Binary classification of sake acidity using excitation-emission matrix fluorescence spectroscopy combined with multivariate analysis models.
EEM fluorescence spectroscopy combined with PLS-DA shows promise for rapidly classifying sake acidity, but further validation on independent samples is necessary.
Where it sits
this study against the rest of the vip (vasoactive intestinal polypeptide) corpusSummary and findings
This study measured the acidity of Japanese sake using excitation-emission matrix fluorescence spectroscopy combined with chemometric modeling. A total of 100 commercial sake samples were classified into lower-acidity (≤ 1.5; n = 43) and higher-acidity (> 1.5; n = 57) groups. The best-performing model achieved an accuracy of 0.880.
Abstract
Acidity is a key sensory and quality attribute of Japanese sake; however, its conventional determination by titration is destructive and time-consuming. This study investigated excitation-emission matrix (EEM) fluorescence spectroscopy combined with chemometric modeling as a rapid, non-destructive alternative for binary classification of sake according to acidity level. A total of 100 commercial sake samples were classified into lower-acidity (≤ 1.5; n = 43) and higher-acidity (> 1.5; n = 57) groups based on industry standards. Six classification pipelines were compared using stratified nested cross-validation: Partial Least Squares Discriminant Analysis (PLS-DA) and Sparse Partial Least Squares Discriminant Analysis (sPLS-DA) applied to the full EEM dataset, and Support Vector Machine (linear kernel) and Random Forest classifiers applied to features extracted by Parallel Factor Analysis (PARAFAC) and Principal Component Analysis (PCA). EEM-PLS-DA achieved the best overall performance (Accuracy 0.880, Macro-F1 0.866, ROC-AUC 0.901) and ranked highest in the Friedman test (p < 0.05). Nemenyi post-hoc comparisons (CD = 2.384) confirmed a significant difference only between EEM-PLS-DA and the worst-performing model in each metric. Variable Importance in Projection (VIP) and regression coefficient analyses identified five fluorescence regions associated with tyrosine-like fluorophores, aromatic amino acids, polyphenols, Maillard reaction products, and riboflavin. These results indicate that acidity-related information is distributed across multiple fluorophore systems. Overall, these results support EEM fluorescence spectroscopy combined with PLS-DA as a promising proof-of-concept for rapid, reagent-free screening, indicating potential for sake quality control pending validation on independent samples.
Background
This paper addresses the challenge of determining acidity in Japanese sake, which is traditionally done through destructive and time-consuming titration methods. Previous methods have not provided a rapid, non-destructive alternative for classifying sake based on acidity levels. The study aims to evaluate the effectiveness of excitation-emission matrix fluorescence spectroscopy combined with multivariate analysis models for this purpose.
Methods
The study utilized a total of 100 commercial sake samples, classified into lower-acidity (≤ 1.5; n = 43) and higher-acidity (> 1.5; n = 57) groups. Six classification pipelines were compared, including EEM-PLS-DA and sPLS-DA applied to the full EEM dataset, as well as Support Vector Machine and Random Forest classifiers applied to features extracted via PARAFAC and PCA. Stratified nested cross-validation was employed for model evaluation.
Results
The EEM-PLS-DA model achieved an accuracy of 88.0% (0.880) in classifying sake samples. The Macro-F1 score was 0.866, and the ROC-AUC was reported as 0.901. Statistical significance was indicated by a Friedman test result of p < 0.05. Nemenyi post-hoc comparisons yielded a critical difference (CD) of 2.384, confirming significant differences between the best and worst-performing models.
Interpretation
The findings suggest that EEM fluorescence spectroscopy combined with PLS-DA is a promising method for rapid classification of sake acidity. While the accuracy and other metrics are statistically significant, the clinical relevance in terms of practical application remains to be validated with independent samples. Limitations include the absence of real-world application data and potential confounding factors related to sample selection.
Key findings
- Accuracy 0.880 for EEM-PLS-DA model.
- Macro-F1 score 0.866.
- ROC-AUC 0.901.
- p < 0.05 for Friedman test indicating significant differences.
- CD = 2.384 for Nemenyi post-hoc comparisons.
Limitations
- Not reported in abstract.
- Independent validation not conducted.
- Potential confounding factors not addressed.
- Sample selection criteria not detailed.