Prediction of glucose level conditions from sequential data

Abstract

In type 1 diabetes management, mobile health applications are becoming a cornerstone to empower people to self-manage their disease. There are many applications addressed to calculate insulin doses based on the current information (e.g. carbohydrates intake) and a few of them are accompanied by modules able to supervise postprandial conditions and recommend corrective actions if the user falls in an abnormal state (i.e. hyperglycaemia or hypoglycaemia). On the other hand, mobile apps favour the gathering of historical data from which machine learning techniques can be used to predict if user conditions will worsen. This work presents the application of k-nearest neighbour on the historical data gathered on patients, so that given the information related to a sequence of meals, the method is able to predict if the patient will fall in an abnormal condition. The experimentation has been carried out with the UVA-Padova type 1 diabetes simulator over eleven adult profiles. Results corroborate that the use of sequential data improve significantly the prediction outcome when forecasts distinguish the type of meal (breakfast, lunch and dinner)