Cross-study validation and guidance for resampling approaches for partial least squares discriminant analysis in spectral machine learning.

Resampling strategies are central to model selection and performance estimation in spectroscopic machine learning (SML), yet their impact is often treated as secondary to model choice. In this tutorial study, we systematically evaluated how common resampling methods influence model selection, predictive performance, and estimation bias within a fixed partial least squares-discriminant analysis (PLS-DA) framework across twelve independent spectroscopic datasets. Model tuning was performed using both the one-standard-error (1SE) rule and highest Macro F1 (HMF1) selection, and performance was assessed at the sample level using repeated external validation. Bootstrap and jackknife methods demonstrated the strongest ability to preserve predictive performance, whereas LOOCV and selected Venetian blinds and some K-fold approaches achieved a more favorable balance between performance stability and model parsimony. However, resampling strategies differed substantially in their ability to estimate true test performance. A Bayesian Bradley-Terry analysis revealed that K-fold cross-validation with five folds consistently produced the least biased estimates of external test Macro F1, outperforming both bootstrap and leave-one-out approaches. In contrast, LOOCV, jackknife, and Venetian blinds methods exhibited increased bias due to small or unrepresentative validation sets under sample-level resampling. These findings demonstrate that resampling strategy is a primary determinant of both model selection behavior and performance estimation accuracy in SML. This work provides a practical framework for selecting resampling methods that improve the reliability and interpretability of spectroscopic machine learning models.