Quality Detection Of Fritillaria Thubergii Using Spectroscopic Techniques And Machine Learning

The traditional Chinese medicine（TCM）industry is an essential component of the Chinese medicine industry.The Outline of the 14th Five-Year Plan（2021-2025）for National Economic and Social Development and the Long-Range Objectives Through the Year 2035clearly calls for‘Strengthen supervision of TCM quality and promote the improvement of TCM quality’,pointing out the direction for the green,digital,intelligent,and sustainable development of China’s TCM industry.Fritillaria thunbergii Miq.is one of the most important herbs in China,which has been widely cultivated in the south-east coastal,south-central and eastern areas of China.The bulbs of Fritillaria species growing in China have been used as antitussive and expectorant herbs in TCM.However,the environmental conditions of herbs existing in a particular area or ecosystem such as physical,chemical,biological,and climatic factors can influence the quality and safety of herbs within that environment.Therefore,it is crucial to evaluate the quality and safety of Fritillaria thunbergii.So far,laboratory-based chemical methods,have been used in the qualitative and quantitative analyses of TCM.These methods are high accuracy and sensitive for chemical composition analysis,but complex sample preparation and time-consuming procedures make them unusable for rapid detection.These methods also fail to support the detection during cultivation,acquisition,and distribution of TCM.Recently,hyperspectral imaging（HSI）,near-infrared spectroscopy（NIRS）and laser-induced breakdown spectroscopy（LIBS）,as a non-or minimal destructive detection with short response time,have broad application potential and development prospects in quantitative and qualitative detection of TCM.With the development of algorithms,such as machine learning（ML）and deep learning（DL）,the combination of multi-source spectrum technology and advanced algorithms create a favorable condition for establishing reliable models for rapid detection.This study focused on the quality and safety of Fritillaria thunbergii using HSI,NIRS and LIBS comprehensively,with ML and DL coupled,for establishing assessment methods and models on varieties discrimination,heavy metal pollution and polysaccharide content predictions.The main research contents are as follows:（1）The characteristics of HSI spectrum of Fritillaria thunbergii were studied,and the DL recognition model of multi-variety were established to realize the high-precision and accurate recognition.A feasibility study is presented that examines the use of hyperspectral imaging integrated with convolutional neural networks（CNN）to distinguish twelve（12）Fritillaria varieties（n=360）.The performance of support vector machines（SVM）and partial least squares-discriminant analysis（PLS-DA）was compared with that of convolutional neural network（CNN）.Principal component analysis（PCA）was used to assess the presence of cluster trends in the spectral data.Cross-validation was used to optimize the performance.Among all the discriminant models,CNN was the most accurate with 96.88%,88.89%in training and test sets,followed by PLS-DA and SVM with 92.59%,81.94%and 99.65%,79.17%,respectively.The results obtained in the study revealed that the application of HSI in conjunction with the deep learning technique can be used for the classification of Fritillaria thunbergii varieties rapidly and non-destructively.（2）The method of variable selection and gradient boosting machine for heavy metal elements in Fritillaria thunbergii was proposed,which achieved the rapid quantitative detection of Cd,Cu and Pb in Fritillaria thunbergii.Heavy metals（Cd,Cu and Pb）in Fritillaria thunbergii were analyzed simultaneously using LIBS coupled with variable selection and chemometrics.A high-accuracy and fast approach for the detection of heavy metals novel tree-based ensemble methods combined with LIBS for heavy metals detection was presented in the study,and inspiring prediction performance has been achieved.Gradient boosting machine（GBM）,extreme gradient boosting machine（XGBoost）and category boosting machine（CatBoost）,learning algorithms were evaluated and compared with two conventional machine learning models（SVR and PLSR）.Three promising wavelength selection methods were also applied for comparison,namely,a competitive adaptive reweighted sampling method（CARS）,a random frog method（RF）,and an uninformative variable elimination method（UVE）.Compared to full wavelengths,the selected wavelengths produced best results.Overall,for GBM combined with variables selection the following R_P² and RMSEP were obtained:0.9403,8.5354 mg kg^-1;0.9665,11.9356 mg kg^-1;0.9686,10.1224 mg kg^-1were obtained for Cd Cu and Pb,respectively.For XGBoost combined with variables selection the following R_P²,and RMSEP were obtained:0.9584,10.5338 mg kg^-1;0.9886,12.5759 mg kg^-1,0.9692,11.5611mg kg^-1 were obtained for Cd,Cu and Pb,respectively and finally for CatBoost combined with variables selection the following were obtained:R_P² and RMSEP were obtained:0.9881,9.3476mg kg^-1;0.9928,10.3316 mg kg^-1;0.9744,8.4403 mg kg^-1 were obtained for Cd Cu and Pb,respectively.The results obtained showed that the novel boosting algorithms are capable of performing variable selection effectively,with CatBoost-RF providing best results for Cd,Cu and CatBoost-UVE for Pb.Therefore,the result suggested that LIBS coupled with boosting algorithms and wavelength selection methods can be used to rapidly and accurately detect heavy metals in Fritillaria by extracting very few variables containing useful information and eliminating non-informative variables.Consequently,the results obtained in the study are excellent benchmarks for heavy-metal predictions in Fritillaria based on novel tree-based ensemble methods.（3）The combination of NIRS and tree-based ensemble method for polysaccharide detection in Fritillaria thunbergii was investigated,and the rapid and quantitative detection of polysaccharide was realized.The performance of near-infrared（NIR）spectroscopy coupled with various chemometrics and machine learning methods was compared in order to detect the polysaccharide content in Fritillaria.The spectral pre-treatment was first carried out using Savitzky-golay（SG）and orthogonal signal correction（OSC）methods.For the purpose of simplifying the models and improving generalization performance,variable selection（VS）methods such as variable importance in projection（VIP）,recursive partial least square（r PLS）,synergy interval partial least square（si-PLS）,and successive projection algorithm（SPA）were used to select the feature wavelengths.Multivariate quantitative analysis of NIR data was carried out using the multiple linear regression（MLR）,principal component regression（PCR）,partial least squares regression（PLSR）,support vector regression（SVR）,and extreme gradient boosting（XGBoost）regression algorithms.It was found that the combination of models using SG-VS-XGBoost and（SG+OSC+MSC）-VS-XGBoost achieved state-of-the-art results on many machine learning challenges.A comparison of SVR,PCR,and PLSR,which are powerful and classic machine learning methods,perform relatively well with both spectral pretreatment methods.However,XGBoost obtained the best results,with the exception of（SG+OSC+MSC）+VIP+PLSR.The prediction result for SG-r PLS-XGBoost obtained is as follows:R_P²=0.6588,In contrast,（SG+OSC+MSC）-VIP-PLSR yielded a relatively better prediction result with R_P²=0.6698.The linear PLSR model achieved the best result with R_P²=0.6811,for the polysaccharide content detection in Fritillaria.It was demonstrated that diverse VS methods for selected feature wavelengths can remove redundant information,reduce modeling complexity,mitigate overfitting,and improve the generalization performance of the model.A series of modeling techniques have been tested in this study,which provides a framework for spectral pretreatment,variables selection,machine learning method selection,and a comparative analysis of model performance for rapid and quantitative detection of polysaccharides in varieties of Fritillaria.Similarly,this approach can be employed to analyze other kinds of TCM quantitatively.