Single Hidden Layer Perceptron GAN for Mixed Meal Continuous Glucose Monitoring Data

Generative Adversarial Networks (GANs) have emerged as valuable solutions for generating realistic data, addressing challenges such as missing values and/or data scarcity. In the context of diabetes management, where data scarcity is an issue, GAN s may offer a method for generating synthetic continuous glucose monitoring (CGM) traces resembling the characteristics of real data. This study focuses on developing a GAN for CGM mixed meal traces generation leveraging on single-macronutrient CGM traces. To this aim, a GAN model, based on a single hidden layer perceptron architecture, was designed using a set of freely-available single-macronutrients data; hyperparameter settings was performed using a random search approach. Reliability of GAN generated mixed meal data was tested through comparison with real mixed meal data taken from a second freely-available dataset. Comparison was performed using the Two One-Sided Test (TOST) for equivalence on a set of ten characteristics features (area under the curve, median, mean, standard deviation, coefficient of variation, maximum, minimum, sample at which maximum and minimum occur, time in range) extracted from generated and real mixed-meal data. The median area under the curve (AUC) of the generated and real data resulted equal to 3.40 [3.13, 3.94]× 103mg/dL (median [1st quartile, 3rd quartile]), and 3.95 [3.60, 4.10]× 103mg/dL, respectively. The results of the TOST showed substantial equivalence between generated and real mixed meal CGM traces for all the features. Indeed, the margin of equivalence δ was no more than 0.4 standard deviation of each feature. In conclusion, the proposed GAN appears promising to generate CGM mixed meal traces leveraging on single-macronutrient CGM traces.