Developing a GAN-Based Blood Glucose T1DM Outcome Prediction Model for Clinical Use

Abstract

This project presents the development and evaluation of a novel outcome prediction model for blood glucose levels in patients with Type 1 Diabetes Mellitus (T1DM) using a Wasserstein Conditional Gener ative Adversarial Network (WCGAN). This Generative Deep Learning (GDL) model is trained on the Re placeBG dataset, which includes time series data from 226 T1DM patients, comprising Plasma Insulin (PI) administration, Rate of Appearance (RA) of carbohydrates, demographic and temporal information. The WCGAN, comprising over 3 million parameters, was trained to iteratively generate synthetic blood glucose profiles. These would allow producing long series of data in order to analyze the effect of differ ent therapy strategies on the patient the model is mimicking, aiding clinical decisions. The training was conducted over 250 thousand steps using the GeForce RTX 4070 Ti GPU. Once adjusted, this condi tional GAN dynamically generated glucose level predictions based on past and present inputs from the three aforementioned variables: insulin, carbohydrates, and time. The effectiveness of the model was tested by assessing the statistical similarity between the synthetic and real glycemic outcomes, with key metrics showing significant results. The model demonstrated physiological glucose-insulin dynamics, a causal relationship between inputs and outputs, and the possibility to control the variability of the latter modifying the latent space (Z ∈ R3) sampling. Additionally, we showed clear overlap between real and generated data distributions (DR and DG), as well as success in reconstructing missing parts of the first. Despite this, the effect of including time showed mixed results in improving the quality of the outputs. This thesis can be seen as a proof of concept on incorporating the moment of the day into the GAN-based outcome prediction model, and further establishes its feasibility with ten-fold more data than previous work.