A Downstream Task-Driven Paradigm for Evaluating Conditional GAN Output Reliability in Ground-Truth-Free Scenarios: A Case Study on Physiological Time Series

Abstract

Conditional generative adversarial networks (cGANs) are routinely used to adapt physiological signals across acquisition domains so that downstream classifiers can be reused without retraining. A persistent obstacle to deploying these models in clinical or consumer settings is the absence of a per-instance reliability indicator that does not require paired ground-truth references. Standard image-quality scores, distributional metrics such as the Fréchet distance, and likelihood-style proxies are either defined globally, depending on visual statistics that do not transfer to one-dimensional waveforms, or rely on assumptions specific to diffusion models. This paper develops a downstream task-driven paradigm in which the predictive entropy of a fixed downstream classifier serves as the per-instance trustworthiness signal for a cGAN-adapted output. The approach is grounded in a decision-theoretic interpretation of misclassification cost, which yields a Bayes-optimal accept/reject rule and a calibration metric that uses no labels for the generated waveforms themselves. The framework is evaluated on a 25-second wearable photoplethysmography dataset for atrial fibrillation detection. A one-dimensional Pix2pix-style cGAN is trained to reverse additive noise on the test side of the domain. Selecting the lowest-uncertainty 75% of adapted waveforms recovers the AUROC of the clean-input upper bound (0.85 vs. 0.84) and reduces the Uncertainty Calibration Error from 0.071 on noisy inputs to 0.027 on adapted inputs. The Pearson correlation between noisy and adapted entropies is 0.71, indicating that the uncertainty signal tracks generator-induced changes rather than residual measurement properties. These results show that a downstream task can act as a principled, label-free reliability oracle for conditional generators in physiological time-series analysis.

Keywords: Conditional generative adversarial networks; Domain adaptation; Uncertainty quantification; Decision theory; Photoplethysmography; Atrial fibrillation; Trustworthy machine learning