Deep learning and regression approaches to forecasting blood glucose levels for type 1 diabetes. (August 2021)

• Record Type: Journal Article
• Authors: Zhang, Meng; Flores, Kevin B.; Tran, Hien T.
• Is Part Of: Biomedical signal processing and control. Volume 69(2021)
• Publication Date: 2021-08
• Subjects: Neural networks -- Regression -- Encoder decoder -- Time series forecasting -- Diabetes; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102923
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Use large-scaled, real-world data collected from type 1 diabetes patients. Apply efficient deep learning and regression models with/without encoder-decoder. A model based on linear regression outperforms state of the art deep learning models. Present the influence of imputation methods, evaluation metrics and visualization. Compare proposed models against previous studies in the literature. Abstract: Objective: Controlling blood glucose in the euglycemic range is the main goal of developing the closed-loop insulin delivery system for type 1 diabetes patients. The closed-loop system delivers the amount of insulin dose determined by glucose predictions through the use of computational algorithms. A computationally efficient and accurate model that can capture the physiological nonlinear dynamics is critical for developing an efficient closed-loop system. Methods: Four data-driven models are compared, including different neural network architectures, a reservoir computing model, and a novel linear regression approach. Model predictions are evaluated over continuous 30 and 60 min time horizons using real-world data from wearable sensor measurements, a continuous glucose monitor, and self-reported events through mobile applications. The four data-driven models are trained on 12 data contributors for around 32 days, 8 days of data are used for validation, with an additional 10 days of data for out-of-sample testing. Model performance was evaluated by the root mean squared error and the mean absolute error. Results: A neural network model using an encoder-decoder architecture has the most stable performance and is able to recover missing dynamics in short time intervals. Regression models performed better at long-time prediction horizons (i.e., 60 min) and with lower computational costs. Significance: The performance of several distinct models was tested for individual-level data from a type 1 diabetes data set. These results may enable a feasible solution with low computational cost for the time-dependent adjustment of artificial pancreas for diabetes patients.