"Personalized Time Series Forecasting of Blood Glucose Levels"
The development of data-driven capabilities for feedback control in the treatment of Type 1 Diabetes (T1D) requires the accurate prediction of future blood glucose (BG) levels. Specifically, the ability to predict BG levels in 30 and 60 minute time horizons could enable the time-dependent adjustment of treatment in response to the ensuing status of the patient, i.e., if hyper/hypo-glycemia occurs. By providing real-time data from continuous BG monitors, wearable sensor measurements, and self-reporting through mobile applications, the BG Level Prediction Challenge has enabled the capability to test whether models calibrated to individual-level data could ultimately be used for making individualized treatment decisions in T1D.
We trained and analyzed several direct prediction strategies, including different neural network architectures, reservoir computing, and linear regression. We found that the use of multiple linear regression models was the most accurate prediction strategy, and that reservoir computing has both the prediction power and the ability to recover the dynamics from missing intervals.