Abstract
BACKGROUND: Simulated data are a powerful tool for research, enabling benchmarking of blood glucose (BG) forecasting and control algorithms. However, expert created models provide an unrealistic view of real-world performance, as they lack the features that make real data challenging, while black-box approaches such as generative adversarial networks do not enable systematic tests to diagnose model performance. METHODS: To address this, we propose a method that learns missingness and error properties of continuous glucose monitor (CGM) data collected from people with type 1 diabetes (OpenAPS, OhioT1DM, RCT, and Racial-Disparity), and then augments simulated BG data with these properties. On the task of BG forecasting, we test how well our method brings performance closer to that of real CGM data compared with current simulation practices for missing data (random dropout) and error (Gaussian noise, CGM error model). RESULTS: Our methods had the smallest performance difference versus real data compared with random dropout and Gaussian noise when individually testing the effects of missing data and error on simulated BG in most cases. When combined, our approach was significantly better than Gaussian noise and random dropout for all data sets except OhioT1DM. Our error model significantly improved results on diverse data sets. CONCLUSIONS: We find a significant gap between BG forecasting performance on simulated and real data, and our method can be used to close this gap. This will enable researchers to rigorously test algorithms and provide realistic estimates of real-world performance without overfitting to real data or at the expense of data collection.