Computer Science and Engineering
 Gothenburg University | Chalmers

Home page

Home page  Contact us  Site map 



Automatic blood glucose prediction with confidence using recurrent neural networks

C. Meijner, S. Persson, A. Schliep, B. Eliasson and O. Mogren

In Sep 2017. Under review.

Low-cost sensors and mobile platforms combined with machine-learning (ML) solutions enable personalized precision health and disease management. The large number of possible sensors, analysis tasks and adaptation to individuals raises the issue of scale, with respect to the number of ML systems which need to be developed. We present one approach for partially automated ML using deep learning in the context of managing diabetes. Continuous glucose monitor (CGM) systems help diabetics to closely follow their blood glucose values by storing a value every 5-15 minutes. They also provide an excellent source of data for predictions. This paper presents a model that can predict blood glucose levels for diabetics for up to two hours into the future. The solution uses a long short-term memory (LSTM) model which works on the raw data from a CGM device. The approach needs no feature engineering or data pre-processing, it obtains accuracy matching or exceeding the state-of-the-art, is computationally inexpensive, and provides along with the predictions an estimate of their variance, helping users to interpret the predicted levels.