Statistical Analysis for Selective Identifications of VOCs by Using Surface Functionalized MoS₂ Based Sensor Array ^†

Abstract

Disease diagnosis through breath analysis has attracted significant attention in recent years due to its noninvasive nature, rapid testing ability, and applicability for patients of all ages. More than 1000 volatile organic components (VOCs) exist in human breath, but only selected VOCs are associated with specific diseases. Selective identification of those disease marker VOCs using an array of multiple sensors are highly desirable in the current scenario. The use of efficient sensors and the use of suitable classification algorithms is essential for the selective and reliable detection of those disease markers in complex breath. In the current study, we fabricated a noble metal (Au, Pd and Pt) nanoparticle-functionalized MoS₂ (Chalcogenides, Sigma Aldrich, St. Louis, MO, USA)-based sensor array for the selective identification of different VOCs. Four sensors, i.e., pure MoS₂, Au/MoS₂, Pd/MoS₂, and Pt/MoS₂ were tested under exposure to different VOCs, such as acetone, benzene, ethanol, xylene, 2-propenol, methanol and toluene, at 50 °C. Initially, principal component analysis (PCA) and linear discriminant analysis (LDA) were used to discriminate those seven VOCs. As compared to the PCA, LDA was able to discriminate well between the seven VOCs. Four different machine learning algorithms such as k-nearest neighbors (kNN), decision tree, random forest, and multinomial logistic regression were used to further identify those VOCs. The classification accuracy of those seven VOCs using KNN, decision tree, random forest, and multinomial logistic regression was 97.14%, 92.43%, 84.1%, and 98.97%, respectively. These results authenticated that multinomial logistic regression performed best between the four machine learning algorithms to discriminate and differentiate the multiple VOCs that generally exist in human breath.

1. Introduction

In the field of medical diagnostic and health care systems, breath analysis has gained a lot of interest for the noninvasive detection of diseases and monitoring of health parameters [1,2]. More than 1000 volatile organic components (VOCs) are present in exhaled breath, but only some of them are considered disease markers [3,4]. In this context, the selective detection of the different VOCs using smart sensor systems has a high demand for efficient breath analysis. Selective detection can also be achieved using suitable pattern recognition algorithms on sensor signals. For the early detection of disease, the combination of a highly selective sensor and an effective machine learning algorithm is required. Diagnostics through breath is less time consuming compared to the clinical process and, at the same time, it is cost-efficient as does not require well-trained professionals and the sensors are less costly [5,6].

Chemiresistive sensors typically recognize a target VOC by changing their resistance depending upon the adsorption-desorption properties of the analyte to the detecting layer surface. An extensive variety of materials are used for VOC sensing, including thin metal films [7], metal oxides [8,9,10], polymers [11], etc. Accessible surface functionalization possibilities, high surface area to volume ratio, increased flexibility, high sensitivity, and tunable bandgap make two-dimensional molybdenum disulfide (MoS₂) an encouraging channel material to sense the VOC [12,13].

A pattern recognition algorithm also plays an essential role in the detection of VOC. A suitable classifier is required to achieve an effective classification rate in VOC sensing based on the sensor data. Different algorithms such as partial least squares discriminant analysis [14], canonical discriminant analysis [15], k-nearest neighbor [4,16], Discriminant function analysis [17], support vector machine [18], random forest [19], logistic regression [20], etc. were reported in the literature. In some of the reported literature, different types of neural network classifiers were used [21,22,23,24].

In the current study, we used principal component analysis (PCA) and linear discriminant analysis (LDA) to visualize the data in lesser dimensions compared to the original extent. Furthermore, four different supervised algorithms, k-nearest neighbor (kNN), decision tree, random forest, and multinomial logistic regression, were implemented to identify the best-suited algorithm based on their performance.

2. Material and Methods

2.1. Preparation of MoS₂ and Noble Metal Nanoparticles Solutions

All materials MoS₂ (Chalcogenides, Sigma Aldrich, St. Louis, MO, USA), gold (III) chloride (AuCl₃, 99%, Sigma Aldrich), palladium chloride (PdCl₂, 60%, Molychem, Mumbai, India), and chloroplatinic acid (H₄PtCl₆xH₂O, 40%, Molychem) were analytical grade and used without any further purification. The 0.2 wt% MoS₂ solution was prepared in deionized water and stirred for 1.5 h at room temperature to maintain homogeneity. Similarly, 0.1 MM aqueous solutions of noble metal nanoparticles (Au, Pd, Pt) were prepared by adding corresponding metal salts in deionized water with continuous stirring and dropwise diluted HCl was also added to obtain stable and uniform nanoparticles at room temperature.

Au, Pd and Pt nanoparticle-loaded MoS₂ samples were prepared using the spray coating technique. Firstly, MoS₂ solution was spray coated on washed SiO₂/Si substrate and dried at room temperature. In the final step, nanoparticle solutions were spray-coated on previously deposited MoS₂ and dried at room temperature.

Thermal annealing was performed for 4 h at 250 °C to provide crystallization and thermal stability in all 4 samples (MoS₂, Au-MOS₂, Pd-MoS₂, and Pt-MoS₂). The MoS₂ flakes were coated by very tiny Au NPs (12 nm), whereas the larger Pd NPs (54 nm) were deposited, maintaining a consistent spacing. Pt nanoparticles on the MoS₂ surface had the biggest size (77 nm) of the three.

2.2. Fabrication of Sensors

Au source and drain electrodes of 150 nm thickness were deposited on all four samples by using an electron beam evaporation unit. Sensors were then placed into a sensor holder, and further sensing performance was studied.

The sensor holder was placed in a glass sealed sensing chamber of 650 mL on a heating plate. The sensing performance of prepared sensors was examined by a static mode sensing setup where VOCs were injected using micro syringes (Hamilton micro syringe). The sensor was recovered by flowing 450 SCCM synthetic air by using a mass flow controller. The amount of injected VOC was calculated by using the formula: C (ppm) = 2.46 × (V1D/VM) × 103, where D (gm/mL), M (gm/mol), and V (L) represent density of the VOC, molecular weight of the VOC and volume of vaporization chamber, respectively [13,25,26]. Seven different VOCs, i.e., acetone, 2-propanol, benzene, ethanol, methanol, toluene, and xylene were tested during the study. Using a Keithley 6487 source meter, sensing performance was recorded applying 1 V constant bias. The sensitivity of the sensor was calculated by formula; Ra − Rv/Ra × 100 where Ra and Rv were the resistances of the sensor in the air and in target VOC.

To read the generated output of sensors stored in CSV file a python script was used. All the algorithms, analysis, and plotting were performed on Python 3.7 and Jupiter notebook as a platform.

3. Results and Discussion

3.1. VOC Sensing

As a reference ambient, synthetic air was used to perform the gas sensing measurements of the four different sensors: pure MoS₂, Au-MoS₂, Pd-MoS_2, and Pt-MoS₂. Figure 1 shows the change in resistance (MΩ) with respect to time at 50 °C. In the presence of VOCs, as the exposer time increased, the resistance offered by the sensor decreased. This decrease in resistance confirms that the sensor is an n-type property. In the presence of seven distinct VOCs, i.e., acetone, 2-propanol, benzene, ethanol, methanol, toluene, and xylene, four different sensors, i.e., pure MoS₂, Au-MoS₂, Pd-MoS₂, and Pt-MoS₂, were observed and stored for further processing of data.

3.2. Data Analysis

Figure 2 describes the influence of the volatile organic components (VOCs) on the outcomes of two-dimensionality reduction techniques: principal component analysis and linear discriminant analysis. The measurement parameters were kept constant during the experiment. The operating temperature was 50 °C, the response was taken up to 600 s. and the sample concentration was 100 ppm.

The response obtained by the four different sensors for seven different VOCs was used for principal component analysis (PCA). The three-dimensional plot between the first principal component (PC1), second principal component (PC2), and third principal component (PC3) is represented in Figure 2. As we have four independent variables (sensor responses), the maximum principal component obtained was four. Therefore, we have considered only the first three principal components contributing the most to the explained variance in this analysis. The total explained variance was 93.58%, in which PC1 contributed 52.52%, PC2 contributed 30.91%, and PC3 contributed 10.14%. All seven VOCs had their compact clusters, and they separated. Still, the separation between the cluster of acetone/2-propanol and benzene/toluene was quite small, which increased the possibility of misclassification.

Taking account of the problem of discrimination between the different VOCs, a linear discrimination analysis was also performed. The same sensor response vector was used in the linear discriminant analysis (LDA). Figure 2b shows that the employment of the classifier allowed the discrimination of all seven VOCs. Thus, LDA is highly efficient for investigating the VOCs based on the sensor response. A three-dimensional plot is shown in Figure 2b, which clearly depicts the performance of the LDA on the raw data (sensor response vector). The different VOCs were densely clustered within their groups, and they were well separated from each other. So, there was a significantly lower probability of misclassification among the VOCs. The 2-propanol was slightly more stretched along the second linear discriminant function (LD2) axis, and xylene was along the third discriminant function (LD3). The three discriminant functions, LD1, LD2, and LD3 contributed 71.22%, 27.42% and 1.21%, respectively, the total resultant explained the variance for the classifier becoming 99.85%.

3.3. VOC Identification

The previously discussed LDA and PCA plot gives only the visual representation of the separation of VOCs based on the sensor response. The goal of the sensor setup is to design a generalized model based on the known data during the training phase and try to predict the class when an unknown data sample is encountered.

The supervised algorithm was performed in the current work to determine the VOCs; four different machine learning algorithms such as k-nearest neighbor (kNN), decision tree, random forest, and multinomial logistic regression were used to identify those seven VOCs. The normalized sensor response was fed to the algorithms, and the whole data set was divided into training testing data with 70% and 30%, respectively. The data set consisted of 4200 measurements of each sensor, with each class containing 600 data vectors and seven classes. So, 2940 vectors were used to train the model, and the remaining 1260 vectors were used to test the model. For identification of VOCs, above 84% was the classification accuracy for every classifier with an accuracy of 97.14%, 92.43%, 84.1%, 98.97% for kNN, decision tree, random forest, and multinomial logistic regression, respectively. A confusion matrix is used to calculate the classification accuracy, and the confusion matrix furnishes the observation into what components were mistakenly classified. Figure 3a shows the confusion matrix of kNN where 11 samples of toluene were classified as xylene and 10 samples of benzene were wrongly predicted as ethanol. Figure 3b is a representation of the confusion matrix obtained from the decision tree algorithm. The confusion matrix of the random forest and multinomial logistic regression are shown in Figure 3c,d, respectively. In multinomial logistic regression, only 12 benzene samples were identified as acetone, and one sample of xylene was identified as toluene.

4. Conclusions

The ability of a surface-functionalized MoS₂ sensor to distinguish between the various VOCs was appraised by PCA and LDA, in which LDA laid out the excellent separation between the VOCs. Further, to evaluate the effectiveness of the sensor output to identify the VOCs, four different machine learning-based (supervised) classification algorithms were implemented, and among them, the k-nearest neighbor and multinomial logistic regression were performed outstandingly with an accuracy of 97.14% and 98.97%, respectively. Thus, high selectivity and accuracy authenticate that the system discriminates and differentiates the multiple VOCs that generally exist in human breath.