Determination of the number of significant components in liquid chromatography nuclear magnetic resonance spectroscopy

In this paper, the effectiveness of methods for determining the number of significant components is evaluated in four simulated and four experimental liquid chromatography nuclear magnetic resonance (LC–NMR) spectrometric datasets. The following methods are tested: eigenvalues, log eigenvalues and eigenvalue ratios from principal component analysis (PCA) of the overall data; error indicator functions [residual sum of squares (rssq), residual standard deviation (RSD), ratio of successive residual standard deviations (RSDRatio), root mean square error (RMS), imbedded error (IE), factor indicator functions, scree test and Exner function], together with their ratio of derivatives (ROD); F-test (Malinowski, Faber–Kowalski and modified FK); cross-validation; morphological score (MS); purity-based approaches including orthogonal projection approach (OPA) and SIMPLISMA; correlation and derivative plots; evolving PCA (EPCA) and evolving PC innovation analysis (EPCIA); subspace comparison. Five sets of methods are selected as best, including several error indicator functions, their ratio of derivatives, the residual standard deviation ratio, orthogonal projection approach (OPA) concentration profiles and evolving PCA using an expanding window (EW). Omitting the dataset with the highest noise level, RSS, Malinowskiʹs F-test, concentration profiles using SIMPLISMA and subspace comparison with PCA score also perform well.