Multi-Task Disentangled Autoencoder for Time-Series Data in Glucose Dynamics.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Data Science

Background:

• Continuous Glucose Monitoring (CGM) data analysis is crucial for managing Type 1 Diabetes Mellitus (T1D).
• Latent representations (LR) and generative models, particularly Variational Autoencoders (VAE), are underexplored for CGM data analysis.
• Existing methods may not fully capture complex glucose dynamics or integrate future information.

Purpose of the Study:

• To propose MD-VAE, a multi-task disentangled variational autoencoder.
• To explore the characteristics of latent representations (LR) derived from CGM data.
• To exploit LR for diverse tasks including glucose forecasting, event detection, and temporal clustering in T1D.

Main Methods:

• Applied MD-VAE to both simulated (T1DMS) and real patient CGM data.
• Utilized a multi-task learning framework with disentangled variational autoencoders.
• Focused on extracting meaningful latent representations from time-series glucose data.

Main Results:

• MD-VAE successfully extracted meaningful LR from CGM data.
• The derived LR were exploitable for various downstream tasks, demonstrating utility.
• MD-VAE differentiated sequence characteristics based on clinical parameters, showing interpretability.
• The model showed potential for integrating current and future glucose dynamics and event detection.

Conclusions:

• MD-VAE offers a novel approach for analyzing CGM data in T1D.
• Latent representations derived from VAEs hold significant potential for diverse applications in diabetes management.
• The model can provide complementary insights into glucose dynamics, unexpected events, and device interactions.