Exploring the potential of paper-based analytical sensors for tea geographical origin authentication

Abstract

Tea (Camellia sinensis (L.) Kuntze—surname of German origin) is a popular beverage consumed worldwide due to its health benefits. Its quality depends on measuring features that may discriminate teas from distinct provenances. Protected designation of origin (PDO) is therefore a very useful label for tea quality evaluation. In the present work, antioxidant activity profiles obtained from microfluidic paper-based analytical devices (µPADs) were analyzed by chemometrics to determine the tea geographic origin. Based on the existing literature, we constructed a database containing chemical data from 26 samples and evaluated it by principal component analysis (PCA) coupled to linear discriminant analysis (LDA). Antioxidant activity was an effective LDA predictor for sample discrimination accomplishing accuracies from 75 to 82%. Modeling performance was favored by an external validation method. The best classification model was found using the first nine PCs as input variables. Training samples achieved a perfect success rate, while the test ones were predicted with 83% specificity, 100% sensitivity, and 90% overall accuracy. The modeling robustness was verified by integrating AUC (0.943) from ROC curve. The PCA-LDA approach taken here demonstrated that the teas coming from different countries can be correctly authenticated through µPADs, thus contributing to certificate samples PDO.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1007/s13197-022-05440-1.

Introduction

Proof of geographical origin is one of the main issues of authenticity around the world for food and beverages. Protected Designation of Origin (PDO), Protected Geographical Indication (PGI) and Traditional Speciality Guaranteed (TSG) labels are often used to ensure the foodstuff safety and quality of European origin, particularly in geographical environment (Dias and Mendes 2018). The information related to the PDO is directly associated with authentication for fighting against fraud or mislabeling. This specific quality declaration is great interest in products manufacturing with a higher market retail price, since they may be target of fraudulent activities (Katerinopoulou et al. 2020). In this sense, authentication process is essential to control food labeling preventing consumers, producers, and traders from financial damage.

Tea is a popular beverage widely consumed worldwide from the Camellia sinensis (L.) Kuntze plant due to its health benefits. The tea quality is determined by its majority components such as caffeine, free amino acids, and polyphenols (Zhu et al. 2017). The latter constitute an important group of natural antioxidants which are molecules that scavenge or quench free radicals by reacting with them, and thus avoiding oxidative damage (Nimse and Pal 2015). These compounds have played an important role in preventing diseases related to aging, coronary artery, and even cancers, in addition to being used as anti-proliferatives, anti-inflammatories, and anti-atherosclerotics (Pastor et al. 2019).

Antioxidant activity in tea samples and other plant supplements has been determined by colorimetry (Everette et al. 2010), electrochemical sensing (Thangaraj et al. 2012), high performance liquid chromatography (HPLC) (Nishitani and Sagesaka 2004), and gas chromatography (GC) (Proestos and Komaitis 2013). Traditionally, colorimetric assays are based on cupric reducing antioxidant capacity (CUPRAC), oxygen radical absorbance capacity (ORAC), ferric reducing antioxidant power (FRAP), 2,2′-azino-bis(3-ethylbenzthiazoline-6-sulfonic acid) (ABTS), 2,2-diphenyl-1-picrylhydrazyl (DPPH), total radical-trapping antioxidant parameter (TRAP), and hydroxyl radical scavenging activity (Bener et al. 2010; Carocho and Ferreira 2013).

The quality control evaluation of Camellia sinensis samples from different botanical and geographical origins has been predominantly studied by chemometric approaches for authentication purposes. The most commonly used analytical techniques include UV–vis spectroscopy, Fourier-transform infrared (FTIR) spectroscopy, HPLC (Aboulwafa et al. 2019), elemental analyzer-isotope ratio mass spectrometry (EA-IR-MS), inductively coupled plasma optical emission spectrometry (ICP-OES), ICP-MS (Liu et al. 2020, 2021), ultra-HPLC–MS (Sánchez et al. 2020), near-IR and mid-IR spectroscopy (Zhu et al. 2019), and DNA-based technologies (Grazina et al. 2020). Although the above techniques have been effective for authentication of tea around the world, they are expensive and require trained personnel to be operated, in addition to time-consuming for analyzes by involving complicated sample pretreatments which generate hazardous waste.

Microfluidic paper-based analytical devices (µPADs) have become an attractive platform for on-site detection, especially for point-of-care (POC) testing (Martinez et al. 2007). They are able to perform qualitative and quantitative determinations of multiple analytes on a single strip. Its scope encompasses applications in several areas including environmental monitoring, clinical analysis, and food safety (Yamada et al. 2017). Compared to the techniques mentioned above, these devices provide various advantages such as wide availability, simple to fabricate and operate, measuring speed, good output, low sample and reagent requirement, low cost per analysis, portability, and low waste generation, which can be easily treatable by incineration. Besides being able to be functionalized via chemical coupling, the devices are compatible with a variety of detection methods, contributing to improve the sensitivity of the analytical measurements (Liu et al. 2015). Although several studies based on µPADs have been proposed to determine the antioxidant capacity of teas (Choleva et al. 2015; Piyanan et al. 2018; Puangbanlang et al. 2019; Hao et al. 2021), there is a lack of reports addressing the geographical origin authentication of tea samples through µPADs.

In view of the above, the aim of the present work was to evaluate compiled antioxidant activity data for the authentication of tea samples according to their geographical origin and thereby prove that the µPADs are reliable to perform this task. Principal component analysis (PCA) was first applied to the whole sample set for discovering clustering trends among classes and computing the principal components (PCs) that will be used as input variables for fitting linear discriminant analysis (LDA) modeling. Later, several validation methods were tested to find a suitable tea classification model.

Methods

Tea samples

Relevant data for this study were obtained by searching through ScienceDirect and Scopus regarding the methods based on microfluidic paper-based analytical devices (µPADs) for the analysis of tea samples. The criteria to build the database encompassed the following requirements: (i) tea samples with clearly identified country provenances; (ii) antioxidant activity expressed as gallic acid content; (iii) analytical data ought to be verified by conventional or reference methods. Table 1 presents the details of the 26 tea samples picked up from 3 articles published between 2015 and 2019. Samples come from two countries, i.e. Greece (Choleva et al. 2015) and Thailand (Piyanan et al. 2018; Puangbanlang et al. 2019), so these were the classes considered for the chemometric analysis.

Table 1.

Features of the tea samples picked up from the literature for geographical origin authentication purposes

No	Analytical Device	Principle	Detection technique	Comparative method	Provenance	References
1		CUPRAC(Cu2+/Cu1+)	Colorimetry	CUPRAC	Thailand	Puangbanlang et al. (2019)
2		Au sensor	Colorimetry	CUPRAC	Greece	Choleva et al. (2015)
3		Nanoceria(Ce4+/Ce3+)	Based on distance	CUPRAC	Thailand	Piyanan et al. (2018)

CUPRAC Cupric reducing antioxidant capacity assay

Devices

The microdevices were designed on white background filter paper and its dimensions were optimized in order to achieve a compromise among easy handling of the devices and minimum consumption of reagents and samples. The hydrophobic barriers were made using the wax-printing method (Carrilho et al. 2009). The details of the fabrication of these devices may be consulted from the original articles (Choleva et al. 2015; Piyanan et al. 2018; Puangbanlang et al. 2019).

Sample preparation and analysis

Tea samples were prepared according to the following procedures: (1) Samples from Choleva et al. (Choleva et al. 2015) were extracted in 5 mL methanol–water (20:80) under the ultrasound irradiation influence for 10 min; (2) Samples from Piyanan et al. (Piyanan et al. 2018) and Puangbanlang et al. (Puangbanlang et al. 2019) were boiled in 200 mL deionized water at 80 °C for 5 min. The extracted solutions were filtered by filter paper before analysis using the µPADs.

Three types of reactions onto the filter paper surface were used for antioxidant analysis: (i) analyte-mediated on-paper formation of gold nanoparticles (Choleva et al. 2015), (ii) partial reduction of cerium ion from Ce⁴⁺ to Ce³⁺ on nanoceria deposited along the detection chanel (Piyanan et al. 2018), and (iii) cupric reducing antioxidant capacity (CUPRAC) assay (Puangbanlang et al. 2019).

The determination of antioxidant activity in the tea samples was carried out using two detection techniques: (1) Colorimetric detection based on digital images (Puangbanlang et al. 2019; Choleva et al. 2015); (2) Detection based on distance (Piyanan et al. 2018). In the former, images of the devices were captured with commercial scanners, and effective intensity of color pattern was measured using an ImageJ software. In the latter, the color distance measurement was performed using a commmon ruler. The analytical response of both detection systems was directly proportional to the analyte concentration. Antioxidant activity content was employed to unveil the geographical origin of teas, and certify therefore their PDOs. Statistical evaluation of the compiled data was performed by two-way ANOVA test and regression analysis. The confidence interval was set at 95% in both cases.

Modeling

The database encompassing antioxidant activity profiles was first organized in a matrix format, where rows corresponded to 26 tea samples compiled from literature and its antioxidant capacities were contained in a column. Then, the data matrix was autoscaled and mean centered to equalize the variability ranges in prediction accuracies, which contributes to reduce misclassification rates.

Aiming at unveiling the geographical origin of tea samples produced in two countries (Greece and Thailand), principal component analysis (PCA) and linear discriminant analysis (LDA) were applied to compiled data matrix. PCA was used to visualize the natural distribution of samples in the space defined by the first two principal components (PCs) and guide in the classification approach to be applied (Bro and Smilde 2014). LDA served to find a suitable model allowing the correct identification of the provenance of the tea samples (Moncayo et al. 2015). To reach this goal, compiled sample set was randomly split up into two subsets (calibration and validation). The calibration set was used for building the classifiers, while the validation set served for testing the built models. In this sense, the methods of hold-out (14 samples for training and 12 samples for testing), external validation (16 samples for training and 10 samples for testing), and five-fold cross-validation (21 samples for training and 5 samples for testing) were applied. The splitting up into subsets was made in a stratified manner, so that the proportion of each class in the original matrix was kept in the new subsets. This scheme guaranteed the random sampling consistency.

The evaluation of classification modeling performance is an essential task in machine learning. Accuracy (ratio between all correct predictions and total number of examined cases), specificity (correct negative predictions divided by the number of negative cases), and sensitivity (correct positive predictions divided by the number of positive cases) were therefore calculated (Lantz 2015). Furthermore, the receiver operating characteristics (ROC) curve was also computed. This curve is an important tool for identifying classification issues at various threshold settings. It is defined as the fraction of true positives out of the positives (true positive rate) over the fraction of false positives out of the negatives (false positive rate). In this case, accuracy was calculated by integrating the area under the curve (AUC), which is one of the most common measures to evaluate predictive ability of the classifiers, in particular, binary ones. The higher AUC, the model will be better at predicting classes (Gigliarano et al. 2014). All calculations for classification meaningful interpretations were made using R-project software version 3.6.3.

Results and discussion

Statistical data management

After a detailed literature search, antioxidant activity profiles collected from tea samples that were analyzed by µPADs and conventional methods are summarized in Table 2, in which a comparison was made. Overall averages were 335.11 µmol g^–1 for µPADs and 261.97 µmol g^–1 for conventional assays. Considering the precision criterion, ANOVA test revealed no significant differences among the data compiled from both analytical approaches, because the calculated F value (F = 0.702) did not exceed the critical value (F = 4.013) at 95% confidence level (P = 0.405). Regarding accuracy, a high Pearson's coefficient (R = 0.9476) was obtained, indicating good correlation between the methods evaluated. These results enabled the data collected from different µPADs were analyzed by chemometrics for authentication purposes.

Table 2.

Statistical evaluation of the antioxidant activity data obtained from tea samples

Region	µPAD based methods		Conventional methods
	Average ± SD	Expressed as:	Average ± SD	Expressed as:
Greece	50.17 ± 72.17	µmol GA g–1	88.68 ± 42.63	µmol CA g–1
Thailand	536.25 ± 431.34	µmol GA g–1	384.28 ± 230.94	µmol GA g–1
ANOVA test
Calculated F value			0.702
Critical F value			4.013
p-value			0.405
Regression analysis
Slope ± SD			0.533 ± 0.035
Intercept ± SD			83.212 ± 18.118
Pearson's coefficient (R)			0.9476

SD Standard deviation, GA Gallic acid, CA Catechin

Prediction of the tea geographical origin

The compiled antioxidant activity data were analyzed by two pattern recognition tools (i.e. PCA and LDA) to explore clustering trends among the classes studied, in addition to proposing a reliable model for the prediction of the geographical origin of tea samples.

Exploratory data analysis

PCA was carried out for discovering the associations and patterns of the input data set. This step consisted of calculating a series of variables (PCs) which are able to maintain the original data information without losing its variance. The maximum information from the sample set is represented by the first PC and each successive PC expresses the cumulative variance not accounted by the previous PCs (Bro and Smilde 2014).

The score plot obtained from PCA is represented by Fig. 1. As can be seen, the extracted first two PCs explained 80.2% of total variance. In addition, 100% of data variability was reached by contributing of PC1–PC9. Samples from Greece exhibited PC1 negative values, whereas samples from Thailand showed predominantly PC2 positive scores. Despite the scattering of scores observed from Thailand samples, the grouping trend between teas of the same provenance is clearly noticeable.

Fig. 1.

Scores projected in the space PC1 vs PC2 obtained by analyzing the antioxidant activity data

Analysis of Fig. 1 indicates that the antioxidant capacity might be an effective predictor for classifying tea sample provenances. In this regard, a supervised method like linear discriminant analysis (LDA) could serve to authenticate the tea samples according to their geographical origin.

Classification approach

The geographical origin of tea samples was determined by carrying out binary classifications based on LDA. This algorithm maximizes the variance among classes and minimizes the variance within each class to differentiate groups of samples (Moncayo et al. 2015). A centroid was therefore created which represents the mean position of all points in all directions. Samples predicting was performed considering the minimal distance to the centroid of each class. Fig. S1 (Supplementary Material) shows samples projection in the space defined by the first two discriminant functions. As an outcome, we can observe a better distribution of the tea samples according to their provenance. LD2 showed mainly negative scores for samples from Greece, while the Thai ones exhibited indistinctly positive and negative values. The two distinct groupings revealed by the above chart indicate a good separation of tea samples coming from different countries, which facilitates geographical origin authentication.

Table 3 shows the results obtained for discrimination of the two classes of tea by evaluating different validation methods. For fitted models, the sensitivities were greater than 86%, whereas specificities were in the range 64–80%. In addition, the prediction accuracies in the test sample set ranged from 75 to 82%. The highest classification rate was obtained when applying an external validation method, which was attributed to model C. These results indicate that the Greece and Thailand samples can be differentiated against each other with a good success rate.

Table 3.

Classification performance by using different validation methods

Model	Validation method	Accuracy (%)	Specificity (%)	Sensitivity (%)
A	Cross-validation	75	64	90
B	Hold-out	79	73	88
C	External validation	82	80	86

Despite the aforementioned findings, the classification performance could be improved by increasing the number of input variables (Hastie et al. 2008). In this sense, the PCs calculated from PCA were used for fitting LDA modeling. Considering that the first nine PCs explained 100% of the data total variance, all these PCs were selected (PC1–PC9). To build the new classifier, the sample set with the new variables was again trained and tested using the above mentioned external validation method.

The modeling performance to predict the geographical origin of tea samples is presented in Table 4. Since binary classifications are based on the estimation of positive and negative classes, the negative prediction rate for Thailand samples was 20% higher than the positive prediction rate for Greece samples. This fact may be due to the lower number of observations considered for the latter group. It is common that the imbalance between groups of samples may affect the classification results. Nevertheless, this drawback was overcome because the detection made here prevailed equally for both classes. Consequently, the new fitted model was trained with a perfect success rate, indicating greater reliability on testing predictions. The correct classifications ranged from 75 to 100%, whereas misclassification rates were less than 25%. The results suggest that the discrimination of tea samples coming from Greece and Thailand may be accomplished with 91.7% balanced accuracy, 83% specificity, and 100% sensitivity. In addition, this model achieved an overall accuracy of 90%, which indicates that the classes evaluated can be predicted successfully.

Table 4.

Correct classifications obtained for the prediction of tea geographical origin

True origin	Prediction of the tea classes (%)
	Calibration		Validation
	Greece	Thailand	Greece	Thailand
Greece	100	0	75	25
Thailand	0	100	0	100
Positive class	Greece
Negative class	Thailand
Specificity (%)			83
Sensitivity (%)			100
Positive prediction value (%)			80
Negative prediction value (%)			100
Detection prevalence (%)			50
Balanced accuracy (%)			91.7
Overall accuracy (%)			90

The receiver operating characteristic (ROC) curve was also computed for assessing classification protocol. This probability curve was generated by plotting the true positive rate (sensitivity) versus the false positive rate (1–specificity) at several threshold settings (Fig. 2). Considering the integration of the area under the curve (AUC) is an effective measure of accuracy, this task was carried out. The probable values that this metric can take ranged from 0 to 1, and its interpretation could serve to predict the discriminative power of any classifier (Gigliarano et al. 2014). A worthless test is represented by an area equal or less than 0.5, while a perfect test is subject to an area of 1. The AUC (0.943) value obtained by fitting input data evidences a high degree of separation from tea classes, indicating good performance of LDA model. This result verifies the aforementioned findings, so the tea samples with different provenances can be correctly identified.

Fig. 2.

ROC curve obtained for fitting LDA model from the PCs

Chemometric processing of antioxidant activity profiles proved that the µPADs can be reliably used for tea authentication. Therefore, tea samples coming from Greece and Thailand could be successfully distinguished using the joint efforts of the PCA and LDA tools, which is relevant for certifying their PDO.

Conclusion

In this work, a tea database collected from existing literature was assessed by chemometrics for demonstrating the µPADs effectiveness in the authentication of teas according to their geographical origin. Antioxidant activity profiling turned out to be a suitable LDA predictor for identifying the provenance of the samples with overall accuracies ranging from 75 to 82% in the test sample set. The highest prediction rate was accomplished by applying an external validation method. The PCA-LDA approach allowed to find the best tea classification model. Consequently, the discrimination of the tea samples was achieved with 91.7% balanced accuracy, 100% sensitivity, and 83% specificity. The robustness of the classification protocol in cases of insufficient data was checked by means of the ROC curve (AUC = 0.943), indicating that the modeling fitted from PCs tolerates the lack of data. The combination of these tools proved that the µPADs could be reliably used for certifying the PDO of tea samples.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, Investigation, Methodology, Validation, Formal analysis, Visualization, Writing—original draft, Writing—review & editing MPR. Resources, Funding acquisition, Project administration, Supervision MPCM.

Funding

The authors have not disclosed any funding.

Data availability

The datasets generated and/or analysed during the current study is not publicly available due to privacy concerns but are available from the corresponding author upon reasonable request.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.