A machine learning-based method for adulteration detection in white rice vinegar using different non-targeted spectroscopy

Technical Summary: Machine Learning-Based Adulteration Detection in White Rice Vinegar Using Non-Targeted Spectroscopy

Problem Statement
White rice vinegar is a widely consumed condiment in Asian markets that is highly susceptible to economically motivated adulteration, primarily through dilution with cheap synthetic acetic acid or substitution with low-value spirit vinegar. While traditional chromatographic methods can detect such fraud, they are often too complex, time-consuming, and infrastructure-heavy for routine, high-throughput industrial screening. Although non-targeted spectroscopic techniques (NIR, FTIR, EEM) offer rapid, non-destructive alternatives, existing literature suffers from critical gaps: a lack of focus on industrial-grade white rice vinegar (as opposed to protected designation of origin products), insufficient systematic comparison of different spectroscopic modalities for the same task, and a failure to adequately address the impact of "brand specificity"—the significant chemical variability between commercial brands—on the generalization and feature selection of machine learning models.

Methodology
This study implemented a comprehensive, data-driven workflow to evaluate and compare the efficacy of three non-targeted spectroscopic techniques: Near-Infrared (NIR), Fourier Transform Infrared (FTIR), and Fluorescence Excitation-Emission Matrix (EEM) spectroscopy.

• Sample Preparation: Four commercial white rice vinegar brands (Hengshun, Qianhe, Shuangyu, and Donghu) were subjected to simulated adulteration using a 4% (v/v) aqueous acetic acid solution. Adulteration levels ranged from 0% (pure) to 50% (v/v) in eight distinct steps, with ten analytical replicates per brand-concentration combination.
• Data Acquisition:
  • NIR: Collected using a handheld spectrometer (1350–2500 nm).
  • FTIR: Acquired via Attenuated Total Reflection (ATR) (400–4000 cm⁻¹).
  • EEM: Recorded using a spectrofluorometer (Excitation: 230–400 nm; Emission: 240–600 nm).
• Data Processing and Modeling:
  • Visualization: t-SNE (t-distributed Stochastic Neighbor Embedding) was used for nonlinear dimensionality reduction to visualize clustering patterns.
  • Preprocessing: A systematic, exhaustive search was conducted across 16 structured preprocessing workflows (combinations of ALS, SNV, SGF, PS, and MC) to objectively identify optimal pipelines based on RMSECV.
  • Feature Selection: Four algorithms—Uninformative Variable Elimination (UVE), Variable Importance in Projection (VIP), Competitive Adaptive Reweighted Sampling (CARS), and Genetic Algorithm (GA)—were employed to select the top 10% of characteristic wavebands.
  • Classification: Five machine learning models (LDA, PLS-DA, DT, RF, and SVM) were trained and optimized via grid search.
  • Evaluation: A rigorous 5×5 nested cross-validation scheme was used to assess performance using Accuracy, Precision, Recall, F1-score, and Cohen's Kappa.

Key Results

1. Spectroscopic Efficacy: FTIR spectroscopy consistently demonstrated the strongest discriminative potential across all four brands. In benchmark tests using full-band data, optimal FTIR models achieved cross-validation accuracies ranging from 95.00% to 100.00%, significantly outperforming NIR (71.25%–95.00%) and EEM (66.25%–90.00%).
2. Impact of Feature Selection: The integration of feature selection with FTIR maintained or improved model performance. Optimal FTIR models based on selected characteristic bands achieved identification accuracies between 85.00% and 100.00%. In contrast, feature selection generally reduced the accuracy of NIR models across all brands and showed mixed results for EEM models.
3. Brand Specificity: A critical finding was the pronounced variability in optimal modeling pathways depending on the vinegar brand.
   • Preprocessing: The optimal preprocessing pipeline (e.g., ALS-PS vs. SNV-PS) varied significantly between brands.
   • Feature Selection: The most effective feature selection algorithm (e.g., VIP, CARS, or GA) and the resulting distribution of characteristic bands differed by brand.
   • Model Architecture: The best-performing classifier (e.g., LDA vs. RF) was not universal; for instance, LDA was optimal for Hengshun FTIR data, while RF was superior for Qianhe, Shuangyu, and Donghu.
4. Spectral Characteristics: While raw spectral curves showed high similarity across brands and adulteration levels, t-SNE visualization revealed that FTIR provided the clearest separation between adulteration levels. EEM spectra showed the weakest clustering, with significant overlap between authentic and adulterated samples for certain brands.

Significance and Claims
The paper claims that the integration of FTIR spectroscopy with a systematic chemometric strategy (comprising rigorous preprocessing, feature selection, and machine learning) provides a robust foundation for detecting adulteration in white rice vinegar. The study highlights that while high-precision authentication tools are feasible, they cannot be "one-size-fits-all." The authors emphasize that brand specificity is a critical factor; the optimal preprocessing, feature selection, and model architecture must be tailored to the specific chemical matrix of each vinegar brand. Consequently, the study advocates for accounting for brand variability when developing generalized authentication tools for the food industry, rather than assuming a single modeling approach can effectively handle diverse commercial products. The findings support the development of rapid, reliable detection methods but caution that future research must address the challenge of creating unified, brand-fused models to handle inter-brand variations effectively.