Effect of preprocessing high-resolution mass spectra on the pattern recognition of Cannabis, hemp, and liquor

Abstract

High-resolution mass spectrometry (HRMS) combined with pattern recognition was used to discriminate among twenty-five Cannabis samples, twenty hemp samples, and eight liquor samples. The effects of preprocessing on multivariate data analysis were evaluated for Orbitrap high-resolution mass spectra. Different root transformations were evaluated with respect to the bin width and the average classification rates. In addition, linear binning and proportional binning with various resolving powers were studied with respect to the average classification rates. The proportional binning with the square root transformation gave the best overall performance for chemical profiling or spectral fingerprinting. Six classification methods, fuzzy rule-building expert system (FuRES), linear discriminant analysis (LDA), super partial least squares discriminant analysis (sPLS-DA), support vector machine (SVM), SVM classification tree type gap (SVMTreeG), and SVM classification tree type entropy (SVMTreeH) had similar trends in prediction rate with respect to the resolving power. The optimal proportional mass bin width may depend on the data set, i.e., for the Cannabis samples is resolving power 10−4, for the hemp samples and the liquor samples are resolving power 10−3. Hence, data preprocessing methods such as different transformations, binning strategies, and resolving powers are important factors to be optimized for HRMS direct infusion measurements combined with pattern recognition to be an authentication and characterization tool for various products.