Detecting Coronary Artery Disease Using Exhaled Breath Analysis

Coronary artery disease (CAD) is the leading cause of morbidity and mortality worldwide [1]. In the USA, the estimated annual incidence of new coronary syndrome is over 600,000, with >350,000 annual deaths attributed to CAD [1]. In light of its high prevalence and associated significant morbidity and mortality, there is an unmet need for a simple, inexpensive, noninvasive tool aimed at detection of CAD.

In the past two decades, the analysis of exhaled breath has drawn scientific and clinical interest due to its potential to noninvasively detect biochemical processes of the human body. In addition to oxygen, nitrogen, and carbon dioxide, the exhaled breath contains low concentrations of various volatile organic compounds (VOCs) [2]. Almost 3,000 VOCs have been reported in the scientific literature, most of which are believed to reflect endogenous metabolic processes, such as inflammation and oxidative stress, in which metabolic products are released to the blood stream, and when reaching the lungs, are excreted by diffusion across the pulmonary alveolar membrane and exhaled through the breath. In addition, exhaled VOCs contain also exogenous components, reflecting exposure to exogenous substances such as tobacco and pollution [3, 4].

A wide range of modified nanomaterials have been used in exhaled breath analysis, in two main approaches: the first approach attempts a selective detection of specific VOCs, usually by gas chromatography, which uses high-selective receptors designed for specific, preidentified VOCs. However, currently, no VOCs were identified to be associated uniquely to a specific disease. The second approach – the one used in this study – is using a cross-reactive nanomaterial-based sensor arrays, which allow pattern recognition rather than identifying and naming a specific VOC. The main technique in this approach is based upon electrical conductivity – the electronic nose (eNose) – which is used in this current study, as detailed below [3]. Although primarily studied in lung diseases [5‒9], exhaled breath analysis was also shown to be useful in the diagnosis and classification of various extrapulmonary diseases [10‒14]. Different strategies have been applied for analyzing breath VOC samples including mass spectrometry or ion mobility-based methods [15, 16] and sensor-based methods [17, 18]. The latter approach advantages are the ability to develop a simple-to-use point-of-care system allowing fast diagnostics [19].

The main pathologic process involved in the development of CAD is atherosclerosis, a chronic inflammatory disease of the arterial wall. It is well established that inflammation plays a major role in the process of atherosclerosis, involving endothelial dysfunction, shear, and oxidative stress [20‒22]. Given these highly active metabolic processes that are part of the atherosclerotic pathogenesis, it can be expected to find a reflection of these processes also in breath contents among patients with atherosclerotic CAD. In an attempt to evaluate the noninvasive, simple tool of breath analysis, we conducted a proof-of-concept study, testing the diagnostic and classificatory abilities of an eNose in the diagnosis of coronary heart disease. The aim of this pilot study was to analyze the respiratory profile of individuals with coronary heart disease and to evaluate the possible use of the eNose system in detecting atherosclerotic CAD.

This prospective study included patients without known history of CAD, who presented with chest pain to the Beilinson hospital’s emergency department (ED) between the years 2018–2019. Exclusion criteria were a need of urgent cardiac intervention, hemodynamic instability, or inability to provide a breath sample.

After signing an informed consent, patients were asked to abstain from smoking, eating, or drinking during the 3 h prior to the breath test. Patients’ breath samples were collected using a designated portable eNose system, designed by the Laboratory for Nanomaterial-Based Devices (Technion, Israel). The eNose is a nanomaterial-based sensor array system with 12 different sensors, and each is composed of gold nanoparticles covered with different organic ligands that absorb headspace VOCs (sensor development has been previously described [23]). The exposure of the sensors to a sample headspace results in a measurable change in their electrical resistance, which is recorded by a designated software on the portable computer. Breath is sampled via a disposable mouthpiece into a 38-mL glass tube. The last fraction of the breath (alveolar part) is then automatically sucked into the sensor chamber at a rate of 300 mL/min. The sampling glass tube and sensors are temperature controlled at 40°C to avoid condensation of the sample. Data from sensors are recorded and stored on a portable computer and are later transformed by statistical or structural algorithms to identify various volatile patterns, leading to the production of a “breath-print.” This method is easy to use and provides highly sensitive results but does not allow quantitative results or identification of specific substances [14].

During their hospital stay, patients were referred to coronary evaluation either by coronary computed tomography angiography (CCTA) or by coronary angiography, according to doctors’ decision and regardless of the study. Based upon the results, data of breath samples were retrospectively divided into one of 3 clinical groups: (1) group 1: obstructive CAD (any coronary stenosis causing narrowing of >50% of the lumen), (2) group 2: nonobstructive CAD (any coronary stenosis causing ≤50% of the lumen), (3) group 3: no-CAD – normal coronary arteries.

To evaluate the diagnostic performance of the system and its ability to recognize obstructive coronary disease, a two-step approach was taken: in the first step of the study, five classification models were built using a discriminant factor analysis (DFA) algorithm. DFA determines the linear combinations of the input variables (features extracted from each sensor), so the variance within each class is minimized and the variance among classes maximized. The DFA output variables (i.e., canonical variables [CVs]) constitute mutually orthogonal dimensions; the first CV is the most powerful discriminating dimension. Leave-one-out cross-validation was used to calculate the success of the classification in terms of the numbers of true-positive, true-negative, false-positive, and false-negative predictions. Given k measurements, the model was computed using k−1 training vectors. All possibilities of leave-one-sample-out were considered, and the classification accuracy was estimated as the average performance over the k tests. For pattern recognition and data classification, Python (Python Software Foundation) was used:

1. Model 1: group 1 (obstructive CAD) compared to group 2 (nonobstructive CAD)

2. Model 2: group 1 (obstructive CAD) compared to group 3 (no-CAD)

3. Model 3: group 2 (nonobstructive CAD) compared to group 3 (no-CAD)

4. Model 4: group 1 (obstructive CAD) compared to group 2 + group 3 (nonobstructive CAD + no-CAD)

5. Model 5: group 1 + group 2 (obstructive CAD + nonobstructive CAD) compared to group 3 (no-CAD)

In the second step of the study, the diagnostic performance of the classification models was examined in a blinded manner on a new set of patients using JMP Pro, version 15.0.0 (SAS Institute Inc., Cary, NC, USA, 1989–2005). The data were compared with the actual results of coronary arteries evaluation. Sensitivity, specificity, and accuracy were calculated for each model, in both study steps. Flowchart of the study protocol is presented in Figure 1. Figure 2 provides a graphic representation of breath sampling and analysis procedure.

Baseline and clinical data were collected by interviews and by institutional medical records. Data are described as percentage (%) for categorical variables and as mean ± 95% confidence interval for continuous variables. The significance of the differences between the means was assessed by Student’s t test and ANOVA. The Fisher test and χ² test were used to compare discrete variables. A result was considered significant if it had p value less than 0.05.

The study was approved by the Ethics Committee of Beilinson Hospital at Rabin Medical Center (0622-16-RMC). Written informed consent was obtained from each patient included in the study, and all procedures were carried out according to the ethical principles established in the Helsinki Declaration.

In the first part of the study, the nanomaterial-based sensor system was used to validate a breath analysis-based CAD detection model in 60 adult patients (42/60 [70%] males, mean age 52.8 ± 11.7 years) who had no previous history of CAD and who presented to the ED due to chest pain (“training set”). Fifty-seven (95%) of patients had at least one cardiovascular risk factor, and 42 (70%) patients presented with atypical chest pain complaints. Other clinical characteristics of the patients are presented in Table 1. Forty-one (68%) patients underwent CCTA as primary coronary imaging modality; in 22 (37%) patients, percutaneous angiography was performed. Twenty (33%) patients were discharged from the hospital without any further treatment; in 24 (40%) patients, medical treatment (either statins alone or statins + aspirin) was initiated as a result of coronary imaging; and 16 (26%) patients underwent an active revascularization. Clinical presentation and management of patients are presented in Table 1.

Breath analysis data were classified based upon patients’ coronary assessment by CCTA or angiography as obstructive CAD (17 patients), nonobstructive CAD (17 patients), and no-CAD (26 patients). As detailed in the method section, five classifier models were built to compare between the clinical groups. All five models reached similar statistical strength, with sensitivity ranging from 65% (model 2) to 82% (model 1), specificity from 59% (model 5) to 70% (model 4), and accuracy ranging from 65% (model 5) to 72% (model 1). Accuracy, sensitivity, and specificity for the models are presented in Table 2.

In the second part of the study, breath samples were collected from a new set of 25 adult patients (“blind set”). Patients in the blind set were older (59.1 ± 11.4 years, p = 0.02) compared with the training set, but the other baseline characteristics did not differ significantly between groups, as presented in Table 1. Compared with the first group, fewer patients in the second group underwent CCTA (68% vs. 48%, p = 0.09, respectively), and more underwent percutaneous angiography (37% vs. 60%, p = 0.05, respectively). There were no significant differences in the nature of finding of coronary evaluation, neither in management decisions between groups, as can be seen in Table 1.

In this part of the study, only two models – 4 and 5 – were used as they included all three optional clinical groups. Using these two models, samples were analyzed and categorized into clinical groups based on the classifiers of the first step (“blinded set”). The data were then compared with the actual results of the coronary assessment of the patients by CCTA or angiography. Model 4, which compared obstructive CAD (group 1) with nonobstructive and no-CAD (groups 2 and 3), reached 89% sensitivity, 31% specificity, 83% negative predictive value (NPV), 42% positive predictive value (PPV), and 52% accuracy. Model 5, comparing both obstructive and nonobstructive CAD (groups 1 and 2) with non-CAD (group 3), reached 69% sensitivity, 67% specificity, 54% NPV, 79% PPV, and 68% accuracy. Models are presented in Table 2.

Table 3 presents baseline characteristics of patients according to their actual coronary assessment. Patients with obstructive CAD (n = 26) were generally older (mean age 60.4 ± 9.8 years) compared to patients with nonobstructive CAD (n = 24) and patients with no-CAD (n = 35) (53.0 ± 11.0 and 51.4 ± 12.6, respectively, p = 0.008), with male dominance (96% males vs. 67% and 57%, respectively). These patients also had significantly more CV risk factors (69% with ≥3 CV risk factors vs. 29% and 37%, respectively, p = 0.01), with significantly higher use of antiplatelets (35% vs. 8% and 9%, respectively, p = 0.01), statin (46% vs. 3% and 9%, respectively, p = 0.003), and antidiabetic drugs (30% vs. 12% and 58%, respectively, p = 0.02).

In the present study, we aimed to analyze the respiratory profile of patients who presented with chest pain and to evaluate the relationship, if any, between various respiratory profiles and the presence of atherosclerotic CAD. To the best of our knowledge, this is the first study to use breath analysis in evaluating the presence of CAD in patients with acute chest pain.

To evaluate the diagnostic performance of the sensor system as a simple, noninvasive tool to recognize patients with obstructive CAD who need quick further medical evaluation and management, a two-step approach was taken. In the first part of the study, we validated breath analysis-based CAD detection models in 60 patients with acute chest pain and no history of CAD. In the second part of the study, we tried to best simulate a real-life scenario, in which a patient with unknown CAD status is presenting with chest pain and management decisions will depend on the probability that patient has a real CAD. Therefore, only two models – 4 and 5 – were used as they allow inclusion of all three clinical groups. In clinical setting, the focus would be on correct identification of obstructive CAD as this will necessitate a faster and more intensive clinical management. In our vision, future use of breath analysis system such as the eNose will not replace any of the currently used diagnostic modalities but will rather be a complimentary tool to assist decision-making in the management of a patient with chest pain. The greatest value of on-the-spot breath analysis could be in patients with low-intermediate probability of CAD, alongside with in-depth history taking, clinical assessment, physical examination, and electrocardiogram (ECG). This could be especially useful in outpatient clinics and community settings, in which a decision on whether a patient is in need of urgent evacuation to a hospital may be paramount. A further assessment of a patient with suspected coronary etiology for chest pain will be acceptable even if eventually proved to have none or nonobstructive CAD. Therefore, these preliminary results support the potential use of breath test analysis in identifying patients with chest pain who require a quicker management and further coronary evaluation.

Among patients presenting with acute chest pain to the ED, in approximately 35–40%, the etiology will be CAD, while others will have distinct cardiac conditions or a noncardiac disease [14]. Current first-line diagnostic tools in patients suspected to have acute coronary syndrome include an ECG and cardiac biomarkers (such as high-sensitive cardiac troponin). When there are no clear markers for an acute coronary syndrome by ECG or cardiac biomarkers, further diagnostic options include stress testing, CCTA, or angiography – with only the two last ones allowing an anatomic assessment of the presence of CAD [24]. Coronary angiography is the gold standard for the diagnosis of CAD; however, as an invasive procedure, it is accompanied by the risk of complications, as well as exposure to contrast agent and radiation. CCTA is a noninvasive option, providing a noninvasive coronary diagnostic assessment of patients with chest pain, but it is also accompanied by exposure to contrast agent and radiation [25]. Breath analysis is not likely to replace any of the current modalities in diagnosing CAD; however, it has the potential of constituting a powerful tool to aid and enhance early detection of disease and therefore affect clinical decision-making. In the present study, breath sampling was done separately from data analysis. Ideally, breath sampling will be performed bed-side, by a simple, portable eNose system which will give quick and reliable data analysis.

Breath analysis was studied mainly in the field of pulmonary pathologies. Searching the medical literature, we have come across two interesting studies using eNose in detecting CAD. Tozlu and colleagues used exhaled breath analysis to differentiate patients who had gone primary percutaneous intervention for acute myocardial infarction from patients with stable CAD with 97% accuracy, and patients with stable CAD from patients without heart disease with 81% accuracy [26]. Segreti and colleagues used an eNose to identify patients with chronic CAD who had an indication for myocardial revascularization with 78% sensitivity and 68% specificity [27]. Along with our study, this is a proof of concept for the possible ability of exhaled breath analysis to noninvasively detect CAD, requiring further evaluation and/or intervention.

There are many methods to analyze eNose data, which include different approaches for dimension reduction, classification, and validation. Recent review by Leopold and colleagues [28] showed that estimating a diagnostic performance of an eNOSE on a training set alone is not sufficient, even after internal validation, and therefore, an external validation set is recommended. In this current study, we indeed followed this approach. In this case, no dimension reduction was needed, and the classification was done using the LDA (DFA) method. To strengthen the model, we used both an internal cross-validation step as part of the training and a separate external validation step with a “blind” set of patients.

This study has several limitations. First, it includes a relatively small cohort group. Having small number of patients makes it difficult to assess possible confounding factors which could have affected the VOCs’ profile. For example, as could be expected, patients with obstructive CAD had higher prevalence of baseline cardiovascular risk factors (Table 3). Accordingly, those patients were using more medications at baseline, including statins, antiplatelets, and antidiabetic drugs. Some of these medications may have a direct or indirect influence on VOCs’ profile – for example, as statins stabilize the atherosclerotic plaque, their use may attenuate the inflammatory process involved in atherosclerosis. Moreover, patients with CAD often suffer from comorbidities related to common risk factors, such as chronic obstructive pulmonary disease, obstructive sleep apnea, and associated malignancies, which also may affect exhaled breath pattern and components [5, 6, 8, 10, 29]. Such confounders could be assessed only in much larger scale future studies. A second limitation is that in the current system, there is lack of control for expiratory flow rate, breath hold, and anatomic dead space which were previously shown to potentially effect VOCs’ pattern with the use of eNose [30]. These important aspects should be addressed in future studies and will ideally be integrated into more advanced eNose systems. Another limitation is that due to technical issues, samples were collected during a rather long period. These obstacles were partially due to the COVID-19 pandemic which halted further recruitment of patients to the second part of the study, due to natural concern of breath sampling during the pandemic. Nonetheless, the use of completely new set of patients to blindly test the classifying models based upon the training set adds a considerable strength to our study.

This is a pilot study using a portable nanomaterial-based sensor array system (eNose) to detect CAD in patients without previous history of cardiac disease, complaining of chest pain. In this proof-of-concept study, the system was shown to have good primary results in correctly identifying patients with CAD. Moreover, taking advantage of the tested portable system allowed to examine at the point-of-care in the ED, potentially shortening diagnostic step. Further research in larger cohorts is needed to increase the statistical power of classifications and provide better validation of the concept suggested in our study.

The research was conducted ethically in accordance with the World Medical Association Declaration of Helsinki. The study was approved by the Ethics Committee of Beilinson Hospital at Rabin Medical Center (0622-16-RMC). Written informed consent was obtained from each patient included in the study, and all procedures were carried out according to the ethical principles established in the Helsinki Declaration.

The authors have no conflicts of interest to declare.

No financial support was received in this study.

Inbar Nardi Agmon – conceptualization and design, collection, and assembly of data; data analysis and interpretation; manuscript writing – original draft, review and editing, and final approval of manuscript. Yoav Y Broza – conception and design, data analysis and interpretation, manuscript writing, and final approval of manuscript. Gharra Alaa and Ashraf Hamdan – data analysis and interpretation and final approval of manuscript. Alon Eisen and Ran Kornowski – conception and design, manuscript writing, and final approval of manuscript. Hossam Haick – conception and design and final approval of manuscript.

The data that support the findings of this study are not publicly available due to their containing information that could compromise the privacy of research participants but are available from Dr. Inbar Nardi Agmon upon reasonable request.

Inbar Nardi Agmon and Yoav Y. Broza contributed equally to the study.