Rapid quantification of six phenolic acids in Pandani Tectorii Radix et Rhizoma using near-infrared spectroscopy and chemometrics.

Abstract

This study evaluated near-infrared (NIR) spectroscopy for quantifying six major phenolic acids in Pandani Tectorii Radix et Rhizoma (PT): neochlorogenic acid (NCA), chlorogenic acid (CGA), cryptochlorogenic acid (CCA), isochlorogenic acid A (IAA), isochlorogenic acid B (IAB), and isochlorogenic acid C (IAC). An adaptive workflow was evaluated through full factorial ablation and 5 × 4 nested cross-validation (nested CV). Three partitioning strategies were compared: random, Kennard-Stone (KS), and joint X-Y distance (SPXY). Nine preprocessing methods included raw spectra (RAW), standard normal variate (SNV), multiplicative scatter correction (MSC), Savitzky-Golay first- and second-order derivatives (SG1 and SG2), SG1-SNV, SG2-SNV, detrended combined with SNV (Detrend-SNV), and orthogonal signal correction (OSC). Four variable selection methods, including full spectrum (Full), competitive adaptive reweighted sampling lite (CARS Lite), variable importance in projection (VIP), and interval stability-VIP, were combined with partial least squares regression (PLSR) and support vector regression (SVR). Under the original nested CV partitioning, Random splitting yielded the lowest overall prediction error for five analytes, whereas KS performed better for IAC. Random seed sensitivity analysis indicated that Random splitting performance was partly dependent on the specific sample partition. SVR-based workflows had numerically lower errors than the compared PLSR workflows, and the best models generally combined VIP-based selection with SG-based preprocessing, often alongside SNV. Final SVR models achieved a coefficient of determination (R2) > 0.95 and a relative predictive deviation (RPD) > 5.0 on independent tests. Nevertheless, no method-family contrast remained significant after variance correction and Holm adjustment across 258 exploratory comparisons. A conditional training-size analysis showed that reducing outer training samples from 48 to 24 increased mean out-of-fold root mean square error (RMSE) to 2.34-4.27 times the 48-sample baseline and lowered mean R2 from 0.963 to 0.988 to 0.714-0.839. Exploratory grouped validation yielded pooled R2 values of 0.465-0.709 and RPD values of 1.38-1.87. These findings demonstrate the feasibility of NIR-based multi-component analysis for PT while defining its performance boundaries under limited-data conditions and providing a methodological reference for rapid quality assessment of medicinal materials.

Abstract as supplied by the indexing source. © the authors / publisher.