Abstract
Introduction Early and accurate diagnosis of interstitial lung diseases (ILDs) remains a major challenge. Better noninvasive diagnostic tools are much needed. We aimed to assess the accuracy of exhaled breath analysis using eNose technology to discriminate between ILD patients and healthy controls, and to distinguish ILD subgroups.
Methods In this cross-sectional study, exhaled breath of consecutive ILD patients and healthy controls was analysed using eNose technology (SpiroNose). Statistical analyses were done using partial least square discriminant analysis and receiver operating characteristic analysis. Independent training and validation sets (2:1) were used in larger subgroups.
Results A total of 322 ILD patients and 48 healthy controls were included: sarcoidosis (n=141), idiopathic pulmonary fibrosis (IPF) (n=85), connective tissue disease-associated ILD (n=33), chronic hypersensitivity pneumonitis (n=25), idiopathic nonspecific interstitial pneumonia (n=10), interstitial pneumonia with autoimmune features (n=11) and other ILDs (n=17). eNose sensors discriminated between ILD and healthy controls, with an area under the curve (AUC) of 1.00 in the training and validation sets. Comparison of patients with IPF and patients with other ILDs yielded an AUC of 0.91 (95% CI 0.85–0.96) in the training set and an AUC of 0.87 (95% CI 0.77–0.96) in the validation set. The eNose reliably distinguished between individual diseases, with AUC values ranging from 0.85 to 0.99.
Conclusions eNose technology can completely distinguish ILD patients from healthy controls and can accurately discriminate between different ILD subgroups. Hence, exhaled breath analysis using eNose technology could be a novel biomarker in ILD, enabling timely diagnosis in the future.
Abstract
Exhaled breath analysis using eNose technology can accurately discriminate between different ILD subgroups and individual diseases. eNose technology could be a novel diagnostic tool in ILD, enabling timely diagnosis in the future. https://bit.ly/3jJs2hf
Introduction
Interstitial lung diseases (ILDs) encompass a diverse group of more than 200 different disorders, which are associated with substantial morbidity and mortality [1, 2]. Idiopathic pulmonary fibrosis (IPF) is the most common ILD and has the worst prognosis [3]. The disease course of ILDs is very heterogeneous; some ILDs are reversible and may be self-limiting, others remain stable, and a subgroup of patients has a progressive phenotype [1]. Moreover, a substantial minority of patients (∼10%) have unclassifiable ILD [4, 5]. Establishing an accurate diagnosis can be challenging, because symptoms such as cough and dyspnoea are nonspecific [6]. Many patients receive one or more misdiagnoses and often undergo invasive diagnostic procedures before the diagnosis of ILD is confirmed [7]. A study in IPF showed that 55% of patients consulted three or more physicians before receiving a final diagnosis. Furthermore, the majority of patients reported a treatment delay of >1 year between initial presentation and confirmed diagnosis [8]. Currently, a multidisciplinary team (MDT) discussion is considered the “gold standard” for diagnosis of ILD [9].
With the availability of different treatment options, it has become increasingly important to achieve an early diagnosis [10, 11]. For IPF, two antifibrotic drugs (nintedanib and pirfenidone) are available that slow down disease progression. More recently, antifibrotic drugs have also shown efficacy in other progressive fibrotic ILDs, which will change the treatment landscape for these patients [12]. Patients with fibrotic ILDs other than IPF can have a predominantly inflammatory phenotype, a more fibrotic phenotype or a combination of both [3]. This emphasises the importance of accurate phenotyping of patients, to decide which treatment should be given (i.e. immunosuppressive medication, antifibrotic medication or a combination). Currently, no biomarkers are available to reliably make this distinction. Thus, there is a major need for noninvasive, widely available, inexpensive tools for diagnosis and monitoring of disease course, especially in the current era with advancing treatment options.
An emerging noninvasive diagnostic technique is the analysis of volatile organic compounds (VOCs) in exhaled breath. Cells create VOCs during metabolic and pathological processes; thousands of these VOCs can be found in exhaled breath [13]. The composition of VOCs in exhaled breath could serve as a biomarker in a wide range of diseases [14, 15]. One method to analyse VOCs is by using electronic nose (eNose) technology. eNoses have several cross-reactive gas sensors, which react to multiple compounds in the VOC mixture. This results in a unique pattern of sensor responses: the breathprint. Individual cases can be classified into disease groups by use of pattern recognition algorithms. The field of breathomics is rapidly evolving and relatively high diagnostic accuracies have been published in other diseases [16–19]. Hence, eNose technology has been proposed as a diagnostic tool, for clinical and inflammatory phenotyping, and to predict response to therapy [16, 17]. In ILD, studies on breathomics are scarce. One study in sarcoidosis showed that the breathprint of patients with untreated pulmonary sarcoidosis could be distinguished from healthy controls; treated sarcoidosis patients could not be discriminated from healthy controls [20]. Another recently published study evaluated the ability of eNose technology to identify different ILD subgroups, and to compare ILD with healthy controls and COPD patients [21]. This study also showed adequate distinction between patients with ILD from healthy controls and COPD patients. However, different ILD subgroups could not be accurately separated from each other.
Here, we aimed to investigate the reliability of exhaled breath analysis using eNose technology to discriminate between ILD patients and healthy controls, and to distinguish ILD subgroups.
Methods
Study design and population
This was a single-centre, cross-sectional study in the Erasmus Medical Center (Rotterdam, The Netherlands). The study was approved by the Medical Ethics Committee (MEC-2019-0230). Between July 2019 and February 2020, consecutive outpatients with a diagnosis of ILD, according to the American Thoracic Society/European Respiratory Society statement criteria [1, 9], or a diagnosis of sarcoidosis (all Scadding stages), according to the World Association of Sarcoidosis and Other Granulomatous Diseases criteria [22], were eligible to participate. For patients with interstitial pneumonia with autoimmune features (IPAF), we used the classification criteria of Fischer et al. [23]. We included prevalent patients, with a diagnosis established in an expert centre MDT. All patients signed written informed consent before inclusion in the study. The healthy controls were hospital staff without a history of lung diseases, who gave informed consent.
Measurements
Measurements were performed using a cloud-connected eNose (SpiroNose; Breathomix, Leiden, The Netherlands). The SpiroNose is an integration between eNose technology and routine spirometry, and has been technically and clinically validated [16, 24]. The SpiroNose has seven different types of cross-reactive metal oxide semiconductor sensors. These sensors are present in duplicate in sensor arrays on both the inside (to measure VOCs in exhaled breath) and on the outside of the SpiroNose (to measure VOCs in ambient air). A SpiroNose measurement consists of five tidal breaths, followed by an inspiratory capacity manoeuvre to total lung capacity, a 5-s breath-hold and subsequently a slow expiration (flow <0.4 L·s−1) to residual volume. All eNose measurements were performed in duplicate. The sensor readings were sent in real-time via a gateway to the online BreathBase analysis platform, which includes the secured online database of Breathomix (ISO27001 and NEN7510 certified). A more detailed description of the methods and set-up can be found in de Vries et al. [16].
Data collection
Participants completed a short survey about factors relevant for the measurement, such as smoking history and food intake in the last 2 h. Data about medication use, lung function tests, pathology results, radiology and recent laboratory parameters were collected from the medical records. Patients were labelled as having pulmonary fibrosis in case of reticulations with traction bronchiectasis and/or honeycombing on the most recent computed tomography (CT) scan, assessed by a thoracic radiologist. Data about forced vital capacity (FVC) and diffusing capacity of the lung for carbon monoxide (DLCO) were collected if available.
Data analysis
The eNose sensor signals were processed and corrected for ambient VOCs as previously described [16, 24]. The peak value of each sensor for exhaled breath was determined and normalised to the most stable sensor (sensor 2). Sensor-to-sensor ratios were used to reduce interarray differences. Lastly, the ratio between peak sensor values and sensor values during breath-hold were calculated. Both the normalised sensor peaks and the ratio between peak sensor values and breath-hold values were used in data analysis. Statistical analysis was done using partial least square discriminant analysis (PLS-DA) and receiver operating characteristic (ROC) analysis. In the ROC analysis, the area under the curve (AUC) values and corresponding 95% confidence intervals were determined. Sensitivity, specificity and accuracy were calculated.
For larger subgroups (i.e. ILD versus healthy controls and IPF versus non-IPF ILDs), training and validation sets by split analysis (2:1) were used, as recommended for metabolomics experiments [25]. After collection of the entire dataset, the dataset was randomly divided into a training set and a validation set. For randomisation, we used the “sample” function in R (www.r-project.org), which generates random samples with a pre-specified size. We used the model obtained in the training set to calculate the PLS-DA variates of the validation set (supplementary figure S1). We compared the training and validation sets based on the AUC of the ROC curves of PLS-DA components 1 and 2. For comparisons between individual diagnoses we did not use training and validation sets, because of the smaller sample sizes. For these groups, only PLS-DA component 1 was used for ROC analysis. We focused on comparing diagnoses that often cause diagnostic dilemmas in clinical practice because of their similarities in clinical presentation and imaging. Descriptive statistics were used to analyse baseline data. Between-group comparisons were done using independent sample t-tests, ANOVA, Chi-squared tests, Kruskal–Wallis tests and Fisher's exact tests. Analyses were done using R version 3.6.2 (using the mixOmics package version 6.1.1) and SPSS Statistics for Windows version 25 (IBM, Armonk, NY, USA).
Results
In total, 322 consecutive patients with ILD and 48 healthy controls were included in this study. The overall mean±sd age for ILD patients was 61.6±13.3 years, 59.9% of patients were male and 5.3% of patients were current smokers. Mean±sd FVC (n=316) of all ILD patients was 3.23±1.1 L or 82.4±19.1% of predicted and mean±sd DLCO (n=305) was 60.6±21.9% of predicted. Mean age and percentage of males were significantly higher in the ILD group compared with healthy controls (table 1). Furthermore, ILD patients had significantly more pack-years than healthy controls, but there was no difference in the percentage of current smokers.
Baseline characteristics of interstitial lung disease (ILD) patients and healthy controls
ILD patients were categorised in seven groups: IPF, sarcoidosis, connective tissue disease-associated ILD (CTD-ILD), chronic hypersensitivity pneumonitis (CHP), IPAF, idiopathic nonspecific interstitial pneumonia (iNSIP) and other ILDs. Diagnoses with less than 10 included patients (cryptogenic organising pneumonia, respiratory bronchiolitis-ILD, asbestosis, drug-induced ILD, granulomatosis with polyangiitis and unclassifiable ILD) were classified as “other ILDs”. Baseline characteristics of these individual groups can be found in supplementary table S1. Details about medication use can be found in supplementary table S2. There were significant differences in age, sex, pack-years, FVC, DLCO and percentage of patients with fibrosis on high-resolution CT between ILD subgroups, but no difference in the percentage of current smokers. Of all sarcoidosis patients, 89% had parenchymal involvement (n=125).
ILD versus healthy controls
The breathprints of 322 ILD patients and 48 healthy controls were compared; groups were divided into a training set and a validation set. The training set consisted of 215 ILD patients and 32 healthy controls; the validation set consisted of 107 ILD patients and 16 healthy controls. The results of PLS-DA for the training set are shown in figure 1, together with the corresponding ROC curve. The eNose completely discriminated between ILD patients and healthy controls with an AUC of 1.00, for both the training and the validation set. Accordingly, the sensitivity, specificity and accuracy of the model were 100% (table 2). Additional analyses in subgroups (i.e. only nonsmokers, only male healthy controls and healthy controls >40 years of age) yielded a similar discriminative ability between ILD and healthy controls, with an AUC of 1.00. These results are shown in the supplementary material.
a) Training set of breathprints from patients with interstitial lung disease (ILD; n=215) compared with breathprints from healthy controls (n=32). b) Receiver operating characteristic curve of partial least square discriminant analysis (PLS-DA) components 1 and 2 for the training set. AUC: area under the curve.
Diagnostic performance of the training and validation sets
IPF versus non-IPF ILDs
Patients with ILD were divided into IPF (n=85) and non-IPF ILDs (n=237), and separated into a training set and a validation set. The training set consisted of 57 IPF patients and 158 patients with non-IPF ILDs; the validation set consisted of 28 IPF patients and 79 patients with non-IPF ILDs. The results of PLS-DA for the training set are shown in figure 2, together with the corresponding ROC curve. In the training set, the AUC was 0.91 (95% CI 0.85–0.96); in the validation set, the AUC reached 0.87 (95% CI 0.77–0.96). In the validation set, the sensitivity was 95%, the specificity was 79% and the accuracy of the model was 91%. Results of the training set are shown in table 2. Results were similar when patients with sarcoidosis were excluded from the non-IPF ILD group (training set AUC 0.92, 95% CI 0.87–0.97; validation set AUC 0.91, 95% CI 0.83–0.98).
a) Training set of breathprints from patients with idiopathic pulmonary fibrosis (IPF; n=57) compared with breathprints from patients with non-IPF ILDs (n=158). b) Receiver operating characteristic curve of partial least square discriminant analysis (PLS-DA) components 1 and 2 for the training set. AUC: area under the curve.
Individual ILDs
Subsequently, breathprints of individual diagnoses were compared with each other. The diagnostic performances of the models for comparison between different ILDs are presented in table 3. All figures can be found in supplementary figure S2. As we included a group of patients that did not have a classifying ILD diagnosis but fulfilled the criteria of IPAF, we performed an exploratory analysis comparing these patient with both IPF and CTD-ILD, as these are the most common clinical differential diagnoses in IPAF (figure 3).
Models for direct comparison between individual diagnoses
Partial least square discriminant analysis (PLS-DA) results for connective tissue disease-associated interstitial lung disease (CTD-ILD), interstitial pneumonia with autoimmune features (IPAF) and idiopathic pulmonary fibrosis (IPF).
Discussion
In this study, we aimed to evaluate the reliability of exhaled breath analysis using eNose technology to discriminate between ILD patients and healthy controls, and to distinguish ILD subgroups. The eNose accurately discriminated between ILD patients and healthy controls, both in the training set and the validation set. Moreover, the eNose adequately discriminated between individual ILDs and between IPF versus non-IPF ILDs.
Until now, one other pilot study in ILD investigated the ability of eNose technology to recognise ILDs [21]. In line with our results, healthy controls could be distinguished from ILDs (IPF, CTD-ILD and cryptogenic organising pneumonia). In the present study, we have confirmed and extended this finding in an independent training and validation cohort. The eNose in the current study could discriminate between individual diagnoses with a high sensitivity, specificity and accuracy. This was not shown by Krauss et al. [21], presumably due to a smaller number of ILD patients in their study (n=174) compared with the current study (n=322), which could have led to insufficient training of the device [21]. The encouraging results in our study warrant further confirmation and external validation in larger (multicentre) cohorts. A larger cohort for individual ILDs coming from different MDTs will allow for comparison of eNose-based diagnosis with “gold standard” MDT diagnosis. Moreover, a larger and more diverse dataset will further increase the accuracy of eNose technology to detect ILD and distinguish between different diagnoses, making this a potentially new tool for rapid, noninvasive diagnosis of ILD.
Surprisingly, patients with IPF, IPAF and CTD-ILD had a distinctive breathprint and could be discriminated with a relatively high accuracy, especially IPF versus IPAF. This raises the question whether IPAF could be a separate disease entity. Until now, the term IPAF has been primarily used as a research concept, but not as a clinical diagnosis. IPAF is thought to have a significant overlap with IPF and CTD-ILD, and is often considered as undifferentiated CTD-ILD [23, 26]. Moreover, IPAF is a very heterogeneous concept, as the classification criteria are based on a combination of features from three different domains (clinical, serological and morphological). This makes the preliminary results of the differences between IPAF, CTD-ILD and IPF in the current study even more interesting. eNose technology could potentially be used to determine whether distinct phenotypic clusters can be identified within the patient group currently classified as IPAF. Obviously, these results need to be extended and confirmed in larger studies.
A potential limitation in this study is the fact that the healthy controls were much younger, largely female and nonsmokers. Previously published studies indicated that age, sex and smoking status seem to have little influence on the breathprint [16, 27, 28]. Moreover, additional analyses in our cohort showed the same discriminative ability between ILD and healthy controls when smokers, younger participants and females were excluded from the analysis. Hence, we believe that these differences in demographics have not impacted our results. Because of the cross-sectional design of the current study we were not able to evaluate whether medication use has an influence on the breathprint in ILD patients.
A possible obstacle for the further development of eNose technology towards a point-of-care tool in ILD is the lack of a “gold standard” for the diagnosis of ILD. Diagnoses are based on MDT meetings and a substantial part of ILD remains unclassifiable [1, 4, 5, 9]. Because there is no real “gold standard”, it is highly likely that a minority of patients has been incorrectly diagnosed. This means that the pattern recognition algorithms receive wrong so-called “gold standard” information and the algorithm is trained based on partly incorrect information. This same limitation has been mentioned by Walsh et al. [29] in a recent study where a deep learning algorithm learned to classify fibrotic lung disease on high-resolution CT. A larger training dataset would result in a better performing algorithm, as the percentage of incorrectly labelled patients will be relatively smaller. This further emphasises the need for future research on eNose technology in ILD, preferably as a multicentre effort, to increase the size of the available datasets and account for differences in diagnoses between MDTs [30]. Moreover, because of the lack of a “gold standard”, longitudinal data about disease progression and survival will be necessary for definite validation of our findings.
Based on our data, we believe that exhaled breath analysis has the potential to enable early and accurate diagnosis of ILD in the future. Results of eNose measurements could give guidance during MDT meetings and enhance diagnostic certainty in clinical practice. In addition, unsupervised cluster analysis can be performed to cluster patients based on their breathprint, irrespective of the underlying diagnosis, similar to a recent study among patients with asthma or chronic obstructive pulmonary disease [16]. This data-driven approach could potentially distinguish different disease phenotypes which have not yet been clinically identified. Further studies should reveal whether patients with distinct eNose-based phenotypes have a different disease behaviour and/or response to therapy. If so, this may be a novel technology to predict disease progression and prognosis in ILD. To gain more insights in the underlying pathobiology of various ILDs, eNose technology could be combined with mass spectrometry techniques. In contrast to eNose technology, which is based on signal pattern recognition, gas chromatography/mass spectrometry can measure individual VOCs in exhaled breath [31].
In conclusion, eNose technology has the potential to become a novel diagnostic tool for ILD. eNose measurements can hopefully be used in the future to increase diagnostic confidence and allow for point-of-care diagnostics in the doctor's office, thereby reducing diagnostic delays and improving care and treatment for patients with ILD.
Supplementary material
Supplementary Material
Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author.
Supplementary material ERJ-02042-2020.SUPPLEMENT
Shareable PDF
Supplementary Material
This one-page PDF can be shared freely online.
Shareable PDF ERJ-02042-2020.Shareable
Acknowledgements
The authors thank the pulmonary function technicians from the Pulmonary Function Dept in the Erasmus Medical Center (Rotterdam, The Netherlands) for their valuable assistance during this study, and C.A.L. van den Berg (Dept of Respiratory Medicine, Erasmus Medical Center) for her help with the eNose measurements.
Footnotes
This article has supplementary material available from erj.ersjournals.com
This study is registered at the Netherlands Trial Register with identifier NL7860.
Conflict of interest: C.C. Moor reports grants from Boehringer Ingelheim, during the conduct of the study; grants and other from Boehringer Ingelheim, outside the submitted work; all fees and grants were paid to the Erasmus MC.
Conflict of interest: J.C. Oppenheimer has nothing to disclose.
Conflict of interest: G. Nakshbandi has nothing to disclose.
Conflict of interest: J.G.J.V. Aerts has nothing to disclose.
Conflict of interest: P. Brinkman has nothing to disclose.
Conflict of interest: A-H. Maitland-van der Zee reports grants from GSK, grants and personal fees from Boehringer Ingelheim, personal fees from AstraZeneca, outside the submitted work; all personal fees were paid to the Amsterdam UMC.
Conflict of interest: M.S. Wijsenbeek reports grants from Boehringer Ingelheim, during the conduct of the study; grants and other from Boehringer Ingelheim and Hoffman la Roche, other from Galapagos and Respivant, outside the submitted work; all fees and grants were paid to the Erasmus MC.
Support statement: The lease for the SpiroNose was financially supported by Boehringer Ingelheim. Funding information for this article has been deposited with the Crossref Funder Registry.
- Received May 28, 2020.
- Accepted July 20, 2020.
- Copyright ©ERS 2021
References
