Prediction of Blood Glucose Concentration Ahead of Time with Feature Based Neural Network

Abstract

Diabetes has become a major health challenge affecting nearly 300 million people around the world. Complications of diabetes can be prevented by proper monitoring and regulation of glucose concentration in blood plasma. Continuous Glucose Monitoring Systems help to track the time course of blood glucose. These devices have the additional feature of giving threshold alert and predictive alert which is needed for an early warning of impending hypoglycemia. However, the accuracy of predictive alerts in currently available CGM devices is not very promising. Various algorithms have been developed in this regard by researchers. Still, a 100% accuracy has not been achieved. In our work, we have approached this prediction by training a simple neural network with the extracted features of continuous glucose monitoring sensor data time series. The data was obtained in three different ways, one set from the Self Monitoring Blood Glucose values, the second set from a diabetes resource and the third one from the patients using continuous glucose monitoring systems. A feed forward neural network with back propagation algorithm is trained with features of input patterns. The network is trained and validated to meet out the performance goal. The Root Mean Square Error between the actual glucose value and the predicted glucose value is used as the performance measure. It is observed that as the length of prediction horizon extends, the error increases. However, tracking of Hypoglycemic and Hyperglycemic trends are superior to the earlier approaches.