Multivariate calibration with slowly responding reference measurements

Abstract

The application of multivariate calibration models, specifically those using partial least squares (PLS) regression to relate near infrared (NIR) spectral data to analyte concentrations, relies upon accurate knowledge of the concentrations during model building. In a physiologic system, such as human skeletal muscle, these concentrations can be measured using invasive sensors which may have material properties that limit diffusion of analytes to the sensing chemistry, thus taking several minutes to fully respond to an analyte change which actually occurs in seconds. This results in a poor time correlation between reference measurements of analyte concentrations and spectral data, which in turn degrades the performance of the PLS model. We mathematically modeled the response of an invasive sensor measurement and used this response to develop a filter to time-match the raw NIR spectra before building the PLS model. PLS models for interstitial pH in exercising human flexor digitorum profundus muscle were developed with and without the time-matching filter. In a single exercising subject, root mean square error of prediction (RMSEP) = 0.05 pH units and r 2 = 0.39 without filtering, but improved to RMSEP = 0.02 pH units with r 2 = 0.91 after the time-matching filter was implemented. The time-matching filter was shown to be effective in improving model performance when spectral response is more rapid than the invasive sensor reference measurement.