The present study was performed using an ASTREE II electronic tongue (Alpha M.O.S, Toulouse, France). The E-tongue consisted of a hexadecimal autosampler, a silver/silver chloride reference electrode, a data acquisition system, a workstation running AlphaSoft software (version 12; Alpha M.O.S, Toulouse, France) and seven sensors. The seven sensors were called ZZ2808-2-512 (ZZ), CA2804-2-440 (CA), DA2808-12-330 (DA), BA2808-2-230 (BA), GA2808-2-361 (GA), BB2011-09-141 (BB) and AB2011-10-010 (AB), and were specifically developed for measuring bitterness. The sensors contain two semiconductor regions composed of a thermal insulation material, each covered with a different molecular membrane, with different adsorption properties and the detection thresholds.

Samples were weighed using an electronic balance (±0.1 mg accuracy) at room temperature and dissolved completely in deionized water. Solutions of berberine hydrochloride (lot no. 101002; Sichuan Province Yuxin Pharmaceutical Co., Ltd., Deyang, China), rhynchophylline (lot no. 20100216; Hubei Tungshun Medicine Science & Technology Co., Ltd., Wuhan, China), leonurine (lot no. SX-091205; Xi'an Hao-Xuan Bio-Tech Co., Ltd., Xi'an, China), matrine (lot no. KS20110725; Xi'an Jiatian Biotechnology Co., Ltd.) and quinine (lot no. 20100510; Shanxi Tianyuan Pharmaceutical Co. Ltd, Yùnchéng, China) were prepared at concentrations of 0, 0.025, 0.1, 0.25, 0.5 and 1 mmol and stored at 4°C until required.

Ten-fold concentrations of 35 Chinese herbs were prepared, relative to the mean of the prescribed dosage in the Pharmacopoeia of the People's Republic of China. The herbs were placed in 2,000 ml water, soaked for 30 min and heated in a microwave (2100 W) until boiling. The power was then reduced to 600 W and the solution heated for a further 20 min. The remaining herb pieces were filtered out, an additional 2,000 ml of water was added, the solution heated until boiling and then boiled for 10 min. This process was repeated for each herb. Then, the filtrates of the first and second decoctions were combined, mixed and cooled to room temperature, followed by centrifugation at room temperature for 15 min at 1,434 × g. The supernatant was collected and the volume adjusted to 4,000 ml. Samples were aliquoted, capped, sterilized and stored at 4°C until required.

Purified water was used as the washing solution and 0.5 mmol caffeine solution was used as the sample. The E-tongue was cleaned with washing solution (6 times, 10 sec/wash). Then, 80 ml of sample solution was placed in a 120 ml beaker for analysis. Each E-tongue measurement used all 7 sensors, which was taken at room temperature and lasted 120 sec. The same sample was measured 10 times consecutively, with the values of last 4 measurements being used to calculate the relative standard deviation (RSD). In addition, RSD was measured in in 10 sec ranges (e.g. 0–9, 1–10 and 2–11 sec).

At the beginning of a measurement the signal is unstable, stabilizing as the number of measurements increase. To investigate how many replicates were required for the response signal to become stable, matrine and berberine hydrochloride samples (described previously) were measured using the E-tongue 10 times, with each measurement lasting for 120 sec. RSD values were calculated for all measurements and then compared.

Berberine hydrochloride (80 ml) was added to a 120 ml beaker for E-tongue measurements. The following two measurement schemes were then performed: i) The E-tongue was washed once in-between each measurement of the same sample; and ii) the same sample was measured without washing in-between. In this experiment, the same sample was measured seven times (120 sec each) and the results of last 4 measurements used to calculate the RSD.

Aqueous solutions of 0.5 mmol rhynchophylline, matrine, quinine and leonurine were prepared in triplicate (12 samples total). E-tongue measurements were conducted under the optimized conditions described above. In this experiment, the same solution was measured seven times (120 sec each) and the results of last 4 measurements used to calculate the RSD. The RSD of each sample prepared in triplicate was then compared to evaluate reproducibility.

The measurement time is <15 min for a single sample, however, a complete experiment requires the measurement of multiple samples. In order to ensure accuracy of the results, the measurement of all samples need to be finished within a measurement cycle (6 h). Aqueous solutions of quinine at 0.1, 0.5 and 1.0 mmol were prepared and measured using the E-tongue. Measurements were repeated four times. All measurements were completed within 6 h, and the results of last 4 measurements used to calculate the RSD.

Aqueous solutions of quinine at 0.025, 0.1, 0.25, 0.5 and 1.0 mmol were measured in triplicate on days 1, 2 and 3 to measure E-tongue data reproducibility on different days. In addition, the RSD of E-tongue measurements over the three days was calculated.

In the present study 35 Chinese herbs (lot no. 20110224; Henan Zhongyi Pharmaceutical Co., Ltd., Zhengzhou, China) were selected from the ‘Pharmacopoeia of the People's Republic of China’ (29). This text describes the TCMs used in the present study by the following characteristics: Light smell, tasteless; light smell, slightly bitter taste; light smell, bitter taste; and light smell, very strong bitter taste. The tastes described were evaluated using a human taster panel according to the degree of bitterness (30–32). TCM samples, prepared as described above, were measured using the E-tongue under the optimized conditions described above.

The results of E-tongue measurements and the human taste panel were evaluated using the robust regression analysis method (33,34). E-tongue measurements of 6 samples were excluded, using standardized residuals and score distance as indicators. These findings were in preparation for a publication elsewhere (35). The remaining 29 samples were grouped into two classes or three classes (Table I), based upon their bitterness levels (I–V) described by Liu et al (30). A two-dimensional bitterness classification model was established by considering level I as a class and grouping levels II, III, IV and V as the second class. The three-dimensional classification models consider level I as a class, levels II and II as a second class, and levels IV and V as a third class.

Least squares support vector machine (LS-SVM), simple support vector machine (SVM) or discriminant analysis (DA) classification algorithms were used for the classification models. MATLAB (release R2011b; Mathworks, Inc., Natick, MA, USA) and LS-SVMlab Toolbox software [version 1.8; www.esat.kuleuven.be/sisita/lssvmlab (accessed 23/01/16)] was used.

Using the LS-SVM method, the accuracy of the classification models compared with the results of the human taste panel measured, this was then used to select the most appropriate function for study, including linear kernels, polynomial kernels and radial basis functions. For each type of kernel, a self-compiled program screened and optimized the model parameters repeatedly, finally selecting the most appropriate kernel function and parameters. Modeling optimization was performed with SVM (36) and DA (37), with classification accuracy rates of cross-validation calculated separately.

PLS regression analysis was conducted on latent variables. Then, the projection scoring factors in principal component space were used to produce the two- and three-dimensional classification results.

The aim of the present study was to identify when E-tongue measurements become stable. This is important to assess the validity of any future measurements, as the signal should not be affected by noise. Firstly, sensor measurements of 0.5 mmol caffeine solution showed the characteristic response curve of sensors to the same solution (Fig. 1) and the signal became more stable over time. Fig. 2 shows the RSD values of the seven different sensors to 0.5 mmol caffeine solution. The RSD decreased over time and reached a minimum by 120 sec. So, 120 sec measurement times were used for all subsequent measurements.

Next, the present study investigated how many sample measurements were required to produce a stable response signal. For each test sample, 10 replicate measurements were taken (Fig. 3). RSD was identified to decrease as the number of measurements taken increased. The RSD from taking 4–7 measurements was not reduced further when >7 measurements were taken. For example, for the berberine hydrochloride sample, the RSD values of 7 repeats and 10 repeats were 1.89 and 3.05 fold that of 3 repeats (Fig. 3A). For matrine, the RSD values of 4–7 repeats and 7–10 repeats were 1.99 and 2.04 times higher, respectively, than that of 3–6 repeats (Fig. 3B). Therefore, the number of measurements taken of each sample was selected to be 7. In addition, this will minimize analysis time and extend the E-tongue lifetime.

To determine the best approach to measurements, two different schemes were tested on berberine hydrochloride solutions (Fig. 4). In scheme 1 each measurement of the same sample was followed by a single clean, whereas in scheme 2 measurements of the same sample were taken in a row. Scheme 2 was identified to be more stable because its RSD values were between 1.5 and 2 fold lower compared with those from scheme 1. The difference between the two schemes, at all respective concentrations, was significant (P<0.05; Fig. 4). In a practical sense, the single cleaning between each measurement used in scheme 1 required the E-tongue to switch back and forth between the sample and its corresponding washing cup. This was more time-consuming than the method used in scheme 2. Thus, scheme 2 was chosen for the remaining analysis.

The RSD values of the same sample in triplicate were typically <5% for the majority of sensors (Fig. 5). However, when the sample was matrine or quinine, the RSD values from sensors DA and GA were ≥11% (Fig. 5). The baseline responses of all seven sensors were normal and the E-tongue passed its self-checking protocol. Thus, this variation is likely due to the samples and not the E-tongue.

Quinine solutions at different concentrations were measured using the E-tongue within a 6 h measurement cycle (Fig. 6). The RSD within 6 h was <4%, indicating that the approach was suitable for sample analysis.

Quinine solutions of five different concentrations were measured using the E-tongue to evaluate the inter-day precision of measurements (Fig. 7). The results identified that measurements from sensors ZZ, BA, BB and AB were relatively stable, with an RSD variation over three days of <10%. Large variations in measurements were found for sensors CA, GA and DA, with RSD values varying by <37%. Therefore, subsequent measurements were taken on the same day and in the same 6 h measurement cycle.

In binary and tertiary classification performed with LS-SVM, the polynomial kernel was selected following optimization. For binary bitterness classification, all 29 samples were correctly classified (human taste panel results), with a cross-validation accuracy rate of 100%. For tertiary bitterness classification, 26 of the 29 samples were correctly classified, with a cross-validation accuracy rate of 89.66%. The misclassified samples were samples 8, 14 and 22.

In binary and tertiary classification using SVM, the polynomial kernel was selected following optimization. For binary classification, this approach correctly classified 28 out of 29 samples, with a cross-validation accuracy rate of 96.55%. Sample 14 was misclassified. For tertiary classification, 25 of the 29 samples were correctly classified with a cross-validation accuracy rate of 86.21%. The misclassified samples were sample 14, 17, 22 and 23.

In binary classification of the 29 samples using DA, 26 out of 29 samples were correctly classified, with a cross-validation accuracy rate of 89.66%. The misclassified samples were 14, 17, and 23. For tertiary classification using DA, 26 out of 29 samples were correctly classified, with a cross-validation accuracy rate of 89.66%. The misclassified samples were sample 14, 17, and 22.

The results of two-dimensional and three-dimensional bitterness classifications results of 29 samples based on PLS analysis are shown in Fig. 8A and B, respectively. The 29 samples could be grouped into two classes or three classes. However, samples 14 and 17 were on the border of the binary classification and samples 8, 14, 17, 22 and 23 were on the border of the tertiary classification, so could easily be misclassified.

In conclusion, the present study optimized the E-tongue measurement protocol to use a 120 sec measurement acquisition time, with 7 replicates and optimized the washing process. The optimized washing process comprised of 6 cleans of 10 sec each. The sample solution was then measured continuously, without cleaning in-between measurements of the same sample. Following completion of measurements, the response values of sensors were used for further analysis. This optimized method used had good reproducibility and high precision within 6 h, but poor inter-day precision. Therefore, measurements should be taken within the same 6 h measurement cycle, rather than on different days or in different measurement cycles.

The optimized protocol was then used to screen and identify the bitterness intensities of 29 TCM decoctions, which were compared to the bitterness intensities previously established by a human taste panel. The results showed that with appropriate data processing, the E-tongue could accurately identify the bitterness intensity of TCM decoctions.