New Method Pairs Terahertz Spectroscopy With AI to Authenticate Premium Ginseng's Age

Determining the age of mountain-cultivated ginseng (MCG) has always been challenging because the current scientific methods have not achieved the necessary accuracy. However, a recent study published in Microchemical Journal proposed a non-destructive technique that could achieve 96% accuracy.¹

What is mountain-cultivated ginseng?

Mountain-cultivated ginseng (MCG) is a functional food and herbal medicine that is used in traditional Chinese medicines.² It has commonly found in China, Korea, and Japan, and it has several positive health benefits. For example, MCG is known to improve the health of human endocrine, immune, and central nervous systems.²

What did the researchers do in this study?

The research team, led by Hongxi Xu and Yan Peng of Shanghai University of Traditional Chinese Medicine and the University of Shanghai for Science and Technology, combined terahertz (THz) spectroscopy with convolutional neural networks (CNN) to classify 54 MCG samples ranging from 5 to 20 years of age.¹ Plant age is commercially significant because ginsenoside concentrations, which are the bioactive compounds associated with ginseng's therapeutic properties, accumulate over time, making older specimens substantially more valuable.^1,3

What were the core findings of this study?

The core finding is that the THz-CNN pairing outperforms high-performance liquid chromatography (HPLC) combined with the same neural network architecture. In direct comparison, the THz-CNN model achieved 92.59% age classification accuracy versus 77.78% for HPLC-CNN across three growth-stage groupings.¹ The researchers attributed HPLC's lower performance to inter-batch variability in chromatographic column silanization, which introduces measurement drift that degrades the machine learning model's reliability.¹

The researchers also noted how they were able to improve accuracy further in their study. They did so through the creation of a feature-augmented input framework.¹ This feature-augmented input framework a method that feeds both raw spectral data and extracted high-importance features simultaneously into the neural network. That approach pushed classification accuracy to 96.30%.¹

What are the implications of this study?

There are both knowledge and practical implications that emerged from this study. For the knowledge aspect, the study revealed how the nonlinear relationship between ginsenoside accumulation and plant age was an important mechanism in making the method work.¹ The terahertz (THz) spectrum captures fingerprint spectral features and multi-year evolution patterns across multiple ginsenosides simultaneously, enabling multi-dimensional classification that single-compound HPLC quantification cannot replicate.¹

Meanwhile, on the practical implications, the study showed that THz spectroscopy is a better alternative method than HPLC for ginsenoside quantification. Using HPLC for this context is standard, but because of its drawbacks, including the need chemical solvents, sample preparation, and column maintenance, the researchers identified an area where another analytical method could be better.¹ The researchers demonstrated that THz spectroscopy is non-destructive, requires no sample preparation, and preserves the intact root, which is a critical advantage when authenticating high-value specimens where physical integrity affects market price.¹

References

1. Wang, J.; Wang, S.; Sun, L.; et al. Age Identification of Mountain-cultivated Ginseng based on Terahertz Precision Spectroscopy and Convolutional Neural Network. Microchem. J. 2026, 221, 116954. DOI: 10.1016/j.microc.2026.116954
2. Xu, X.-F.; Cheng, X.-l.; Lin, Q.-h.; et al. Identification of Mountain-cultivated Ginseng and Cultivated Ginseng Using UPLC/oa-TOF MSE with a Multivariate Statistical Sample-profiling Strategy. J. Ginseng Res. 2015, 40 (4), 344–350. DOI: 10.1016/j.jgr.2015.11.001
3. Chen, W.; Balan, P.; Popovich, D. G. Analysis of Ginsenoside Content (Panax ginseng) from Different Regions. Molecules 2019, 24 (19), 3491. DOI: 10.3390/molecules24193491