Nowadays, the concentration of some biomarkers in breath has been proven to be related with certain diseases. For example, lung diseases may alter volatile organic compounds (VOCs) in breath because both Mycobacteria and oxidative stress resulted from Mycobacterial infection generate distinctive VOCs. And digestive system diseases will reflect in hydrogen of ones’ exhaled breath because sugar could not be fully digested and would be decomposed or fermented and produce more hydrogen. By selecting proper sensors that can respond to the components, it is possible to analyze a person’s breath odor and thus his or her health state. A few examples will further prove these points. The level of nitric oxide can be used as a diagnostic for asthma [20]. Patients with renal disease have higher concentrations of ammonia [21]. The concentration of VOCs, such as cyclododecatriene, benzoic acid, and benzene is much higher in lung cancer patients [22]. Table 2 lists the relationship between biomarkers and some typical diseases.

Currently, the measurement of exhaled breath is usually performed by two common gas analysis apparatuses, gas chromatography (GC) [26] or electronic nose (e-nose) [27]. GC can separate and identify molecules that are responsible for typical odors occurring in specific diseases, which is very accurate for

disease identification. But this kind of apparatus is expensive and not portable. Its sampling and assaying processes are complicated and time consuming (about one hour for one sample), and its results require expert’s interpretation [28]. Therefore, it is hard to use such apparatus as a domestic or clinical tool.

Electronic noses, or e-noses, are devices that “smell” or detect odor. An e-nose consists of a mechanism for chemical detection, such as an array of electronic sensors, and a mechanism for processing. Using e-nose is a less expensive and more portable way for breath analysis. Recently, e-nose has gradually been used in medicine for the diagnosis of renal disease [29], diabetes [30], lung cancer [31], and asthma [32]. Though all of these methods work satisfactorily in breath analysis, their results could possibly be improved. That is because, commercial e-noses, with their marketing concerns, have to provide versatility in applications, such as coffee, wine, and fragrances identification. The versatility, however, limits their performance in disease detection since their sensor selection has to match broad applications.

The idea of e-nose is inspired by the mechanisms of human olfaction. In general, basic elements of an e-nose system include an “odor” sensor array, a data preprocessor, and a pattern recognition engine [33]. Among them, the sensor array, like signal receptors, is the key part of e-nose. The application of sensor array on odor recognition was demonstrated firstly by Persaud and Dodd in 1982 [34]. Currently, e-nose has been much developed and used to fulfill a large number of industrial needs, such as food, chemistry, fragrances, security, and environment [35]. Recently, the feasibility of using e-noses for monitoring the health of human and diagnosing diseases in an early stage has been demonstrated [36].

As early as 1997, Wang et al. [37] designed an e-nose with one SnO₂ thin film sensor for diabetes diagnosis. The authors tested their device by using the breath samples collected from 18 patients and 14 healthy persons. The concentration of blood sugar of the subjects was used as reference. The results showed that the e-nose was able to diagnose diabetes with a sensitivity of 77.8% and a specificity of 35.7%.

In 2001, Lin et al. [36] reported a study about the application of e-nose with six quartz crystal sensors to detect renal diseases. Discriminant Analysis (DA) was carried out to analyze the sensor signals. The clinical test result showed that the e-nose could discriminate the breath samples from 30 normal subjects, 83 uremia patients, and 61 chronic renal disease patients with a total correct classification of 86.78%.

In 2009, Ogorodnik et al. [38] analyzed VOCs from a breath sample of a patient with different lung diseases by using an e-nose with ten MOSFET sensors and four SnO₂ sensors. In total, 66 individuals ? 23 with asthma, 3 with chronic obstructive pulmonary disease (COPD), 12 with pneumonia, 13 with lung cancer, 4 in the past operation state (removed lung cancer), and 11 healthy volunteers were tested at two different times and ANN analysis was employed to classify the samples of cancer and other lung diseases. The results showed that the e-nose could identify lung cancer with 100% accuracy, healthy subjects with 100% accuracy, and asthma with 82.6% accuracy.

In 2010, Guo et al. [39] proposed a method of monitoring the blood glucose levels of diabetics via measuring the concentration of breath acetone. A SVM classifier was used to evaluate the accuracy of classifying the samples into the groups with different blood glucose levels. The results indicated that the system was not only able to distinguish between breath samples from patients with diabetes and healthy subjects, but also to represent the fluctuation of blood glucose of diabetics. In the same year, Guo et al. [40] improved accuracy of diabetes condition monitoring by using a SRC method. Coupled with SRC, the system was able to classify these levels with a much better accuracy than the accuracy reported in [41].

In 2017, Yan et al. [42] proposed a domain learning method using domain features and independence maximization. A logic regression classifier was used to evaluate the accuracy on diabetes, chronical kidney disease and cardiopathy. The results showed that the method can improve the classifying performance by reducing the instrumental variation and time-varying drift.

Table 3 concludes some existing trials on medical breath analysis with e-noses.

To evaluate the performance of our device, a large-scale breath dataset was collected. We cooperated with Guangzhou Hospital of Traditional Chinese Medicine and collected data from inpatient volunteers with a specific e-nose device [43].

In the device, there were 12 different odor sensors selected based on the relationship between breath biomarkers and diseases into consideration. The sensors in the device should be sensitive to the VOCs, carbon dioxide, humidity, and temperature. Thus, a sensor array with 11 sensors was optimized for the purpose of detecting one’s breath. The sensor array included six ordinary metal oxide semiconductor (MOS) sensors, three temperature modulated MOS sensors, a carbon dioxide sensor, and a temperature-humidity sensor. Specifically, the temperature-humidity sensor had two input channels for temperature and humidity respectively. Therefore, there were, in total, 12 input channels. The model, manufacturer and function of the sensors are listed in Table 3. The suffix “-TM” indicates a temperature-modulated sensor.

For each sample, we first collected the patient’s breath and recorded the signals. The diagnosis was then given by an authoritative doctor as the classification labels. Then some biochemical indicators, such as blood glucose, blood pressure and blood lipids were collected. Finally, in this dataset, there were in total over 10,000 samples of 47 classes, including 1491 healthy samples and samples of 46 different kinds of diseases. In this paper, a subset of healthy samples and samples for six kinds of diseases are selected for experiments, including breast disease, cardiopathy, diabetes, lung disease, kidney disease and gastritis. Table 4 shows the number of each class used in selected subset.

All the samples were collected from hospitals in Guangzhou. However, since most of the healthy samples were provided by medically-examined young people while disease samples by elder patients, it is difficult to make age-matched subsets, which is a limitation of this dataset.

Before analyzing data, original signals should be preprocessed so as to be transformed into standard samples. Four steps were taken: faulty signal removal, de-noising, baseline manipulation and normalization.

A faulty signal is a common problem in devices with sensors. In our system, causes of faulty signals were complicated, including mis-operation, bad connection and device damage. In order to make the system more robust, these signals should be removed before analysis.

De-noising aims to remove the noise from the original signals by utilizing a low-pass filter since the signal is mainly interfered by high-frequency noise.

The purpose of baseline manipulation is to compensate baseline drift. The baseline value is the average response in the baseline stage of each sensor. The value is then subtracted from the whole response curve to eliminate the interference of background noise of the sensors [44]. Assume that for each sensor transient of each sample, there are k dimensions, where k = 1, …, Nk, and b dimensions in the baseline stage, where b = 1, …, Nb. The response at time t_k is denoted as R ( t k ) . The baseline response is B ( t b ) . Then baseline manipulation can be computed as:

R B ( t k ) = R ( t k ) − 1 N b ∑ t b = 1 N b B ( t b ) (1)

Normalization is used to compensate for sample-to-sample variations caused by analyte concentration. R B ( t k ) is a sample after the baseline manipulation step, and the normalized response R B N ( t k ) can be defined as:

R B N ( t k ) = R B ( t k ) max ( R B ( t k ) ) (2)

To reduce the dimension of the origin features, principal component analysis (PCA) can be used. PCA projects high-dimensional data into a low-dimensional subspace while keeping most of the data variance.

Some low-dimensional geometric features can also be extracted from the origin response curves. One of the traditional features of gas sensors is steady state response. When a gas sensor is used to sense a gas sample, its response will reach a steady state in a few minutes. The steady state response has a close relationship with the concentration of the measured gas. Therefore, the 9D feature vector contains most of the information needed for disease screening.

However, additional useful information is carried in the transient responses [45]. Transient responses are often related to the change of gas flow (injection/purge) or temperature (for TM sensors). The feature set includes magnitude, difference, derivative, second derivative, integral, slope and phase features, as well as features in frequency domain such as fast Fourier transformation (FFT) and wavelet. The extracted features in both space domain and frequency domain are described in Table 5.

For each class, the first 50 samples collected by the first device were chosen as the labeled training sets and the rest were test samples. Logistic regression method was adopted as the classifier. Sequential forward selection (SFS) method was used to optimize the features. SFS method is a greedy strategy. In each iteration, one feature was selected from all features that could achieve the best classification accuracy together with the features already selected. Figure 1 shows the results of forward selection in different disease diagnosis tasks. For each graph, the horizontal axis represents the number of features used and the vertical axis represents the classification accuracy. We will find the peak point of each curve and record which features are selected in the peak points.

In Table 6, we concluded the best combination of features and sensors selected for each task (the peak points in each curve of Figure 1). It can be found that wavelet features contributed most in all the tasks. Pca feature, MeanMag feature, phase feature and derivative feature differed in different tasks. Other features did not improve the performance of the system. The sensors that contributed most to meeting the relationship between diseases and breath biomarkers in each task are listed in Table 2 and Table 7. S7 (TGS2602) contributes most to metabolic diseases.

To further evaluate the performance of the system, we built a classification tree to deal with the multi-disease diagnosis problem. Synthetic Minority Over-sampling Technique (SMOTE) [46] was used in the experiment. SMOTE method is an over-sampling approach in which the minority class is over-sampled by creating “synthetic” examples rather than over-sampling with replacement.

Figure 2 shows the result of the classification tree. When an unknown sample comes, it will first be tested if it is diabetes. Then check if it is healthy or kidney disease in turn. Finally, if the sample belongs to none of these classes, it may be disease in lung, stomach or respiratory system. Accuracy of each level of the classification tree is 90%, 75%, 63% and 76%, respectively. It can be seen that metabolic diseases are easier to be separated out than diseases of respiratory system or digestive system.

However, this is only for one-label samples. For those patients with more than one disease, multi-labeled methods should be considered.