A multitask Transformer-VAE framework for robust PPG denoising and atrial fibrillation screening

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Photoplethysmography (PPG) is a noninvasive biosignal used for atrial fibrillation (AF) screening with wearable devices.
• Challenges in PPG-based AF detection include motion artifacts, noise, and intersubject variability, which distort waveforms and affect diagnostic consistency.

Purpose of the Study:

• To develop an end-to-end dual-task model for simultaneous PPG signal denoising and AF detection.
• To improve the reliability and accuracy of AF screening using wearable PPG devices.

Main Methods:

• A transformer-based encoder with two task branches and an alignment constraint was employed.
• The model was trained and evaluated on extensive public and internal PPG datasets.
• Performance was assessed using AUC and accuracy, with robustness tests under varying noise conditions and an A-Test for error stability.

Main Results:

• The proposed dual-task model achieved an AUC of 0.9097 and 88.4% accuracy, outperforming baseline models.
• Robustness experiments confirmed stable performance across different signal-to-noise ratios.
• The model preserved physiologically relevant rhythm features and demonstrated stable error behavior.

Conclusions:

• The Transformer-VAE framework enables accurate, noise-resilient AF detection from PPG signals by integrating signal reconstruction and diagnostic learning.
• The model can support wearable AF screening by providing risk indicators and signal-quality cues for potential ECG confirmation and clinical follow-up.