Similarity and dissimilarity measures for HPLC fingerprints in quality control of herbal extracts - Case study of sage adulteration by olive, hazelnut, and cistus leaves.

Abstract

Quality control of herbal products, whether medicines or spices, requires the continual development of robust analytical frameworks. Chromatographic fingerprinting, while providing comprehensive chemical information, is frequently used in similarity analysis for comparison with reference chromatograms and building reliable classifiers. However, selecting optimal similarity or distance measures and understanding the impact of signal preprocessing on such classification remains challenging. This study evaluates four similarity measures: congruence (CNG), Pearson (PEA), Spearman (SPE), and Kendall (KND) correlation coefficients, and three distances: Euclidean (EUC), Manhattan (MNH), and Chebyshev (CHE), applied to 115 HPLC profiles of sage samples and samples adulterated with cistus (CS), hazelnut (HZ), and olive leaves (OL) as a model system. Measures were assessed across different adulteration levels (5-100%) and chromatographic signal preprocessing in two ways: a) in the context of their behavior in a consensus comparison using the sum of ranking differences (SRD), and b) according to their classification performance. SRD analysis revealed that CNG and PEA are closest to the consensus, while non-parametric measures, such as SPE, are the most resilient to preprocessing. However, multi-factor ANOVA demonstrated that correlation-optimized warping (COW) and standard normal variate (SNV) of HPLC signals increase the misclassification rates (MR) of several similarity and distance measures at low adulteration levels (5-20%). Crucially, CHE behaved as an SRD outlier while demonstrating the highest discriminatory power for all adulterants, even at low adulteration levels (5-10%), across both signal preprocessing methods. For highly overlapping fingerprints, such as those adulterated with OL or HZ samples, rank-based measures (SPE and KND) applied directly to untreated signals yielded the lowest misclassification rate (21.43%). Therefore, targeted screening workflow using CHE or non-parametric similarity measures provides the most efficient and sensitive quality control framework.