Combined effects of PPG preprocess and dynamic spectrum extraction on predictive performance of non-invasive detection of blood components based on dynamic spectrum

The non-invasive detection accuracy of blood components based on dynamic spectrum (DS) not only depends on the accuracy of each step individually, but also the proper cooperation between these steps. Therefore, combined effects of PPG preprocess and DS extraction on the non-invasive detection of hemoglobin (Hb) were investigated in this paper: DS extraction method itself has some but insufficient ability of noise suppression, and therefore PPG preprocess method should be chosen properly to complement DS extraction method's deficiency in noise suppression and further improve the detection accuracy. In this study, zero phase-shift bandpass filter, wavelet filter and Empirical Mode Decomposition (EMD) filter, were applied for PPG preprocess and single-trial estimation was applied for DS extraction. Partial Least Square Regression (PLSR) was applied for model establishment between DS and Hb content. The results showed that in comparison with non-preprocess, the root mean square error of prediction set (RMSEP) is reduced from 14.1003 g/L to 10.7270 g/L, 11.1018 g/L, 11.2768 g/L respectively with zero phase-shift bandpass filter, wavelet filter, EMD filter. As a result, zero phase-shift bandpass filter cooperates with single-trial estimation better in noise suppression to improve the prediction accuracy of Hb content. By studying the combined effects, we could know how to combine various methods of each step together to obtain the best calibration model. This study also has referential significance to quantitative analysis with NIR spectroscopy.