Utilizing Brain Activity to Non-Invasively Predict Blood Glucose Levels

Abstract

Diabetes Mellitus is a chronic condition that must be routinely self-monitored to prevent life threatening complications, diabetic coma, and to mitigate long-term complications, such as neuropathy and kidney failure. Despite these risks, many people with diabetes do not properly self-monitor. An often cited reason for not properly self-monitoring is the discomfort associated with the lancing procedure. To alleviate the need for finger-sticks, this research has focused on a non-invasive blood glucose monitoring technique by combining electroencephalography (EEG) with artificial neural networks (ANN). Eleven volunteers participated in three, hour-long, recording sessions. Not all volunteers completed all sessions. During this time the volunteer's blood glucose level (BGL) was monitored in fifteen-minute increments, with the exception of the fifteenth minute. At fifteen minutes into the data collection, the volunteers drank a 52g carbohydrate load. EEG data was recorded continuously over the data collection period. From the EEG data, frequency features were extracted and used to train an ANN that would predict the users BGL based upon EEG input. This data was used to determine which channels that correlated most strongly with glucose levels, determined the optimal number of neurons used for the training algorithm, and finally trained/tested an ANN. The channels that corresponded most consistently and strongly are channels AF4, F4, F8, FC5, O1, O2, and P8. The optimal number of neurons for the network was determined to be 99 neurons. Ultimately, the method of training an ANN to non-invasively predict BGLs has moderate success, but needs more research to be conclusive.