Frequency-range optimized preprocessing methods for quantitative analysis of glucose in blood serum from broadband dielectric spectra.

This paper presents the results of combining scatter correction and spectral derivation preprocessing methods with frequency-range optimization to improve the accuracy of blood glucose concentration prediction when using a Partial Least Square Regression model. Using broadband dielectric spectroscopy, absorption spectrums in the frequency range of 500 MHz to 50 GHz were gathered from blood serums. Partial Least Square Regression models were trained using data gathered from samples of varying glucose concentrations and temperatures, and the quality of the predictions were evaluated by performing leave-one-out cross validation. Potential improvements in prediction accuracy were assessed by finding the optimal frequency ranges of dielectric absorption spectrums through iteration, in addition to treatments using common preprocessing methods. The most effective combination of preprocessing method and its corresponding characteristic frequency range was determined from validation using several samples. Finally, among all the preprocessing methods and the frequency ranges explored, performing Savitzky-Golay filtering without derivations lowered the average root mean squared error amongst all the samples from 172 mg/dL without preprocessing, to 143 mg/dL. Additionally, by focusing on a specific frequency range, the root mean squared error dropped to 80.2 mg/dL.