Integration of SWIR hyperspectral imaging, UHPLC-Q-TOF/MS and optimized machine learning for non-destructive authentication of Pseudostellaria heterophylla

Accurate authentication of Pseudostellaria heterophylla (TZS) is essential for quality control in traditional Chinese medicine (TCM). This study pioneers a non-destructive authentication framework by systematically comparing machine vision and hyperspectral imaging (HSI) across visible-near-infrared and short-wave infrared (SWIR) spectral ranges. Multiple classification algorithms, including support vector machines, partial least squares discriminant analysis, extreme learning machines, and backpropagation neural networks (BPNN), were implemented. Wavelength selection was performed using competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and their hybrid approach, while model parameters were optimized through particle swarm optimization (PSO). Results revealed that the SWIR HSI system achieved 96.4 % classification accuracy using a CARS-SPA-PSO-BPNN model with only 16 effective wavelengths in 13.9 s, accomplishing 95.9 % spectral dimensionality reduction while maintaining optimal performance. Complementary untargeted metabolomic profiling via UHPLC-Q-TOF/MS identified 41 key differential metabolites, including fatty acids, lipids, cyclic peptides, and terpenoids, correlated with the spectral features, mechanistically validating the model’s decisions. This work lays the foundation for a streamlined hyperspectral sorting system to detect the authenticity of TZS, advancing TCM standardization and industrialization. • EWs were extracted by CARS, SPA, and CARS-SPA algorithms. • Particle swarm optimization (PSO) algorithm optimizes SVM and BPNN parameters. • SWIR HSI outperformed machine vision and Vis/NIR HSI in detecting TZS varieties. • CARS-SPA-PSO-BPNN model detects TZS varieties with simplicity and high precision. • Explain detection mechanism by analyzing TZS metabolomic data from UHPLC-Q-TOF/MS