A Proof-of-Concept Study of Meat Broth Classification Using UV–Vis Spectroscopy and Explainable Decision Tree Analysis

Abstract

Background: Rapid discrimination of animal-derived food products is relevant to food authenticity, consumer protection, and dietary compliance. UV–Vis spectroscopy provides a simple analytical approach for obtaining optical fingerprints of food matrices; however, spectral interpretation may benefit from transparent machine-learning methods. Objective: This proof-of-concept study investigated the feasibility of combining UV–Vis spectrophotometry with an interpretable Decision Tree classifier to discriminate chicken, pork, and beef broths and to identify wavelength variables contributing to classification. Methods: Broths prepared from chicken, pork, and beef were analyzed by UV–Vis spectrophotometry in the 244.5–249.0 nm spectral region, yielding 10 wavelength-specific absorbance variables. Two predesignated datasets were used: 15 training observations (five per class) and 15 test observations (five per class). Model development used only the training dataset. Because 10-fold stratification is not feasible with only five observations per class, shuffled 10-fold KFold cross-validation (random_state = 42) was applied to the training data. A Decision Tree classifier using Gini impurity was evaluated by accuracy, macro-precision, macro-recall, macro-F1, out-of-fold (OOF) confusion matrix, and multiclass one-vs-rest ROC-AUC. The final tree was then fitted to all training observations and evaluated once on the untouched test dataset. Feature importance and SHAP values were used for model interpretation. No feature scaling was applied to Decision Tree modeling. Results: All 10 validation folds were classified correctly, producing mean cross-validation accuracy, macro-precision, macro-recall, and macro-F1 of 1.000 ± 0.000. Pooled OOF predictions also yielded accuracy, macro-precision, macro-recall, macro-F1, and multiclass OvR ROC-AUC of 1.000. The final Decision Tree had a depth of two and three terminal leaves. On the separately designated 15-observation test dataset, all observations were correctly classified (confusion matrix diagonal: 5, 5, and 5), with accuracy, macro-precision, macro-recall, macro-F1, and macro OvR ROC-AUC all equal to 1.000. Feature importance was concentrated at 245.5 and 248.5 nm (0.50 each), and SHAP analysis likewise identified these wavelengths as the dominant model contributors. These perfect metrics must be interpreted cautiously because the observations are technical replicates from a very limited number of broth preparations. Conclusion: The findings demonstrate the feasibility of integrating UV–Vis spectroscopy with an interpretable Decision Tree for preliminary discrimination of chicken, pork, and beef broth spectra. Under the present proof-of-concept dataset, the model achieved perfect 10-fold OOF and separately designated test-set classification, with 245.5 and 248.5 nm driving the fitted decision rules. However, the technical-replicate structure and very small number of independent broth preparations preclude claims of biological generalization. Validation using larger numbers of independently prepared samples, multiple biological sources and batches, and genuinely external test sets is required before practical food-authentication use can be established.