Therapy-driven Deep Glucose Forecasting. (January 2020)

• Record Type: Journal Article
• Authors: Aiello, Eleonora Maria; Lisanti, Giuseppe; Magni, Lalo; Musci, Mirto; Toffanin, Chiara
• Is Part Of: Engineering applications of artificial intelligence. Volume 87(2020)
• Publication Date: 2020-01
• Subjects: Diabetes -- Forecasting -- Prediction -- Deep learning -- LSTM; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2019.103255
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: The automatic regulation of blood glucose for Type 1 diabetes patients is the main goal of the artificial pancreas, a closed-loop system that exploits continue glucose monitoring data to define an optimal insulin therapy. One of the most successful approaches for developing the artificial pancreas is the model predictive control, which exhibits promising results on both virtual and real patients. The performance of such controller is highly dependent on the reliability of the glucose–insulin model used for prediction purpose, which is usually implemented with classic mathematical models. The main limitation of these models consists in the difficulties of modeling the physiological nonlinear dynamics typical of this system. The availability of big amount of in silico and in vivo data moved the attention to new data-driven methods which are able to easily overcome this problem. In this paper we propose Deep Glucose Forecasting, a deep learning approach for forecasting glucose levels, based on a novel, two-headed Long-Short Term Memory implementation. It takes in input the previous values obtained through continue glucose monitoring, the carbohydrate intake, the suggested insulin therapy and forecasts the interstitial glucose level of the patient. The proposed architecture has been trained on 100 virtual adult patients of the UVA/Padova simulator, and tested on both virtual and real patients. The proposed solution is able to generalize to new unseen data, outperforms classical population models and reaches performance comparable to classical personalized models when fine-tuning is exploited on real patients.