Custom Electronic Nose design for Sherry Wines: Optimizing feature selection for machine learning-based classification

Custom Electronic Nose design for Sherry Wines: Optimizing feature selection for machine learning-based classification | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Custom Electronic Nose design for Sherry Wines: Optimizing feature selection for machine learning-based classification Itziar Mengual, Xavier Marimon, Ferran Esquinas, Marina Batllo, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6421640/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 19 You are reading this latest preprint version Abstract The Electronic Nose (e-nose) is a growing tool widely used in food applications. This study classifies Sherry wines using an e-nose equipped with a matrix of gas sensors. By extracting features from identified phases of the sensor responses (slope, plateau, and ventilation) and applying the Elastic Net technique for feature selection, we improved machine learning (ML) algorithm performance while reducing dataset dimensionality. Our novel data acquisition approach incorporates a ventilation phase to enhance volatile detection, boosting wine classification accuracy. We analyzed features extracted from different phases of sensor response using ML models like Support Vector Machines, Discriminant Analysis, and Random Forests. High classification accuracy exceeding 95% was achieved, with Linear SVM demonstrating exceptional performance on test data, indicating strong generalization capabilities. Our findings demonstrate that the proposed e-nose design equipped with ML algorithms and efficient feature extraction methods offers a reliable solution for Sherry wine classification. Physical sciences/Engineering Physical sciences/Engineering/Chemical engineering Physical sciences/Physics/Techniques and instrumentation Sherry Wines Electronic nose Metal-oxide sensors Volatile organic compounds Gas sensors Electronic design Feature extraction Feature selection Classification techniques Machine learning Deep Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 1. Introduction Spain is one of the main wine producers over the world, with a yearly production of 44 million hectoliters (Ruggieri et al., 2009 ). Amongst the different producers in the country, Jerez (Sherry) and San Lucar (Manzanilla) Region, in the south of Spain, produces one of the most highly appreciated wines due to the area conditions and their characteristic aging methods (Pardo-Calle et al., 1970 ). The aging process of these wines is characterized by a dynamic system called “ criaderas y solera ” or “ soleraje ”. Here, the wines are stored in White American Oak (Quercus alba) butts and are organized in groups deppreviending on their aging stage, a method called “ escalas ”. The bottled wine comes from the oldest group which receives the “solera” denomination; then, they begin with the extracting process, which is called “ saca ”. After this, an equal amount of wine is extracted from the previous level of the “ escala ” containing the second oldest wine “1st Criadera”. This way, each level is replenished by the previous level (the 2nd Criadera is replenished by the 3rd, the 3rd by the 4th and so on), in a process called “ rocio ” until the level that contains the youngest wine called “ Sobretabla ”. The final product is often a combination of related “ soleras ” to obtain the desired oenological characteristics of the producer (Saldaña, 2022 ). The objective of this system is to preserve and ensure the oenological properties and quality of the Sherry wine. The constant mixture of new and old wine during the “ sacas ” and “ rocios ” mitigates the effects of the different “ añadas ” that constitute the “ soleraje ” providing a mean age for the resultant wine (Sanz et al., 2019 ). This special treatment gives Sherry wines sensory characteristics whose variations and nuances are rarely matched in other parts of the wine world. Three different grape varieties are used in the production of this type of wine (Palomino, Pedro Ximénez and Moscatel), resulting in a wide variety of wines that clearly differ in aroma, flavor, texture, and color, having each one their own individual characteristics which need to be maintained and assured. One of the most used indicators to assess wine type, quality and production process is, precisely, aroma. Aroma comes from a vast number of volatile compounds which have been closely related to different wine properties (Vera et al., 2010 ). Only on Sherry Wines, more than 150 different aromas have been identified (Durán-Guerrero et al., 2021). For this reason, distinguishing wines through this characteristic is a highly complex task, normally developed by sommeliers, trained wine professionals capable of detecting wine characteristics through their aroma, color, taste, and density. However, this training is highly complex and costly, so depending on them to classify or assess wine quality can be an expensive task. Apart, other methods such as gas chromatography or mass spectrometry can be used to analyze wine volatiles, but they are time-consuming and labor-intensive. The electronic nose (e-nose) is composed by a set of volatile sensible sensors that pretends to mimic the olfactory perception of humans (Lozano et al., 2006 ).This tool has demonstrated to be especially useful in the food industry (Ghasemi-Varnamkhasti et al., 2018; Preedy, 2016 ), having a special impact on the wine sector. The e-nose functions by detecting different volatiles present in the environment and coding those presences into signals that can be then processed. Using different classifying techniques, this tool allows to distinguish the different characteristics from the products, resulting in an easy, fast, and cheap way to solve classification problems. Previous work has been done in the wine field, presenting this tool as a successful approach for many problems in quality assessment, wine classification, and aromatic compound detection (Aleixandre et al., 2008 ; Chilo et al., 2016 ; Fuentes et al., 2020 ; Summerson et al., 2021 ; Tan & Xu, 2020 ). However, until now, many of these classification techniques have required many samples or sensors to obtain useful results, making its applicability in real-world production complex (Lozano et al., 2008 ). In this study, our goal is to classify varieties of Sherry wines using a simple and cost-effective artificial nose, also known as an e-nose. We aim to achieve this using a small number of samples and comparing various classification methods. The objective is to evaluate the accuracy of different data processing techniques to determine the most effective approach for this specific family of wines. 2. Materials and Methods 2.1. Sherry Wines We selected three distinctive types of Sherry wines, namely Amontillado, Oloroso, and Pedro Ximénez. Each of these wines was chosen for its unique qualities, a result of their special production process, and wide availability. The use of three grape varieties—Palomino, Pedro Ximénez, and Moscatel—ensures a diverse range of wines with distinct aromas, flavors, textures, and colors. Amontillado is a dry wine produced with Palomino grape first undergoing a biological process and finishing an oxidative environment (Valcárcel-Muñoz et al., 2022 ). During the early aging stages, the decaying layer prevents the wine from oxidation and the through the fermentation process ethanol is combined with oxygen to produce acetaldehyde (Echave et al., 2021 ). In the later stages of the aging process the yeast layer disappears, and the wine is exposed to oxidation. The result is a wine with a color between topaz and amber with a smell reminiscent of aromatic herbs. Oloroso is a dry wine produced with a full-bodied Palomino must and its ageing process revolves around oxidation. Different from Amontillado, there is no yeast layer (Valcárcel-Muñoz et al., 2022 ). This entails a wine with a higher degree of alcohol (around 18.5%) and a dark gold color. Lastly, Pedro Ximénez is a sweet wine produced with an overly ripe grape of the same name. Its aging process is oxidative but, unlike Oloroso, due to its density and sugar content the ethanol decreasing rate by evaporation is faster than the water loss due to osmosis (Macedo et al., 2023 ). 2.2. The e-nose Different types of volatile sensors can be used on an e-nose based on its application (Aguilera et al., 2012 ). In our case, we aimed to obtain significant results using a selected matrix of sensors which produce a change on the output when exposed to the selected samples of Sherry wines. We conducted an analysis of the volatile components in Sherry wines to identify the most suitable sensor matrix, although it is expected none of them has a specific response to these components. Table 1 provides a summary of the selected sensors. Table 1 Characteristics of each gas sensor. Id Gas Sensor Target Gases Detection Range (ppm) U0 MP-5 Propane, LPG 300 ~ 10,000 U1 MP-4 Methane, Natural gas 300 ~ 10,000 U2 MP-2 Propane, smoke 200 ~ 10,000 U3 MP-7 CO 50 ~ 1,000 U4 MP-9 CO, Methane 50 ~ 1,000 CO 300 ~ 10,000 CH4 U6 MP-901 VOC, alcohol, smoke 1 ~ 50 U8 TGS-2600 Hydrogen 1 ~ 30 U9 TGS-2602 VOCs, ammonia, Thiols (H 2 S) 1 ~ 30 ppm EtOH U10 TGS-2620 Alcohol, Solvent vapors 50 ~ 5,000 ppm EtOH To acquire signals from the specified matrix, we have designed an electronic circuit (Sensing board), which can accommodate up to 16 MOX sensors. The block diagram shown in Fig. 1 also includes two multiplexer analog switches (CD74HC4051E, Texas Instruments, USA). These switches allow selective data collection from a smaller set of sensors through configurable firmware running on a control board connected through the IO interface. The sensing board additionally includes switching components, specifically a pair of solid-state relays (G6K-2P-Y-DC5, Omron Electronics, Japan), to control up to two external pumps (fresh air and vacuum). This provides additional flexibility, enabling two different configurations for handling the airflow of the sample’s headspace. The board contains an integrated circuit for measuring temperature and humidity (SHT30-ARP-B, Sensirion, Switzerland). The design includes components to ensure noise rejection and high-precision voltage regulators for supplying correspondingly the heating element and the reference voltage of the MOX sensors. The proposed e-nose also requires an embedded control board which has been implemented using an Arduino-UNO board containing an ATmega328P microprocessor (Microchip Technology, USA) which allows the acquisition through an embedded 10-bits A/D converter of the multiplexed gas sensor outputs. The microcontroller board is connected via USB a to a custom-made MATLAB script (R2023b MathWorks, USA) running on a PC, which allows the storing and post-processing of the data acquired from the e-nose. Both the sensing board and microprocessor board are placed inside a custom mechanical chassis, which also incorporates a fan (02510SS-12P-AT-00, NMB Technologies, USA) mounted at the top and controlled through a switch module (IR520, Vishay Siliconix, USA). The fan enables primarily flushing the sample´s headspace exposed to the sensors after every measurement cycle. The fan serves as dual purpose: to desaturate the sample and facilitating the cleaning process. Contrasting with studies by Aguiar et al. (Aguiar et al., 2023 ) and Markelov et al. (Santos & Rodrigues, 2020 ) which showed that fan-assisted extraction enhances volatile detection and increases spectrometer peaks. While Aguiar et al. demonstrated that the fan accelerates volatile compound transport and improves headspace flushing, our application benefits from the fan's ability to promote dynamic headspace, reducing sample saturation and facilitating a cleaner analysis environment for subsequent measurements. To address these challenges, we designed a 3D-printed funnel that allows to expose the sample headspace uniformly to the sensors placed on the PCB board and when closed, it isolates the airspace exposed to the sensors by flushing it with fresh air. When working with volatile sensors, two primary challenges arise: avoiding or detecting possible contaminants and ensuring adequate detection of volatiles of interest. To address these challenges, we designed a 3D-printed funnel that allows for uniform exposure of the sample headspace to the sensors placed on the PCB board. When closed, it isolates the airspace exposed to the sensors by flushing it with fresh air enabling. This capsule not only isolates the sensors but also facilitates the concentration of volatiles, such as those present in a glass of wine. This concentration process generates a distinctive peak in the sensor signal, enhancing sample identification. The funnel incorporates a shutter mechanism controlled by the firmware programmed on the Arduino board that enable the exposure of the sample headspace at specific points during the acquisition process. The CAD design of the capsule is depicted in Fig. 2 . All these components were embedded in a 3D-printed structure, designed, and fabricated to support the electronic boards, the fan and other possible components as pumps that could be employed on different configurations. The design shown Fig. 3 , has been developed using a CAD design software (SolidWorks 2023, USA) and 3D-printed (HP Jet Fusion MJF 5200, USA) using PA12 Nylon-12 powder (Sinterit, Poland). The design allows to accommodate a wine glass cup Riedel Ouverture (Riedel, Austria) and a shutter to control the flow of the headspace of the sample toward the Sensing board. 2.3. Sample acquisition During the sample acquisition process, a MATLAB code was implemented to interface with the microprocessor, enabling data collection. Figure 4 presents a graphic representation of the sample preparation, detailing the process starting with the e-nose connection to the computer, to the collection of raw data from each wine measurement cycle. Samples were obtained by pouring 20 ml of each type of wine into a wine glass. The 3D-printed device was then sealed on top of the glass using Parafilm M sealing film (Bemis, USA). The sample was maintained at a constant temperature of 5°C using a temperature-controlled bath (Gilson Co., USA). The measurement procedure began with the shutter closed. Approximately 100 s later, the gate was opened, exposing the e-nose to wine volatiles for 300 s. After 200 s of the shutter opening, the fan was turned on until the end of the measurement cycle. Following the activation of the fan, the shutter remained open for another 100 s before closing. The measurement cycle concluded 100 s after the shutter's closure, resulting in a total duration of 600 s, with a sampling frequency of F s = 4.7 Hz. Each sensor sample was taken five times to enhance value stability, with a repetitive frequency of F rep = 0.2 Hz. A total of 10 samples were collected for each type of wine, with each sample comprising five distinct phases. The experiment begins with the pre-phase, during which the shutter is closed, allowing volatile compounds in the wine to accumulate in the headspace. Next, the slope phase initiates when the gate opens, resulting in a huge rise in sensor values as volatiles reach the sensor board. This is followed by the plateau phase, where volatile saturation causes the sensor readings to stabilize. The ventilation phase then begins, with the fan activating to desaturate the sample while the shutter remains open. Finally, the post-phase concludes the experiment, closing the shutter and initiating the cleaning process. The pre and post phases served as baselines. Figure 5 illustrates the sensor values for each wine sample, along with the corresponding step signals from the shutter gate and fan activation. Following the acquisition of samples, they were saved as .csv files and subsequently imported into MATLAB to construct a signal dataset. This dataset comprised 30 wine samples along with their respective labels. Each sample was generated from the outputs obtained from the matrix of 9 sensors. Figure 6 summarizes the data from 10 cycle repetitions for each wine type, excluding the headspace flush signal between exposures, with all cycles for one wine type completed before proceeding to the next. During the repeating cycles, a slight increase trend was observed in the sensor baseline, which was particularly distinguishable for sensors U4 (MP-9) and U3 (MP-7). The target gas for both sensors is CO in the range of 50 to 1,000 ppm. The increase in both sensors emphasizes the sensitivity to CO. Using the initial trial values as a baseline, we corrected the offset drift for each sensor by calculating the difference between these baseline values and the subsequent measurements. 2.4. Feature Extraction and Selection Figure 6 demonstrates the similarity in the sensor responses across the three types of wine samples, highlighting the challenge in distinguishing patterns. Conventional statistical methods are insufficient because they struggle to manage large, intricate datasets and identify subtle patterns. This demands the use of Machine Learning (ML) techniques, which excel at handling such complex data and uncovering critical patterns for accurate classification. When using traditional Machine Learning for classification, the first step is to extract different features from the signal. The extracted features can vary depending on the type of signal, although there are some typical ones that are commonly used. In our case, we decided to extract three different types of features, using the most common ones inside each group. All the extracted features can be seen on Table 2 . In Table 2 , nineteen features are listed for each sensor. The experiment comprises three main phases: slope, plateau, and ventilation. These phases yield variations on sensor output signals, which requires separated analysis to extract valuable insights for each phase. Each sensor and phase provide 19 variables, leading to 171 variables per phase and a total of 513 variables per wine sample. This helps to prevent the model from capturing noise and random fluctuations in the data, a problem known as overfitting, and optimize the results analysis by reducing inefficiencies caused by highly correlated and redundant features. To address this, regression techniques offer penalties that serve dual purposes: reducing overfitting by assigning very low values close to zero to some features ( L 2 penalty, Ridge Regression) and promoting feature selection by multiplying some feature coefficients by zero ( L 1 penalty, Lasso Regression). In our analysis, we utilized Elastic Net, which combines both penalties to leverage their strengths optimally. This approach reduces the weight of certain variables ( L 2 ) while also eliminating some of them to mitigate the impact of highly correlated features ( L 1 ), ensuring a balanced feature selection process. Figure 7 provides a schematic overview of the methodology employed for feature selection. The hyperparameters used were the L 1 ratio and the alpha coefficient ( α ). The L 1 ratio determines the balance between L 1 (Lasso) and L 2 (Ridge) penalties, where a value of 0 indicates purely Ridge Regression and a value of 1 indicates purely Lasso Regression. The α coefficient, on the other hand, controls the degree of regularization applied, with common values typically close to zero. We utilized the ElasticNetCV function(Friedman et al., 2010 ; Kim et al., 2007 ) from the Scikit-Learn library in Python (Python 3.12.2, Python Software Foundation, USA). The α parameter was set to [0.01, 0.1, 1.0, 10.0], and the L 1 ratio was set to [0.5, 0.7, 0.9, 1]. Lasso Regression was emphasized due to its effectiveness in reducing dataset dimensionality by eliminating irrelevant variables. The Elastic Net feature selection was applied separately to each of the three phases and collectively to the combined set of features from all phases. This process reduced the number of variables from 513 to 31 overall: three temporal features from the slope phase, 10 from the plateau phase (nine temporal and one spectral), and 18 from the ventilation phase (14 temporal, two spectral, and two non-linear). The optimal hyperparameters for the combined analysis were an α = 0.1 and an L 1 ratio of 0.5, indicating equal use of both penalties. Figure 8 shows the correlation matrix using a heatmap, illustrating the state of the data before (A) and after (B) dimensionality reduction. The heatmap before the reduction (A) reveals a pattern in the correlations, indicating a repetitive set of values that convey similar information. It is important to note that the relevance lies not in the specific variables but in the overall pattern and insights provided through the colors. For the individual phases, the feature selection results and optimal hyperparameters were as follows: Slope phase: 16 features selected (14 temporal and two spectral) with an α = 0.1 and an L1 ratio of 0.5. Plateau phase: 37 features selected (17 temporal, 10 spectral, and 10 non-linear) with an α = 0.01 and an L1 ratio of 0.7. Ventilation phase: 9 features selected (seven temporal and two non-linear) with an α = 0.1 and an L1 ratio of 1.0. 2.5. Classification Techniques We evaluated the classification accuracy of various ML algorithms for differentiating wine samples. The algorithms included decision trees, linear discriminant analysis, logistic regression, Naïve Bayes, support vector machines, k-nearest neighbors, ensemble models, and traditional neural networks. After feature extraction and selection, we normalized features using z -scores for models that required scaling transformation. For instance, decision trees do not require transformation due to their node-based structure, whereas support vector machines do, as they rely on geometrical distances to construct hyperplanes. Prior to each analysis, the samples were randomly divided into training (80%) and testing (20%) datasets, stratifying the target variable (type of wine) to preserve the same proportions of examples in each wine class as observed in the original data. All analyses were conducted using five-fold cross-validation. Accuracy was selected as the metric for comparing results between the test and training data, as well as across different sets of features. Four different analyses were conducted: initially, each algorithm was applied separately to individual groups of phases (slope, plateau, or ventilation); subsequently, an integrated analysis incorporating all phases was conducted. Another method explored for signal classification was deep learning, which allows the classification of different samples without needing prior feature extraction. Neural networks are computational models inspired by the human brain, consisting of interconnected nodes (neurons) organized in layers. They process input data to produce outputs and learn by adjusting weights and biases using techniques like backpropagation and gradient descent. A growing type of neural network used to classify signals is the Long-Short-Term Memory Network (LSTM), which are a specialized type of recurrent neural networks (RNNs). Unlike traditional RNNs, LSTMs effectively manage long-term dependencies by incorporating memory cells that maintain information over extended periods and are commonly used for sequential data classification. In this study, an LSTM network architecture was used, composed of an input layer (No. of sensors 9), a bidirectional LSTM layer, a fully connected layer (100 units), an exponential normalized function (SoftMax) layer, and a classification layer. The LSTM network was trained with the following settings: a maximum of 25 epochs, a mini-batch size of 10, an initial learning rate of 0.01, and a gradient threshold of 1. Of the 30 signals from each type of acquisition, 25 were used for training. Several evaluation metrics: accuracy, F1-score, recall, and reproducibility were computed for the classification methods. 3. Results and Discussion As depicted in Table 3 several algorithm evaluation metrics were obtained for the validation and test datasets across the different training groups. These results provide a comprehensive overview of the performance of each classification method throughout the various phases of the experiment. The validation and test metrics were presented to demonstrate the efficacy of the algorithms. An increase in test results compared to validation results suggests that the model generalizes well to all data rather than overfitting to the training data. The highest values observed in the test metrics can be attributed to the small number of samples used ( n = 7). In future studies, expanding the size of the training and test datasets may enhance the validation metrics (Ghorbanzadeh et al., 2019 ; Hestness et al., 2017). The features derived from the slope phase performed consistently across various algorithms, including Decision Trees and SVMs. The accuracy of the test set ranged from 83–100%, and the accuracy of the validation set ranged from 88–96%. The features from the plateau phase also showed strong performance, particularly with Decision Trees and Ensemble Bagged Trees, achieving accuracy over 95% in validation and between 83% and 100% in test results. Features from the ventilation phase performed better with some classical ML algorithm namely SVMs (Linear and Quadratic), KNNs and NNs. When analyzing the set with features from all phases together, NN demonstrated the strongest performance, achieving validation accuracy scores ranging from 95.6–97.8%. However, this performance was not maintained for all NN when evaluated using the test set, indicating some overfitting to the validation data. The LSTM network achieved an accuracy of 90%, which, although lower than traditional ML methods, is still notable given the small sample size. Deep Learning techniques and Neural Networks may excel with more complex and larger datasets, suggesting their potential value in future research. Table 3 Accuracy results from the Validation Model across phases using the corresponding extracted features. Machine Learning Algorithms Accuracy (%) Type Phases Slope Features Plateau Features Ventilation Features All Features Datasets Validation Test Validation Test Validation Test Validation Test Tree Fine 96,0% 100% 96,0% 83,3% 84,0% 83,3% 80,0% 100% Medium 96,0% 100% 96,0% 83,3% 84,0% 83,3% 80,0% 100% Coarse 96,0% 100% 96,0% 83,3% 84,0% 83,3% 80,0% 100% Discriminant Analysis Linear 88,0% 83,3% 84,0% 83,3% 88,0% 100% 80,0% 83,3% Logistic Regression Efficient 72,0% 100% 52,0% 33,3% 32,0% 50,0% 40,0% 66,7% Kernel 84,0% 100% 56,0% 83,3% 80,0% 100% 80,0% 100% Naïve Bayes Kernel 80,0% 83,3% 76,0% 83,3% 88,0% 83,3% 72,0% 100% Support Vector Machine Linear 88,0% 100% 84,0% 100% 92,0% 100% 84,0% 83,3% Quadratic 88,0% 100% 84,0% 83,3% 92,0% 100% 88,0% 100% Cubic 92,0% 83,3% 80,0% 83,3% 88,0% 100% 80,0% 100% Fine Gaussian 64,0% 50,0% 32,0% 33,3% 40,0% 50,0% 36,0% 33,3% Medium Gaussian 88,0% 100% 72,0% 100% 80,0% 100% 84,0% 100% Coarse gaussian 56,0% 66,7% 44,0% 33,3% 60,0% 66,7% 52,0% 100% Kernel 88,0% 83,3% 72,0% 100% 80,0% 100% 88,0% 83,3% k-Nearest Neighbors Fine 84,0% 83,3% 76,0% 100% 84,0% 100% 96,0% 83,3% Medium 48,0% 83,3% 64,0% 66,7% 88,0% 83,3% 84,0% 100% Coarse 36,0% 33,3% 36,0% 33,3% 36,0% 33,3% 36,0% 33,3% Cosine 68,0% 83,3% 80,0% 50,0% 92,0% 100% 92,0% 100% Cubic 44,0% 83,3% 68,0% 50,0% 84,0% 83,3% 84,0% 100% Weighted 84,0% 83,3% 68,0% 100% 96,0% 100% 96,0% 83,3% Ensemble Boosted Trees 36,0% 33,3% 36,0% 33,3% 36,0% 33,3% 36,0% 33,3% Bagged Trees 92,0% 100% 96,0% 100% 88,0% 83,3% 76,0% 83,3% Subspace Discriminant 96,0% 83,3% 96,0% 100% 92,0% 83,3% 92,0% 100% Subspace KNN 72,0% 50,0% 52,0% 83,3% 68,0% 50,0% 44,0% 100% RUSBoosted Trees 36,0% 33,3% 36,0% 33,3% 40,0% 33,3% 48,0% 100% Traditional Neural Networks Narrow (10 layers) 96,0% 100% 84,0% 100% 92,0% 100% 96,0% 83,3% Medium (25 layers) 92,0% 83,3% 88,0% 100% 96,0% 100% 92,0% 83,3% Wide (100 layers) 92,0% 83,3% 80,0% 100% 96,0% 100% 96,0% 83,3% Bilayered (10, 10 layers) 84,0% 83,3% 84,0% 100% 92,0% 83,3% 96,0% 100% Trilayered (10, 10, 10 layers) 92,0% 100% 84,0% 83,3% 96,0% 33,3% 88,0% 100% LSTM - - - - - - - 93,3% 93,4% While the Ensemble Subspace Discriminant achieved over 90% validation accuracy, its test accuracy dropped to 83.3%, suggesting overfitting. Linear SVMs, however, achieved 100% test accuracy (84.0%-92.0% validation), making it the top performer. This improvement suggests that the hyperplane constructed to distinguish between wine varieties based on their features is robust and generalizes well to new, unseen test data. Consequently, future focus will be directed towards deepening our understanding of this algorithm. Looking deeply into the Linear SVM algorithm, we can identify which features played a significant role in wine variety discrimination. For each phase, ElasticNetCV was used to select relevant features. In the slope phase, where volatile compounds reach the sensors, 16 features were identified. The plateau phase, characterized by the saturation of volatile compounds, included 37 features. Finally, during the ventilation phase, 9 features were selected to desaturate the sample. Figure 9 highlights the top features from each Linear SVM model, ranked according to their weight in determining the model's hyperplane. While not all features have been described in detail, the most relevant ones have been highlighted to emphasize their importance and explain why they are effective in differentiating the wine varieties. The extracted features highlighted the importance of temporal features, which played a significant role in the analysis. Most of the features shown in the bar plots are temporal, such as RMS, Standard Deviation, Peak Value, and Mean. Non-linear and spectral features, like Band Power, Approximation Entropy, and Correlation Dimension, were used less frequently, likely due to the nature of the signals in this study and the larger number of temporal features computed. Nevertheless, all features were essential for the effective implementation of the algorithm. The most relevant features identified include Peak Value, determined as the maximum sensor value; RMS, defined as the square root of the sensor's mean square; and Mean, defined as the average sensor value. The analysis of Sherry wines using gas sensors MP-9 (U4) and TGS-2620 (U10) has revealed significant insights into the differentiation of Amontillado, Oloroso, and Pedro Ximénez. Most of the key features identified are from these sensors, as indicated by their higher importance in the bar plot. The MP-9 sensor (U4) targets CO and methane, gases that can be present in trace amounts due to the fermentation and oxidative processes involved in wine aging. The presence of CO is linked to various VOCs produced during these processes. The TGS-2620 sensor (U10) targets alcohol and solvent vapors, which are prominent in wines due to the ethanol and other aromatic compounds formed during fermentation and oxidation. During the slope phase, where vapor concentrations are increasing, the peak value of the MP-9 sensor is particularly important as it captures the highest concentration of gases like CO and methane. These gases are released at varying rates by different Sherry wines due to their oxidative properties and specific volatiles emitted during the aging process. Additionally, the standard deviation of the MP-5 sensor (U0) which targets propane gas reflects the variability in hydrocarbon concentration, which is influenced by the different grape varieties and aging methods, making it a crucial feature for distinguishing between the wines. During the plateau phase, the TGS-2620 sensor's peak value, mean, and RMS are essential metrics, reflecting the maximum concentration, average presence, and overall variability of alcohol and aromatic compounds, respectively, which are critical for understanding the consistent presence of these compounds in the vapor phase. The Pedro Ximénez wine, known for its high residual sugar content, has an alcohol level of 15.5%, which may enhance the volatilization of compounds detectable by the TGS-2620 sensor. In contrast, Amontillado, a dry wine with 17.5% alcohol, and Oloroso, with higher alcohol content, 20%, likely have fewer detectable volatiles. The higher concentration of volatile aromatic compounds, such as esters and aldehydes, in the sweet Pedro Ximénez wine makes these compounds more easily detectable by the sensor compared to the dry wines. In the ventilation phase, characterized by decreasing vapor concentrations, the skewness of the MP-9 -Skew(U4)- sensor measures the asymmetry of the vapor distribution, providing insight into the rate and manner in which each wine releases its volatile compounds. The dense and sweet Pedro Ximénez, with its high sugar content, shows variability in skewness, exhibiting both positive and negative values due to differences in air fluctuations and evaporation dynamics. In contrast, the drier Amontillado, with 17.5% alcohol content and low residual sugar, consistently exhibits negative skewness, reflecting a rapid initial burst of alcohol vapors followed by a quick decline. Oloroso, with the highest alcohol content at 20%, demonstrates positive skewness, indicating a significant initial burst of alcohol vapors. These patterns, depicted in Fig. 10 with Amontillado points in the negative skewness region, Oloroso points in the positive skewness region, and Pedro Ximénez points spread across both regions, highlight the influence of residual sugar and alcohol content on the volatile compound release during ventilation. Moreover, the TGS-2620 sensor's peak value ( y -axis from Fig. 10 A) and RMS ( y -axis from Fig. 10 B) provide additional critical metrics for distinguishing between the wines. The sweet and dense Pedro Ximénez, despite having a lower alcohol content, exhibits higher values in both RMS and peak due to its high sugar content, which enhances the volatilization of compounds. In contrast, Amontillado and Oloroso, both dry wines with higher alcohol content, show lower values in RMS and peak, reflecting their fewer detectable volatiles compared to Pedro Ximénez. As demonstrated in Fig. 10 , the decision boundaries in the 2D plots illustrate the effectiveness of the Linear Support Vector Machine models using features from the ventilation phase. This visualization confirms the model's capability to distinguish between different wine varieties based on the extracted features. However, it also highlights some misclassified samples, indicating limitations when using only two-dimensional feature spaces. The results table further emphasizes the necessity of a sensor matrix and the inclusion of various relevant features to significantly improve classification accuracy. This finding reinforces the quantitative results and underscores the practical applicability of these advanced classification techniques in accurately identifying and distinguishing between wine varieties. The data presented highlights the effectiveness of our chosen machine learning models in classifying wine samples based on their chemical compositions. Building on these findings, it is important to consider the broader implications and potential applications of these technologies in the food industry. These techniques have proven particularly relevant for quality, assessment (Tan & Xu, 2020 ) flavor production, and overall product manufacturing (Chilo et al., 2016 ). Consequently, it is essential to thoroughly explore the capabilities and performance of these technologies to maximize their potential in various food industry applications. Wines are indeed a significant part of the food industry, and the need for producers to maintain high-quality standards makes this product an ideal candidate for e-nose studies. Previous research has demonstrated various applications of machine learning in the wine industry, including detecting smoke contamination (Fuentes et al., 2020 ), identifying specific aromatic compounds, (Summerson et al., 2021 ) and classifying different types of wine (Aleixandre et al., 2008 ). However, these devices often focus on detecting very specific compounds or, in the case of classification, achieve relatively low performance with complex components and rely on a single type of classification technique. Therefore, there is a need for more comprehensive approaches that integrate multiple classification techniques to enhance the accuracy and reliability of wine quality assessments. Our study introduces a novel approach by incorporating a ventilation phase with the use of a fan, which enhances the distinction between wine varieties. This method effectively leverages the differences in evaporation dynamics and volatile compound release, significantly improving classification accuracy. The ability to distinguish between wines more accurately using this enhanced technique underscores the practical applicability and potential of using advanced e-nose systems with integrated ventilation phases in the wine industry. Future research should focus on further refining these models and exploring their applicability in real-world settings, (Rodríguez-Méndez et al., 2016 ) ultimately contributing to the continuous improvement of food quality and safety standards (Lu et al., 2022 ). 5. Conclusion Our study highlights the significant advancements achieved by using machine learning models to classify wine samples based on their chemical compositions. This research underscores the broader implications and potential applications of these technologies in the food industry. Wines, as a crucial component of the food sector, are particularly well-suited for e-nose studies. While previous research has shown various applications of machine learning in the wine industry—such as detecting smoke contamination, identifying specific aromatic compounds, and classifying different types of wine—these studies often rely on single classification techniques and achieve relatively low performance due to their complexity and narrow focus. Our study introduces a novel approach by incorporating a ventilation phase with the use of a fan, which significantly enhances the distinction between wine varieties. This method effectively leverages the differences in evaporation dynamics and volatile compound release, leading to improved classification accuracy. The dense and sweet Pedro Ximénez, despite having lower alcohol content, exhibited higher values in both RMS and peak due to its high sugar content. In contrast, the drier Amontillado and Oloroso wines, with higher alcohol content, showed lower values in RMS and peak, reflecting their fewer detectable volatiles. This enhanced technique underscores the practical applicability and potential of using advanced e-nose systems with integrated ventilation phases in the wine industry. Effective classification was facilitated by the feature selection technique, which reduced redundant information and dimensionality. Utilizing the most promising methods, we achieved accuracy exceeding 95% with several algorithms, including Support Vector Machines, Discriminant Analysis, and Random Forests. The ability to distinguish between wines more accurately using this innovative method highlights the necessity of comprehensive approaches that integrate multiple classification techniques. These findings reinforce the quantitative results and underscore the practical applicability of these advanced classification techniques in accurately identifying and distinguishing between wine varieties. Future research should focus on further refining these models and exploring their applicability in real-world settings, ultimately contributing to the continuous improvement of food quality and safety standards. Declarations Funding: We would like to acknowledge the financial support provided by Fundació Clarós, which made this research possible. Conflicts of Interest : The authors declare no conflict of interest. Author Contribution Conceptualization and methodology, X.M. and A.P., software development M.B., I.M., X.M., F.E. Data curation, X.M., I.M. and A.P. Writing of main manuscript text by I.M., A.P., F.E. and X.M. and I.M and X.M. prepared figures 1-10. Project administration and supervision X.M., A.P. ,R.P, P.C.; funding acquisition X.M. , A.P. and R.P. All authors reviewed the manuscript. All authors have read and agreed to the published version of the manuscript.” Data Availability The dataset build by the authors and used for analysis of the Sherry Wines on this manuscript is available at: 10.5281/zenodo.15190410 References Aguiar, Mariana S., Coelho, André F. S. M. R., Almeida, Paulo J., & Santos, João Rodrigo. (2023). Fan Assisted Extraction of Volatile Carbonyl Compounds from Coffee Brews Based on the Full Evaporation Technique. Foods , 12 (18), 3389. https://doi.org/10.3390/foods12183389 Aguilera, Teodoro, Lozano, Jesús, Paredes, José A., Álvarez, Fernando J., & Suárez, José I. (2012). Electronic Nose Based on Independent Component Analysis Combined with Partial Least Squares and Artificial Neural Networks for Wine Prediction. Sensors , 12 (6), 8055–8072. https://doi.org/10.3390/s120608055 Aleixandre, M., Lozano, J., Gutiérrez, J., Sayago, I., Fernández, M. J., & Horrillo, M. C. (2008). Portable e-nose to classify different kinds of wine. Sensors and Actuators B: Chemical , 131 (1), 71–76. https://doi.org/10.1016/j.snb.2007.12.027 Chilo, Jose, Pelegri-Sebastia, Jose, Cupane, Maria, & Sogorb, Tomas. (2016). E-nose application to food industry production. IEEE Instrumentation & Measurement Magazine , 19 (1), 27–33. https://doi.org/10.1109/MIM.2016.7384957 Durán-Guerrero, Enrique, Castro, Remedios, de Valme García-Moreno, María, del Carmen Rodríguez-Dodero, María, Schwarz, Mónica, & Guillén-Sánchez, Dominico. (2021). Aroma of Sherry Products: A Review. Foods , 10 (4), 753. https://doi.org/10.3390/foods10040753 Echave, Javier, Barral, Marta, Fraga-Corral, Maria, Prieto, Miguel A., & Simal-Gandara, Jesus. (2021). Bottle Aging and Storage of Wines: A Review. Molecules , 26 (3), 713. https://doi.org/10.3390/molecules26030713 Friedman, Jerome, Hastie, Trevor, & Tibshirani, Rob. (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software , 33 (1), 1–22. Fuentes, Sigfredo, Summerson, Vasiliki, Viejo, Claudia Gonzalez, Tongson, Eden, Lipovetzky, Nir, Wilkinson, Kerry L., Szeto, Colleen, & Unnithan, Ranjith R. (2020). Assessment of Smoke Contamination in Grapevine Berries and Taint in Wines Due to Bushfires Using a Low-Cost E-Nose and an Artificial Intelligence Approach. Sensors , 20 (18), 5108. https://doi.org/10.3390/s20185108 Ghasemi-Varnamkhasti, Mahdi, Apetrei, Constantin, Lozano, Jesus, & Anyogu, Amarachukwu. (2018). Potential use of electronic noses, electronic tongues and biosensors as multisensor systems for spoilage examination in foods. Trends in Food Science & Technology , 80 , 71–92. https://doi.org/10.1016/j.tifs.2018.07.018 Ghorbanzadeh, Omid, Blaschke, Thomas, Gholamnia, Khalil, Meena, Sansar Raj, Tiede, Dirk, & Aryal, Jagannath. (2019). Evaluation of Different Machine Learning Methods and Deep-Learning Convolutional Neural Networks for Landslide Detection. Remote Sensing , 11 (2), 196. https://doi.org/10.3390/rs11020196 Hestness, Joel, Narang, Sharan, Ardalani, Newsha, Diamos, Gregory, Jun, Heewoo, Kianinejad, Hassan, Patwary, Md. Mostofa Ali, Yang, Yang, & Zhou, Yanqi. (2017). Deep Learning Scaling is Predictable, Empirically . Kim, Seung-Jean, Koh, K., Lustig, M., Boyd, Stephen, & Gorinevsky, Dimitry. (2007). An Interior-Point Method for Large-Scale -Regularized Least Squares. IEEE Journal of Selected Topics in Signal Processing , 1 (4), 606–617. https://doi.org/10.1109/JSTSP.2007.910971 Lozano, J., Arroyo, T., Santos, J. P., Cabellos, J. M., & Horrillo, M. C. (2008). Electronic nose for wine ageing detection. Sensors and Actuators B: Chemical , 133 (1), 180–186. https://doi.org/10.1016/j.snb.2008.02.011 Lozano, J., Santos, J. P., Aleixandre, M., Sayago, I., Gutierrez, J., & Horrillo, M. C. (2006). Identification of typical wine aromas by means of an electronic nose. IEEE Sensors Journal , 6 (1), 173–178. https://doi.org/10.1109/JSEN.2005.854598 Lu, Lin, Hu, Zhanqiang, Hu, Xianqiao, Li, Dan, & Tian, Shiyi. (2022). Electronic tongue and electronic nose for food quality and safety. Food Research International , 162 , 112214. https://doi.org/10.1016/j.foodres.2022.112214 Macedo, Leandro Levate, Corrêa, Jefferson Luiz Gomes, da Silva Araújo, Cintia, & Cardoso, Wilton Soares. (2023). Use of Ethanol to Improve Convective Drying and Quality Preservation of Fresh and Sucrose and Coconut Sugar-impregnated Strawberries. Food and Bioprocess Technology , 16 (10), 2257–2271. https://doi.org/10.1007/s11947-023-03066-5 Pardo-Calle, C., Segovia-Gonzalez, M. M., Paneque-Macias, P., & Espino-Gonzalo, C. (1970). An approach to zoning in the wine growing regions of “Jerez-Xérès-Sherry” and “Manzanilla-Sanlúcar de Barrameda” (Cádiz, Spain). Spanish Journal of Agricultural Research , 9 (3), 831–843. https://doi.org/10.5424/sjar/20110903-343-10 Preedy, Victor. (2016). Electronic noses and tongues in food science. (María Luz Rodríguez Méndez, Ed.; Elsevier). Rodríguez-Méndez, María L., Saja, José A. De, González-Antón, Rocio, García-Hernández, Celia, Medina-Plaza, Cristina, García-Cabezón, Cristina, & Martín-Pedrosa, Fernando. (2016). Electronic Noses and Tongues in Wine Industry. Frontiers in Bioengineering and Biotechnology , 4 . https://doi.org/10.3389/fbioe.2016.00081 Ruggieri, Luz, Cadena, Erasmo, Martínez-Blanco, Julia, Gasol, Carles M., Rieradevall, Joan, Gabarrell, Xavier, Gea, Teresa, Sort, Xavier, & Sánchez, Antoni. (2009). Recovery of organic wastes in the Spanish wine industry. Technical, economic and environmental analyses of the composting process. Journal of Cleaner Production , 17 (9), 830–838. https://doi.org/10.1016/j.jclepro.2008.12.005 Saldaña, César. (2022). El libro de los vinos de jerez . Editorial Almuzara. Santos, João Rodrigo, & Rodrigues, José A. (2020). Characterization of volatile carbonyl compounds in defective green coffee beans using a fan assisted extraction process. Food Control , 108 , 106879. https://doi.org/10.1016/j.foodcont.2019.106879 Sanz, JB, Liem, P., Sierra, ÁG, & Toste, ES. (2019). Jerez, manzanilla y montilla: vinos tradicionales de Andalucía . Abalon Books. Summerson, Vasiliki, Viejo, Claudia Gonzalez, Pang, Alexis, Torrico, Damir D., & Fuentes, Sigfredo. (2021). Assessment of Volatile Aromatic Compounds in Smoke Tainted Cabernet Sauvignon Wines Using a Low-Cost E-Nose and Machine Learning Modelling. Molecules , 26 (16), 5108. https://doi.org/10.3390/molecules26165108 Tan, Juzhong, & Xu, Jie. (2020). Applications of electronic nose (e-nose) and electronic tongue (e-tongue) in food quality-related properties determination: A review. Artificial Intelligence in Agriculture , 4 , 104–115. https://doi.org/10.1016/j.aiia.2020.06.003 Valcárcel-Muñoz, Manuel J., Guerrero-Chanivet, María, del Carmen Rodríguez-Dodero, María, de Valme García-Moreno, María, & Guillén-Sánchez, Dominico A. (2022). Analytical and Chemometric Characterization of Fino and Amontillado Sherries during Aging in Criaderas y Solera System. Molecules , 27 (2), 365. https://doi.org/10.3390/molecules27020365 Vera, L., Mestres, M., Boqué, R., Busto, O., & Guasch, J. (2010). Use of synthetic wine for models transfer in wine analysis by HS-MS e-nose. Sensors and Actuators B: Chemical , 143 (2), 689–695. https://doi.org/10.1016/j.snb.2009.10.027 Table 2 Table 2 is available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table2.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 18 May, 2025 Reviews received at journal 17 May, 2025 Reviewers agreed at journal 11 May, 2025 Reviewers agreed at journal 09 May, 2025 Reviews received at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 07 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers invited by journal 06 May, 2025 Editor assigned by journal 01 May, 2025 Submission checks completed at journal 14 Apr, 2025 First submitted to journal 10 Apr, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6421640","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":454008218,"identity":"5b1d9f9e-e6c3-4bdb-aa1d-956dd4b45d81","order_by":0,"name":"Itziar Mengual","email":"","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":false,"prefix":"","firstName":"Itziar","middleName":"","lastName":"Mengual","suffix":""},{"id":454008220,"identity":"eaaada91-b463-4f74-a351-62a1f5fb82cf","order_by":1,"name":"Xavier Marimon","email":"","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":false,"prefix":"","firstName":"Xavier","middleName":"","lastName":"Marimon","suffix":""},{"id":454008224,"identity":"839034d8-246d-40f6-874d-110b34426dd3","order_by":2,"name":"Ferran Esquinas","email":"","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":false,"prefix":"","firstName":"Ferran","middleName":"","lastName":"Esquinas","suffix":""},{"id":454008225,"identity":"d5e42ede-3f92-4742-b1fa-a8fe30acb258","order_by":3,"name":"Marina Batllo","email":"","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":false,"prefix":"","firstName":"Marina","middleName":"","lastName":"Batllo","suffix":""},{"id":454008226,"identity":"925603a6-5ced-4a01-8aee-3ae358dde87e","order_by":4,"name":"Albert Coll","email":"","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":false,"prefix":"","firstName":"Albert","middleName":"","lastName":"Coll","suffix":""},{"id":454008227,"identity":"fb33e080-9642-47e8-a5af-5ccf37abb489","order_by":5,"name":"Josep Roca","email":"","orcid":"","institution":"Celler de Can Roca","correspondingAuthor":false,"prefix":"","firstName":"Josep","middleName":"","lastName":"Roca","suffix":""},{"id":454008228,"identity":"0a932e9a-261d-4c76-bc93-b370bb337641","order_by":6,"name":"Roman Perez","email":"","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":false,"prefix":"","firstName":"Roman","middleName":"","lastName":"Perez","suffix":""},{"id":454008229,"identity":"a9b88105-cb84-4aac-b0f7-655066b5a12b","order_by":7,"name":"Heloise Vilaseca","email":"","orcid":"","institution":"Celler de Can Roca","correspondingAuthor":false,"prefix":"","firstName":"Heloise","middleName":"","lastName":"Vilaseca","suffix":""},{"id":454008230,"identity":"cc928fb5-371f-4625-b44f-c381f4b234fa","order_by":8,"name":"Pedro Claros","email":"","orcid":"","institution":"Fundacion Claros","correspondingAuthor":false,"prefix":"","firstName":"Pedro","middleName":"","lastName":"Claros","suffix":""},{"id":454008231,"identity":"632f2fba-5529-4986-a70e-8a81009f6f8f","order_by":9,"name":"Alejandro Portela","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvklEQVRIiWNgGAWjYBADGQb2BoYDJGhIYOBh4DlAshaJBCIV87f3Hnxc+cOGh3/mG8MDPxhs7AlqkThzLtnwTEIaj8TtHIODPQxpiQ0E9dzIMZNsSDjMwwDUcpiB4TBh58nffwPS8p9H/uYZkJb/hB1mcIMHpOUAD5AB0nKAkaDDDM/kJRs2pCXzGJ5JKzjYY5BM2C9yx88efNhgYycnd/zw5g8/KuwIO4wBGCPI7iRCA5qWUTAKRsEoGAVYAABiZjui2eJRuQAAAABJRU5ErkJggg==","orcid":"","institution":"Universitat Internacional de Catalunya","correspondingAuthor":true,"prefix":"","firstName":"Alejandro","middleName":"","lastName":"Portela","suffix":""}],"badges":[],"createdAt":"2025-04-10 15:38:21","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6421640/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6421640/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82459024,"identity":"c852f323-5e0d-44df-95fe-410e895f257c","added_by":"auto","created_at":"2025-05-11 13:36:58","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":83956,"visible":true,"origin":"","legend":"\u003cp\u003eBlock diagram of the Sensing board contained in the prototype e-nose. Two multiplexer ICs enable the acquisition of up to 16 sensors through the A/D conversion module embedded into the microcontroller board (not shown).\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/81d7087e5967d2c36a1d4ff6.jpg"},{"id":82458784,"identity":"e824f05d-94fb-4bc3-8368-3059a0979ef2","added_by":"auto","created_at":"2025-05-11 13:28:58","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":52708,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eA\u003c/strong\u003e. The e-nose, showcasing the internal components and wiring with a wine glass. \u003cstrong\u003eB\u003c/strong\u003e. 3D CAD model of the shutter designed to expose the sensing board to wine volatiles and the simulated scenario of the shutter coupled to the sensing board.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/5cfbbf7d6ebef9f92169ca8f.jpg"},{"id":82459022,"identity":"5694bfb4-4935-4c35-8818-bbd13e9c090a","added_by":"auto","created_at":"2025-05-11 13:36:58","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":45299,"visible":true,"origin":"","legend":"\u003cp\u003eE-nose 3D printed chassis supporting the electronic boards and enabling the volatile fingerprints of Sherry wines under analysis with N = 30 samples.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/9aa2cd3d11f49f26adec051e.jpg"},{"id":82458786,"identity":"1edc43fe-5eed-4c98-9b19-586fbd9d227f","added_by":"auto","created_at":"2025-05-11 13:28:58","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":42202,"visible":true,"origin":"","legend":"\u003cp\u003eProtocol followed to capture sensors signals from the setup to the raw signal data.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/b8e660442e4dcf8db40422f5.jpg"},{"id":82458790,"identity":"bc82c2bd-2b18-4fb0-b0a1-aed504a6c9c4","added_by":"auto","created_at":"2025-05-11 13:28:58","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":95651,"visible":true,"origin":"","legend":"\u003cp\u003eSensor output signals obtained from each wine type, along with the activation signals for the shutter gate and fan during the protocol. The protocol is divided into five distinct phases: (1) Pre, (2) Slope, (3) Plateau, (4) Ventilation, (5) Post. The Pre and Post phases are not included in the analysis.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/4d43b65f6fd266d24f240f9f.jpg"},{"id":82458802,"identity":"8592b946-3830-4979-a846-6978cb4ffce5","added_by":"auto","created_at":"2025-05-11 13:28:59","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":199148,"visible":true,"origin":"","legend":"\u003cp\u003eData acquisition from each wine type\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/24719b0d2f48641cf3e5f8a8.jpg"},{"id":82458807,"identity":"79933467-c1ee-4f38-8b5c-bf5ed5da1c4a","added_by":"auto","created_at":"2025-05-11 13:28:59","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":61998,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of the feature selection methodology using an Elastic-Net Model.\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/19b430bd007822f31b8ce3cd.jpg"},{"id":82458795,"identity":"bccd59a2-4a14-42c9-8cc8-61eb90b20e11","added_by":"auto","created_at":"2025-05-11 13:28:58","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":143495,"visible":true,"origin":"","legend":"\u003cp\u003eCorrelation Matrixes before (\u003cstrong\u003eA\u003c/strong\u003e) and after (\u003cstrong\u003eB\u003c/strong\u003e) the feature selection methodology using an Elastic-Net Model.\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/9e02c1a9b8f4f20df39bcf4c.jpg"},{"id":82458810,"identity":"f5a6e3f2-993f-4d91-80de-d201b9b47519","added_by":"auto","created_at":"2025-05-11 13:28:59","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":77098,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance derived from the Linear Support Vector Machine model for each set of features. Colors relate to the phase: Slope (blue), Plateau (green), Ventilation (orange).\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/2f4ede59b9751dc15bff34da.jpg"},{"id":82458794,"identity":"84c377e7-f1a2-4252-b2a0-dcef369f8c5f","added_by":"auto","created_at":"2025-05-11 13:28:58","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":67066,"visible":true,"origin":"","legend":"\u003cp\u003eDecision boundaries generated by a Linear Support Vector Machine model, illustrating the classification of wine varieties based on pairs of features from the Ventilation Phase.\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/2943564da24e6ff8ef1477ac.jpg"},{"id":82459600,"identity":"82f70bd2-1b23-4db6-ab83-962161aa9176","added_by":"auto","created_at":"2025-05-11 13:52:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1851340,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/2fef699d-beb3-430e-8d6a-df8bed71ef3c.pdf"},{"id":82458783,"identity":"4783fdbe-e4fa-4ddb-bf2f-73c54d7ba011","added_by":"auto","created_at":"2025-05-11 13:28:58","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":52193,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-6421640/v1/f36b347ba3f656b3ec7cbda9.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Custom Electronic Nose design for Sherry Wines: Optimizing feature selection for machine learning-based classification","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSpain is one of the main wine producers over the world, with a yearly production of 44\u0026nbsp;million hectoliters (Ruggieri et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Amongst the different producers in the country, Jerez (Sherry) and San Lucar (Manzanilla) Region, in the south of Spain, produces one of the most highly appreciated wines due to the area conditions and their characteristic aging methods (Pardo-Calle et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e1970\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe aging process of these wines is characterized by a dynamic system called \u0026ldquo;\u003cem\u003ecriaderas y solera\u003c/em\u003e\u0026rdquo; or \u0026ldquo;\u003cem\u003esoleraje\u003c/em\u003e\u0026rdquo;. Here, the wines are stored in White American Oak (Quercus alba) butts and are organized in groups deppreviending on their aging stage, a method called \u0026ldquo;\u003cem\u003eescalas\u003c/em\u003e\u0026rdquo;. The bottled wine comes from the oldest group which receives the \u0026ldquo;solera\u0026rdquo; denomination; then, they begin with the extracting process, which is called \u0026ldquo;\u003cem\u003esaca\u003c/em\u003e\u0026rdquo;. After this, an equal amount of wine is extracted from the previous level of the \u0026ldquo;\u003cem\u003eescala\u003c/em\u003e\u0026rdquo; containing the second oldest wine \u0026ldquo;1st Criadera\u0026rdquo;. This way, each level is replenished by the previous level (the 2nd Criadera is replenished by the 3rd, the 3rd by the 4th and so on), in a process called \u0026ldquo;\u003cem\u003erocio\u003c/em\u003e\u0026rdquo; until the level that contains the youngest wine called \u0026ldquo;\u003cem\u003eSobretabla\u003c/em\u003e\u0026rdquo;. The final product is often a combination of related \u0026ldquo;\u003cem\u003esoleras\u003c/em\u003e\u0026rdquo; to obtain the desired oenological characteristics of the producer (Salda\u0026ntilde;a, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe objective of this system is to preserve and ensure the oenological properties and quality of the Sherry wine. The constant mixture of new and old wine during the \u0026ldquo;\u003cem\u003esacas\u003c/em\u003e\u0026rdquo; and \u0026ldquo;\u003cem\u003erocios\u003c/em\u003e\u0026rdquo; mitigates the effects of the different \u0026ldquo;\u003cem\u003ea\u0026ntilde;adas\u003c/em\u003e\u0026rdquo; that constitute the \u0026ldquo;\u003cem\u003esoleraje\u003c/em\u003e\u0026rdquo; providing a mean age for the resultant wine (Sanz et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis special treatment gives Sherry wines sensory characteristics whose variations and nuances are rarely matched in other parts of the wine world. Three different grape varieties are used in the production of this type of wine (Palomino, Pedro Xim\u0026eacute;nez and Moscatel), resulting in a wide variety of wines that clearly differ in aroma, flavor, texture, and color, having each one their own individual characteristics which need to be maintained and assured.\u003c/p\u003e \u003cp\u003eOne of the most used indicators to assess wine type, quality and production process is, precisely, aroma. Aroma comes from a vast number of volatile compounds which have been closely related to different wine properties (Vera et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Only on Sherry Wines, more than 150 different aromas have been identified (Dur\u0026aacute;n-Guerrero et al., 2021). For this reason, distinguishing wines through this characteristic is a highly complex task, normally developed by sommeliers, trained wine professionals capable of detecting wine characteristics through their aroma, color, taste, and density. However, this training is highly complex and costly, so depending on them to classify or assess wine quality can be an expensive task. Apart, other methods such as gas chromatography or mass spectrometry can be used to analyze wine volatiles, but they are time-consuming and labor-intensive.\u003c/p\u003e \u003cp\u003eThe electronic nose (e-nose) is composed by a set of volatile sensible sensors that pretends to mimic the olfactory perception of humans (Lozano et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2006\u003c/span\u003e).This tool has demonstrated to be especially useful in the food industry (Ghasemi-Varnamkhasti et al., 2018; Preedy, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), having a special impact on the wine sector. The e-nose functions by detecting different volatiles present in the environment and coding those presences into signals that can be then processed. Using different classifying techniques, this tool allows to distinguish the different characteristics from the products, resulting in an easy, fast, and cheap way to solve classification problems. Previous work has been done in the wine field, presenting this tool as a successful approach for many problems in quality assessment, wine classification, and aromatic compound detection (Aleixandre et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Chilo et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Fuentes et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Summerson et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Tan \u0026amp; Xu, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). However, until now, many of these classification techniques have required many samples or sensors to obtain useful results, making its applicability in real-world production complex (Lozano et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2008\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn this study, our goal is to classify varieties of Sherry wines using a simple and cost-effective artificial nose, also known as an e-nose. We aim to achieve this using a small number of samples and comparing various classification methods. The objective is to evaluate the accuracy of different data processing techniques to determine the most effective approach for this specific family of wines.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1. Sherry Wines\u003c/h2\u003e\n \u003cp\u003eWe selected three distinctive types of Sherry wines, namely Amontillado, Oloroso, and Pedro Xim\u0026eacute;nez. Each of these wines was chosen for its unique qualities, a result of their special production process, and wide availability. The use of three grape varieties\u0026mdash;Palomino, Pedro Xim\u0026eacute;nez, and Moscatel\u0026mdash;ensures a diverse range of wines with distinct aromas, flavors, textures, and colors.\u003c/p\u003e\n \u003cp\u003eAmontillado is a dry wine produced with Palomino grape first undergoing a biological process and finishing an oxidative environment (Valc\u0026aacute;rcel-Mu\u0026ntilde;oz et al., \u003cspan class=\"CitationRef\"\u003e2022\u003c/span\u003e). During the early aging stages, the decaying layer prevents the wine from oxidation and the through the fermentation process ethanol is combined with oxygen to produce acetaldehyde (Echave et al., \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e). In the later stages of the aging process the yeast layer disappears, and the wine is exposed to oxidation. The result is a wine with a color between topaz and amber with a smell reminiscent of aromatic herbs.\u003c/p\u003e\n \u003cp\u003eOloroso is a dry wine produced with a full-bodied Palomino must and its ageing process revolves around oxidation. Different from Amontillado, there is no yeast layer (Valc\u0026aacute;rcel-Mu\u0026ntilde;oz et al., \u003cspan class=\"CitationRef\"\u003e2022\u003c/span\u003e). This entails a wine with a higher degree of alcohol (around 18.5%) and a dark gold color. Lastly, Pedro Xim\u0026eacute;nez is a sweet wine produced with an overly ripe grape of the same name. Its aging process is oxidative but, unlike Oloroso, due to its density and sugar content the ethanol decreasing rate by evaporation is faster than the water loss due to osmosis (Macedo et al., \u003cspan class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2. The e-nose\u003c/h2\u003e\n \u003cp\u003eDifferent types of volatile sensors can be used on an e-nose based on its application (Aguilera et al., \u003cspan class=\"CitationRef\"\u003e2012\u003c/span\u003e). In our case, we aimed to obtain significant results using a selected matrix of sensors which produce a change on the output when exposed to the selected samples of Sherry wines. We conducted an analysis of the volatile components in Sherry wines to identify the most suitable sensor matrix, although it is expected none of them has a specific response to these components. Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e provides a summary of the selected sensors.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003e\u003cem\u003eCharacteristics of each gas sensor.\u003c/em\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eId\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eGas Sensor\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTarget Gases\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDetection Range (ppm)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMP-5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePropane, LPG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e300\u0026thinsp;~\u0026thinsp;10,000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMP-4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMethane, Natural gas\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e300\u0026thinsp;~\u0026thinsp;10,000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMP-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePropane, smoke\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e200\u0026thinsp;~\u0026thinsp;10,000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU3\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMP-7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u0026thinsp;~\u0026thinsp;1,000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMP-9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCO, Methane\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u0026thinsp;~\u0026thinsp;1,000 CO\u003c/p\u003e\n \u003cp\u003e300\u0026thinsp;~\u0026thinsp;10,000 CH4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU6\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMP-901\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVOC, alcohol, smoke\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u0026thinsp;~\u0026thinsp;50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU8\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTGS-2600\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHydrogen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u0026thinsp;~\u0026thinsp;30\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU9\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTGS-2602\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVOCs, ammonia, Thiols (H\u003csub\u003e2\u003c/sub\u003eS)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u0026thinsp;~\u0026thinsp;30 ppm EtOH\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eU10\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTGS-2620\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAlcohol, Solvent vapors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u0026thinsp;~\u0026thinsp;5,000 ppm EtOH\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eTo acquire signals from the specified matrix, we have designed an electronic circuit (Sensing board), which can accommodate up to 16 MOX sensors. The block diagram shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e also includes two multiplexer analog switches (CD74HC4051E, Texas Instruments, USA). These switches allow selective data collection from a smaller set of sensors through configurable firmware running on a control board connected through the IO interface.\u003c/p\u003e\n \u003cp\u003eThe sensing board additionally includes switching components, specifically a pair of solid-state relays (G6K-2P-Y-DC5, Omron Electronics, Japan), to control up to two external pumps (fresh air and vacuum). This provides additional flexibility, enabling two different configurations for handling the airflow of the sample\u0026rsquo;s headspace. The board contains an integrated circuit for measuring temperature and humidity (SHT30-ARP-B, Sensirion, Switzerland). The design includes components to ensure noise rejection and high-precision voltage regulators for supplying correspondingly the heating element and the reference voltage of the MOX sensors.\u003c/p\u003e\n \u003cp\u003eThe proposed e-nose also requires an embedded control board which has been implemented using an Arduino-UNO board containing an ATmega328P microprocessor (Microchip Technology, USA) which allows the acquisition through an embedded 10-bits A/D converter of the multiplexed gas sensor outputs. The microcontroller board is connected via USB a to a custom-made MATLAB script (R2023b MathWorks, USA) running on a PC, which allows the storing and post-processing of the data acquired from the e-nose. Both the sensing board and microprocessor board are placed inside a custom mechanical chassis, which also incorporates a fan (02510SS-12P-AT-00, NMB Technologies, USA) mounted at the top and controlled through a switch module (IR520, Vishay Siliconix, USA).\u003c/p\u003e\n \u003cp\u003eThe fan enables primarily flushing the sample\u0026acute;s headspace exposed to the sensors after every measurement cycle. The fan serves as dual purpose: to desaturate the sample and facilitating the cleaning process. Contrasting with studies by Aguiar et al. (Aguiar et al., \u003cspan class=\"CitationRef\"\u003e2023\u003c/span\u003e) and Markelov et al. (Santos \u0026amp; Rodrigues, \u003cspan class=\"CitationRef\"\u003e2020\u003c/span\u003e) which showed that fan-assisted extraction enhances volatile detection and increases spectrometer peaks. While Aguiar et al. demonstrated that the fan accelerates volatile compound transport and improves headspace flushing, our application benefits from the fan\u0026apos;s ability to promote dynamic headspace, reducing sample saturation and facilitating a cleaner analysis environment for subsequent measurements.\u003c/p\u003e\n \u003cp\u003eTo address these challenges, we designed a 3D-printed funnel that allows to expose the sample headspace uniformly to the sensors placed on the PCB board and when closed, it isolates the airspace exposed to the sensors by flushing it with fresh air.\u003c/p\u003e\n \u003cp\u003eWhen working with volatile sensors, two primary challenges arise: avoiding or detecting possible contaminants and ensuring adequate detection of volatiles of interest. To address these challenges, we designed a 3D-printed funnel that allows for uniform exposure of the sample headspace to the sensors placed on the PCB board. When closed, it isolates the airspace exposed to the sensors by flushing it with fresh air enabling.\u003c/p\u003e\n \u003cp\u003eThis capsule not only isolates the sensors but also facilitates the concentration of volatiles, such as those present in a glass of wine. This concentration process generates a distinctive peak in the sensor signal, enhancing sample identification. The funnel incorporates a shutter mechanism controlled by the firmware programmed on the Arduino board that enable the exposure of the sample headspace at specific points during the acquisition process. The CAD design of the capsule is depicted in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eAll these components were embedded in a 3D-printed structure, designed, and fabricated to support the electronic boards, the fan and other possible components as pumps that could be employed on different configurations. The design shown Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, has been developed using a CAD design software (SolidWorks 2023, USA) and 3D-printed (HP Jet Fusion MJF 5200, USA) using PA12 Nylon-12 powder (Sinterit, Poland). The design allows to accommodate a wine glass cup Riedel Ouverture (Riedel, Austria) and a shutter to control the flow of the headspace of the sample toward the Sensing board.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3. Sample acquisition\u003c/h2\u003e\n \u003cp\u003eDuring the sample acquisition process, a MATLAB code was implemented to interface with the microprocessor, enabling data collection. Figure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e presents a graphic representation of the sample preparation, detailing the process starting with the e-nose connection to the computer, to the collection of raw data from each wine measurement cycle.\u003c/p\u003e\n \u003cp\u003eSamples were obtained by pouring 20 ml of each type of wine into a wine glass. The 3D-printed device was then sealed on top of the glass using Parafilm M sealing film (Bemis, USA). The sample was maintained at a constant temperature of 5\u0026deg;C using a temperature-controlled bath (Gilson Co., USA). The measurement procedure began with the shutter closed. Approximately 100 s later, the gate was opened, exposing the e-nose to wine volatiles for 300 s. After 200 s of the shutter opening, the fan was turned on until the end of the measurement cycle. Following the activation of the fan, the shutter remained open for another 100 s before closing. The measurement cycle concluded 100 s after the shutter\u0026apos;s closure, resulting in a total duration of 600 s, with a sampling frequency of \u003cem\u003eF\u003c/em\u003e\u003csub\u003e\u003cem\u003es\u003c/em\u003e\u003c/sub\u003e= 4.7 Hz. Each sensor sample was taken five times to enhance value stability, with a repetitive frequency of \u003cem\u003eF\u003c/em\u003e\u003csub\u003e\u003cem\u003erep\u003c/em\u003e\u003c/sub\u003e= 0.2 Hz.\u003c/p\u003e\n \u003cp\u003eA total of 10 samples were collected for each type of wine, with each sample comprising five distinct phases. The experiment begins with the pre-phase, during which the shutter is closed, allowing volatile compounds in the wine to accumulate in the headspace. Next, the slope phase initiates when the gate opens, resulting in a huge rise in sensor values as volatiles reach the sensor board. This is followed by the plateau phase, where volatile saturation causes the sensor readings to stabilize. The ventilation phase then begins, with the fan activating to desaturate the sample while the shutter remains open. Finally, the post-phase concludes the experiment, closing the shutter and initiating the cleaning process. The pre and post phases served as baselines. Figure \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e illustrates the sensor values for each wine sample, along with the corresponding step signals from the shutter gate and fan activation.\u003c/p\u003e\n \u003cp\u003eFollowing the acquisition of samples, they were saved as .csv files and subsequently imported into MATLAB to construct a signal dataset. This dataset comprised 30 wine samples along with their respective labels. Each sample was generated from the outputs obtained from the matrix of 9 sensors. Figure \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e summarizes the data from 10 cycle repetitions for each wine type, excluding the headspace flush signal between exposures, with all cycles for one wine type completed before proceeding to the next.\u003c/p\u003e\n \u003cp\u003eDuring the repeating cycles, a slight increase trend was observed in the sensor baseline, which was particularly distinguishable for sensors U4 (MP-9) and U3 (MP-7). The target gas for both sensors is CO in the range of 50 to 1,000 ppm. The increase in both sensors emphasizes the sensitivity to CO. Using the initial trial values as a baseline, we corrected the offset drift for each sensor by calculating the difference between these baseline values and the subsequent measurements.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4. Feature Extraction and Selection\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e demonstrates the similarity in the sensor responses across the three types of wine samples, highlighting the challenge in distinguishing patterns. Conventional statistical methods are insufficient because they struggle to manage large, intricate datasets and identify subtle patterns. This demands the use of Machine Learning (ML) techniques, which excel at handling such complex data and uncovering critical patterns for accurate classification.\u003c/p\u003e\n \u003cp\u003eWhen using traditional Machine Learning for classification, the first step is to extract different features from the signal. The extracted features can vary depending on the type of signal, although there are some typical ones that are commonly used. In our case, we decided to extract three different types of features, using the most common ones inside each group. All the extracted features can be seen on Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eIn Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, nineteen features are listed for each sensor. The experiment comprises three main phases: slope, plateau, and ventilation. These phases yield variations on sensor output signals, which requires separated analysis to extract valuable insights for each phase. Each sensor and phase provide 19 variables, leading to 171 variables per phase and a total of 513 variables per wine sample. This helps to prevent the model from capturing noise and random fluctuations in the data, a problem known as overfitting, and optimize the results analysis by reducing inefficiencies caused by highly correlated and redundant features.\u003c/p\u003e\n \u003cp\u003eTo address this, regression techniques offer penalties that serve dual purposes: reducing overfitting by assigning very low values close to zero to some features (\u003cem\u003eL\u003c/em\u003e\u003csub\u003e2\u003c/sub\u003e penalty, Ridge Regression) and promoting feature selection by multiplying some feature coefficients by zero (\u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e penalty, Lasso Regression). In our analysis, we utilized Elastic Net, which combines both penalties to leverage their strengths optimally. This approach reduces the weight of certain variables (\u003cem\u003eL\u003c/em\u003e\u003csub\u003e2\u003c/sub\u003e) while also eliminating some of them to mitigate the impact of highly correlated features (\u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e), ensuring a balanced feature selection process.\u003c/p\u003e\n \u003cp\u003eFigure 7 provides a schematic overview of the methodology employed for feature selection. The hyperparameters used were the \u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e ratio and the alpha coefficient (\u003cem\u003e\u0026alpha;\u003c/em\u003e). The \u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e ratio determines the balance between \u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e (Lasso) and \u003cem\u003eL\u003c/em\u003e\u003csub\u003e2\u003c/sub\u003e (Ridge) penalties, where a value of 0 indicates purely Ridge Regression and a value of 1 indicates purely Lasso Regression. The \u003cem\u003e\u0026alpha;\u003c/em\u003e coefficient, on the other hand, controls the degree of regularization applied, with common values typically close to zero.\u003c/p\u003e\n \u003cp\u003eWe utilized the ElasticNetCV function(Friedman et al., \u003cspan class=\"CitationRef\"\u003e2010\u003c/span\u003e; Kim et al., \u003cspan class=\"CitationRef\"\u003e2007\u003c/span\u003e) from the Scikit-Learn library in Python (Python 3.12.2, Python Software Foundation, USA). The \u003cem\u003e\u0026alpha;\u003c/em\u003e parameter was set to [0.01, 0.1, 1.0, 10.0], and the \u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e ratio was set to [0.5, 0.7, 0.9, 1]. Lasso Regression was emphasized due to its effectiveness in reducing dataset dimensionality by eliminating irrelevant variables.\u003c/p\u003e\n \u003cp\u003eThe Elastic Net feature selection was applied separately to each of the three phases and collectively to the combined set of features from all phases. This process reduced the number of variables from 513 to 31 overall: three temporal features from the slope phase, 10 from the plateau phase (nine temporal and one spectral), and 18 from the ventilation phase (14 temporal, two spectral, and two non-linear). The optimal hyperparameters for the combined analysis were an \u003cem\u003e\u0026alpha;\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.1 and an \u003cem\u003eL\u003c/em\u003e\u003csub\u003e1\u003c/sub\u003e ratio of 0.5, indicating equal use of both penalties.\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e shows the correlation matrix using a heatmap, illustrating the state of the data before (A) and after (B) dimensionality reduction. The heatmap before the reduction (A) reveals a pattern in the correlations, indicating a repetitive set of values that convey similar information. It is important to note that the relevance lies not in the specific variables but in the overall pattern and insights provided through the colors.\u003c/p\u003e\n \u003cp\u003eFor the individual phases, the feature selection results and optimal hyperparameters were as follows:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003e\n \u003cp\u003eSlope phase: 16 features selected (14 temporal and two spectral) with an \u003cem\u003e\u0026alpha;\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.1 and an L1 ratio of 0.5.\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003ePlateau phase: 37 features selected (17 temporal, 10 spectral, and 10 non-linear) with an \u003cem\u003e\u0026alpha;\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.01 and an L1 ratio of 0.7.\u003c/p\u003e\n \u003c/li\u003e\n \u003cli\u003e\n \u003cp\u003eVentilation phase: 9 features selected (seven temporal and two non-linear) with an \u003cem\u003e\u0026alpha;\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.1 and an L1 ratio of 1.0.\u003c/p\u003e\n \u003c/li\u003e\n \u003c/ul\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e2.5. Classification Techniques\u003c/h2\u003e\n \u003cp\u003eWe evaluated the classification accuracy of various ML algorithms for differentiating wine samples. The algorithms included decision trees, linear discriminant analysis, logistic regression, Na\u0026iuml;ve Bayes, support vector machines, k-nearest neighbors, ensemble models, and traditional neural networks. After feature extraction and selection, we normalized features using \u003cem\u003ez\u003c/em\u003e-scores for models that required scaling transformation. For instance, decision trees do not require transformation due to their node-based structure, whereas support vector machines do, as they rely on geometrical distances to construct hyperplanes.\u003c/p\u003e\n \u003cp\u003ePrior to each analysis, the samples were randomly divided into training (80%) and testing (20%) datasets, stratifying the target variable (type of wine) to preserve the same proportions of examples in each wine class as observed in the original data. All analyses were conducted using five-fold cross-validation. Accuracy was selected as the metric for comparing results between the test and training data, as well as across different sets of features. Four different analyses were conducted: initially, each algorithm was applied separately to individual groups of phases (slope, plateau, or ventilation); subsequently, an integrated analysis incorporating all phases was conducted.\u003c/p\u003e\n \u003cp\u003eAnother method explored for signal classification was deep learning, which allows the classification of different samples without needing prior feature extraction. Neural networks are computational models inspired by the human brain, consisting of interconnected nodes (neurons) organized in layers. They process input data to produce outputs and learn by adjusting weights and biases using techniques like backpropagation and gradient descent. A growing type of neural network used to classify signals is the Long-Short-Term Memory Network (LSTM), which are a specialized type of recurrent neural networks (RNNs). Unlike traditional RNNs, LSTMs effectively manage long-term dependencies by incorporating memory cells that maintain information over extended periods and are commonly used for sequential data classification. In this study, an LSTM network architecture was used, composed of an input layer (No. of sensors 9), a bidirectional LSTM layer, a fully connected layer (100 units), an exponential normalized function (SoftMax) layer, and a classification layer. The LSTM network was trained with the following settings: a maximum of 25 epochs, a mini-batch size of 10, an initial learning rate of 0.01, and a gradient threshold of 1. Of the 30 signals from each type of acquisition, 25 were used for training. Several evaluation metrics: accuracy, F1-score, recall, and reproducibility were computed for the classification methods.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3. Results and Discussion","content":"\u003cp\u003eAs depicted in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e several algorithm evaluation metrics were obtained for the validation and test datasets across the different training groups. These results provide a comprehensive overview of the performance of each classification method throughout the various phases of the experiment.\u003c/p\u003e\n\u003cp\u003eThe validation and test metrics were presented to demonstrate the efficacy of the algorithms. An increase in test results compared to validation results suggests that the model generalizes well to all data rather than overfitting to the training data. The highest values observed in the test metrics can be attributed to the small number of samples used (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;7). In future studies, expanding the size of the training and test datasets may enhance the validation metrics (Ghorbanzadeh et al., \u003cspan class=\"CitationRef\"\u003e2019\u003c/span\u003e; Hestness et al., 2017).\u003c/p\u003e\n\u003cp\u003eThe features derived from the \u003cem\u003eslope phase\u003c/em\u003e performed consistently across various algorithms, including Decision Trees and SVMs. The accuracy of the test set ranged from 83\u0026ndash;100%, and the accuracy of the validation set ranged from 88\u0026ndash;96%. The features from the \u003cem\u003eplateau phase\u003c/em\u003e also showed strong performance, particularly with Decision Trees and Ensemble Bagged Trees, achieving accuracy over 95% in validation and between 83% and 100% in test results. Features from the\u0026nbsp;\u003cem\u003eventilation phase\u003c/em\u003e performed better with some classical ML algorithm namely SVMs (Linear and Quadratic), KNNs and NNs. When analyzing the set with features from all phases together, NN demonstrated the strongest performance, achieving validation accuracy scores ranging from 95.6\u0026ndash;97.8%. However, this performance was not maintained for all NN when evaluated using the test set, indicating some overfitting to the validation data. The LSTM network achieved an accuracy of 90%, which, although lower than traditional ML methods, is still notable given the small sample size. Deep Learning techniques and Neural Networks may excel with more complex and larger datasets, suggesting their potential value in future research.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAccuracy results from the Validation Model across phases using the corresponding extracted features.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"15\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 26.6576%;\"\u003e\n \u003cp\u003eMachine Learning Algorithms\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"12\" style=\"width: 36.0192%;\"\u003eAccuracy (%)\u0026nbsp;\u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003ePhases\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 9.6861%;\"\u003e\n \u003cp\u003eSlope Features\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\" style=\"width: 9.6861%;\"\u003e\n \u003cp\u003ePlateau Features\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 11.5903%;\"\u003e\n \u003cp\u003eVentilation Features\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 9.6861%;\"\u003e\n \u003cp\u003eAll Features\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eDatasets\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003eValidation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003eValidation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003eValidation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003eValidation\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"3\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eTree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eFine\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eMedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eCoarse\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eDiscriminant Analysis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eLinear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"2\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eEfficient\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e72,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e52,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e32,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e40,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e66,7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eKernel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e56,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eKernel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e76,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e72,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"7\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eSupport\u003c/p\u003e\n \u003cp\u003eVector\u003c/p\u003e\n \u003cp\u003eMachine\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eLinear\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eQuadratic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eCubic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eFine Gaussian\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e64,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e32,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e40,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eMedium Gaussian\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e72,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eCoarse gaussian\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e56,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e66,7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e44,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e60,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e66,7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e52,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eKernel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e72,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"6\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003ek-Nearest Neighbors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eFine\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e76,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eMedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e48,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e64,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e66,7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eCoarse\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eCosine\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e68,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eCubic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e44,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e68,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eWeighted\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e68,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"5\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eEnsemble\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eBoosted Trees\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eBagged Trees\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e76,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eSubspace Discriminant\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eSubspace KNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e72,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e52,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e68,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e50,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e44,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eRUSBoosted Trees\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e36,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e40,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e48,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" rowspan=\"5\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eTraditional Neural Networks\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eNarrow\u003c/p\u003e\n \u003cp\u003e(10 layers)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eMedium\u003c/p\u003e\n \u003cp\u003e(25 layers)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eWide (100 layers)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e80,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eBilayered\u003c/p\u003e\n \u003cp\u003e(10, 10 layers)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003eTrilayered\u003c/p\u003e\n \u003cp\u003e(10, 10, 10 layers)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e92,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e84,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e83,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e96,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e33,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e88,0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e100%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 14.7362%;\"\u003e\n \u003cp\u003eLSTM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 12.0042%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.2853%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.3877%;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6.0435%;\"\u003e\n \u003cp\u003e93,3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 3.6427%;\"\u003e\n \u003cp\u003e93,4%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eWhile the Ensemble Subspace Discriminant achieved over 90% validation accuracy, its test accuracy dropped to 83.3%, suggesting overfitting. Linear SVMs, however, achieved 100% test accuracy (84.0%-92.0% validation), making it the top performer. This improvement suggests that the hyperplane constructed to distinguish between wine varieties based on their features is robust and generalizes well to new, unseen test data. Consequently, future focus will be directed towards deepening our understanding of this algorithm. Looking deeply into the Linear SVM algorithm, we can identify which features played a significant role in wine variety discrimination. For each phase, ElasticNetCV was used to select relevant features. In the slope phase, where volatile compounds reach the sensors, 16 features were identified. The plateau phase, characterized by the saturation of volatile compounds, included 37 features. Finally, during the ventilation phase, 9 features were selected to desaturate the sample.\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e highlights the top features from each Linear SVM model, ranked according to their weight in determining the model\u0026apos;s hyperplane. While not all features have been described in detail, the most relevant ones have been highlighted to emphasize their importance and explain why they are effective in differentiating the wine varieties.\u003c/p\u003e\n\u003cp\u003eThe extracted features highlighted the importance of temporal features, which played a significant role in the analysis. Most of the features shown in the bar plots are temporal, such as RMS, Standard Deviation, Peak Value, and Mean. Non-linear and spectral features, like Band Power, Approximation Entropy, and Correlation Dimension, were used less frequently, likely due to the nature of the signals in this study and the larger number of temporal features computed. Nevertheless, all features were essential for the effective implementation of the algorithm. The most relevant features identified include Peak Value, determined as the maximum sensor value; RMS, defined as the square root of the sensor\u0026apos;s mean square; and Mean, defined as the average sensor value.\u003c/p\u003e\n\u003cp\u003eThe analysis of Sherry wines using gas sensors MP-9 (U4) and TGS-2620 (U10) has revealed significant insights into the differentiation of Amontillado, Oloroso, and Pedro Xim\u0026eacute;nez. Most of the key features identified are from these sensors, as indicated by their higher importance in the bar plot. The MP-9 sensor (U4) targets CO and methane, gases that can be present in trace amounts due to the fermentation and oxidative processes involved in wine aging. The presence of CO is linked to various VOCs produced during these processes. The TGS-2620 sensor (U10) targets alcohol and solvent vapors, which are prominent in wines due to the ethanol and other aromatic compounds formed during fermentation and oxidation.\u003c/p\u003e\n\u003cp\u003eDuring the slope phase, where vapor concentrations are increasing, the peak value of the MP-9 sensor is particularly important as it captures the highest concentration of gases like CO and methane. These gases are released at varying rates by different Sherry wines due to their oxidative properties and specific volatiles emitted during the aging process. Additionally, the standard deviation of the MP-5 sensor (U0) which targets propane gas reflects the variability in hydrocarbon concentration, which is influenced by the different grape varieties and aging methods, making it a crucial feature for distinguishing between the wines.\u003c/p\u003e\n\u003cp\u003eDuring the plateau phase, the TGS-2620 sensor\u0026apos;s peak value, mean, and RMS are essential metrics, reflecting the maximum concentration, average presence, and overall variability of alcohol and aromatic compounds, respectively, which are critical for understanding the consistent presence of these compounds in the vapor phase. The Pedro Xim\u0026eacute;nez wine, known for its high residual sugar content, has an alcohol level of 15.5%, which may enhance the volatilization of compounds detectable by the TGS-2620 sensor. In contrast, Amontillado, a dry wine with 17.5% alcohol, and Oloroso, with higher alcohol content, 20%, likely have fewer detectable volatiles. The higher concentration of volatile aromatic compounds, such as esters and aldehydes, in the sweet Pedro Xim\u0026eacute;nez wine makes these compounds more easily detectable by the sensor compared to the dry wines.\u003c/p\u003e\n\u003cp\u003eIn the ventilation phase, characterized by decreasing vapor concentrations, the skewness of the MP-9 -Skew(U4)- sensor measures the asymmetry of the vapor distribution, providing insight into the rate and manner in which each wine releases its volatile compounds. The dense and sweet Pedro Xim\u0026eacute;nez, with its high sugar content, shows variability in skewness, exhibiting both positive and negative values due to differences in air fluctuations and evaporation dynamics. In contrast, the drier Amontillado, with 17.5% alcohol content and low residual sugar, consistently exhibits negative skewness, reflecting a rapid initial burst of alcohol vapors followed by a quick decline. Oloroso, with the highest alcohol content at 20%, demonstrates positive skewness, indicating a significant initial burst of alcohol vapors. These patterns, depicted in Fig. \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003e with Amontillado points in the negative skewness region, Oloroso points in the positive skewness region, and Pedro Xim\u0026eacute;nez points spread across both regions, highlight the influence of residual sugar and alcohol content on the volatile compound release during ventilation.\u003c/p\u003e\n\u003cp\u003eMoreover, the TGS-2620 sensor\u0026apos;s peak value (\u003cem\u003ey\u003c/em\u003e-axis from Fig. \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003eA) and RMS (\u003cem\u003ey\u003c/em\u003e-axis from Fig. \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003eB) provide additional critical metrics for distinguishing between the wines. The sweet and dense Pedro Xim\u0026eacute;nez, despite having a lower alcohol content, exhibits higher values in both RMS and peak due to its high sugar content, which enhances the volatilization of compounds. In contrast, Amontillado and Oloroso, both dry wines with higher alcohol content, show lower values in RMS and peak, reflecting their fewer detectable volatiles compared to Pedro Xim\u0026eacute;nez.\u003c/p\u003e\n\u003cp\u003eAs demonstrated in Fig. \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003e, the decision boundaries in the 2D plots illustrate the effectiveness of the Linear Support Vector Machine models using features from the ventilation phase. This visualization confirms the model\u0026apos;s capability to distinguish between different wine varieties based on the extracted features. However, it also highlights some misclassified samples, indicating limitations when using only two-dimensional feature spaces. The results table further emphasizes the necessity of a sensor matrix and the inclusion of various relevant features to significantly improve classification accuracy. This finding reinforces the quantitative results and underscores the practical applicability of these advanced classification techniques in accurately identifying and distinguishing between wine varieties.\u003c/p\u003e\n\u003cp\u003eThe data presented highlights the effectiveness of our chosen machine learning models in classifying wine samples based on their chemical compositions. Building on these findings, it is important to consider the broader implications and potential applications of these technologies in the food industry. These techniques have proven particularly relevant for quality, assessment (Tan \u0026amp; Xu, \u003cspan class=\"CitationRef\"\u003e2020\u003c/span\u003e) flavor production, and overall product manufacturing (Chilo et al., \u003cspan class=\"CitationRef\"\u003e2016\u003c/span\u003e). Consequently, it is essential to thoroughly explore the capabilities and performance of these technologies to maximize their potential in various food industry applications.\u003c/p\u003e\n\u003cp\u003eWines are indeed a significant part of the food industry, and the need for producers to maintain high-quality standards makes this product an ideal candidate for e-nose studies. Previous research has demonstrated various applications of machine learning in the wine industry, including detecting smoke contamination (Fuentes et al., \u003cspan class=\"CitationRef\"\u003e2020\u003c/span\u003e), identifying specific aromatic compounds, (Summerson et al., \u003cspan class=\"CitationRef\"\u003e2021\u003c/span\u003e) and classifying different types of wine (Aleixandre et al., \u003cspan class=\"CitationRef\"\u003e2008\u003c/span\u003e). However, these devices often focus on detecting very specific compounds or, in the case of classification, achieve relatively low performance with complex components and rely on a single type of classification technique. Therefore, there is a need for more comprehensive approaches that integrate multiple classification techniques to enhance the accuracy and reliability of wine quality assessments. Our study introduces a novel approach by incorporating a ventilation phase with the use of a fan, which enhances the distinction between wine varieties. This method effectively leverages the differences in evaporation dynamics and volatile compound release, significantly improving classification accuracy. The ability to distinguish between wines more accurately using this enhanced technique underscores the practical applicability and potential of using advanced e-nose systems with integrated ventilation phases in the wine industry. Future research should focus on further refining these models and exploring their applicability in real-world settings, (Rodr\u0026iacute;guez-M\u0026eacute;ndez et al., \u003cspan class=\"CitationRef\"\u003e2016\u003c/span\u003e) ultimately contributing to the continuous improvement of food quality and safety standards (Lu et al., \u003cspan class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eOur study highlights the significant advancements achieved by using machine learning models to classify wine samples based on their chemical compositions. This research underscores the broader implications and potential applications of these technologies in the food industry. Wines, as a crucial component of the food sector, are particularly well-suited for e-nose studies. While previous research has shown various applications of machine learning in the wine industry\u0026mdash;such as detecting smoke contamination, identifying specific aromatic compounds, and classifying different types of wine\u0026mdash;these studies often rely on single classification techniques and achieve relatively low performance due to their complexity and narrow focus.\u003c/p\u003e \u003cp\u003eOur study introduces a novel approach by incorporating a ventilation phase with the use of a fan, which significantly enhances the distinction between wine varieties. This method effectively leverages the differences in evaporation dynamics and volatile compound release, leading to improved classification accuracy. The dense and sweet Pedro Xim\u0026eacute;nez, despite having lower alcohol content, exhibited higher values in both RMS and peak due to its high sugar content. In contrast, the drier Amontillado and Oloroso wines, with higher alcohol content, showed lower values in RMS and peak, reflecting their fewer detectable volatiles. This enhanced technique underscores the practical applicability and potential of using advanced e-nose systems with integrated ventilation phases in the wine industry. Effective classification was facilitated by the feature selection technique, which reduced redundant information and dimensionality. Utilizing the most promising methods, we achieved accuracy exceeding 95% with several algorithms, including Support Vector Machines, Discriminant Analysis, and Random Forests.\u003c/p\u003e \u003cp\u003eThe ability to distinguish between wines more accurately using this innovative method highlights the necessity of comprehensive approaches that integrate multiple classification techniques. These findings reinforce the quantitative results and underscore the practical applicability of these advanced classification techniques in accurately identifying and distinguishing between wine varieties. Future research should focus on further refining these models and exploring their applicability in real-world settings, ultimately contributing to the continuous improvement of food quality and safety standards.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eWe would like to acknowledge the financial support provided by Fundaci\u0026oacute; Clar\u0026oacute;s, which made this research possible.\u003c/p\u003e \u003cp\u003e \u003cb\u003eConflicts of Interest\u003c/b\u003e: The authors declare no conflict of interest.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization and methodology, X.M. and A.P., software development M.B., I.M., X.M., F.E. Data curation, X.M., I.M. and A.P. Writing of main manuscript text by I.M., A.P., F.E. and X.M. and I.M and X.M. prepared figures 1-10. Project administration and supervision X.M., A.P. ,R.P, P.C.; funding acquisition X.M. , A.P. and R.P. All authors reviewed the manuscript. All authors have read and agreed to the published version of the manuscript.\u0026rdquo;\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe dataset build by the authors and used for analysis of the Sherry Wines on this manuscript is available at: 10.5281/zenodo.15190410\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAguiar, Mariana S., Coelho, Andr\u0026eacute; F. S. M. R., Almeida, Paulo J., \u0026amp; Santos, Jo\u0026atilde;o Rodrigo. (2023). Fan Assisted Extraction of Volatile Carbonyl Compounds from Coffee Brews Based on the Full Evaporation Technique. \u003cem\u003eFoods\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(18), 3389. https://doi.org/10.3390/foods12183389\u003c/li\u003e\n\u003cli\u003eAguilera, Teodoro, Lozano, Jes\u0026uacute;s, Paredes, Jos\u0026eacute; A., \u0026Aacute;lvarez, Fernando J., \u0026amp; Su\u0026aacute;rez, Jos\u0026eacute; I. (2012). Electronic Nose Based on Independent Component Analysis Combined with Partial Least Squares and Artificial Neural Networks for Wine Prediction. \u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(6), 8055\u0026ndash;8072. https://doi.org/10.3390/s120608055\u003c/li\u003e\n\u003cli\u003eAleixandre, M., Lozano, J., Guti\u0026eacute;rrez, J., Sayago, I., Fern\u0026aacute;ndez, M. J., \u0026amp; Horrillo, M. C. (2008). Portable e-nose to classify different kinds of wine. \u003cem\u003eSensors and Actuators B: Chemical\u003c/em\u003e, \u003cem\u003e131\u003c/em\u003e(1), 71\u0026ndash;76. https://doi.org/10.1016/j.snb.2007.12.027\u003c/li\u003e\n\u003cli\u003eChilo, Jose, Pelegri-Sebastia, Jose, Cupane, Maria, \u0026amp; Sogorb, Tomas. (2016). E-nose application to food industry production. \u003cem\u003eIEEE Instrumentation \u0026amp; Measurement Magazine\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(1), 27\u0026ndash;33. https://doi.org/10.1109/MIM.2016.7384957\u003c/li\u003e\n\u003cli\u003eDur\u0026aacute;n-Guerrero, Enrique, Castro, Remedios, de Valme Garc\u0026iacute;a-Moreno, Mar\u0026iacute;a, del Carmen Rodr\u0026iacute;guez-Dodero, Mar\u0026iacute;a, Schwarz, M\u0026oacute;nica, \u0026amp; Guill\u0026eacute;n-S\u0026aacute;nchez, Dominico. (2021). Aroma of Sherry Products: A Review. \u003cem\u003eFoods\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(4), 753. https://doi.org/10.3390/foods10040753\u003c/li\u003e\n\u003cli\u003eEchave, Javier, Barral, Marta, Fraga-Corral, Maria, Prieto, Miguel A., \u0026amp; Simal-Gandara, Jesus. (2021). Bottle Aging and Storage of Wines: A Review. \u003cem\u003eMolecules\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(3), 713. https://doi.org/10.3390/molecules26030713\u003c/li\u003e\n\u003cli\u003eFriedman, Jerome, Hastie, Trevor, \u0026amp; Tibshirani, Rob. (2010). Regularization Paths for Generalized Linear Models via Coordinate Descent. \u003cem\u003eJournal of Statistical Software\u003c/em\u003e, \u003cem\u003e33\u003c/em\u003e(1), 1\u0026ndash;22.\u003c/li\u003e\n\u003cli\u003eFuentes, Sigfredo, Summerson, Vasiliki, Viejo, Claudia Gonzalez, Tongson, Eden, Lipovetzky, Nir, Wilkinson, Kerry L., Szeto, Colleen, \u0026amp; Unnithan, Ranjith R. (2020). Assessment of Smoke Contamination in Grapevine Berries and Taint in Wines Due to Bushfires Using a Low-Cost E-Nose and an Artificial Intelligence Approach. \u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e(18), 5108. https://doi.org/10.3390/s20185108\u003c/li\u003e\n\u003cli\u003eGhasemi-Varnamkhasti, Mahdi, Apetrei, Constantin, Lozano, Jesus, \u0026amp; Anyogu, Amarachukwu. (2018). Potential use of electronic noses, electronic tongues and biosensors as multisensor systems for spoilage examination in foods. \u003cem\u003eTrends in Food Science \u0026amp; Technology\u003c/em\u003e, \u003cem\u003e80\u003c/em\u003e, 71\u0026ndash;92. https://doi.org/10.1016/j.tifs.2018.07.018\u003c/li\u003e\n\u003cli\u003eGhorbanzadeh, Omid, Blaschke, Thomas, Gholamnia, Khalil, Meena, Sansar Raj, Tiede, Dirk, \u0026amp; Aryal, Jagannath. (2019). Evaluation of Different Machine Learning Methods and Deep-Learning Convolutional Neural Networks for Landslide Detection. \u003cem\u003eRemote Sensing\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(2), 196. https://doi.org/10.3390/rs11020196\u003c/li\u003e\n\u003cli\u003eHestness, Joel, Narang, Sharan, Ardalani, Newsha, Diamos, Gregory, Jun, Heewoo, Kianinejad, Hassan, Patwary, Md. Mostofa Ali, Yang, Yang, \u0026amp; Zhou, Yanqi. (2017). \u003cem\u003eDeep Learning Scaling is Predictable, Empirically\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eKim, Seung-Jean, Koh, K., Lustig, M., Boyd, Stephen, \u0026amp; Gorinevsky, Dimitry. (2007). An Interior-Point Method for Large-Scale -Regularized Least Squares. \u003cem\u003eIEEE Journal of Selected Topics in Signal Processing\u003c/em\u003e, \u003cem\u003e1\u003c/em\u003e(4), 606\u0026ndash;617. https://doi.org/10.1109/JSTSP.2007.910971\u003c/li\u003e\n\u003cli\u003eLozano, J., Arroyo, T., Santos, J. P., Cabellos, J. M., \u0026amp; Horrillo, M. C. (2008). Electronic nose for wine ageing detection. \u003cem\u003eSensors and Actuators B: Chemical\u003c/em\u003e, \u003cem\u003e133\u003c/em\u003e(1), 180\u0026ndash;186. https://doi.org/10.1016/j.snb.2008.02.011\u003c/li\u003e\n\u003cli\u003eLozano, J., Santos, J. P., Aleixandre, M., Sayago, I., Gutierrez, J., \u0026amp; Horrillo, M. C. (2006). Identification of typical wine aromas by means of an electronic nose. \u003cem\u003eIEEE Sensors Journal\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e(1), 173\u0026ndash;178. https://doi.org/10.1109/JSEN.2005.854598\u003c/li\u003e\n\u003cli\u003eLu, Lin, Hu, Zhanqiang, Hu, Xianqiao, Li, Dan, \u0026amp; Tian, Shiyi. (2022). Electronic tongue and electronic nose for food quality and safety. \u003cem\u003eFood Research International\u003c/em\u003e, \u003cem\u003e162\u003c/em\u003e, 112214. https://doi.org/10.1016/j.foodres.2022.112214\u003c/li\u003e\n\u003cli\u003eMacedo, Leandro Levate, Corr\u0026ecirc;a, Jefferson Luiz Gomes, da Silva Ara\u0026uacute;jo, Cintia, \u0026amp; Cardoso, Wilton Soares. (2023). Use of Ethanol to Improve Convective Drying and Quality Preservation of Fresh and Sucrose and Coconut Sugar-impregnated Strawberries. \u003cem\u003eFood and Bioprocess Technology\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(10), 2257\u0026ndash;2271. https://doi.org/10.1007/s11947-023-03066-5\u003c/li\u003e\n\u003cli\u003ePardo-Calle, C., Segovia-Gonzalez, M. M., Paneque-Macias, P., \u0026amp; Espino-Gonzalo, C. (1970). An approach to zoning in the wine growing regions of \u0026ldquo;Jerez-X\u0026eacute;r\u0026egrave;s-Sherry\u0026rdquo; and \u0026ldquo;Manzanilla-Sanl\u0026uacute;car de Barrameda\u0026rdquo; (C\u0026aacute;diz, Spain). \u003cem\u003eSpanish Journal of Agricultural Research\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(3), 831\u0026ndash;843. https://doi.org/10.5424/sjar/20110903-343-10\u003c/li\u003e\n\u003cli\u003ePreedy, Victor. (2016). \u003cem\u003eElectronic noses and tongues in food science. \u003c/em\u003e(Mar\u0026iacute;a Luz Rodr\u0026iacute;guez M\u0026eacute;ndez, Ed.; Elsevier).\u003c/li\u003e\n\u003cli\u003eRodr\u0026iacute;guez-M\u0026eacute;ndez, Mar\u0026iacute;a L., Saja, Jos\u0026eacute; A. De, Gonz\u0026aacute;lez-Ant\u0026oacute;n, Rocio, Garc\u0026iacute;a-Hern\u0026aacute;ndez, Celia, Medina-Plaza, Cristina, Garc\u0026iacute;a-Cabez\u0026oacute;n, Cristina, \u0026amp; Mart\u0026iacute;n-Pedrosa, Fernando. (2016). Electronic Noses and Tongues in Wine Industry. \u003cem\u003eFrontiers in Bioengineering and Biotechnology\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e. https://doi.org/10.3389/fbioe.2016.00081\u003c/li\u003e\n\u003cli\u003eRuggieri, Luz, Cadena, Erasmo, Mart\u0026iacute;nez-Blanco, Julia, Gasol, Carles M., Rieradevall, Joan, Gabarrell, Xavier, Gea, Teresa, Sort, Xavier, \u0026amp; S\u0026aacute;nchez, Antoni. (2009). Recovery of organic wastes in the Spanish wine industry. Technical, economic and environmental analyses of the composting process. \u003cem\u003eJournal of Cleaner Production\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(9), 830\u0026ndash;838. https://doi.org/10.1016/j.jclepro.2008.12.005\u003c/li\u003e\n\u003cli\u003eSalda\u0026ntilde;a, C\u0026eacute;sar. (2022). \u003cem\u003eEl libro de los vinos de jerez\u003c/em\u003e. Editorial Almuzara.\u003c/li\u003e\n\u003cli\u003eSantos, Jo\u0026atilde;o Rodrigo, \u0026amp; Rodrigues, Jos\u0026eacute; A. (2020). Characterization of volatile carbonyl compounds in defective green coffee beans using a fan assisted extraction process. \u003cem\u003eFood Control\u003c/em\u003e, \u003cem\u003e108\u003c/em\u003e, 106879. https://doi.org/10.1016/j.foodcont.2019.106879\u003c/li\u003e\n\u003cli\u003eSanz, JB, Liem, P., Sierra, \u0026Aacute;G, \u0026amp; Toste, ES. (2019). \u003cem\u003eJerez, manzanilla y montilla: vinos tradicionales de Andaluc\u0026iacute;a\u003c/em\u003e. Abalon Books.\u003c/li\u003e\n\u003cli\u003eSummerson, Vasiliki, Viejo, Claudia Gonzalez, Pang, Alexis, Torrico, Damir D., \u0026amp; Fuentes, Sigfredo. (2021). Assessment of Volatile Aromatic Compounds in Smoke Tainted Cabernet Sauvignon Wines Using a Low-Cost E-Nose and Machine Learning Modelling. \u003cem\u003eMolecules\u003c/em\u003e, \u003cem\u003e26\u003c/em\u003e(16), 5108. https://doi.org/10.3390/molecules26165108\u003c/li\u003e\n\u003cli\u003eTan, Juzhong, \u0026amp; Xu, Jie. (2020). Applications of electronic nose (e-nose) and electronic tongue (e-tongue) in food quality-related properties determination: A review. \u003cem\u003eArtificial Intelligence in Agriculture\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e, 104\u0026ndash;115. https://doi.org/10.1016/j.aiia.2020.06.003\u003c/li\u003e\n\u003cli\u003eValc\u0026aacute;rcel-Mu\u0026ntilde;oz, Manuel J., Guerrero-Chanivet, Mar\u0026iacute;a, del Carmen Rodr\u0026iacute;guez-Dodero, Mar\u0026iacute;a, de Valme Garc\u0026iacute;a-Moreno, Mar\u0026iacute;a, \u0026amp; Guill\u0026eacute;n-S\u0026aacute;nchez, Dominico A. (2022). Analytical and Chemometric Characterization of Fino and Amontillado Sherries during Aging in Criaderas y Solera System. \u003cem\u003eMolecules\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e(2), 365. https://doi.org/10.3390/molecules27020365\u003c/li\u003e\n\u003cli\u003eVera, L., Mestres, M., Boqu\u0026eacute;, R., Busto, O., \u0026amp; Guasch, J. (2010). Use of synthetic wine for models transfer in wine analysis by HS-MS e-nose. \u003cem\u003eSensors and Actuators B: Chemical\u003c/em\u003e, \u003cem\u003e143\u003c/em\u003e(2), 689\u0026ndash;695. https://doi.org/10.1016/j.snb.2009.10.027\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table 2","content":"\u003cp\u003eTable 2 is available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-science-of-food","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjscifood","sideBox":"Learn more about [npj Science of Food](http://www.nature.com/npjscifood/)","snPcode":"41538","submissionUrl":"https://submission.springernature.com/new-submission/41538/3","title":"npj Science of Food","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Sherry Wines, Electronic nose, Metal-oxide sensors, Volatile organic compounds, Gas sensors, Electronic design, Feature extraction, Feature selection, Classification techniques, Machine learning, Deep Learning","lastPublishedDoi":"10.21203/rs.3.rs-6421640/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6421640/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe Electronic Nose (e-nose) is a growing tool widely used in food applications. This study classifies Sherry wines using an e-nose equipped with a matrix of gas sensors. By extracting features from identified phases of the sensor responses (slope, plateau, and ventilation) and applying the Elastic Net technique for feature selection, we improved machine learning (ML) algorithm performance while reducing dataset dimensionality. Our novel data acquisition approach incorporates a ventilation phase to enhance volatile detection, boosting wine classification accuracy. We analyzed features extracted from different phases of sensor response using ML models like Support Vector Machines, Discriminant Analysis, and Random Forests. High classification accuracy exceeding 95% was achieved, with Linear SVM demonstrating exceptional performance on test data, indicating strong generalization capabilities. Our findings demonstrate that the proposed e-nose design equipped with ML algorithms and efficient feature extraction methods offers a reliable solution for Sherry wine classification.\u003c/p\u003e","manuscriptTitle":"Custom Electronic Nose design for Sherry Wines: Optimizing feature selection for machine learning-based classification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-11 13:28:53","doi":"10.21203/rs.3.rs-6421640/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-18T20:37:17+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-17T21:01:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"153323974768641365275153166739574316390","date":"2025-05-11T15:32:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"14581221911532657515412432098423347466","date":"2025-05-09T05:34:22+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-08T23:47:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"191428062738306907850137539183246963851","date":"2025-05-08T18:00:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"15030258779381629161208380435306705922","date":"2025-05-08T06:31:49+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"256619030907505550162294002412772494435","date":"2025-05-08T05:01:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"193941053542094146659220775709379052503","date":"2025-05-08T03:08:17+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"34791776428652007017268469919389747226","date":"2025-05-07T02:45:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"288222595837256415511146736294982348339","date":"2025-05-07T01:22:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"273426240107090041224013980854837055343","date":"2025-05-07T00:30:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"14064258706805892056652600495352819473","date":"2025-05-06T19:28:53+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"218028055181553581311513672149495948780","date":"2025-05-06T18:37:40+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"332725681812058346832931588837526266402","date":"2025-05-06T18:25:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-06T17:31:23+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-01T17:49:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-14T15:20:39+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Science of Food","date":"2025-04-10T15:29:14+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-science-of-food","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjscifood","sideBox":"Learn more about [npj Science of Food](http://www.nature.com/npjscifood/)","snPcode":"41538","submissionUrl":"https://submission.springernature.com/new-submission/41538/3","title":"npj Science of Food","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7c4a575a-577f-460b-a198-3d65aa564a09","owner":[],"postedDate":"May 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":48280542,"name":"Physical sciences/Engineering"},{"id":48280545,"name":"Physical sciences/Engineering/Chemical engineering"},{"id":48280546,"name":"Physical sciences/Physics/Techniques and instrumentation"}],"tags":[],"updatedAt":"2025-05-18T20:53:10+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-11 13:28:53","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6421640","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6421640","identity":"rs-6421640","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}