Noninvasive Glucose Monitoring With Low-Order Electrical Impedance Sensing Arrays: An Information Boosting Approach

Noninvasive blood glucose (BG) monitoring using bioelectrical impedance technology holds promise for enhanced BG management. Array electrodes enhance the sensing capability in complex tissues but increase data acquisition time. However, the acquisition time varies significantly across array sizes, with the proposed method showing the acquisition time of a $2\times 2$ array is 1.04% of that of a $4\times 4$ array. To tackle the challenge, this work proposes an upscaling and reconstruction model (URM) for low-order array impedance data based on electrical impedance tomography (EIT) and a U-shaped autoencoder, to boost the information of low-order electrode array data that can be acquired rapidly. An end-to-end glucose concentration classification model based on principal neighborhood aggregation is also introduced, which incorporates a natural feature extraction (NFE) module that supplements the feature information of low-order array data and a multiscale constraint (MSC) method to balance the intermediate training process. The proposed model is validated through EIT simulations, achieving an image similarity score of 0.93 for the URM. Additionally, validation using a biological tissue simulator yielded a classification accuracy of 90.82% precise of 90.85%, recall of 90.79%, and $F1$ -score of 90.77%.