A dynamic spectrum optimized single-trail extraction method for noninvasive analysis of human blood components

Abstract

Dynamic spectroscopy (DS) is a new noninvasive detection method of human blood components. Theoretically, it can eliminate the individual differences of static tissues such as skin and muscle and the influence of measurement conditions on measurement accuracy, and realize high-precision noninvasive blood component analysis. Obtaining the dynamic spectrum with a high signal-to-noise ratio is the key to modeling and analysis. To further improve the quality of the dynamic spectrum and ensure the robustness of the model, this paper proposes an optimized single-trail extraction method, which is improved for the main single-trail extraction method at present. In this method, the sampling points deviating from the fitting line in the rising edge signal of photoplethysmography (PPG) are eliminated by statistical method many times and using the superposition average of the slope between each sampling point on the rising edge to obtain the high-quality DS. To verify the effectiveness of the method proposed in this paper, we collected the spectral PPG data of the near-infrared transmission of the fingers of 231 volunteers, the wavelength range was 591.8 nm-1100 nm, and the hemoglobin concentration range of the total sample was 90–176 g / L, then used the optimized single-trail extraction method and the single-trail extraction method to extract DS from the spectral PPG data of all samples and divided them into an experimental group and a control group according to the different extraction methods. Two methods of spectral smoothness and modeling analysis were selected to evaluate the effectiveness of the spectrum. The experimental data are as follows: when the DS is evaluated with spectral smoothness as an index, the mean square error of the experimental group's spectrum is reduced from 0.011 to 0.0059, which is 53.6 % of the control group; and when using the least partial squares method to establish a hemoglobin model for prediction, compared with the control group, the related coefficient of the correction set (Rc) in the experimental group reached 0.91, an increase of 7.96 %, and the root mean square error of the correction set (RMSEC) of the calibration set decreased to 5.63g/L, decreased by 22.17 %, the related coefficient of the prediction set (Rp) reached 0.86, increased by 10.90 %, and the root mean square error of the prediction set (RMSEP) decreased to 6.27g/L, decreased by 19.76 %. The experimental results show that the optimized single-trail extraction method proposed in this paper improves the spectral smoothness and modeling effect compared with the original single-trail extraction method effectively improving the quality of DS. This method provides a new way for further improving the accuracy of noninvasive blood component quantitative analysis based on dynamic spectrum and helps to promote the application process of noninvasive detection of blood components.