Machine Learning-Aided Microwave Sensing for Blood Glucose Detection: Data Information Enhancement from Reduced Measurement Set

A combined approach of microwave sensor and machine learning procedure is proposed in this paper to realize accurate prediction of blood glucose concentration in terms of classification/regression model, by exploiting the full information (energy) content coming from a reduced set of scattering data provided by a compact fractal-shaped microwave sensor working in the transmission mode. Starting from an experimental baseline of six complex permittivity curves measured from blood-mimicking phantoms in the physiological spectrum 72 ÷ 216 mg/dL, six full-wave simulated transmission S21 curves are derived and adopted as starting points for the training process in the machine learning algorithm. Then, a frequency-dependent quadratic fitting model coupled with non-stationary stochastic noise perturbation is adopted to properly enrich the dataset, thus avoiding the need for a huge number of measurements. The proposed methodology is evaluated in-silico and further validated on an independent experimental scenario including 15 VNA-measured S21 transmission spectral responses. These are acquired from blood-mimicking gel phantoms on a text fixture (5 measurement curves used for training set augmentation, and 10 adopted for test validation). The discussed results demonstrate the preliminary effectiveness of the proposed approach as a promising proof-of-concept for non-invasive blood glucose monitoring, offering promising perspectives for future clinical investigations.