-sided transformations.
• Fine-tuned pretrained encoder weights on small labeled datasets (MIMIC III, UMass, DeepBeat) for AF detection.

Main Results:

• Identified a preferred combination of single-sided transformation with the Drift operation for PPG data.
• Achieved superior F1 scores compared to supervised learning with only 1-20% of training data.
• Demonstrated a clear advantage over supervised learning even on a 0.01% DeepBeat training set.

Conclusions:

• Self-supervised contrastive pretraining effectively utilizes unlabeled PPG data for AF detection.
• This approach significantly reduces reliance on large labeled datasets.
• Offers a viable solution for AF detection challenges posed by limited labeled data.