Risk Minimization for Trustworthy GAN-Assisted Atrial Fibrillation Screening: A Clinical Confidence Modeling Approach
Main article
Abstract
Atrial fibrillation is a leading cause of stroke and sudden cardiac incidents, and consumer-grade wearable photoplethysmography devices are increasingly used for opportunistic screening. However, ambulatory recordings are corrupted by motion artifacts, ambient light fluctuations, and skin-contact variability, which compromise the diagnostic accuracy of pretrained classifiers. Generative adversarial networks have emerged as a popular tool for restoring noisy wave forms, yet their tendency to produce subtle hallucinations raises a serious obstacle to clinical deployment: physicians cannot tell whether a denoised signal is faithful or has been silently rewritten. This study reframes that obstacle as a Bayesian decision problem and proposes a clinical confidence modeling pipeline that explicitly minimizes expected misclassification cost rather than reconstruction error. A one-dimensional Pix2pix-style conditional generator restores noisy photoplethysmography waveforms, after which a pretrained AlexNet-1D classifier produces both a posterior estimate and a normalized predictive entropy. The entropy is treated as a clinical confidence score that drives selective rejection at the screening level. Calibration is assessed with the Uncertainty Calibration Error and is externally grounded against downstream classification accuracy on a held-out cohort drawn from a balanced split of the Deepbeat dataset. Across the full retained cohort, denoising recovers an area under the receiver operating characteristic curve of 0.83 from 0.75 on noisy inputs, while rejecting the lowest-confidence quartile lifts the area under the curve to 0.86 and the Matthews correlation coefficient at fixed sensitivity from 0.43 to 0.53. The framework offers a deployment-ready strategy in which generator hallucinations are screened out before they reach the cardiologist.
