Blood glucose forecasting in type 1 diabetes mellitus patients using machine learning techniques

Abstract

Diabetes Mellitus is a metabolic disease that causes the body to lose control over blood glucose regulation. Patients with Type 1 diabetes completely rely on insulin therapy by themselves or using some automated insulin delivery systems. In both the cases, it is pertinent to have good estimate of future blood glucose levels. An efficient diabetes management demands accurate prediction of future blood glucose levels, failure of which results in short and long term health complications. With the modern exordium of quantified-self such as continuous glucose monitoring(CGM) systems, a patient can have access to their personalized glycemic profile which can be utilized for accurate prediction of future blood glucose levels. In recent years, machine learning methodologies have sparked a lot of interest in predicting glucose levels in diabetic patients, leading to the development of a variety of methods and techniques. However, the prediction accuracies of these methods are not good enough to be declared them as reliable predictors for evaluating glycemic conditions. In this research work we utilized multi-layered Long Short Term Memory(LSTM) network a famous deep learning technique based on recurrent neural network(RNN) for making prediction of blood glucose levels in patients with type 1 diabetes. The proposed framework predicts the future blood glucose level using Ohio T1DM dataset at prediction horizon(PH) of 30 and 60 minutes. Experimentation was also carried out on better feature representation to model in order to achieve higher prediction accuracy. The effect of different input feature sets, towards improvement of prediction accuracy was also been investigated. The results on Ohio T1DM Dataset (2018), that contain eight weeks’ worth of data shows that our method achieves the lowest RMSE score of 14.76mg/dL and 25.48mg/dL for prediction horizon of 30min and 60min respectively. The obtained results are the best known as per our knowledge using this dataset. The proposed methodology can be utilized in closed loop systems for precise insulin delivery to patient for their better glycemic control.