Principal Component Regression for Broadband Microwave-Microfluidic Chemometrics on Small Sample Counts

While broadband dielectric spectroscopy enables label-free analysis of biological and chemical materials, extracting multiple concentrations from the data has remained a challenge. This work is one of the first demonstrations of simultaneous concentration extraction of three liquid constituents in a solution using broadband microwave spectroscopic data. Furthermore, the used methods eliminate the need for de-embedding or characterizing the measurement setup, simplifying the process. Advanced regression techniques such as Principal Component Regression (PCR) and Principal Least Squares (PLS) are applied to determine the concentrations of sodium chloride, glucose, and ethanol in water. As input data, $S$-parameters are measured between 0.5 and 26.5 GHz on a broadband coplanar waveguide sensor with a microfluidic container to transport the liquids. The training datasets consist of 27 and 34 samples, respectively. The mean absolute percentage error for sodium chloride predictions ranged from 3-5% for the different methods, while the minimal errors for glucose and ethanol predictions were 6-7% and 4-6%, respectively.