Insulin is central in metabolic health, but direct assessment of its dynamics remains costly and impractical. While physiological models based on differential equations, such as the Bergman minimal model, capture glucose-insulin dynamics, they usually require both glucose and insulin measurements for parameter estimation. Physics-informed neural networks (PINNs) estimate parameters from sparse, nosiy data, even when some variables are latent. In a simulation study, we applied PINNs to estimate the parameters of the Bergman minimal model from noisy glucose data alone, simulating an intravenous glucose tolerance test. To incorporate plausible parameter ranges, we extended the PINN loss function with an additional term. The results show that PINNs can capture key glucose-insulin dynamics from sparse, noisy glucose data and plausible ranges on the parameters. With further refinement, this approach holds promise for inferring hidden insulin dynamics during routine metabolic health assessments, enabling improved monitoring based solely on glucose measurements.