Introduction

Codonopsis Radix is a family of plants used mainly in China and East Asian countries, such as Korea and Japan, to replenish vital energy, or qi1,2. The dried roots of Codonopsis pilosula (Franch.) Nannf, Codonopsis pilosula Nannf. var. modesta (Nannf.) L. T. Shen and Codonopsis tangshen Oliv. are all used as Codonopsis3. Traditionally, doctors of Chinese medicine prescribe Codonopsis Radix to strengthen spleen, tonify lung, nourish blood and generate fluid. Besides, Codonopsis Radix is utilized as a nutraceutical component in conventional foods, functional foods, and dietary supplements4. The roots are also used in Chinese tonic teas and soups, as well as being roasted with millet, eaten raw, baked, or pickled in miso5. However Codonopsis with different sources has different flavor. The roots of Codonopsis pilosula Nannf. var. modesta (Nannf.) L. T. Shen have tight skin, thick striate, firm root strips, and sweet-tasting. What’s more, Codonopsis is widely cultivated in Gansu, Shanxi, Shaanxi, Hubei, Sichuan, Jilin province of China, which further diversifies its flavor6. Thus, it is necessary to figure out the botanical origin of Codonopsis to keep the taste of foods stable.

The majority of botanical origin identification currently relies on experiential detection, or used in conjunction with laboratory physicochemical detection7. However, the subjectivity and insufficient quantitative characterization limit the spread of experiential methods. Although the thin layer chromatography, liquid chromatography, and other methods of laboratory physicochemical detection are frequently used to identify the botanical origin of traditional Chinese medicine (TCM), these methods often come with complex pre-processing processes and protracted detection times8.

Thanks to the advancements in sensor manufacturing, it is possible to identify the botanical origin of foods using electronic sensory instruments such as electronic nose and electronic tongue9,10,11,12,13. Electronic nose is also known as odor fingerprint technology14,15. It functions by simulating the human nose's ability to detect, evaluate, and assess scents. The advantages of simple, fast, and easy pretreatment make it useful for rapid evaluation of complex volatile gas mixtures. The electronic tongue, also known as taste fingerprint technology, is a multi-sensor detection system that analyzes and identifies "taste" by simulating human taste organs16. It can accurately and specifically identify the taste, namely sour, bitter, salty, fresh, astringent, and sweet. All the aforementioned electronic sensory equipment have been reported to identify the geographical origin, source, grade of honey16, Kiwifruit17, Pseudostellariae Radix18,19 and Fritillariae Cirrhosae Bulbus20. But it remains unclear whether data fusion can improve the classification performance of common classification methods on imbalanced data, such as the botanical origin of Codonopsis Radix.

Upsampling or downsampling the minority or majority class are two techniques to handle imbalanced data for a classification problem. But these two types of methods are unfriendly to the majority of analysts, since they all need some programming skills. In this work, both electronic tongue and nose techniques were employed to gather sensory information about the smell and taste of the Codonopsis samples. To accurately identify the botanical origin of Codonopsis Radix, three classification methods were separately trained on the individual data and the fused one. Data preprocessing method were also employed to eliminate the negative influence cause by data magnitude. The optimal model built on the e-tongue data performs comparable to that on the e-nose data. The results also showed that combining the two different sensors increased the prediction performance on imbalanced data. Even so, there is still much work to be done before a model is ready for usage in industry.

Materials and methods

Codonopsis Radix samples

The samples of Codonopsis Radix were gathered from a number of cities in China such as Longnan (Gansu Province), Dingxi (Gansu Province), Changzhi (Shanxi) provinces. A total of 88 samples were dried, crushed and then sifted using a 60 mesh screen. The original plants of the samples were identified by Prof. Hu Fangdi from Lanzhou University. The detailed information of each Codonopsis Radix sample were shown in Supplement Table 1. The experiments for this study were conducted in conformity with the pertinent guidelines and regulations established by the Ministry of Agriculture and Rural Affairs of the People's Republic of China.

Electronic nose

The PEN3 Electronic Nose system (Airsense Analytics GmBH, Schwerin, Germany) was used in this investigation to detect the smell of the sample. The E-nose system comprise a sampling device, a sensor array and a system software for data acquisition and storage. The sensor array, which is the most important component, is made up of 10 MOS sensors, each of which is sensitive to particular volatile substances (Table 1). The article17 defines the primary applications and detection limitations of the 10 MOS sensors.

Table 1 Characteristics of MOS sensors utilized in PEN3 electronic nose.
Full size table

Electronic tongue

Electronic tongue (e-tongue) measurements were performed with Taste-Sensing System SA 402B (Intelligent Sensor Technology Co., Ltd., Atsugi, Japan), which was designed for the determination of taste of liquid samples. Similar to e-noses, signals from nonspecific sensors were used to mimic the functions of human taste receptors. The five basic flavors of sour, sweet, bitter, salty, and fresh as well as astringent were measured through detecting the changes of membrane potential caused by electrostatic or hydrophobic interaction between flavor substance and artificial lipid membrane. The sensor set includes six sensors, namely AAE, CT0, C00, AE1, CA0, GL1, which stand for umami, saltiness, bitterness, astringency, sourness, and sweetness, respectively21. Table 2 details the relationship between the sensors of SA 402B and their physiological meaning.

Table 2 Sensors of SA 402B electronic tongue and typical food substances.
Full size table

Electronic nose detection

For the method, 3.0 g of each sample was used and put in sample chamber before being tested with the PEN3 electronic nose. Prior to measurement, a sensor check was done to make sure the sensors were operating within the proper voltage range. Each sample was then incubated at 30 ℃ for 60 min in order to reach headspace equilibrium. To normalize the sensor signal, the gas chamber was first cleaned with gas filtered by active charcoal. After each injection and data collection, the sensor self-cleaning time was extended to 120 s to re-establish a stable instrument baseline. The operation parameters were set as following: Sample interval [s]: 1.0; Flush time [s]: 80.0; Zero point trim time: 5.0; Presampling time [s]: 5.0; Measurement time: 80.0; Chamber flow (mL/min): 400; Initial injection flow: 400.

To reduce the sampling error, each sample was measured in triplicate. The data collected at 70 s of the profile were used for the statistical analysis to ensure that the electronic nose reached the adsorption equilibrium. Finally, each sample was represented by a 10-bit vector.

Electronic tongue detection

For the measurements, 100 mL of each sample filtrate was utilized. The electrodes were conditioned in accordance with the developers' instructions before commencing the experiment. In the beginning, it was necessary to perform a sensor check to make sure all the sensors were operating within the proper voltage range. Every sample measurement began with a cleaning procedure: the sensors were washed for 90 s with the positive and negative cleaning solutions, and then for 120 s with a standard cleaning solution. By measuring the potential of the reference solution after cleaning, the stability of the lipid membrane potential was recorded (Vr). When the sensor response was stable (deviation less than 0.5 mV) during the measurement, the sample solution was examined for 30 s (Vs). The sensor output for taste was formed by the difference between Vr and Vs. After a quick cleaning procedure, the membrane potential was measured once more in standard solution to assess the samples' aftertaste, or CPA (change in membrane potential owing to adsorption). The five fundamental taste signals were measured four times, with the final three results typically being used to guarantee data stability. For the sweetness, one more time was measured.

In addition to the five fundamental taste, three aftertastes of umami, bitterness, and astringency were also recorded. As a result, in a duplication, 9 values will be generated for each sample.

Data analysis

For the e-nose data, the vectors of duplicate measurements were averaged to reduce measurement variance. Meanwhile, the vector of the last three observations was averaged for each sample of the e-tongue data.

Linear Discriminant Analysis (LDA), or discriminant function analysis is a generalization of Fisher's linear discriminant, a method used in statistics and other fields, to find a linear combination of features that characterizes or separates two or more classes of objects22. For the plain LDA method, there is no parameter that need to be optimized, LDA models were directly developed to classify the botanical origin of Codonopsis.

Classification and Regression Trees (CART) is a common classification algorithm that is required to build a decision tree on the basis of Gini’s impurity index. Specifically, the CART algorithm split the nodes into subnodes by searching for the best homogeneity for the subnodes with the help of the Gini index23. By default, the decision tree model is allowed to grow to its full depth, which will run the risk of overfitting the training data. In order to prevent this from happening, the decision tree must be pruned. In this study, ten-fold inner cross-validation was used to find the best pruning level.

Partial Least Square-Discriminant Analysis (PLS-DA) is a versatile algorithm commonly used for predictive and descriptive modelling. Over the past two decades, PLS-DA has demonstrated great success in modelling high-dimensional datasets24. Despite this, success can still be achieved in low- to middle-dimensional datasets, provided collinearity exists. There are several parameters needs to be optimized before reaching reliable and valid outcomes. One of the most important parameter is the number of latent variables. To provide a trustworthy estimation on the impact of the number of latent variables, eight-fold cross-validation was used.

For binary classification, evaluation metrics including accuracy, precision, sensitivity, specificity, and F1 score are frequently utilized25. However, classification accuracy fails for imbalanced class distributions. Thus, the error rate redefined as Eq. (1) was used.

$${\text{error rate }} = { 1} - {\text{ mean}}\left( {{\text{sensitivity}}} \right)$$
(1)

All the calculation was performed in MATLAB with classification toolbox (Ver. 6.0) developed by Milano Chemometrics and QSAR Research Group26.

Results

Identification the botanical origins of Codonopsis with electronic nose

Figure 1 displays the magnitude and range of different sensors of e-nose (a) and e-tongue (b). The response values of both techniques range from a few tenths to tens. Thus, pre-processing techniques must be utilized to reduce the negative effect brought on by the magnitude of various sensors. Scaling to unit length and Z-score Normalization are two commonly used methods in the literature. The aim of scaling to unit length is to scale the components of a feature vector such that the complete vector has length one. This means dividing each component by the Euclidean length of the vector (Unitlength). Standardization, a.k.a autoscaling, makes the values of each feature in the data have zero mean and unit variance (Z-score). Figure 2 shows the distribution of data preprocessed by the above two methods. Both preprocessing method align the responses of each sensor to the same level. But the magnitude of data preprocessed by Z-score is ten times larger than that of Unitlength.

Figure 1
figure 1

An overview of raw data before variable preprocessing: (a) the magnitude and range of different sensors of e-nose; (b) the magnitude and range of different sensors of e-tongue; (c) the overlay plot of 88 samples.

Full size image
Figure 2
figure 2

Data distributions after variable preprocessing.

Full size image

From the overlay plot in Fig. 1c, it can be observed that it is hard to clearly distinguish the two group of samples on any sensor of E-nose and E-tongue. Thus, after data pre-processing, three binary classification models were built using CART, LDA and PLS-DA, separately. The results of the optimal model are showed in Table 3. All the metrics were estimated by five-fold cross-validation. For CART in the classification toolbox, the best pruning level was determined automatically by an inner ten-fold cross-validation for each train set in the five-fold cross-validation. But the best pruning level of each train set was not output in the final results. While the number of latent variables in the PLSDA model was optimized by an independent eight-fold cross-validation. The extra cross-validation was used to alleviate the concerns about the potential overly optimistic estimation of the prediction performance.

Table 3 The cross validation results of botanical origin identification based on electronic nose.
Full size table

The error rate in this table illustrate that the prediction performance can be improved obviously following pretreatment. The same conclusion can be obtained in terms of accuracy. Based on error rate, the comparison of the models' performances reveals that PLSDA performs better than LDA and CART. However, with respect to their accuracy, the conclusion is reversed. From the confusion matrix (Fig. 3), it can be seen that CART tends to assign the sample to Codonopsis pilosula (Franch.) Nannf (the majority class), while PLSDA model tends to classify the sample as Codonopsis pilosula Nannf. var. modesta (Nannf.) L. T. Shen (the minority class). In fact, the sample size of the two groups is unbalanced. If all the samples were classified as Codonopsis pilosula (Franch.) Nannf, the model accuracy rate could reach 83%.

Figure 3
figure 3

The confusion matrix of CART and PLSDA on the electronic nose data.

Full size image

Identification the botanical origin of Codonopsis Radix with electronic tongue

Similar to the electronic nose data, data preprocessing was executed on the electronic tongue data to eliminate the influence caused by data range across sensors. Undoubtedly, the prediction performance of these models can be improved more or less after pretreatment. As is shown in Table 4, no method is always performed better than the others. The LDA model outperformed CART model in terms of both error rate and accuracy. Compared with LDA, the error rate of PLSDA was reduced significantly after preprocessing. While the accuracy of PLSDA model on the z-score data exhibited a different trend.

Table 4 The cross validation results of botanical origin identification based on electronic tongue.
Full size table

To appreciate how these diverse results arose, the prediction details were summarized in Table 5. There is a huge difference between the number of samples not assigned to any group predicted by PLSDA and LDA. But the number of samples not assigned could not explain the performance of PLSDA on the Z-score data. Taking the results presented in Table 3 into consideration, the reason can be attributed to the preference of the two classification methods.

Table 5 The classification results of LDA and PLSDA on electronic tongue data.
Full size table

Identification the botanical origin of Codonopsis Radix with electronic nose and electronic tongue

The results obtained on the electronic tongue data and electronic nose data both demonstrated that preprocessing could improve the prediction performance of classification methods. However, since the number of samples not assigned to the data preprocessed by Unitlength is larger than that of z-score in Table 5, data fusion was directly executed on the data preprocessed by z-score.

From the raw data extracted from the electronic nose (X1) and electronic tongue (X2), it can be observed that there are few variables extracted (X1 has 10 variables and X2 has 9 variables). Thus, the two datasets were augmented first to pursue possible improvement in the classification model. The hyperparameters of CART, LDA, and PLSDA were tuned using cross-validation. The final result was presented in Table 6.

Table 6 The classification performance of CART, LDA and PLSDA on data fused with two different strategies.
Full size table

Taking the cross-validation accuracy as an index, the three classification models rank as follows: CART > PLSDA > LDA. While ranked with respect to the error rate of cross-validation, the three models are ordered in the opposite sequence. In fact, it is evident from the confusion matrix that CART outperforms PLSDA by a small margin, but with a stronger trend to classify the sample of Codonopsis pilosula Nannf. var. modesta (Nannf.) L. T. Shen as Codonopsis pilosula (Franch.) Nannf. While, the PLSDA model not only maintains high prediction accuracy but also strikes a balance between the need to correctly distinguish the two groups of Codonopsis Radix. Therefore, based on the aforementioned findings, PLSDA should be chosen as the final modeling approach on the augmented data.

Comparing the results of PLSDA model in Tables 3, 4 and 6, it can be observed that augmenting the electronic nose data with the electronic tongue data could enhance the classification performance. Previous studies27 have demonstrated that PCA is an effective data presentation learning method and the features learned could meet the general requirements of mid-level data fusion methods. But it remains unclear whether PCA works for the data when the number of samples is larger than the number of variables. Thus, principle components (PCs) were learned from the electronic nose data and the electronic tongue data separately. Then, the PCs were used in conjunction to train a classification model.

The number of PCs was chosen as the PCs whose cumulating contribution rate was greater than 90%. For the e-tongue data, the first three PCs explained 93.43% of total variance. Meanwhile, the first two PCs explained 98.57% of total variance of e-nose data. The PCs were merged into a matrix Xpc of size (88, 5). Then, the CART, LDA, and PLSDA models were trained on the feature matrix Xpc.

The variance of the explained of the first three principal components of e-tongue was 93.43%, and the cumulative contribution of the first two principal components of e-nose was 98.57%. The new matrix Xpc (88*5) was obtained by combining the principal component scores of the two types of data, and then classification models were established with CART, LDA and PLSDA respectively. The results were shown in Table 6.

As shown in Table 6, the models built on Xpc performed worse than their corresponding model on the raw data. Both CART and LDA still prone to identify the samples as Codonopsis pilosula (Franch.) Nannf. Although the performance of PLSDA was worse than CART and LDA, it reserve the ability to discern sample of Codonopsis pilosula Nannf. var. modesta (Nannf.) L. T. Shen. Even though, there are no samples not assigned to any group. In short, the data fusion strategy based on PCA is not suitable for our data. However, whether this conclusion can be extended to the data that the number of samples is larger than the number of variables remains to be verified.

Discussion

The results of models trained with different classification methods before and after data preprocessing showed that data pretreatment could enhance the classification performance of models built on the data of both e-nose and e-tongue. Although the classification methods used in this study are diverse in their theoretical background, none performed better than the other two in all metrics. From the boxplot in Fig. 2, it was observed that both z-score and Unitlength aligned the responses of different sensors to the same level. By doing so, the contribution of sensors with slower responses will be amplified. That might explain the improvement of data pretreatment on classification models.

Generally, PLSDA is generally used in cases with more variables than samples28. The PLSDA methods were used in this study since collinearity problem existed in e-nose and e-tongue data. Collinearity introduces the risk of over-fitting when using the classification method based on the inverse of the Matrix. When applying LDA to e-tongue and e-nose separately and fused data, the prediction error rate and accuracy of the resubmission test are much better than that of cross validation.

Although merging the data of multi-sensors is simple, it is an effective data fusion method29. The results obtained in this study demonstrated that augmenting the e-tongue data with e-nose data could improve the model’s ability to distinguish the samples of Codonopsis pilosula Nannf. var. modesta (Nannf. ) L. T. Shen from Codonopsis pilosula (Franch.) Nannf. However, in the Codonopsis Radix dataset, the majority class is much larger than the minority class. Both CART and LDA failed to train an impartial model; they all tended to assign samples to the majority class. While PLSDA keeps a good balance between the two classes. Further research is required to determine whether PLSDA can fulfill the requirement of minority class classification for other imbalanced data.

Conclusions

In this study, 88 batches of Codonopsis Radix sample were measured by electronic nose and electronic tongue. The collected data was used separately or in conjunction to create a classification model capable of correctly determine the botanical origin of each sample. The results showed that by merging the electronic nose data and electronic tongue data into a single data set could assist the PLSDA method in building a usable classification model even if the dataset is highly imbalanced. Specifically, the error rate reduced to 14% from 28% on e-nose or 16% on e-tongue. Moreover, it was found that both pretreatment methods used in this study could improve the prediction performance of the models. In other words, the magnitude and range of electronic sensors have a great influence on the prediction ability of the model. But data fusion based on features extract by PCA failed to further improve the prediction performance. More effort should be made to make it clear what the cause of this phenomenon is.