Machine-learning-assisted ultraviolet-visible spectroscopy for low-level multi-component adulteration identification...

Low-level adulteration of botanical ingredients is difficult to detect because weak adulterant signals are often masked by the dominant matrix spectrum. Here, Ultraviolet-Visible (UV-Vis) spectroscopy combined with machine learning was investigated for identification of low-level multi-component adulteration in Coptis chinensis. Scutellaria baicalensis, Rheum palmatum, and Curcuma longa were used as adulterants to construct 29 classes, including pure, single-, binary- and ternary-adulteration samples, at a fixed total adulterant content of 3 wt%. Twenty independently prepared samples were included in each class, resulting in a total of 580 independent samples for model development and evaluation. The spectra showed strong overlap, nearly unchanged peak positions, and subtle intensity variations, making visual inspection and single-wavelength analysis insufficient for detailed discrimination. Multiple spectral preprocessing methods and machine-learning models were systematically compared using stratified five-fold cross-validation at the independent sample level. Following this systematic comparison, moving-average preprocessing combined with either random forest or extreme gradient boosting achieved the highest overall classification accuracy of 95.9%. Spectral-window analysis further showed that broader wavelength regions provided higher accuracy than narrow peak-centered windows, indicating that the discriminative information is distributed across the spectral profile rather than confined to isolated absorption peaks. Even though the study was limited to three model adulterants under the present controlled laboratory conditions, these results demonstrate the feasibility of extracting weak composition-dependent information from highly overlapping UV-Vis spectra and highlight the need for further validation using more diverse adulterants and sample sources.