Evaluation and comparison of machine learning techniques for the prediction of blood glucose value in patients with type 1 diabetes mellitus

World Health Organization (WHO) recognizes diabetes as one major health issue due to its severe complications and steadily increasing prevalence. Blood glucose oscillation related to diabetes can lead to severe acute and chronic conditions. While type 2 diabetes can be prevented with appropriate lifestyle choices, type 1 diabetes can not be prevented, and patients affected by this disease need a plan to integrate their insulin deficiency. Pharmacological therapy is extremely effective but imposes to the patient to be strictly adherent to a fixed schedule for food intake and physical activity. To improve patient quality of life and to reduce dangerous oscillation of the sugar level in the blood, researchers and industry are developing forecasting methods able to predict future values of the blood glucose level. State of the art methods can be classified in multivariate and univariate systems. Multivariate systems use different algorithms to analyze past values of activity, food intake, sugar level, and any other available and relevant data to estimate future blood glucose levels, while univariate systems are based only on past data collected from continuous glucose monitors. The system of this study is a univariate system, developed with machine learning techniques. First, an extensive study of the available dataset is done. The aim of this preliminary study is the identification of the optimal subset of data to use for training to avoid any bias and polarization of the training dataset. Then, a customized data filtering process is used to avoid the identified problems without decreasing the quantity of considered data excessively. After this first phase, further studies are done to evaluate the application of state of the art architectures to the specific context of blood glucose prediction. The analysis is performed with a method proposed to evaluate the clinical impact of the quality of forecast, including consequences on pathological events prediction. The main neural networks that are implemented and analyzed are combinations of recurrent and convolutional layers. This study provides a baseline for future assessments of similar techniques, thanks to the evaluation of the analysis of the predicted measurements with the proposed method. Furthermore, the study identifies the strengths and drawbacks of different methods, suggesting a direction for future developments toward the implementation of systems of adequate quality for real use.