Improving the accuracy of mid-infrared prediction models by selecting the most informative wavelengths through uninformative variable elimination

Mid-infrared spectroscopy (MIRS) is used to collect milk phenotypes at population level. The aim of this study was to test the ability of uninformative variable elimination (UVE) approach to select informative wavelengths before multivariate analysis. Reference values (n = 386) of milk titratable acidity were randomly selected from an existing database. The dataset was randomly divided into calibration (80%) and validation (20%) sets, and partial least squares (PLS) analysis was carried out before and after UVE procedure. After UVE procedure, 244 informative wavelengths were retained for subsequent PLS analysis. The elimination of uninformative variables before PLS regression led to an improvement of the accuracy of MIRS prediction models and it substantially reduced the computational time. Finally, dealing with much less variables would enhance the efficiency of MIRS models to predict phenotypes at population level