Comparative machine learning-assisted terahertz and ultraviolet-visible spectroscopy for rapid authentication and...

Reliable authentication and quantitative assessment of adulteration levels in traditional Chinese medicinal herbs are essential yet remain analytically challenging due to their complex chemical profiles, morphological similarities, and frequent adulteration. Here, we comparatively evaluate ultraviolet-visible (UV-Vis) spectroscopy, terahertz time-domain spectroscopy (THz-TDS), and their feature-level fusion combined with machine learning for rapid herbal authentication and adulteration-level prediction. Using Coptis chinensis as a representative target, the method was applied to three key analytical tasks: species identification, adulterant-type discrimination and regression-based adulteration-level prediction. Classification models were used for species identification and adulterant-type discrimination, while regression models were used to predict adulteration levels as continuous variables. After spectral preprocessing, supervised machine-learning models achieved high classification accuracies, exceeding 90% for THz-TDS data and reaching 100% for UV-Vis data. The regression analysis demonstrated proof-of-concept evaluation of adulteration-level prediction within the five predefined mixture levels, with lower internal cross-validation errors obtained from the UV-Vis spectra. Paired THz-TDS and UV-Vis spectra from the same samples were further used to directly compare THz-TDS alone, UV-Vis alone, and feature-level fusion. Under the standard methanol-extraction condition, feature fusion matched the classification performance of UV-Vis alone and showed generally comparable regression performance, while providing improved performance compared with THz-TDS alone. These results provide a comparative proof-of-concept assessment of the three spectral input strategies and establish a basis for future validation using broader sample sources and independent external mixtures.