Sensors & Actuators: B. Chemical 357 (2022) 131345 Contents lists available at ScienceDirect Sensors and Actuators: B. Chemical journal homepage: www.elsevier.com/locate/snb Resolution of opiate illicit drugs signals in the presence of some cutting agents with use of a voltammetric sensor array and machine learning strategies Dionisia Ortiz-Aguayo a, Xavier Cetó a, Karolien De Wael b, Manel del Valle a, * a Sensors and Biosensors Group, Department of Chemistry, Universitat Autònoma de Barcelona, Edifici Cn, 08193 Bellaterra, Spain b AXES Research Group, Department of Bioscience Engineering, University of Antwerp, Groenenborgerlaan 171, 2020 Antwerp, Belgium A R T I C L E I N F O A B S T R A C T Keywords: In the present work, the resolution and quantification of mixtures of different opiate compounds in the presence Electronic tongue of common cutting agents using an electronic tongue (ET) is evaluated. More specifically, ternary mixtures of Voltammetric sensors heroin, morphine and codeine were resolved in the presence of caffeine and paracetamol. To this aim, an array of Opioids three carbon screen-printed electrodes were modified with different ink-like solutions of graphite, cobalt (II) Partial-least squares regression phthalocyanine and palladium, and their responses towards the different drugs were characterized by means of square wave voltammetry (SWV). Developed sensors showed a good performance with good linearity at the µM level, LODs between 1.8 and 5.3 µM for the 3 actual drugs, and relative standard deviation (RSD) ca. 2% for over 50 consecutive measurements. Next, a quantitative model that allowed the identification and quantification of the individual substances from the overlapped voltammograms was built using partial least squares regression (PLS) as the modeling tool. With this approach, quantification of the different drugs was achieved at the μM level, with a total normalized root mean square error (NRMSE) of 0.084 for the test subset. 1. Introduction mixtures is investigated herein as a proof-of-concept of what can be achieved. The consumption and trafficking of illicit drugs have increased In the USA, the Controlled Substances Act (CSA) defines five classes significantly over the last years, creating a negative impact in people’s of drugs: narcotics, depressants, stimulants, hallucinogens and anabolic health and in the economy, while contributing to an increase of crimi­ steroids [4]. Among those, narcotics (also known as “opioids”) represent nality [1]. The increase in consumption and trafficking has also enriched one of the biggest health and economic burden, accounting 63% of the illicit drug markets, which are powerful systems of production and deaths by drug overdose in the USA in 2015 and nearly 70% in 2018 [5, distribution that generate large amount of unwanted activities [2]. In 6]. A particular class of opioids are opiates, which originate or are this direction, the rapid detection of illicit drugs to disarticulate such derived from naturally occurring alkaloids found in certain poppy spe­ markets and safeguard the public still remains a challenge for cies, specifically Papaver somniferum [7]. Common opiates include authorities. opium, heroin, morphine and codeine. The main drawbacks posed by currently used on-site methods for the Several alternate methods for the individual and simultaneous detection of illicit drugs and their precursors are the low accuracy of determination of opiates have been reported in the literature. Those are color tests, or the high cost and low portability of spectroscopic tests. In based on common analytical techniques such as chromatography [8,9], the light of the pressing need for better drug test systems at border capillary electrophoresis [10,11], chemiluminescence [12,13], diffuse controls, BorderSens project [3] aims to establish the basis for the reflectance near-infrared spectroscopy [14] or surface plasmon reso­ development of a portable device capable to quickly test for different nance (SPR) based immunosensors [15]. Despite being powerful, these drugs, precursors and cutting agents, with outstanding accuracy and techniques present some disadvantages that hinder their application for reduced false positives and false negatives. However, given the chal­ the on-site drug monitoring, like requiring a sample pre-treatment step lenge that such a task represents, the quantitative analysis of opiates and/or laboratory facilities (low portability), being time-consuming and * Corresponding author. E-mail address: manel.delvalle@uab.cat (M. del Valle). https://doi.org/10.1016/j.snb.2021.131345 Received 30 July 2021; Received in revised form 13 December 2021; Accepted 29 December 2021 Available online 31 December 2021 0925-4005/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 Fig. 1. Scheme of the experimental procedure for the electrode surface modification. Firstly, an ink-like solution was prepared incorporating the corresponding modifier. Then, 1 µL was dropped on the surface and dried at 40 ◦ C. quite costly (both from the equipment and reagents side). from the measured voltammograms. In this regard, the use of electrochemical sensors represents a promising alternative which may allow to overcome these limitations. 2. Experimental These devices can be used as powerful analytical tools for the analysis of illicit substances from street and biological samples as they offer low- 2.1. Reagents and samples cost measurement systems with rapid response, simple usage and high portability; all of them required characteristics for point-of-use forensic Standards of heroin hydrochloride and codeine were purchased from applications. Despite the advantages that electrochemical methods may Chiron Chemicals, Australia. Morphine hydrochloride, potassium offer, the simultaneous determination of the aforementioned com­ monophosphate, potassium chloride and potassium hydroxide were pounds can be challenging given their similar electrochemical response purchased from Sigma-Aldrich (St. Louis, MO, USA). and the complexity of the samples [16]. Precisely, the main drawback is Cobalt(II) phthalocyanine (CoPc) and palladium powder (< 1 µm, ≥ that, when using bare electrodes, there might be some overlapping be­ 99.9%; Pd), which were used as modifiers, as well as mesitylene and tween peaks, or even peak suppression [17]. Actually, the most polystyrene, which were used for the preparation of the ink composite, complicated situation is the discrimination between heroin and were also obtained from Sigma-Aldrich (St. Louis, MO, USA). Graphite morphine, since both molecules present the same functional groups in powder (particle size < 50 µm) was received from BDH (BDH Laboratory their skeleton and consequently, their corresponding fingerprint is quite Supplies, Poole, UK). similar. To overcome this difficulty, modification of the electrochemical Samples were prepared in 20 mM phosphate buffer (PBS) at pH 7.0 sensors may be required, but it might also be necessary to couple these containing 100 mM KCl as supporting electrolyte for the electrochemical techniques with other strategies such as the use of chemometrics; a measurements. All aqueous solutions were prepared using MilliQ water combination that is known as electronic tongue (ET) [18,19]. (ρ > 18.2 MΩ cm). All reagents were of analytical grade and used According to the IUPAC [20], an ET is defined as “a multisensor without further purification. Fresh stock solutions were prepared daily system, which consists of a number of low selective sensors and uses in order to prevent its degradation. advanced mathematical procedures for signal processing based on pattern recognition and/or multivariate data analysis”. These bio­ 2.2. Apparatus and voltammetric measurements mimetic systems, in opposition to classical approaches, are based on the combination of low selective and/or cross-responsive sensors to obtain SWV measurements were performed using a multi-channel poten­ rich and complementary analytical information. Next, the coupling with tiostat/galvanostat/impedance analyzer MultiPalmSens4 (PalmSens, chemometric tools to analyse the data allows to deconvolute complex Houten, The Netherlands) controlled by MultiTrace software. ItalSens overlapping electrochemical responses and achieve the simultaneous screen-printed carbon electrodes (SPCE) containing a graphite working quantitative determination of several analytes. Thus, such an approach electrode (3 mm diameter), a carbon counter electrode and a pseudo- allows to extract meaningful chemical information from these complex silver reference electrode (PalmSens, The Netherlands) were used for data. the measurements. SWV measurements were performed by placing 50 In the present work, the capabilities of ET-based systems in forensic µL of the sample onto the SPCE. The single scan SWV parameters were as applications will be demonstrated by attempting the simultaneous follows: potential range from − 0.2 to 1.5 V, step potential of 5 mV, determination of three opiates (heroin, morphine and codeine) in the amplitude of 25 mV and frequency of 10 Hz. presence of two common cutting agents (paracetamol and caffeine). The chosen voltammetric sensor array consisted of three screen-printed 2.3. Modification of the electrode surface carbon electrodes (SPCE) modified with graphite, cobalt (II) phthalo­ cyanine and palladium inks, and square wave voltammetry (SWV) was SPCE were modified with standard catalysts employed in electro­ the measuring technique. Firstly, the behavior of the sensors towards analysis, as are CoPc and Pd [21,22], incorporated using a each of the compounds was evaluated individually, characterizing its self-formulated graphite-polystyrene ink. More in detail, the mixture analytical response . Secondly, a partial least square regression (PLS) contained the following mass fractions: 58% of graphite, 32% of model for their simultaneous quantification at the µM level was built powdered polystyrene and 10% of modifier, in this case graphite, cobalt 2 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 (II) phthalocyanine or palladium. subset were prepared based on a tilted factorial experimental design 33 The corresponding modifier, graphite and polystyrene were thor­ (27 samples) [23]. With this approach it is possible to get a better dis­ oughly mixed with 250 µL of mesitylene for 2 h. After that, the mixture tribution of the samples that avoids the repetition of concentrations as was sonicated for 2 min in order to obtain a medium thick solution. per the selected levels. The concentrations ranges considered for each of Finally, 1 µL of the ink-like composite was dropped onto the working the compounds were in the range 0–750 µM, with 15 extra samples electrode surface of a SPCE and dried at 40 ◦ C for at least 1 h in order to forming the test subset. remove the solvent (Fig. 1). After that, the electrodes were ready to use, In the second case (mixtures of the three drugs plus the two cutting not requiring the usage of any activation step. agents), samples for the train subset were prepared based on a central composite face-centered (CCF) experimental design with 3 levels of concentration (27 samples). As already stated, this was preferred as the 2.4. Characterization of the electrode by scanning electron microscopy number of samples required to complete a full factorial design would be just too high (35 = 243 samples). In this case the concentrations for each The morphological characterization of the modified SPCE electrodes of the compounds ranged from 0 to 750 µM, with 17 extra samples was performed by Field Emission Gun-Scanning Electron Microscope forming the test subset. (FEG-SEM) of Zeiss, model MERLIN SM0087 and Energy Dispersive X- As an extra precaution to control drifts or periodic trends as the same Ray Analysis (EDX). Imaging was performed based on secondary, back- sensing units were used for the whole series, all this samples were scattered electrons. analysed in random order and alternating their measurement with the measurement of a blank solution (PBS), which served as cleaning stage 2.5. Samples under study of the electrode surfaces, but also as control. In the present work, two different scenarios were evaluated with the 2.6. Data analysis proposed ET. Firstly, the analysis of ternary mixtures of heroin, codeine and morphine was considered to assess the potential of the ET to achieve Initial analysis of the voltammetric signals was done with MultiTrace their individual quantification. Secondly, the quantification of the same software (PalmSens, Houten, The Netherlands), which allowed to drugs in the presence of two common cutting agents (viz. paracetamol calculate the peak heights and areas for the different stock samples. and caffeine) was attempted to assess the potential of the ET in a more From those, the calibration plots were build using the data of the realistic scenario. replicate measurements (n = 4) with the aid of Sigmaplot (Systat Soft­ To this aim, two different sets of samples were prepared (one for each ware Inc., San Jose, CA, USA), and analytical parameters such as of the above-mentioned scenarios). Each set of samples consisted of a sensitivity, linear range, LOD, etc. were calculated. train subset, which is used to build the model, and a test subset, which is Chemometric analysis was done in Matlab R2018b (MathWorks, used to assess its performance. The concentrations of the compounds Natick, MA, USA), making use of its Statistics and Machine Learning mixtures of the train subset were defined by an experimental design, Toolbox, by specific routines written by the authors. Briefly, upon while for the test subset, some extra samples with concentrations for measurement of the set of samples described in Section 2.5 with the each of the compounds randomly distributed along the experimental sensor array, voltammetric responses of the three electrodes were domain were also prepared. Besides, in order to keep the number of combined into a single vector. Genetic algorithms (GAs) were then used samples required down to a reasonable level, two different experimental as feature selection tool to reduce the number of inputs to be fed to the designs were employed as the number of samples required increases chemometric model given the large dimensionality of the voltammetric exponentially with the number of compounds considered. data [24]. Next, quantitative models to individually quantify each of the In the first case (mixtures of the three drugs), samples for the train analytes were built by partial least squares regression (PLS-1) [25], and their performance was then assessed towards the samples that formed the test subset to obtain more realistic performance indicators. 3. Results and discussion 3.1. Characterization of the sensor array Modified electrodes can be prepared by several different techniques [23,26]. In the present work, the approach used for the modification of the electrodes is based on the use of a composite material (containing the different modifiers) through the formation of an ink-like paste, which is then casted onto the electrode, generating a new surface highly suitable to carry out electrochemical measurements. Besides, the preparation and modification with these inks is extremely easy and has a very low-cost, making this methodology an interesting approach for the fabrication of chemically-modified transducers. In our case, after an initial screening of different electrochemical modifiers considered in prior studies involving ETs [27–31], an array of three electrodes was prepared using bare graphite, Pd and CoPc as the modifiers. The selection was based on the suitability of the different electrodes to obtain the discrimination of the different drugs (data not shown) [32]. The inclusion of graphite provided somehow a control point that allowed to actually evaluate the advantages derived from the incorporation of the other modifiers. Pd is well-known for its good Fig. 2. Electrochemical fingerprint of 300 µM solutions of the five substances (electro)catalytic activity towards a wide range of reactions and com­ under study: 1) caffeine, 2) codeine, 3) heroin, 4) morphine and 5) paracetamol pounds, while the usage of nanoparticles has demonstrated to be an with the proposed array in this work: Carbon (red), CoPc (black) and Pd (blue). attractive alternative to the respective bulk metals given its higher 3 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 layers (Fig. S1, supplementary material), and more importantly, confirm the presence of the modifiers inside the ink-like composite. Next, the evaluation of the voltammetric responses of each of the modified sensors towards each of the compounds individually was car­ ried out to assure that distinguished signals are generated . That is, to ensure that the electrodes respond to the different analytes, and that differentiated responses are also obtained between them. To carry out the measurements, SWV was chosen given its high sensitivity and fast scan rates, which in combination with the compact low-power instru­ mentation required for electrochemical measurements, offers particular Fig. 3. Chemical structure of (A) heroin, (B) morphine and (C) 6-MAM (a promise for decentralized security screening applications [33,34]. heroin metabolite). As can be observed in Fig. 2, different overlapping peaks can be remarked. In the case of heroin, an irreversible oxidation split peak is surface/mass ratio and improved electrochemical properties. Lastly, shown around 0.97 V, corresponding to the oxidation of the tertiary phthalocyanines are reported to be efficient electrocatalysts in the amine group (Fig. 3A), resulting in a secondary amine which is then determination of many important inorganic, organic, or biological further oxidized (Figs. 3 and S2) [35]. In addition, a smaller extra peak compounds. appeared at a lower potential (ca. 0.40 V, which is more evident in Before tackling the resolution of the mixtures with their corre­ Fig. 4), corresponding to the oxidation of the phenol group of 6-mono­ sponding cutting agents, which is the main goal of this work, the acetylmorphine (6-MAM, Fig. 3C) present in a 3% w/t in the sample. modified electrodes were first physically and electrochemically 6-MAM is an impurity, commonly found in heroin synthesis, that comes characterized. both from the incomplete acetylation of morphine as well as from the Upon modification of the SPCE as described in Section 2.3, those hydrolysis of heroin (as it is a product of its hydrolysis). were characterized by SEM imaging. Microscopy studies show that the Similarly, morphine also shows the oxidation peak ca. 0.40 V two modifiers are distributed quasi-homogeneously along the graphite Fig. 4. Square wave voltammograms obtained for (A-C; top row) heroin, (D-F; middle row) morphine and (G-I; bottom row) codeine using (A,D,G; left column) graphite, (B,E,H; middle column) CoPc and (C,F,I; right column) Pd, respectively. Series of plots correspond to increasing concentrations from 10 to 1000 µM. Insets correspond to the linear regressions of peak height (at the observed potential maximum) vs. concentration, excluding the point 1000 µM as saturation of the vol­ tammetric signal was reached for certain compounds. 4 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 Table 1 Calibration data (y vs. x) for the separate determination of heroin, morphine, codeine, paracetamol and caffeine employing the proposed sensor array. Compound Potential (V) Sensitivity (nA µM-1) Intercept (μM) r LOD (µM) Linear Range (µM) Sensor 1: Carbon Heroin 0.49 6.1 0.28 0.986 3.33 25–200 2.5 1.13 0.985 200–750 1.16 4.7 0.21 0.997 31.8 25–750 Morphine 0.43 3.3 0.10 0.996 5.33 25–200 1.3 0.51 0.997 200–750 0.99 4.2 0.16 0.995 8.65 25–200 1.4 0.76 0.994 200–750 Codeine 1.11 21.7 0.72 0.998 1.80 25–750 Paracetamol 0.39 22.3 0.12 0.999 0.82 25–750 Caffeine 1.33 15.6 4.02 0.993 44.0 50–750 Sensor 2: CoPc Heroin 0.43 7.2 0.15 0.999 3.95 25–200 3.1 1.03 0.996 200–750 1.14 5.5 0.22 0.994 83.3 25–750 Morphine 0.40 2.8 0.071 0.994 2.88 25–200 1.5 0.32 0.995 200–750 1.11 3.4 0.11 0.994 96.6 25–200 1.5 0.53 0.991 200–750 Codeine 1.09 18.9 0.21 0.999 4.29 25–750 Paracetamol 0.40 17.4 -0.16 0.999 0.75 25–750 Caffeine 1.36 10.8 0.0069 0.998 65.0 50–750 Sensor 3: Pd Heroin 0.27 14.3 0.41 0.988 5.31 25–200 3.5 1.59 0.991 200–750 1.01 34.8 0.12 0.975 14.8 25–200 4.2 1.57 0.966 200–750 Morphine 0.46 3.2 0.086 0.985 25.9 25–200 1.5 0.44 0.993 200–750 1.01 6.9 0.62 0.954 60.7 25–200 1.7 1.68 0.961 200–750 Codeine 1.11 19.8 2.31 0.997 11.7 25–750 Paracetamol 0.38 19.0 -0.11 0.999 3.33 25–750 Caffeine 1.35 11.6 17.2 0.983 104 200–750 corresponding to the phenol group of 6-MAM [36], but being more only to evaluate the response of the sensors, but also to determine the prominent in this case. The second peak corresponds to the oxidation of concentration ranges which they can operate and that will be used to do the tertiary amine group (Fig. 3B), which in this case is not further the analysis with the electronic tongue approach. As can be observed in oxidized (Fig. S3). In the case of codeine, only one broad peak can be Fig. 4s and S7, the responses obtained were linear for the vast majority observed, which again corresponds to the oxidation of the tertiary amine of cases in the studied ranges (r > 0.99), showing one peak or two (Fig. S4), but also showing a small shoulder, which is almost super­ depending on the compound evaluated. imposed and that is attributed to the oxidation of the 6-hydroxy groups In the case of heroin, two peaks were shown. C and CoPc modified [37]. Finally, in the case of paracetamol, only a well-defined Gaussian inks presented two linear ranges for the peak 1 (black), corresponding to peak corresponding to the oxidation of the amide group (Fig. S5) is the oxidation of the phenol group of 6-MAM. The lower linear range observed ca. 0.40 V [38,39], while in the case of caffeine, also a single goes from 25 to 200 µM and the high range from 200 to 750 µM. For peak (Fig. S6) is obtained, which corresponds to the oxidation of the C-8 peak 2 (red), one linear range is presented (from 25 to 750 µM), corre­ to N-9 bond to give the substituted uric acid [39,40]. In this direction, sponding to the oxidation of the amine group (Fig. 4AB). Pd modified ink the proposed mechanisms for the electrochemical oxidation of the showed a different performance displaying two linear ranges for peak 1 different evaluated substances are provided in Figs. S7 and Figs. S2 to and peak 2 (Fig. 4C). S6. The next compound analysed was morphine. In this occasion, a It can also be seen that the oxidation of the phenol group is over­ similar response was given for the three sensors of the array. Two peaks lapped with the oxidation peak of paracetamol. Similarly, the same were observed, with two linear ranges for each of them (Fig. 4D–F). The situation occurs with the second oxidation peak of heroin and morphine, low range goes from 25 to 200 µM and the high from 200 to 750 µM, which also will overlap with the one from codeine. However, this is not similar to heroin as could be expected. an issue given distinguishable voltammetric profiles are obtained; i.e. Codeine and paracetamol showed a single peak with good linearity different peak shapes and sensitivities are still obtained for each of the over the whole range (from 25 to 750 µM) with the three modifiers compounds. Lastly, in the case of caffeine, a single oxidation peak at tested (Fig. 4G–I and Fig. S7J–L). Lastly also a single peak is obtained for higher potential is obtained (ca. 1.33 V). caffeine, but with different linear range based on the electrode consid­ After this first general overview of the voltammetric responses, the ered. Employing C and CoP, the linear range goes from 50 to 750 µM, calibration curves for the five compounds under study were constructed whereas with Pd modified ink, the linear range narrows from by measuring solutions of increasing concentration from 25 to 750 µM. 200–750 µM (Fig. S7M–O). This step is relevant for further quantification models as it is important Based on the previous results, for the multi-determination of the to identify the proper working ranges. All of the electrochemical mea­ drugs mixtures, the concentration working ranges were streamlined surements were performed in PBS at pH 7.0. The selection of this neutral from 0 to 750 µM for heroin, morphine, codeine, paracetamol and pH is due to the fact that heroin and morphine suffer hydrolysis re­ caffeine. The analytical parameters derived from the calibration curves actions at basic pH, as is described in the literature [41,42]. for each sensor are summarized in Table 1. In all cases, the peak height which corresponds to the maximum of the oxidation signal was taken. This characterization is essential not 5 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 Fig. 5. Modeling ability of the optimized GA-PLS model for the 3 compounds case. Comparison graphs of obtained vs. expected concentrations for (A) heroin, (B) morphine and (C) codeine, for both the train (•, solid line) and test subsets (○, dotted line). The dashed line corresponds to the ideal comparison line (y = x). 3.2. Repeatability and reproducibility studies In the previous section, the sensitivities of each of the electrodes towards each of the compounds have been shown different (Table 1), a Upon completing the calibration, a stability study was done in order situation highly desirable when developing an ET application. Thus, the to demonstrate that the sensors were capable to withstand the large next step prior to build the quantitative model that allows to dete­ number of measurements necessary when developing ET applications. termine the individual substances from the overlapped voltammograms To this aim, codeine was selected as the substance to evaluate the was the selection of the chemometric tool to be used. In this case, given variation on its voltammetric response upon successive measurements, the ultimate goal of this research project is developinga device to detect assuming a similar behavior for the rest of the drugs. More specifically, a different illicit drugs, PLS-1 was chosen as the modeling tool given it is stock solution of codeine of 450 µM was measured for 25 consecutive one of the simplest (e.g. in comparison to artificial neural networks, times employing the same electrodes and measuring also a blank (PBS ANNs) and widely used multivariate calibration techniques to choose solution) in between each measurement to evaluate the repeatability of [44]. the sensors. Thus, each sensor was used for 50 consecutive measure­ Lastly, although the use of a pre-processing stage to reduce the ments; a significant number given the disposable nature of SPEs. number of input variables is not required when PLS is being used, it has Voltammetric data were collected and analyzed, from which the demonstrated that even in such cases this data reduction stage improves relative standard deviation (% RSD) was calculated. The values for the model’s prediction and generalization ability [43]. Again, as the aim carbon, CoPc and Pd modified ink-sensors were 1.84%, 2.01% and is to develop the simplest model possible, the use of GAs as feature se­ 2.14%, respectively. Significantly, the three modified sensors presented lection tool was chosen given upon identification of the most relevant better stability than the unmodified sensor, which showed a RSD value features, no further computing processing will be required for each new of 9.87%. Thus, from this study it was concluded that no fouling or drift measurement that is being performed. Thus, in this manner, the most effects are observed with the proposed sensor array. This, in essence, relevant features from each of the voltammograms were selected with means that any possible adsorption of the oxidized forms of the studied the aid of GAs and used as input into the PLS model. The outcome of GAs compounds did not affect appreciably the electrodes’ performance or optimization is shown in Fig. S10, where the raw voltammetric re­ stability, given no differences were observed along consecutive mea­ sponses for 300 µM solutions of each of the considered compounds is surements; equivalently, the same could be said for the applied poten­ plotted with cross marks underneath corresponding to the selected tials, if there is any doubt on the relatively high values used. features. On the contrary to the straighforward idea that the algorithm Complementary to the previous study, the reproducibility of con­ will select essentially the points corresponding to the peaks’ maxima, it struction of the ink-modified SPCE was also assessed. The experiment finally uses regions with less overlap and where the differences between was done preparing each modified ink by triplicate (n = 3) and signals are more pronounced. (that is, the points corresponding to the measuring consecutively with a heroin stock solution. The results for front and back of the peaks). each sensor present a good construction reproducibility with RSD values As already stated in Section 2.5, two different sets of samples were of 3.97%, 6.95% and 5.67% for carbon, CoPc and Pd inks, respectively. prepared: the first one in which ternary mixtures of the three considered drugs were considered, while in the second one also the presence of two different cutting agents was examined. The aim of the first set was to 3.3. Quantitative analysis of drug mixtures using PLS regression confirm the potential of the proposed ET to carry out the individual quantification of the opiates, whereas the second one aims to confirm Despite different voltammetric profiles are obtained for each of the that the ET is able to counterbalance the interferences of the cutting compounds when analysed individually (Fig. 4), even with different agents and successfully carry out the quantification of the drugs. In both response for the different considered electrodes, it is clear that there will cases, GA-PLS models were built using the data for the train subset, and be an overlap on the voltammetric responses when mixtures of those are its performance assessed towards the samples of the test subset, selecting to be analysed at pH 7 (Figs. S8 and S9). Thus, in order to achieve the the number of latent variables (LVs) that lead to the lowest root mean individual quantification of each of the compounds, the use of chemo­ square error (RMSE). metric methods is required as such quantification cannot be achieved via univariate regression (taking either the peak height or area). In this 3.3.1. Mixtures of the three drugs direction, ET approach relies on the combination of an array of sensors From the raw voltammetric responses of the three electrodes (318 that show complimentary responses towards the compounds of interest, current values x 3 sensors), a total of 103 features were selected by use of with a multivariate calibration method that allows to build a model that GAs and used to build the PLS-1 models (Fig. S10). The selected number relates the responses of the different sensors with the concentration of of LVs were 7 for heroin, 15 for codeine and 8 for morphine. Next, the each of these compounds [19,43]. 6 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 Table 2 Fitted regression lines for the comparison between obtained vs. expected values for the different sets of samples and the three considered APIs. Compound Slope Intercept (μM) r RMSE (μM) NRMSE Total NRMSE train subset (n = 27) Heroin 0.998 ± 0.019 0.7 ± 7.5 0.999 9.39 0.014 0.014 Codeine 1.000 ± 0.007 0.1 ± 2.7 1.000 3.35 0.005 Morphine 0.996 ± 0.026 1.3 ± 10.0 0.998 12.5 0.019 test subset (n = 15) Heroin 1.035 ± 0.064 -22.5 ± 24.3 0.995 20.6 0.030 0.026 Codeine 1.020 ± 0.026 -5.1 ± 9.4 0.999 10.1 0.016 Morphine 1.026 ± 0.030 -11.3 ± 11.5 0.999 11.1 0.016 Intervals are calculated at the 95% confidence level. RMSE: root mean square error; NRMSE: normalized root mean square error. Fig. 6. Modeling ability of the optimized GA-PLS model for the 5 Compounds case. Comparison graphs of obtained vs. expected concentrations for (A) heroin, (B) morphine, (C) codeine, (D) paracetamol and (E) caffeine, for both the train (•, solid line) and test subsets (○, dotted line). The dashed line corresponds to the ideal comparison line (y = x). Table 3 Fitted regression lines for the comparison between obtained vs. expected values for the different sets of samples and the five considered compounds. Compound Slope Intercept (μM) r RMSE (μM) NRMSE Total NRMSE train subset (n = 27) Heroin 0.972 ± 0.068 8.7 ± 27.3 0.986 43.5 0.061 0.053 Codeine 0.990 ± 0.041 3.1 ± 16.6 0.995 26.2 0.037 Morphine 0.990 ± 0.040 3.0 ± 16.2 0.995 25.5 0.036 Paracetamol 0.969 ± 0.071 9.7 ± 28.8 0.984 45.9 0.065 Caffeine 0.976 ± 0.063 7.5 ± 25.5 0.987 40.6 0.057 test subset (n = 17) Heroin 0.968 ± 0.169 -0.7 ± 69.3 0.953 76.3 0.107 0.077 Codeine 0.974 ± 0.169 19.2 ± 58.0 0.954 63.3 0.089 Morphine 1.042 ± 0.087 -5.9 ± 28.5 0.989 30.1 0.043 Paracetamol 0.997 ± 0.124 -8.4 ± 38.8 0.975 41.8 0.059 Caffeine 0.981 ± 0.130 -4.5 ± 39.7 0.972 50.1 0.070 Intervals are calculated at the 95% confidence level. RMSE: root mean square error; NRMSE: normalized root mean square error. 7 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 Fig. 7. Histogram comparing the success of PLS models with raw data (red) and GAs-PLS (green) to that of shuffled models (1000 iterations) for (A) codeine and (B) paracetamol, for the set of samples corresponding to mixtures of the 5 compounds. For the shuffled models, the data was fitted to a 3 parameter Gaussian curve and the RMSE values compared to the former. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.) comparison graphs of predicted vs. expected concentrations were built In regards to the latter, herein we have demonstrated how making (Fig. 5) and the linear regressions of the comparison lines fitted (Table 2) use of a proper set of samples, the model is able to correctly quantify for the three drugs. As can be seen from the plot, a satisfactory trend is both the drugs and cutting agents. Thus, the same approach can be obtained in all the cases, with regression lines close to the ideal one applied to the identification and quantification of others mixtures. (y = x). Moreover, from the regression parameters in the table we see Moreover, if multi-way processing methods are to be used instead of these are very close to the ideal ones (i.e. 1 slope and correlation coef­ two-way PLS models, the correction of the presence of an interfering ficient, and 0 intercept), being all of them within the intervals calculated species, even if not initially considered in the response model, might be at the 95% confidence level. Fig. S11 displays how obtained values are possible thanks to their “second order advantage” [45,46]. within the expected joint confidence intervals for all cases. Therefore, taking all this into account, it’s clear that the use of the Despite the satisfactory trend, it has to be reckoned that slightly herein proposed voltammetric ET shows huge potential to carry the better performance is obtained for the train subset, but this is due to the identification and quantification of seized drug samples, either those fact that the data of the test subset is not used at all during the modeling being pure or already mixed with other drugs and/or cutting agents. stage, and thus it provides a more realistic metric of the model perfor­ Lastly, despite already taking the precaution of using a separate mance. However, if comparing the RMSE values obtained for both validation subset of data (the test subset), a permutation test or “target subsets, it can be seen how the differences are actually not that high, shuffling process” was carried out to demonstrate that neither the high which confirms the goodness of the model and the capability of the dimensionality of the data nor the use of GAs and PLS-1 is resulting in proposed ET to achieve the simultaneous quantification of the three over-fitted models. Such test allows the identification of incorrectly drugs. Thus, the next step was to confirm whether using the same perceived cause-and-effect relationships in modeling (“chance correla­ approach we are able not to only quantify mixtures of the pure drugs, tion”) by taking as null hypothesis that samples labels are exchangeable. but also to detect and quantify the cutting agents considered, which Briefly, this test involves repeatedly and randomly reordering of the would provide a more realistic use of the developed strategy. responses variables (Y), followed by the building of a new model upon shuffling of the data labels. In other words, a new model is built upon 3.3.2. Mixtures of the three drugs and the two cutting agents assignment of an “incorrect” y-value to each sample corresponding to Under the same conditions as above, but making use of a different the one from another sample. This process is repeated several times to experimental design given the larger number of analytes to be consid­ ensure that the statistics calculated are significant (up to 1000 times in ered, a new set of samples was prepared as described in Section 2.5 and our case). For each of the permutations, the different performance measured employing the sensor array. As before, GA-PLS models were metrics were calculated and compared to the actual model with the built to determine each analyte in the mixture, but in this case selecting proper labels. As an example, a histogram summarizing the RMSE values a total of 114 features from the raw voltammetric responses (Fig. S10). of the different models for the three drugs plus codeine and paracetamol The selected number of LVs were 13 for heroin, 18 for codeine, 10 for are shown in Fig. 7, from which the significance of the obtained results is morphine, 6 for paracetamol and 12 for caffeine. Again, the comparison evident. graphs of predicted vs. expected concentrations were built for each of the analytes (Fig. 6), and the linear regression parameters calculated 4. Conclusions (Table 3). Again, a satisfactory trend is obtained for both subsets, with slightly better behavior for the train subset as already discussed, but The potential of ETs for the simultaneous determination and quan­ with RMSE values of the same order of magnitude. tification of different opiates in the presence of common cutting agents Although the model performance metrics are slightly worst for the 5 has been demonstrated. More specifically, a voltammetric sensor array compounds case than for the 3 compounds one (total NRMSE for the consisting in three SPCE modified with graphite, cobalt (II) phthalocy­ three drugs of 0.084 vs. 0.026 for the test subset), it has to be considered anine and palladium inks were employed to extract the electrochemical the higher complexity of the case. That is, on the one side, the use of a fingerprints of heroin, morphine, codeine, caffeine and paracetamol by fractional experimental design (as is the CCF) to keep the number of means of SWV. Despite the advantages that electrochemical methods samples required to build the model reasonable. On the other side, and may offer, the simultaneous determination of the aforementioned as already reported, the presence of certain cutting agents can influence compounds can also be challenging given their similar electrochemical the voltammetric response, up to the point that the observed peak for the response; especially when attempting the discrimination between heroin pure compound might not be seen in the presence of the adulterant [16, and morphine, since both molecules present the same functional groups 17]. in their skeleton. Thus, a partial least square regression (PLS) model for 8 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 the quantification of heroin, morphine, codeine, caffeine and paraceta­ [3] Bordersens: Border Detection of Illicit Drugs and Precursors by Highly Accurate Electrosensors, 2021. 〈https://bordersens.eu/〉. mol mixtures at the µM level was built employing a central composite [4] Drug Enforcement Administration. Drugs of abuse: A DEA resource guide, Drug face-centered (CCF) experimental design. A very satisfactory perfor­ Enforcement Administration. US Department of Justice, 2017. mance was obtained, demonstrating that the use of the herein proposed [5] N.E. Hagemeier, Introduction to the opioid epidemic: the economic burden on the voltammetric ET shows huge potential to carry the identification and healthcare system and impact on quality of life, Am. J. Manag. Care 24 (10) (2018) S200. quantification of seized drug street samples, either those being pure or [6] J.K. O’Donnell, R.M. Gladden, P. Seth, Trends in deaths involving heroin and already mixed with other drugs and/or cutting agents. synthetic opioids excluding methadone, and law enforcement drug product reports, Overall, the advantages of ETs to deconvolute complex overlapping by census region — United States, 2006–2015, MMWR Morb. Mortal. Wkly. Rep. 66 (34) (2017) 897–903. electrochemical responses and achieve the simultaneous quantitative [7] J.M.P.J. Garrido, C. Delerue-Matos, F. Borges, T.R.A. Macedo, A.M. Oliveira-Brett, determination of several analytes have been shown. Moreover, the use of Electrochemical analysis of opiates – an overview, Anal. Lett. 37 (5) (2004) a properly formulated graphite-polystyrene ink has been demonstrated 831–844. [8] M.Y. Salem, S.A. Ross, T.P. Murphy, M.A. ElSohly, GC-MS determination of heroin as a simple approach to obtain an array of modified electrodes with metabolites in meconium: evaluation of four solid-phase extraction cartridges, different responses towards the compounds under study. J. Anal. Toxicol. 25 (2) (2001) 93–98. As for the data processing, GAs allowed to reduce the number of [9] C. Meadway, S. George, R. Braithwaite, A rapid GC–MS method for the determination of dihydrocodeine, codeine, norcodeine, morphine, normorphine inputs fed to the model through the identification of its most relevant and 6-MAM in urine, Forensic Sci. Int. 127 (1–2) (2002) 136–141. features, what in turn reduced its complexity , and at the same time [10] Z. Zhang, B. Yan, K. Liu, Y. Liao, H. Liu, CE-MS analysis of heroin and its basic improved its performance. Lastly, by conducting a permutation test or impurities using a charged polymer-protected gold nanoparticle-coated capillary, Electrophoresis 30 (2) (2009) 379–387. “target shuffling process”, it was also demonstrated that neither the high [11] R.B. Taylor, A.S. Low, R.G. Reid, Determination of opiates in urine by capillary dimensionality of the data or the use of GAs and PLS resulted in over- electrophoresis, J. Chromatogr. B Biomed. Appl. 675 (2) (1996) 213–223. fitted models. In fact, this verification provided a high significance for [12] Y. Zhuang, X. Cai, J. Yu, H. Ju, Flow injection chemiluminescene analysis for the obtained RMSE values, higher than 99.99%, both for train and test highly sensitive determination of noscapine, J. Photochem. Photobiol. A Chem. 162 (2–3) (2004) 457–462. subsets (P value lower than 3⋅10-5), as illustrated in Fig. 7. [13] Y. Zhuang, D. Zhang, H. Ju, Sensitive determination of heroin based on In conclusion, the results presented herein suggest the potential of electrogenerated chemiluminescence of tris(2,2′ -bipyridyl)ruthenium(II) these devices to be used as analytical tools for the detection of illicit immobilized in zeolite Y modified carbon paste electrode, Analyst 130 (4) (2005) 534–540. substances from street samples offering low-cost measurement systems [14] J. Moros, N. Galipienso, R. Vilches, S. Garrigues, M. De La Guardia, Nondestructive with rapid response, simple usage and high portability; all of them ideal direct determination of heroin in seized illicit street drugs by diffuse reflectance characteristics for point-of-use forensic or law-enforcement near-infrared spectroscopy, Anal. Chem. 80 (19) (2008) 7257–7265. [15] G. Sakai, K. Ogata, T. Uda, N. Miura, N. Yamazoe, A surface plasmon resonance- applications. based immunosensor for highly sensitive detection of morphine, Sens. Actuators B Chem. 49 (1–2) (1998) 5. CRediT authorship contribution statement [16] A. Florea, J. Schram, M. De Jong, J. Eliaerts, F. Van Durme, B. Kaur, N. Samyn, K. De Wael, Electrochemical strategies for adulterated heroin samples, Anal. Chem. 91 (12) (2019) 7920–7928. Dionisia Ortiz-Aguayo: Methodology, Investigation, Writing – [17] M. de Jong, A. Florea, J. Eliaerts, F. Van Durme, N. Samyn, K. De Wael, Tackling original draft. Ceto Xavier Cetó: Methodology, Software, Writing – poor specificity of cocaine color tests by electrochemical strategies, Anal. Chem. 90 (11) (2018) 6811–6819. original draft. Karolien De Wael: Resources, Writing – review & edit­ [18] P. Ciosek, W. Wróblewski, Sensor arrays for liquid sensing – electronic tongue ing, Funding acquisition. Manel del Valle: Conceptualization, Meth­ systems, Analyst 132 (10) (2007) 963–978. odology, Writing – review & editing, Funding acquisition. [19] M. del Valle, Electronic tongues employing electrochemical sensors, Electroanalysis 22 (14) (2010) 1539–1555. [20] Y. Vlasov, A. Legin, A. Rudnitskaya, C. Di Natale, A. D’Amico, Nonspecific sensor Declaration of Competing Interest arrays (“electronic tongue”) for chemical analysis of liquids: (IUPAC technical report), Pure Appl. Chem. 77 (11) (2005) 1965–1983. [21] F. Arduini, L. Micheli, D. Moscone, G. Palleschi, S. Piermarini, F. Ricci, G. Volpe, The authors declare that they have no known competing financial Electrochemical biosensors based on nanomodified screen-printed electrodes: interests or personal relationships that could have appeared to influence recent applications in clinical analysis, TrAC Trends Anal. Chem. 79 (2016) 114–126. the work reported in this paper. [22] G. Maduraiveeran, M. Sasidharan, V. Ganesan, Electrochemical sensor and biosensor platforms based on advanced nanomaterials for biological and Acknowledgments biomedical applications, Biosens. Bioelectron. 103 (2018) 113–129. [23] X. Cetó, F. Céspedes, M.I. Pividori, J.M. Gutiérrez, M. del Valle, Resolution of phenolic antioxidant mixtures employing a voltammetric bio-electronic tongue, This research was funded by the Spanish Ministry of Science and Analyst 137 (2) (2012) 349–356. Innovation, MCINN (Madrid) through project PID2019-107102RB-C21. [24] E. Richards, C. Bessant, S. Saini, Optimisation of a neural network model for calibration of voltammetric data, Chemom. Intell. Lab. Syst. 61 (1–2) (2002) This project has received funding from the European Union’s Horizon 35–49. 2020 research and innovation programme under Grant agreement No. [25] V. Esposito Vinzi, W.W. Chin, J. Henseler, H. Wang. Handbook of partial least 833787 BorderSens (Border detection of illicit drugs and precursors by squares: concepts, methods and applications, Springer, Berlin, 2010. [26] A.J. Bard, Chemical modification of electrodes, J. Chem. Educ. 60 (4) (1983) 302. highly accurate electrosensors). Karolien De Wael acknowledges the [27] D. Ortiz-Aguayo, M. Bonet-San-Emeterio, M. del Valle, Simultaneous voltammetric support from the University of Antwerp (IOF)UAntwerp) and BELSPO. determination of acetaminophen, ascorbic acid and uric acid by use of integrated Dionisia Ortiz-Aguayo was funded by Universitat Autònoma de Barce­ array of screen-printed electrodes and chemometric tools, Sensors 19 (15) (2019) 3286. lona through a PIF fellowship. M. del Valle thanks the support from the [28] J.M. Gutiérrez, L. Moreno-Barón, M.I. Pividori, S. Alegret, M. del Valle, program ICREA Academia. A voltammetric electronic tongue made of modified epoxy-graphite electrodes for the qualitative analysis of wine, Microchim. Acta 169 (3–4) (2010) 261–268. [29] M.L. Rodriguez-Mendez, C. García-Hernandez, C. Medina-Plaza, C. García- Appendix A. Supplementary material Cabezón, J.A. de Saja, Multisensor systems based on phthalocyanines for monitoring the quality of grapes, J. Porphyr. Phthalocyanines 20 (08n11) (2016) Supplementary data associated with this article can be found in the 889–894. [30] J.P. Metters, R.O. Kadara, C.E. Banks, New directions in screen printed online version at doi:10.1016/j.snb.2021.131345. electroanalytical sensors: an overview of recent developments, Analyst 136 (6) (2011) 1067–1076. References [31] V. Tsakova, R. Seeber, Conducting polymers in electrochemical sensing: factors influencing the electroanalytical signal, Anal. Bioanal. Chem. 408 (26) (2016) 7231–7241. [1] A. Florea, M. de Jong, K. De Wael, Electrochemical strategies for the detection of [32] D. Ortiz-Aguayo, K. De Wael, X. Cetó, Valle, M. del Valle, Voltammetric sensing forensic drugs, Curr. Opin. Electrochem. 11 (2018) 34–40. using an array of modified SPCE coupled with machine learning strategies for the [2] C. Guiney, EU drug markets report 2019, Drugnet Ireland (2020) 18–20. 9 D. Ortiz-Aguayo et al. Sensors and Actuators: B. Chemical 357 (2022) 131345 improved identification of opioids in presence of cutting agents, Journal of [46] A. Mimendia, J.M. Gutiérrez, L.J. Opalski, P. Ciosek, W. Wróblewski, M. del Valle, Electroanalytical Chemistry 902 (2021), 115770. SIA system employing the transient response from a potentiometric sensor [33] M. Galik, A.M. O’Mahony, J. Wang, Cyclic and square-wave voltammetric array—correction of a saline matrix effect, Talanta 82 (3) (2010) 931–938. signatures of nitro-containing explosives, Electroanalysis 23 (5) (2011) 1193–1204. Dionisia Ortiz-Aguayo received her M.Sc. degree in Chemistry in 2015 from the Uni­ [34] A.M. O’Mahony, J.R. Windmiller, I.A. Samek, A.J. Bandodkar, J. Wang, “Swipe and verstitat Autònoma de Barcelona, where she is at the moment completing her Ph.D. in scan”: integration of sampling and analysis of gunshot metal residues at screen- Analytical Chemistry. Her main research topics deal with the development of voltametric printed electrodes, Electrochem. Commun. 23 (1) (2012) 52–55. sensors and the application of electronic tongues as a tool in the pharmaceutical and [35] J.M.P.J. Garrido, C. Delerue-Matos, F. Borges, T.R.A. Macedo, A.M. Oliveira-Brett, forensic analysis. Voltammetric oxidation of drugs of abuse III. Heroin and metabolites, Electroanalysis 16 (18) (2004) 1497–1502. [36] J.M.P.J. Garrido, C. Delerue-Matos, F. Borges, T.R.A. Macedo, A.M. Oliveira-Brett, Xavier Cetó completed his M.Sc. degree in chemistry in 2009 at the Universtitat Autònoma Voltammetric oxidation of drugs of abuse: I. Morphine and metabolites, de Barcelona, and received his Ph.D. in analytical chemistry in 2013 from the same uni­ Electroanalysis 16 (17) (2004) 1419–1426. versity. From 2014 to 2017, he worked as Research Associate at the University of South [37] J.M.P.J. Garrido, C. Delerue-Matos, F. Borges, T.R.A. Macedo, A.M. Oliveira-Brett, Australia. At present, he is a postdoctoral researcher of the BorderSens project at Uni­ Voltammetric oxidation of drugs of abuse II. Codeine and metabolites, verstitat Autònoma de Barcelona. His main research topics deal with the development of Electroanalysis 16 (17) (2004) 1427–1433. new voltammetric (bio)sensors and the application of electronic tongues and chemometric [38] D. Nematollahi, H. Shayani-Jam, M. Alimoradi, S. Niroomand, Electrochemical tools for data analysis, mainly employing voltammetric (bio)sensors, albeit also potenti­ oxidation of acetaminophen in aqueous solutions: kinetic evaluation of hydrolysis, ometric ones. hydroxylation and dimerization processes, Electrochim. Acta 54 (28) (2009) 7407–7415. Karolien De Wael received her Ph.D. in Chemistry from UGent as national science [39] M. Khairy, B.G. Mahmoud, C.E. Banks, Simultaneous determination of codeine and foundation (FWO) fellow. After a prestigious FWO postdoctoral fellowship at the same its co-formulated drugs acetaminophen and caffeine by utilising cerium oxide university, she started as tenure track research professor at University of Antwerp in 2011. nanoparticles modified screen-printed electrodes, Sens. Actuators B Chem. 259 In 2018, she was appointed as full professor. Currently, she is spokesperson of the AXES (2018) 142–154. Research Group performing fundamental, methodological and application-oriented [40] Y. Tadesse, A. Tadese, R.C. Saini, R. Pal, Cyclic voltammetric investigation of research involving a wide range of analytical techniques. Her strategic vision aims at a caffeine at anthraquinone modified carbon paste electrode, Int. J. Electrochem. 1 portfolio of sensor technologies (ful-filling relevant criteria) that can be applied in (2013) 2013–2017. different markets/sectors embracing the idea of responsible research and innovation. The [41] J. Broséus, N. Gentile, P. Esseiva, The cutting of cocaine and heroin: a critical AXES research group is part of the Nanolab Center of Excellence and the IOF Enviromics review, Forensic Sci. Int. 262 (2016) 73–83. consortium in the University of Antwerp. [42] J.R.B. Rodríguez, V.C. Díaz, A.C. Garcia, P.T. Blanco, Voltammetric assay of heroin in illicit dosage forms, Analyst 115 (2) (1990) 209–212. [43] X. Cetó, F. Céspedes, M. del Valle, Comparison of methods for the processing of Manel del Valle received his Ph.D. in Chemistry in 1992 from the Universtitat Autònoma de voltammetric electronic tongues data, Microchim. Acta 180 (5–6) (2013) 319–330. Barcelona, where he is full professor in Analytical Chemistry since 2018. He is a member of [44] E. Richards, C. Bessant, S. Saini, Multivariate data analysis in electroanalytical the Sensors & Biosensors Group where he is a specialist for instrumentation and electro­ chemistry, Electroanalysis 14 (22) (2002) 1533–1542. chemical sensors. He has initiated there the research lines of sensor arrays and electronic [45] G.M. Escandar, H.C. Goicoechea, A. Muñoz de la Peña, A.C. Olivieri, Second- and tongues. Other interests of his work are the use of impedance measurements for sensor higher-order data generation and calibration: a tutorial, Anal. Chim. Acta 806 development, biosensors and the design of automated flow systems. (2014) 8–26. 10