Geographical origin discrimination of Astragalus membranaceus water decoctions using machine learning based on laser-induced breakdown spectroscopy characteristics

Astragali Radix is a bulk traditional Chinese medicinal material whose clinical efficacy is closely associated with geographical origin. Active constituent contents vary significantly across producing regions, yet decoction pieces from different origins are highly similar in appearance, hindering rapid visual identification. This work combines laser-induced breakdown spectroscopy (LIBS) with machine learning to verify the origin of Astragalus membranaceus by measuring the filtered liquid decoction directly. Spectra were collected from 15 independently prepared decoctions (five subareas per origin across Gansu, Inner Mongolia and Xinjiang), giving 900 valid spectra in total. High-resolution spectra (290-540 nm) were preprocessed by airPLS baseline correction, SNV normalization and variable screening, yielding eight characteristic lines of four elements. The 630 spectra from the first four subareas of each origin were used for training and cross-validated hyperparameter selection. The 270 spectra from the fifth subarea (90 per origin) formed a completely held-out test set. PCA-SVM, PCA-LDA and PCA-GNN baselines were also tested. PCA-ResNet-18 reached 99.63% test accuracy (269/270), with recalls of 98.89%, 100% and 100% for Gansu, Inner Mongolia and Xinjiang. All three baselines scored below 98%. The Ca, Mg, K and Al fingerprints are consistent with geochemical variation in mineral uptake and offer an interpretable basis for LIBS origin discrimination. This rapid, reagent-free approach with minimal sample consumption offers a basis for future at-line or online quality control of origin-labelled herbal medicines.