Generative Cleaning Pipelines for Consumer Sensor Cardiology: From Artifact Removal to Clinical-Grade Screening Photoplethysmography; wearable sensors; generative adversarial networks; signal denoising; atrial fibrillation screening; uncertainty quantification; deep learning; trustworthy AI
Main article
Abstract
Background: Consumer wearables now record photoplethysmography (PPG) almost continuously, opening a route to population-scale cardiac screening; yet signals acquired outside the clinic are routinely corrupted by motion, ambient light, and poor sensor contact, and a large share of recorded data is discarded before any analysis. Objective: This review examines how generative machine-learning models—particularly conditional adversarial and related deep networks—are reshaping the cleaning of wearable cardiac signals, and how that cleaning can be connected, end to end, to clinically meaningful screening decisions. Methods: We synthesize evidence across four layers of the processing chain: signal-quality assessment, generative denoising and reconstruction, downstream cardiac inference, and trustworthiness evaluation, drawing on recent methodological and clinical studies. Results: The literature indicates that learned reconstruction can recover diagnostic morphology that classical filtering removes, that downstream task performance is the most informative measure of cleaning quality, and that uncertainty-aware gating markedly lowers the cost of acting on unreliable outputs. However, generative models can also fabricate plausible but spurious waveform features, making calibrated confidence estimates indispensable before clinical deployment. Conclusion: Treating artifact removal and screening as a single, uncertainty-aware pipeline—rather than as isolated steps—offers a credible path from consumer-grade sensing to clinical-grade decision support, provided that evaluation, calibration, and prospective validation keep pace with model capability.
