<![CDATA[Fourier transform infrared (FTIR) spectroscopy and gas chromatography–mass spectrometry (GC-MS), combined with chemometrics and machine learning, are increasingly applied to food authentication. However, model reliability and generalizability depend on preprocessing, feature selection, and validation strategies, particularly when handling high-dimensional analytical data susceptible to overfitting and information leakage. This systematic critical review examined these strategies in FTIR- and GC-MS-based food-authentication models. A systematic search of ScienceDirect was conducted for publications from 2016 to 2026 following PRISMA 2020 guidelines. Of 274 identified records, 74 studies met the eligibility criteria, comprising 45 FTIR-based studies (60.8%) and 29 GC-MS-based studies (39.2%). Common FTIR preprocessing methods included derivatives, standard normal variate, Savitzky–Golay smoothing, mean-centering, and scatter correction, whereas GC-MS studies more frequently employed scaling, normalization, alignment, and peak processing. Feature-selection approaches varied considerably, with variable importance in projection (VIP), loading-based methods, and recursive feature elimination among the most frequently reported. External or independent validation was reported in 21 studies (28.4%). Explicit information-leakage prevention was documented in only three studies (4.1%), while six (8.1%) raised potential leakage concerns and 65 (87.8%) provided insufficient or unclear reporting. Tree-based algorithms were used in 31 studies, predominantly Random Forest, whereas standalone Decision Tree models were less frequently reported. Overall, no preprocessing or feature-selection strategy demonstrated universal superiority across analytical techniques and authentication tasks. Reliable food-authentication models require an integrated, leakage-aware workflow in which data partitioning precedes preprocessing, feature selection, and data-driven model optimization, followed by evaluation using independent samples whenever possible. Future research should prioritize methodological transparency, reproducibility, interpretability, robustness, and external generalizability rather than focusing primarily on internal predictive performance.]]>
