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

End-to-end dual-task model that jointly denoises PPG and detects AF: Transformer encoder with two task branches plus a lightweight alignment constraint so reconstruction and diagnosis stay consistent under motion artifacts. Trained/evaluated on public PPG (internal ~30,773 segments / 91 patients) plus external MIMIC-III-derived set. Reported AUC 0.9097 and accuracy 88.4%, outperforming convolutional and single-task baselines; robustness tests across SNRs; A-Test (repeated balanced k-fold) for split stability.

• Form factor: other

Background

Photoplethysmography (PPG) is a noninvasive biosignal widely used for atrial fibrillation (AF) screening via wearable devices. However, reliable PPG-based detection remains challenging because motion artifacts, noise contamination, and intersubject variability distort waveform morphology and compromise diagnostic consistency.

Methods

We propose an end-to-end dual-task model that denoises PPG signals and detects AF simultaneously. The model uses a transformer-based encoder with two task branches and an additional lightweight alignment constraint to encourage the two tasks to learn consistent representations under motion artifacts. We trained and evaluated the framework on public PPG datasets, including an internal cohort of 30,773 segments from 91 patients and an external dataset derived from the MIMIC-III waveform database. Performance was evaluated via the AUC and accuracy, together with robustness tests across different noise conditions. We also applied an A-Test procedure that repeats balanced k-fold validation with different k values to assess error stability across data splits.

Results

The proposed model achieved an AUC of 0.9097 and an accuracy of 88.4%, outperforming conventional convolutional and single-task baselines. Robustness experiments demonstrated stable performance across varying signal‒to‒noise ratios. The dual-task architecture preserved physiologically relevant rhythm features, including heart rate variability. A-Test evaluation indicated stable error behavior across data splits, suggesting reduced partition sensitivity and improved learning consistency.

Conclusions

This Transformer–VAE framework integrates signal reconstruction and diagnostic learning to enable accurate and noise-resilient AF detection from PPG signals. Within a healthcare workflow, the model supports wearable screening by generating AF risk indicators and signal-quality cues that can trigger confirmatory single-lead ECG assessment and appropriate clinical follow-up.

Trial registration

Not applicable.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.