Continuous Glucose, Insulin and Lifestyle Data Augmentation in Artificial Pancreas Using Adaptive Generative and...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Data Science

Background:

• Artificial pancreas systems rely on diverse data for accurate insulin dosing.
• Data gaps from device errors or patient non-compliance can hinder closed-loop diabetes management.
• High-quality synthetic data generation is crucial for augmenting real-world datasets.

Purpose of the Study:

• To develop and evaluate a generative adversarial network (GAN) model for high-quality synthetic data generation in artificial pancreas systems.
• To investigate the impact of output sequence length on generative and discriminative model selection.
• To improve data augmentation strategies for enhancing glucose forecasting accuracy.

Main Methods:

• A novel GAN-based architecture was designed to automatically select generator and discriminator models based on desired output sequence length.
• The model was trained to generate synthetic data including glucose levels, physical activity, and meal information for individual patients.
• Generated data was combined with real data to train and evaluate glucose forecasting models.

Main Results:

• The proposed GAN model demonstrated effective generation of synthetic glucose, physical activity, and meal data.
• Discriminative scores on the Ohio T1DM and Ohio T1D datasets indicated high-quality synthetic data generation.
• Glucose forecasting models utilizing a mixture of real and synthetic data achieved improved RMSE and MARD scores compared to models using only real data.

Conclusions:

• Sequence length-based architecture selection in GANs leads to superior synthetic data generation for multiple output sequences.
• The developed model effectively augments datasets, enhancing the accuracy of glucose forecasting in artificial pancreas applications.
• This approach offers a promising solution for addressing data scarcity in diabetes management systems.