Abstract
Studies have found differences in the concentration of volatile organic compounds in the breath of diabetics and healthy people, prompting attention to the use of devices such as electronic noses to detect diabetes. In this study, we explored the design of a non-invasive diabetes preliminary screening system that uses a homemade electronic nose sensor array to detect respiratory gas markers. In the algorithm part, two feature extraction methods were adopted, gradient boosting method was used to select promising feature subset, and then particle swarm optimization algorithm was introduced to extract 24 most effective features, which reduces the number of sensors by 56% and saves the system cost. Respiratory samples were collected from 120 healthy subjects and 120 diabetic subjects to assess the system performance. Random forest algorithm was used to classify and predict electronic nose data, and the accuracy can reach 93.33%. Experimental results show that on the premise of ensuring accuracy, the system has low cost and small size after the number of sensors is optimized, and it is easy to install on in-car. It provides a more feasible method for the preliminary screening of diabetes on in-car and can be used as an assistant to the existing detection methods.
Introduction
Diabetes is a group of metabolic diseases characterized by hyperglycemia. In recent years, diabetes and its related diseases have become a worldwide health problem. In traditional medicine, patients prick their fingers to get a blood sample and measure glucose levels in the blood to diagnose diabetes, which is very painful. In recent years, respiratory analysis has attracted the attention of a large number of researchers because of its non-invasive, real-time, and wide application to various patients and diseases.
Studies have shown that different diseases have different characteristics of volatile organic compounds (VOCs). By analyzing the concentration of biomarkers in breath, we were able to achieve a preliminary detection of disease types (1). Related technologies such as gas chromatography-mass spectrometry and ion mobility spectroscopy can be used to analyze the components in patients' breath (2). However, the high cost, inconvenience to carry, and complex operation limit the large-scale application of some measuring instruments.
Electronic nose, as an instrument to imitate human smell, is low-cost and non-invasive and has attracted extensive attention in the detection of gas volatile biomarkers (3, 4). These technologies integrate artificial neural networks and improve existing clinical disease detection methods (5). Studies have shown that the concentration of acetone (6, 7) and some VOCs (8) in diabetic patients is abnormal. The relationship between blood glucose and acetone concentration in breath prompted the research team to design a specific electronic nose sensor (9) and develop a blood glucose value detection system in combination with traditional pattern recognition methods (10, 11, 12), which successfully distinguished chronic kidney disease, diabetes mellitus, and healthy subjects (13).
There are still some problems with the systems that have been developed. First, the devices are too bulky and difficult to carry around, especially in places with limited space such as the automobile cabin. In the application scenario of detecting driver's diabetes in the car, it is necessary to determine the installation position of the electronic nose in the car. According to the analysis of the driver's expiratory flow field and the optional location provided by the car factory, the electronic nose is best installed in the main driver's side door trim panel (14). The nose should be as small as possible because of the tight space in the door. On the premise of ensuring classification accuracy (15, 16), the sensor array should be further optimized to select sensors sensitive to diabetes respiratory detection.
According to the abovementioned problems, in order to make a cheap and reliable electronic nose respiratory analysis system that can be used indoors/in cars, this paper mainly starts from the following aspects: optimize the sensor array in a data-driven way, selection of the best feature values related to diabetes gas. The particle swarm optimization (PSO) algorithm is used to predict diabetes patients. Under the premise of ensuring accuracy, the detection equipment is further miniaturized and the detection energy consumption is reduced so that low-income families can afford and prolong the use of the equipment. In order to make the model more reliable, we enlarged the sample size to obtain the feature data of exhaled gas of people with multiple characteristics, making the model more applicable to a wide range of people. Compared with other papers, the home-use and vehicle-mounted small preliminary screening system proposed in this paper is portable and affordable, which can greatly benefit the patient population.
The rest of this paper is organized as follows: The second section introduces the materials and methods in detail, the third section is the results and discussion, and the last section concludes the paper.
Materials and methods
Electronic nose measuring device developed
In recent years, the electronic nose system developed in the laboratory has been widely used in the agricultural field (17, 18). Based on these studies, the respiratory sample collection system in this paper was developed for diabetes detection, as shown in Fig. 1.
Physical view of electronic nose system.
Citation: Endocrine Connections 12, 3; 10.1530/EC-22-0437
The system consists of 32 commercial gas sensors and their target gases are shown in Table 1. The main component sensor of the electronic nose uses a metal oxide semiconductor sensor. Due to the diversity of respiratory gas composition in diabetic patients, such as ethanol (19), carbon monoxide (20), alkanes (21), and methyl nitrate (22), sensors of different measurement ranges and different companies are used. These different sensors can form a complementary array that can help identify the disease being studied. By calculating the cost of purchasing the corresponding equipment, we can see that the total cost is about $674. The volume is about 7728 cubic centimeters. Before using the detection device, put the electronic nose device into the fume hood first and allow the sensor to warm up for 30 min. Then, the gas to be measured was fed into the bionic chamber attached with a sensor through an air pump with a flow rate of 1.2 L/min. When the gas entered the chamber, the data were sampled immediately, and the collection time was 1 min. When one set of gas collection is completed, the air pump draws fresh air to clean the residual gas in the chamber for 3–5 min to restore the sensor to the baseline level before the next set of sampling.
Summary of the sensor array.
| Number | The gas sensor | The response characteristics |
|---|---|---|
| S1 | TGS2612 | Butane, methane, propane. |
| S2 | TGS2611 | Methane, natural gas |
| S3 | TGS2620 | Vapors of organic solvents, ethanol |
| S4 | TGS2603 | Gaseous air contaminants, trimethylamine, methyl thiol, etc. |
| S5 | TGS2602 | Gaseous air contaminants, VOCs, ammonia, hydrogen sulfide, etc. |
| S6 | TGS2610 | Propane, butane |
| S7 | TGS2600 | Gaseous air contaminants, hydrogen, alcohol, etc |
| S8 | GSBT11 | Formaldehyde, oluene, butyric acid, butane, hydrocarbons |
| S9 | MS1100 | Formaldehyde, toluene, xylene |
| S10 | MP135 | Hydrogen, alcohol, carbonic oxide |
| S11 | MP901 | Alcohol, smoke, formaldehyde, toluene, acetone, benzene |
| S12 | MP-9 | Carbonic oxide, methane |
| S13 | MP-3B | Alcohol |
| S14 | MP-4 | Methane, natural gas, biogas |
| S15 | MP-5 | Liquefied petroleum gas |
| S16 | MP-2 | Propane, smoke |
| S17 | MP503 | Alcohol, smoke |
| S18 | MP801 | Benzene, toluene, formaldehyde, alcohol, smoke |
| S19 | MP905 | Benzene, toluene, formaldehyde, alcohol, smoke |
| S20 | MP402 | Methane, natural gas, biogas |
| S21 | WSP1110 | Nitrogen dioxide |
| S22 | WSP2110 | Toluene, benzene, formaldehyde, alcohol, etc. |
| S23 | WSP7110 | Sulfuretted hydrogen |
| S24 | MP-7 | Carbonic oxide |
| S25 | TGS2612 | Butane, methane, propane |
| S26 | TGS2611 | Methane, natural gas |
| S27 | TGS2620 | Vapors of organic solvents, combustible gases, methane, carbon monoxide, isobutane, hydrogen, ethanol |
| S28 | MP-3B | Alcohol |
| S29 | MP702 | Ammonia gas |
| S30 | TGS2610 | Propane, butane |
| S31 | TGS2600 | Gaseous air contaminants, methane, carbon monoxide, isobutane, ethanol, hydrogen |
| S32 | TGS2618-COO | Butane, LP gas |
Manufacturers: S1–S7, S25–S27, S30–S32 Figaro Engineering Inc, Minoh, Japan; S8 Orgam Technologies, Gwangju, Korea; S9–S24, S28–S29, Winsen Electronics Technology, Zhengzhou, China.
Breath sample collection
A total of 240 volunteer breath samples were collected in this experiment, including 120 diabetic samples and 120 non-diabetic samples. The volunteers' breath samples were taken in the morning when they had not eaten or exercised. Volunteers were asked to take deep breaths through a disposable suction nozzle into a 1 L Tedlar collection bag, with a one-way valve attached to the bag to prevent external air pollution. After the breath sample was collected, the collection bag was immediately sent to the electronic nose equipment for detection. The signals were processed by the electronic nose low-pass filter circuit and then collected by the data acquisition card. Finally, they were digitized and sent to the computer. The specific sample collection process is shown in Fig. 2. For each sample, we obtained a data matrix represented by 32 response curves, and each response curve had 60 s × 100 Hz = 6000 data points.
Breath sample classification
After the volunteers had fasted for at least 8 h, fasting blood samples were taken and venous plasma glucose concentrations were analyzed by glucose oxidase. Gas detection is also carried out. According to the diagnostic criteria recommended by the American Diabetes Association, blood glucose concentration was divided into two groups: diabetic group (fasting blood glucose ≥ 7.0 mmol/L) and non-diabetic group (fasting blood glucose < 7.0 mmol/L). The gas samples measured by the electronic nose were matched with the blood glucose values of the volunteers, respectively.
Based on the purpose of this study, a small non-invasive preliminary screening system for diabetes was established. The system aims to realize the detection of diabetes in family cars, more quickly and more convenient for the preliminary screening of diabetes. The system consists of three main parts: an electronic nose device to collect breath samples from patients, a data analysis model, and in-car deployment. A homemade sample collection device has been described in the section ‘electronic nose measuring device developed’, which collects patient breathing air through an array of electronic nose sensors. Due to the large number of sensors, on the premise of ensuring classification accuracy, we first need to optimize the sensor array through data processing to miniaturize the equipment and save costs.
Sensor array optimization
The data processing method is very important when building the diabetes non-invasive preliminary screening system. By extracting important features of the electronic nose response signal, a specific diabetes fingerprint can be formed, which can effectively distinguish the two groups of people. In this paper, 32 electronic nose sensors were used. In order to save costs, improve the recognition accuracy, and realize the miniaturization application on the car, after the feature extraction, it is necessary to carry out feature selection, remove redundant features and corresponding sensors, select the best sensor combination, improve the execution time and accuracy of the classifier, and realize pattern recognition. The specific sensor optimization steps were as follows.
Gas feature extraction
where x(t) is the original signal, u is the scaling parameter, w is the shifting parameter, ψ is the mother wavelet.
After signal pretreatment, the responses of 32 chemical sensors were spliced into a feature vector. Due to the high feature dimension (600 × 32 = 192,000), it is necessary to extract the features of each sensor's response separately. Feature extraction is to operate the data to obtain more information and ensure the effectiveness of the subsequent pattern recognition algorithm (24). In this paper, two feature extraction methods were used. The first method was to extract five steady-state and transient features from the original response of the sensor. They were mean value, standard deviation, kurtosis, skewness, and form factor, as shown in Table 2. The second method was feature extraction in the transform domain, which took the transform coefficients as the feature to distinguish the sample, and the feature extracted from the transform domain can stably reflect the inherent feature of the sample. In this paper, Fourier transform (25), which is commonly used in electronic nose data feature processing, was used to extract features.
Five time-domain features extracted from sensor response curves.
| Feature | Formula |
|---|---|
| Mean value (x̄ ) |  |
| Standard deviation (σ) |  |
| Skewness (Skew) |  |
| Kurtosis value (KV) |  |
| Form factor (FF) |  |
Using two methods, five features were extracted from each sensor response signal, and a total of 32 × 5 × 2 = 160 × 2 = 320 features were extracted to form 2 feature matrices of respiratory samples with 160 features.
Feature selection
This paper used PSO for feature extraction. Different from traditional feature selection methods, the PSO algorithm proposed by Kennedy and Eberhart (26), as a population-based algorithm, has been widely used in feature selection problems. There were a variety of VOCs in the breath of diabetic patients, and the signal features extracted from cross-sensitive sensors may have complex bidirectional and three-directional interactions. Traditional feature selection methods, which evaluate features independently, have separate redundant features and can be very useful when combined with other features, and vice versa. Compared with feature sorting, the feature subset selection method of PSO can deal with feature interaction better. The algorithm also has the advantages of easy implementation and strong global search ability.
where Vidt and Xidt represent the velocity and position of the d-dimension of the particle at time t, w represents inertia weight, Pidt and Pgdt represent pbest and gbest of the d-dimension of the particle at time t, c1, and c2 are acceleration constants, and r1 and r2 are random numbers uniformly distributed on (0,1).
Since the sensing of compounds by sensors used in the electronic nose has a partial crossover, the features obtained after feature extraction have a certain degree of redundancy. However, PSO is a global search algorithm, and the computational complexity of searching the optimal feature subset in the feature space composed of all features is high. In order to accelerate the feature selection process, this paper fuses PSO with the traditional ranking algorithm. First, the traditional ranking algorithm is used to rank the importance of features and select the promising feature subset. Then, PSO is used to select the selected feature subset twice and finally determine the best feature subset. Specifically, three sorting algorithms are used in this paper: EXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Extremely Randomized Trees (ExtraTrees) sort the features and select the promising subset of features. XGBoost is an open-source project based on Gradient Boosting Decision Tree (GBDT) proposed by Chen et al. (27), which introduces regularization and reduces the complexity of the Tree. LightGBM (28) is an improved GBDT framework model, which uses histogram segmentation algorithm to replace the traditional pre-sorting traversal algorithm, with faster parallel training speed and higher accuracy, and can effectively prevent over-fitting. ExtraTrees is an integrated learning algorithm, which contains many decision trees and the classification result is determined by the vote of many decision trees. This algorithm has a strong generalization ability and the ability to resist noise.
Classification
In the classification stage, features extracted from 32 sensors were divided into two data sets: training set and test set. The training set was used to build the classification model, and the test set was used to verify. In order to accurately predict diabetes patients and select the best gas recognition and classification method, classification methods were adopted in most studies (29, 30, 31): support vector machine (SVM), random forest (RF), K-nearest neighbor (KNN) were used in this paper for comparative analysis and evaluation of electronic nose performance.
The self-made electronic nose respiratory acquisition system was used to conduct model training on the collected respiratory samples, and the redundant and irrelevant features in the original feature vector were removed during training to optimize the sensor array. Finally, the optimized sensor was used to predict diabetes, which constituted the data processing part of the system. The algorithm framework diagram is shown in Fig. 3 After miniaturization, the electronic nose device can be placed in the family car, and the machine learning diagnostic model is combined with the on-board deployment to arrange the on-board screening system.
Results and discussion
The non-invasive diabetes preliminary screening system is essentially a classification problem, where the response signal was quantified into a respiratory fingerprint feature and then the classifier predicted the presence or absence of diabetes. Feature extraction and feature selection have a significant impact on the accuracy and efficiency of classification, and the classification accuracy depends on whether appropriate explanatory features are extracted. Feature selection can remove redundant features from respiratory fingerprint and delete redundant sensors, which can improve the robustness and reduce the cost of equipment. Optimized sensor devices are suitable for deployment in homes and even car parts.
Sensor array optimization results
Sensor array is a key component of the preliminary screening system of the electronic nose. A set of appropriate cross-sensitive VOC sensors can have good detection performance for gas markers of diabetes. The original average response radar diagram of 32 sensors is shown in Fig. 4, indicating differences in respiratory imprints between diabetic volunteers and non-diabetic volunteers.
Respiratory fingerprints for diabetes and non-diabetes.
Citation: Endocrine Connections 12, 3; 10.1530/EC-22-0437
The sensor array was further optimized, two features were extracted from the pre-processed data, and the extracted features were fused. Three sorting algorithms, XGBoost, LightGBM, and ExtraTrees, were used to remove redundant and irrelevant features from the original feature vectors, extract the top 60 common features of the three algorithms, and select feature vectors with statistical significance between the two types of samples. By analyzing the top 30 features of the three algorithms, it is found that the features extracted from sensor S13 account for the largest proportion, which may be related to the sensor's sensitivity to volatile compound ethanol. Figure 5 shows the top three sensors for each algorithm.
The number of features contained by the top three sensors in each algorithm.
Citation: Endocrine Connections 12, 3; 10.1530/EC-22-0437
PSO is used for feature selection. Data were divided into 2/3 training set and 1/3 test set, which had never been used in feature selection. After feature selection, the training set and test set were converted according to selected features, and classification algorithm was used to evaluate the performance of classification. Table 3 shows the parameter settings used in the experiment. The population size was set to 50 and the maximum number of iterations was set to 100 to avoid high computational costs and maintain the efficiency of PSO. The threshold value of selected features is usually set to 0.5 or slightly larger (0.5, 0.7) (32), which will not significantly affect the selection process of PSO. This paper chose a threshold of 0.6 for features in order to use a small number of features at an early stage of evolution (33). In PSO, RF is used to evaluate the selected feature subset, and grid search algorithm is used to find the optimal hyperparameters.
Parameter setting.
| Parameters | Setting |
|---|---|
| Population size | 50 |
| Maximum iterations | 200 |
| C1=C2 | 2 |
| Threshold for selected feature | 0.6 |
| ω | 0.9 |
| Communication topology | Fully connected (PSO) |
After a series of data processing procedures on collected breath samples, 14 sensors with high diabetes screening accuracy were selected for the final device. The performance of the selected sensors in distinguishing between healthy and diabetic samples is shown in Figure 6, which depicts mean maximum response of the 14 chemical sensors for each subset; error bars represent standard deviations. For most VOC sensors, the mean response of the diabetic samples was greater than that of the healthy samples, indicating that these sensors could distinguish between the concentration of VOCs in the breath of diabetic and healthy volunteers. The average response radar diagram of 14 sensors after optimization is shown in Fig. 7. It can be seen that the respiratory marks of diabetes and non-diabetes can still be distinguished.
Average response for each sensor in the two classes. The X-axis is the sensor indicator, the Y-axis is the mean of the maximum value of the pretreatment response, and the error bars represent the standard deviations.
Citation: Endocrine Connections 12, 3; 10.1530/EC-22-0437
Optimized respiratory fingerprints of diabetic and non-diabetic patients.
Citation: Endocrine Connections 12, 3; 10.1530/EC-22-0437
After optimization, the number of sensor arrays changed, and the number of sensors was reduced from 32 to 14, reducing the number of sensors by 56%. Due to the self-designed electronic nose preliminary screening system, when the number of sensors in the system is reduced, the number of acquisition cards and other related instruments and equipment will also be reduced, and the corresponding price will be reduced. By calculating the cost of purchasing corresponding equipment, it can be seen that the total cost is about $361, a decrease of about 42.98%. In addition, when the number of sensors is reduced, the spatial layout of the original sensors can be changed. A reasonable bionic spatial layout design of the sensor array can improve sensor sensitivity and detection performance of the electronic nose system and reduce the size of the electronic nose respiratory system (34). The optimized sensor array has a total volume of about 3400 cubic centimeters, reducing the volume by about 56% and miniaturizing the noninvasive detection system. Easy to fit in the car.
Accuracy of preliminary screening system for diabetes
Diabetes screening is done by distinguishing between healthy and diabetic samples. We first collected digital breath samples and then fused the two feature extraction methods through the data processing steps in the section ‘Materials and methods’. After signal pretreatment, a feature matrix of 320 features was constructed. After feature sorting and feature selection, the most representative 24 features were finally extracted. A pattern recognition method based on principal component analysis (PCA) was used to identify gaseous VOCs in 120 healthy volunteers and 120 diabetic patients to study the ability of features to distinguish health status. PCA results showed that 60.34% of the total variance of the dataset could be explained by PC 1, 2 and 3 (Fig. 8).
Three-dimensional principal component analysis was performed on breath samples from 240 volunteers using data collected by the electronic nose system.
Citation: Endocrine Connections 12, 3; 10.1530/EC-22-0437
Classification algorithm was further used to evaluate the accuracy of the preliminary screening system. Three commonly used classification algorithms were used for comparison, namely SVM with Gaussian kernel (35), RF and KNN (K = 3), and randomly select 30% data as the test set and 70% data as the training set. The accuracy of the classification algorithm further confirms that there are obvious differences between the two types of samples. The results presented in this section show all the methods and the tuning of the classifier to evaluate the best possible results. Run 30 times to calculate the average values of accuracy, precision, recall, and F1 score.The advantages and disadvantages of each algorithm are evaluated from different angles.
The original features were selected by using PSO, and the optimal feature subset was selected. Three classification algorithms were used to compare the results, as shown in Table 4. It can be seen that the RF algorithm using the PSO algorithm can achieve higher classification accuracy, higher recall rate, and higher F1 score, which has a significant advantage compared with other algorithms. Meanwhile, three classification algorithms are also used to evaluate the features without feature selection, as shown in Table 5. By comparing Tables 4 and 5, it can be seen that the highest accuracy is obtained by the RF algorithm before and after feature selection, and the accuracy increases from 88.7 to 93.33%. The classification algorithm showed that the maximum accuracy of the diabetes screening system could reach 93.33%, and the size of feature subset was reduced from 320 to 24. On the premise of ensuring accuracy, further miniaturization of detection equipment and reduction of detection energy consumption made it possible for home and vehicle diabetes screening.
Comparison of classification accuracy of respiratory fingerprint of healthy and diabetic patients after feature selection using PSO.
| Algorithm | Accuracy (%) | Precision (%) | Recall (%) | F1 score (%) |
|---|---|---|---|---|
| RF | 93.33 | 97.05 | 89.90 | 92.80 |
| SVM | 85.28 | 83.20 | 87.93 | 85.38 |
| KNN (K = 3) | 84.63 | 87.10 | 82.37 | 84.42 |
Comparison of classification accuracy of respiratory fingerprint of healthy and diabetic patients before feature selection using PSO.
| Algorithm | Accuracy (%) | Precision (%) | Recall (%) | F1 score (%) |
|---|---|---|---|---|
| RF | 88.7 | 94.8 | 82.10 | 87.86 |
| SVM | 73.47 | 91.13 | 53.68 | 66.70 |
| KNN (K = 3) | 83.43 | 90.46 | 74.77 | 81.58 |
Conclusion
In order to save the system cost and reduce the volume, it is easy to install on the car to realize the on-board application. This paper presents an electronic nose system for diabetes screening based on in-car. The system, which is portable and affordable, includes a homemade electronic nose breath collection device and data analysis algorithms.
In order to save the system cost and reduce the volume, it is different from the previous electronic nose data processing method. In this paper, two feature extraction methods were fused to search for promising feature subsets, and PSO algorithm was introduced to extract the best feature subset for predicting diabetes. The optimal cross-sensitive VOCs sensor array was selected, resulting in a 56% reduction in the number of sensors and an overall cost reduction of $361 USD. In addition, the machine learning diagnostic model is proposed to be deployed in the family car. This model can be used to analyze the preliminary screening results of patients with diabetes and remind patients in time. Respiratory samples from 240 volunteers were analyzed and the classification algorithms were used to evaluate the performance of the preliminary screening system. The classification accuracy of healthy people and diabetic people is the highest at 93.33%. Based on these results, it can be said that the screening system is a promising method for diabetes prediction, as a cheaper and more portable diagnostic tool that can help the average family to detect the disease early and develop appropriate treatment plans.
Compared with previous papers, the self-made gas detection device is adopted in this paper, which can bring higher acquisition accuracy and accuracy, and the amount of data is more than that used in existing papers, so more useful features can be extracted. Different feature extraction and feature sorting methods are used in the model, which makes the classification accuracy higher. Due to individual differences, future research will be carried out according to individual physical characteristics, and cloud storage and other technologies will be combined with smart devices.
Declaration of interest
The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of this review.
Funding
This work was supported by the National Natural Science Foundation of China (51875245), the Science-Technology Development Plan Project of Jilin Province (20200501013GX, 20210401139YY, 20210101248JC, 20220401087YY), the Special Project of Industrial Technology Research and Development of Jilin Province (2022C045-6), the Changchun Science and Technology Project of Changchun (21ZGY02), the ‘13th Five-Year Plan’ Scientific Research Foundation of the Education Department of Jilin Province (JJKH20200870KJ).
Ethics approval
The study protocols were approved by the Medical Ethics Committee of the China-Japan Union Hospital of Jilin University. All participants were informed about the research and were requested to provide written consent.
References
- 1↑
Horváth I, Barnes PJ, Loukides S, Sterk PJ, Högman M, Olin AC, Amann A, Antus B, Baraldi E & Bikov A et al.A European Respiratory Society technical standard: exhaled biomarkers in lung disease. European Respiratory Journal 2017 49 1600965. (https://doi.org/10.1183/13993003.00965-2016)
- 2↑
Phillips M Method for the collection and assay of volatile organic compounds in breath. Analytical Biochemistry 1997 247 272–278. (https://doi.org/10.1006/abio.1997.2069)
- 3↑
Arasaradnam RP, Covington JA, Harmston C, Nwokolo CU. Review article: next generation diagnostic modalities in gastroenterology-gas phase volatile compound biomarker detection. Alimentary Pharmacology and Therapeutics 2014 39 780–789. (https://doi.org/10.1111/apt.12657)
- 4↑
Bikov A, Lázár Z, Horvath I. Established methodological issues in electronic nose research: how far are we from using these instruments in clinical settings of breath analysis? Journal of Breath Research 2015 9 034001. (https://doi.org/10.1088/1752-7155/9/3/034001)
- 5↑
Behera B, Joshi R, Vishnu GA, Bhalerao S, Pandya HJ. Electronic-nose: a non-invasive technology f-or breath analysis of diabetes and lung cancer patients. Journal of Breath Research 2019 13 024001. (https://doi.org/10.1088/1752-7163/aafc77)
- 6↑
Deng C, Zhang J, Yu X, Zhang W, Zhang X. Determination of acetone in human breath by gas chr-omatography–mass spectrometry and solid-phase microextraction with on-fiber derivatiza-tion. Journal of Chromatography. B, Analytical Technologies in the Biomedical and Life Sciences 2004 810 269–275. (https://doi.org/10.1016/j.jchromb.2004.08.013)
- 7↑
Righettoni M, Schmid A, Amann A, Pratsinis SE. Correlations between blood glucose and breath components from portable gas sensors and PTR-TOF-MS. Journal of Breath Research 2013 7 037110. (https://doi.org/10.1088/1752-7155/7/3/037110)
- 8↑
Ghimenti S, Tabucchi S, Lomonaco T, Di Francesco F, Fuoco R, Onor M, Trivella MG. Monitoring breath during oral glucose tolerance tests. Journal of Breath Research 2013 017115. (https://doi.org/10.1088/1752-7155/7/1/017115)
- 9↑
Kalidoss R, Umapathy S, Kothalam R, Sakthivelu U. Adsorption kinetics feature extraction from breathprint obtained by graphene based sensors for diabetes diagnosis. Journal of Breath Research 2020 15 016005. (https://doi.org/10.1088/1752-7163/abc09b)
- 10↑
Yan K, Zhang D, Wu D, Wei H, Lu G. Design of a breath analysis system for diabetes screening and blood glucoselevel prediction. IEEE Transactions on Bio-Medical Engineering 2014 61 2787–2795. (https://doi.org/10.1109/TBME.2014.2329753)
- 11↑
Esfahani S, Wicaksono A, Mozdiak E, Arasaradnam RP, Covington JA. Non-invasive diagnosis of diabetes by volatile organic compounds in urine using FAIMS and Fox4000 electronic nose. Biosensors 2018 8 121. (https://doi.org/10.3390/bios8040121)
- 12↑
Kalidoss R, Umapathy S. A comparison of online and offline measurement of exhaledbreath for diabetes pre-screening by graphene-based sensor; from powder processing to cli-nical monitoring prototype. Journal of Breath Research 2019 13 036008. (https://doi.org/10.1088/1752-7163/ab09ae)
- 13↑
Saidi T, Zaim O, Moufid M, El Bari N, Ionescu R, Bouchikhi B. Exhaled breath analysis using electronic nose and gas chromatography - mass spectrometry for non-invasive diagnosis of chronic kidney disease, diabetes mellitus and healthy subjects. Sensors and Actuators. Part B 2018 257 178–188. (https://doi.org/10.1016/j.snb.2017.10.178)
- 14↑
Ljungblad J, Hök B, Allalou A, Pettersson H. Passive in-vehicle driver breath alcohol detection using advanced sensor signal acquisition and fusion. Traffic Injury Prevention 2017 18 S31–S36. (https://doi.org/10.1080/15389588.2017.1312688)
- 15↑
Guo D, Zhang D, Zhang L, Lu G. Non-invasive blood glucose monitoring for diabetics by means of breath signal analysis. Sensors and Actuators. Part B 2012 173 106–113. (https://doi.org/10.1016/j.snb.2012.06.025)
- 16↑
Hariyanto, Sarno R, Wijaya DR. Detection of diabetes from gas analysis of human breath using e-Nose. In In 11th international conference on Information & Communication Technology and system (ICTS), 241–246. Surabaya, Indonesia: IEEE, 2017. (https://doi.org/10.1109/ICTS.2017.8265677)
- 17↑
Kong C, Zhao S, Weng X, Liu C, Guan R, & Chang Z. Weighted summation: feature extraction of farm pigsty data for electronic nose. In 11th International Conference on Information & Communication Technology and System (ICTS), 241–246.Surabaya, Indonesia:IEEE,2017. (https://doi.org/10.1109/ACCESS.2019.2929526)
- 18↑
Weng X, Luan X, Kong C, Chang Z, Li Y, Zhang S, Al-Majeed S, Xiao Y. A comprehensive method for assessing meat freshness using fusing electronic nose, computer vision, and artificial tactile technologies. Journal of Sensors 2020 2020 1–14. (https://doi.org/10.1155/2020/8838535)
- 19↑
Galassetti PR, Novak B, Nemet D, Rose-Gottron C, Cooper DM, Meinardi S, Blake DR. Breath ethanol and acetone as indicators of serum glucose levels: an initial report. Diabetes Technology and Therapeutics 2005 115–123. (https://doi.org/10.1089/dia.2005.7.115)
- 20↑
Paredi P, Biernacki W, Invernizzi G, Kharitonov SA, Barnes PJ. Exhaled carbon monoxide levels elevated in diabetes and correlated with glucose concentration in blood: a new test for monitoring the disease? Chest 1999 116 1007–1011. (https://doi.org/10.1378/chest.116.4.1007)
- 21↑
Phillips M, Cataneo RN, Cheema T, Greenberg J. Increased breath biomarkers of oxidative stress in diabetes mellitus. Clinica Chimica Acta; International Journal of Clinical Chemistry 2004 344 189–194. (https://doi.org/10.1016/j.cccn.2004.02.025)
- 22↑
Novak BJ, Blake DR, Meinardi S, Rowland FS, Pontello A, Cooper DM, & Galassetti PR. Exhaled methyl nitrate as a noninvasive marker of hyperglycemia in type 1 diabetes. PNAS 2007 104 15613–15618. (https://doi.org/10.1073/pnas.0706533104)
- 23↑
Feng J, Tian F, Yan J, He Q, Shen Y, Pan L. A background elimination method based on wavelet transform in wound infection detection by electronic nose. Sensors and Actuators. Part B 2011 157 395–400. (https://doi.org/10.1016/j.snb.2011.04.069)
- 24↑
Carmel L, Levy S, Lancet D, Harel D. A feature extraction method for chemical sensors in electronic noses. Sensors and Actuators. Part B 2003 93 67–76. (https://doi.org/10.1016/S0925-4005(0300247-8)
- 25↑
Deng C, Lv K, Shi D, Yang B, Yu S, He Z, Yan J. Enhancing the discrimination ability of a gas sensor array based on a novel feature selection and fusion framework. Sensors 2018 18 1909. (https://doi.org/10.3390/s18061909)
- 26↑
Eberhart R, & Kennedy J. A new optimizer using particle swarm theory. In Proceedings of the Sixth International Symposium on Micro Machine and Human Science, 39–43. Nagoya, Japan:Minnesota Historical Society, 1995. (https://doi.org/10.1109/MHS.1995.494215)
- 27↑
Chen T, & Guestrin C. XGBoost: a scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. New York, NY, USA:Association for Computing Machinery,2016. (https://doi.org/10.1145/2939672.2939785)
- 28↑
Pan Q, Tang W, Yao S. The application of LightGBM in Microsoft malware detection. Journal of Physics: Conference Series 2020 1684 012041. (https://doi.org/10.1088/1742-6596/1684/1/012041)
- 29↑
Rahman MM, Charoenlarpnopparut C, & Suksompong P. Signal processing for multi-sensor E-nose system: acquisition and classification. In 10th International Conference on Information, Communications and Signal Processing (ICICS), 1–5. Singapore: IEEE, 2015. (https://doi.org/10.1109/ICICS.2015.7459865)
- 30↑
Boumali S, Benhabiles MT, Bouziane A, Kerrour F, Aguir K. Acetone discriminator and concentration estimator for diabetes monitoring in human breath. Semiconductor Science and Technology 2021 36 085010. (https://doi.org/10.1088/1361-6641/ac0c63)
- 31↑
Yan K, & Zhang D. A novel breath analysis system for diabetes diagnosis. In International Conference on Computerized Healthcare (ICCH), 166–170. Hong Kong, China:IEEE Publications, 2012. (https://doi.org/10.1109/ICCH.2012.6724490)
- 32↑
Xue B, Zhang M, Browne WN. Particle swarm optimisation for feature selection in classification: novel initialisation and updating mechanisms. Applied Soft Computing 2014 18 261–276. (https://doi.org/10.1016/j.asoc.2013.09.018)
- 33↑
Azevedo GLFBG, Cavalcanti GDC, Carvalho Filho ECB. An approach to feature selection for keystroke dynamics systems based on PSO and feature weighting. In IEEE Congress on Evolutionary Computation, 3577–3584: Singapore: IEEE, 2007. (https://doi.org/10.1109/CEC.2007.4424936)
- 34↑
Weng X, Sun Y, Xie J, Deng S, Chang Z. Bionic layout optimization of sensor array in electronic nose for oil shale pyrolysis process detection. Journal of Bionic Engineering 2021 18 441–452. (https://doi.org/10.1007/s42235-021-0022-2)
- 35↑
Chang CC, Lin CJ. LIBSVM: a library for support vector machines. ACM Transactions on Intelligent Systems and Technology 2011 2 1–27. (https://doi.org/10.1145/1961189.1961199)
