Simulating Realistic Continuous Glucose Monitor Time Series By Data Augmentation.

Area of Science:

• Biomedical Engineering
• Data Science
• Diabetes Technology

Background:

• Simulated data are crucial for benchmarking blood glucose (BG) forecasting and control algorithms.
• Expert-created models and black-box approaches like GANs offer limited realism and diagnostic capabilities for real-world performance.
• Existing simulation methods lack the complex features of real continuous glucose monitor (CGM) data, hindering accurate algorithm evaluation.

Purpose of the Study:

• To develop a novel method for augmenting simulated BG data with realistic missingness and error properties derived from real CGM data.
• To improve the fidelity of simulated CGM data for more rigorous testing and benchmarking of BG forecasting algorithms.
• To reduce the performance gap between algorithms tested on simulated data versus real-world CGM data.

Main Methods:

• Learned missingness and error characteristics from real CGM datasets (OpenAPS, OhioT1DM, RCT, Racial-Disparity).
• Augmented simulated BG data with these learned properties to mimic real-world data challenges.
• Evaluated BG forecasting performance using the augmented simulated data against standard simulation practices (random dropout, Gaussian noise, CGM error model).

Main Results:

• The proposed method demonstrated the smallest performance difference compared to real data versus random dropout and Gaussian noise for missing data and error effects individually.
• The combined approach significantly outperformed Gaussian noise and random dropout across most datasets, except OhioT1DM.
• The developed error model notably enhanced results on diverse datasets, indicating improved simulation realism.

Conclusions:

• A significant performance gap exists between BG forecasting on simulated versus real data.
• The proposed method effectively closes this gap, enabling more realistic performance estimates.
• Researchers can now rigorously test algorithms and obtain reliable real-world performance insights without overfitting or extensive data collection.