Precise Noninvasive Estimation of Glucose Using UWB Microwave With Improved Neural Networks and Hybrid Optimization
Xia Xiao, Qun Yu, Qinwei Li, Hang Song, Takamaro Kikkawa
Abstract
A precise noninvasive estimation method for blood glucose concentration is proposed using an improved neural network model and hybrid optimization (INNHO) method. This method consists of a modified back propagation neural network model and a hybrid least squares-random sample consensus algorithm for accuracy optimization. The S-parameter and frequency data measured by ultrawide band (UWB) microwave are employed as the input data for glucose concentration estimation. Before model training, a preprocess technique is put forward to select the data range and generate relative S-parameters from a reference baseline. In this study, a three-layer earlobe model is built to theoretically analyze the wave propagation and elucidate the relationship between the S-parameters and the glucose concentration. Based on the model, the sensitivity and uncertainty of the system are investigated and the suitable frequency range is determined. During the measurement, the S-parameters of UWB signals in the 0.2-4 GHz range are measured and the concentration of glucose water solutions from 20 to 500 mg/dL is investigated. The results show that using the proposed preprocessing method and neural networks, the estimation accuracy can be significantly improved. Then, the hybrid optimization algorithm is able to further enhance the estimation accuracy. The final estimations for solutions of unknown concentrations agree well with the true concentration values. The relative error ranges from 0.31% to 4.64% and the root-mean-square error of the final results is 5.5237 mg/dL, demonstrating that the proposed method is effective and precise for glucose concentration estimation.