Multiple Band-Pass Filtering Method for Improvement on Prediction Accuracy of Linear Multivariate Analysis

Abstract

An approach coupling signal processing and partial least-squares regression analysis (PLS) is described in which raw spectral data are processed with a multiple band-pass filter and the filtered spectra are used in a PLS to build a calibration model for the analyte of interest. The multiple band-pass filter is specifically designed for a desired analyte based on the Fourier frequency characteristics of the pure spectrum of the desired analyte and the spectra of the interference background. It maximizes the ratio of signal to background. This combined multiple band-pass filtering and PLS method (MFPLS) was evaluated by determining clinically relevant levels of glucose, urea, ethanol, and acetaminophen in simulated human sera, in which triglyceride was simulated with triacetin; bovine serum albumin and globulin were used to model protein molecules in the serum. The results demonstrate that MFPLS produces better accuracy of prediction than PLS in all instances.