Non-destructive discrimination of fresh, aged, and frozen-thawed beef using portable near-infrared spectroscopy combined with explainable artificial intelligence

Abstract

Food fraud involving the mislabeling of frozen-thawed beef as fresh threatens supply chain integrity. While portable visible–shortwave near-infrared (Vis-SWNIR) spectroscopy (700–1100 nm) offers rapid, non-destructive screening, the "black-box" nature of many machine learning models hinders regulatory acceptance. This study developed a Vis–SWNIR framework with explainable AI (XAI) to discriminate fresh muscle, 24-h aged and frozen–thawed beef. A systematic benchmark of 22 classifiers on 6000 spectra identified Linear Discriminant Analysis (LDA) as optimal for discriminating aged from frozen–thawed beef, achieving 86.5 % accuracy (95 % CI: [85.23, 87.77]) with superior stability over complex models. XAI methods (SHAP, LIME) integrated with two-dimensional correlation spectroscopy (2D-COS) identified mechanistically coherent biomarkers: myoglobin redox states (∼759 nm), water-protein interactions (∼806 nm) and hydration features (∼970–1060 nm). 2D-COS revealed that myoglobin oxidation and water redistribution occurred sequentially during aging but simultaneously during freezing. Post-hoc power analysis confirmed statistical adequacy, with aged vs. frozen showing mean effect size d = 0.292 ± 0.747 and 99.4 % power, while univariate maxima reached d = ±3.51. This work establishes a transparent, statistically powerful and mechanistically validated framework, providing a deployable, regulatory-compliant solution for real-time beef authentication. • LDA achieved 86.5 % accuracy with robust cross-validation stability. • Freezing explains >80 % PCA variance; altered 750, 805, 970–1060 nm bands. • SHAP and LIME revealed myoglobin, water-protein, and hydration biomarkers. • 2D-COS confirmed sequential aging vs. parallel freeze-thaw mechanisms. • Framework demonstrates promising potential for portable beef authentication.

Topics & Concepts