Rapid identification of lingnan traditional Chinese medicinal materials via headspace-gas chromatography-ion mobility spectrometry combined with deep learning

Objective To develop a rapid, accurate, and non-destructive authentication method for Lingnan traditional Chinese medicinal materials (TCMMs), this study integrated headspace-gas chromatography-ion mobility spectrometry (HS-GC-IMS) with deep learning technology. Methods A total of 4755 HS-GC-IMS fingerprint samples were collected from 11 types of medicinal materials, including 9 types of Lingnan TCMMs (e.g., Amomum villosum , agarwood, Citrus reticulata ) and 2 types of non-Lingnan TCMMs ( Cordyceps sinensis and its weight-increased adulterants), to construct a dataset in Portable Network Graphics (PNG) and comma-separated values (CSV) formats. Three convolutional neural network models (ResNet18, MobileNetV2, and EfficientNet-B0) were trained for multi-class classification, with the ArcFace loss function applied to enhance inter-class separability. Model performance was evaluated using accuracy, precision, recall, and F1-score. Results The EfficientNet-B0 model exhibited the optimal performance, achieving an average accuracy of 99.56%, precision of 97.62%, recall of 99.42%, and F1-score of 98.49%. It successfully distinguished geographical origins of Lingnan TCMMs, verified authenticity of genuine vs. counterfeit samples, and achieved high-accuracy identification of non-Lingnan Cordyceps sinensis . Notably, PNG-format spectral images showed 140-fold higher training efficiency and 12-fold higher inference efficiency than CSV data. Conclusion This study first established an HS-GC-IMS fingerprint dataset covering Lingnan and non-Lingnan TCMMs. The proposed method not only provides a reliable and efficient solution for quality control of Lingnan TCMMs but also demonstrates strong extended application potential, offering a technical innovation for traditional Chinese medicine inspection and promoting its modernization.