An intelligent process analysis method for rapidly evaluating the quality of Chinese medicine with near-infrared non-contact hyperspectral imaging: A case study of Weifuchun concentrate.

Abstract

The quality of Chinese medicine preparations can be greatly influenced by the quality of the intermediates such as extracts or concentrates. However, it is highly challenging to evaluate the quality in a rapid and non-contact manner during manufacturing. Here, we introduce an intelligent hyperspectral analysis method integrating a self-built abnormal region removal algorithm with machine learning and demonstrate its utility using the concentrate of Weifuchun (WFC), a traditional Chinese medicine preparation made from Ginseng Radix et Rhizoma Rubra, Rabdosia Amethystoides, and Aurantii Fructus. To rapidly and non-destructively detect quality attributes of the intermediates in the manufacturing processes of Chinese medicine, an intelligent hyperspectral analysis method was developed for simultaneously quantifying the contents of naringin, neohesperidin, rosmarinic acid, and relative density of WFC concentrates. Samples were evenly spread on solid white flat bottom containers, which were batch placed on a horizontal sample stage. Subsequent to the acquisition of near-infrared (NIR) hyperspectral images, abnormal pixels such as large/small bubbles and fine solids were first removed according to the differential pixel values in the binary grayscale map and the Mahalanobis distance metric. Then, partial least squares (PLS) and support vector machine (SVM) algorithms were used to construct hyperspectral quantitative calibration models for quality attributes. The hyperspectral images were reconstructed based on these models to visually evaluate the quality of the concentrates during manufacturing. As a case study, quality attributes of the WFC concentrates including contents of naringin, neohesperidin, rosmarinic acid, and relative density were determined simultaneously, and coefficients of determination of these quantitative correction models were 0.900, 0.891, 0.851, and 0.920, respectively. The method proposed in this study favors real-time determination of multiple attributes in viscous samples with industrial application prospects.