ABSTRACT This study investigates the performance of near‐infrared (NIR)‐based partial least squares (PLS) regression models for quantifying hyperforin in Hypericum perforatum L., with particular emphasis on temporal validation and robustness analysis. Dried and ground Hypericum samples were analyzed using three different NIR spectrometers: one benchtop device (NIRFlex N‐500) and two handheld devices (NeoSpectra, TrinamiX). Models were calibrated on an initial model development dataset and then applied to a temporally independent screening dataset from the next year's harvest, allowing assessment of predictive stability over time. All devices successfully identified samples with exceptionally high and low hyperforin content in the screening dataset. The overall predictive performance across the entire dataset confirmed temporal validity. Especially, the handheld NeoSpectra showed strong predictive performance, with a root mean squared error of prediction (RMSEP) of 0.36% across the working range of the calibration model from 0.58% to 5.72%. This demonstrates that compact, lower‐cost devices can achieve high analytical performance, while offering additional advantages such as portability and the potential for on‐site analysis directly in the field or at production sites. Analysis of model performance across different numbers of PLS components further provided insights into temporal robustness, illustrating the importance of conservative component selection to maintain predictive reliability over time. In addition to PLS, Ridge and Lasso regressions were evaluated to test alternative regularization strategies that can improve robustness to overparameterization, with Lasso in particular showing comparably good performance.
Schröter et al. (Sun,) studied this question.