Statistical Calibration of Predictive Entropy as a Surrogate Quality Metric for Generative Models on Large-Scale Wearable Health Time Series
Main article
Abstract
Generative models are increasingly used to restore, simulate, and harmonize large-scale wearable health time series. However, their outputs may contain hallucinated morphology, attenuated clinical events, or artifact patterns that remain visually plausible. This study develops a statistical calibration framework in which predictive entropy from a downstream classifier is treated as a surrogate quality metric for generated wearable signals. The proposed approach is motivated by decision-theoretic uncertainty quantification in wearable photoplethysmography analysis, where a generated signal is useful only when it preserves the information needed for a downstream clinical decision. We design an end-to-end pipeline for noisy photoplethysmography windows, generative denoising, atrial-fibrillation-oriented classification, entropy calibration, selective acceptance, and deployment monitoring. Using a large-scale simulated evaluation based on 136,882 wearable windows, the calibrated entropy score reduces uncertainty calibration error from 0.083 to 0.029 and improves the balanced accuracy of accepted generated samples from 0.716 to 0.779. The results show that entropy is not a universal quality measure by itself; rather, it becomes informative when calibrated against downstream decision loss, stratified by signal quality, and monitored for temporal drift. The article contributes a practical data-science framework for quality governance in generative wearable analytics and clarifies how entropy-based acceptance rules can support safer large-scale deployment without requiring reference clean signals for every generated instance.
