Enhancing the Physiological Plausibility of GAN-Generated Blood Glucose in Type 1 Diabetes with Monotonicity Constraints

Generative adversarial networks (GANs) show promise in synthesizing clinically relevant data for diabetes management, yet their limited transparency and physiological plausibility continue to challenge a broader adoption in healthcare contexts. This work compares a standard-GAN and a monotonic-GAN, which enforces physiological monotonicity in the generator to ensure more realistic blood glucose (BG)-generated dynamics. Using simulated data, we first assessed models' ability to generate realistic BG distributions, via t-distributed stochastic neighbor embedding, Wasserstein distance and comparisons of key glycemic summary metrics. Secondly, we evaluated their responses to conditional inputs (CIs), i.e., plasma insulin and carbohydrate (CHO) rate of appearance in plasma, by quantifying monotonicity through Spearman rank-order correlation coefficients. Results indicate that the monotonic-GAN produces more physiologically sound responses to CIs, with BG decreasing monotonically with insulin and increasing with CHO intake, while replicating the BG distribution of the original data without significant differences. These findings suggest that integrating physiological constraints into the GAN architecture enhances the physiological consistency of generated data, without affecting generated data distributions, thus increasing model trustworthiness.