Smartphone-based Multi-parametric Glucose Prediction using Recurrent Neural Networks

Abstract

Diabetes Mellitus causes many deadly diseases, including pancreatic cancer as irregularity of glucose level triggers dysfunctions like unchecked cell growth. The critical stages in glucose irregularity are categorized as hyperglycemia (high blood glucose) and hypoglycemia (low blood glucose) which needs to be detected in advance for the quality of human life. In that sense, many tools have been developed based on artificial intelligence (AI) systems which mostly consider the glucose measurement as a prediction parameter. However, in this study, we propose to employ multi-parameter in glucose prediction based on a Recurrent Neural Network (RNN), a subset of AI, to enhance predictability. The proposed system utilizes a Long-Short Term Memory (LSTM) based RNN to handle complex memory operations caused by multi-parametric prediction. Training and validation scores on the OhioT1DM dataset show the advantage of our proposed system over the baseline systems for predicting glucose levels with a significantly reduced error. The system is later integrated with our custom-designed Android application, BffDiabetes PRO, capable of reading the glucose levels from the sensors via Bluetooth. The BffDiabetes PRO transfers the current glucose level, acceleration, and baseline skin temperature to the server via a cloud system to predict the next level. It receives the prediction result to evaluate whether the glucose level tends to reach the critical stages. In case of this tendency is detected, the BffDiabetes PRO alerts the user for necessary precautions.