Exhaled breath volatile organic compounds (VOCs) detection methods: GC-MS versus eNose in COPD diagnosis-a systematic review and meta-analysis

Abstract

Background

Volatile organic compounds (VOCs) derived from exhaled breath have been studied for their diagnostic potential in chronic obstructive pulmonary disease (COPD). However, the diagnostic efficacy of detection technologies of electronic nose (eNose) and gas chromatography-mass spectrometry (GC-MS) remains unclear. This study aims to systematically compare the diagnostic performance of these two methods in COPD diagnosis.

Methods

This review was conducted in accordance with PRISMA guidelines. Relevant studies were retrieved from databases including PubMed, EMBASE, Cochrane, SciFinder and Web of Science, with a cutoff date of April 30, 2025. Two researchers screened the literature, extracted data, and evaluated the quality of the studies using the QUADAS-2 tool. A bivariate model was used to perform meta-analyses of sensitivity, specificity and heterogeneity for the eNose and GC-MS detection methods.

Results

A total of 39 studies were included in the systematic review, involving 2325 COPD patients and 1574 healthy controls. Among these, 18 studies were incorporated into the meta-analysis, with the pooled sensitivity and specificity of VOCs for differentiating COPD patients from controls being 83% and 78%, respectively. For specific VOCs detection methods, both GC-MS and eNose showed comparable pooled sensitivity of 0.83 and 0.82 respectively. However, eNose demonstrated a significantly higher pooled specificity (0.87; 95% CI: 0.82–0.92) and area under the sROC curve (AUC: 0.9281) than GC-MS with the specificity of 0.68 (95% CI: 0.62–0.74) and an AUC of 0.8291.

Conclusion

This study revealed that VOCs can distinguish COPD patients from healthy controls. While both GC-MS and eNose demonstrated comparable sensitivity, GC-MS - considered the gold standard method for VOCs detection - showed lower specificity and AUC values than eNose in differentiating COPD patients from controls, suggesting that eNose has a lower misdiagnosis rate for COPD. These result may be influenced by the heterogeneity due to the variations in COPD disease stages or differences in breath portion across studies. Future research is recommended to standardize case inclusion criteria and conduct multicenter head-to-head validation studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12890-025-04040-x.

Background

Chronic Obstructive Pulmonary Disease (COPD) ranks as the third leading cause of death worldwide [1]. Epidemiological data show that its prevalence ranges from approximately 38% to 54%, posing a substantial burden on global public health [2]. Currently, pulmonary function tests (PFT) are regarded as the gold standard for diagnosing COPD, however, many cases of COPD are not diagnosed until irreversible lung damage has occurred [3, 4]. Diab et al. revealed that up to 70% of cases may be underdiagnosed, and 30%−60% may be overdiagnosed [3], this issue is especially acute in economically disadvantaged and resource-constrained regions, where the underdiagnosis rate can reach as high as 95.9% [5]. In addition, as the primary population affected by COPD, elderly patients often have physical difficulties to perform the necessary full breathing maneuvers during PFT, resulting in inaccurate test results [6]. Consequently, there is an urgent need for a simple, non-invasive, and reliable screening method for COPD.

As an emerging diagnostic approach, exhaled breath analysis offers significant benefits such as rapidity, convenience, and non-invasive, thereby showing great potential in disease diagnosis [7–9]. The human body produces distinct metabolites under normal physiological and pathological conditions, some of which are excreted through the lung via exhalation in the form of volatile organic compounds (VOCs) [10]. Pathological processes in the lungs of patients with COPD, such as inflammatory responses, oxidative stress, and abnormal cell metabolism, may alter the types and concentrations of VOCs in exhaled breath [11, 12]. By detecting these differences, early diagnosis and disease monitoring of COPD are expected to be achievable.

Numerous studies using various VOCs detection methods have explored the feasibility of VOCs as diagnostic biomarkers for COPD. Gas Chromatography-Mass Spectrometry (GC-MS), electronic nose (eNose) and Real-time mass spectrometry (RT-MS) are the primary methods for VOCs detection. Unlike PFT, which require patients to cooperate in performing specific actions such as forced exhalation, exhaled breath tests only need patients to collect samples through natural exhalation. The process is non-invasive and free of discomfort, making it more acceptable to patients. VOCs detection technologies can be categorized into online and offline methods. Online methods, including RT-MS, eNose and other spectroscopic detection methods. RT-MS rapidly ionizes VOCs into charged ions in the ion source, which then enter the mass analyzer for quick separation and detection based on mass-to-charge ratio differences; eNose captures VOCs response signals via a multi-chemical-sensor array and analyzes the overall signal pattern with machine learning algorithm. Both methods enable rapid exhaled breath detection, but lack the ability to identify and validate specific VOCs [13]. GC-MS, an offline and the gold standard method for VOCs analysis [14], first separates mixed VOCs via a gas chromatographic column, then ionizes the separated VOCs into charged ions in the mass spectrometer, and identifies/quantifies compounds by detecting ion mass-to-charge ratios. It provides more precise, sensitive, and specific VOCs information compared to the aforementioned online methods [15]. Currently, two-dimensional gas chromatography (GC×GC) coupled with MS, as an advanced GC-MS method, has been widely adopted for exhaled VOCs detection, significantly enhancing the separation of chemicals, peak capacity, and resolution [16, 17]. However, this advanced method has not been applied in the field of COPD research.

Various VOCs detection methods have been widely studied in COPD diagnosis, and some studies have developed COPD diagnostic models using VOCs profiles. However, heterogeneities -particularly those related to VOCs detection methods - pose significant challenges to comparing and evaluating results across different studies. To address this, our study conducts a systematic review and meta-analysis of existing research to evaluate the overall diagnostic performance of VOCs analysis methods in COPD, with specific comparison of GC-MS and eNose results.

Materials and methods

Eligibility criteria

This systematic review included trials evaluating VOCs analysis in patients with COPD. The trials were included if their inclusion criteria for study participants involved: patients diagnosed with COPD based on clinical symptoms and PFT (according to the GOLD guidelines, post-bronchodilator lung function with FEV1/FVC < 70%), detectable VOCs in the subjects’ exhaled breath, and studies reported adherence to ethical procedures. The studies were excluded if they had the following reasons: (1) Studies primarily focusing on environmental VOCs; (2) VOCs not derived from exhaled breath; (3) Studies lacking a healthy control group or failing to compare with healthy controls; (4) Focus on the development of VOCs detection technology rather than its application in COPD diagnosis; (5) Protocol publications; (6) Sample size of less than 10 participants; (7) Inaccessibility to the full text of the original study; (8) Studies not written in English or unpublished. The selection process for eligible studies is shown in Fig. 1.

Fig. 1.

Flowchart of eligible studies selection

Search and study selection

This meta-analysis was carried out strictly in accordance with the PRISMA guidelines. Two experienced researchers searched the PubMed, Cochrane, EMBASE, SciFinder and Web of Science databases for literature published up to April 30, 2025. The search keywords included “chronic obstructive pulmonary disease”, “volatile organic compounds”, “exhaled breath diagnosis”, and the full search strategy was added in the supplementary information. The research protocol was registered on the PROSPERO platform (Registration No.: CRD420251031946). The retrieved literature was first de-duplicated, and then the titles and abstracts of the literature were independently screened by two researchers to exclude those that did not meet the inclusion criteria. The full texts of the remaining literature were read to further determine whether they would be finally included in the study.

Data extraction

Relevant information was extracted from the included literature as comprehensively as possible, including the authors, publication years, subject data of the experimental and control groups, VOCs detection methods, exhaled breath sample collection methods, detection results (sensitivity, specificity), and differential VOCs markers. For studies that did not directly report true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), these values were calculated based on the provided sensitivity, specificity, and sample size of the experimental and control groups. All extracted data were documented in Microsoft Excel.

Study quality assessment

The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool was used to evaluate the quality of included studies by four key domains: patient selection, index tests, reference standards, and study flow and timing. For each domain, the risk of bias was assessed, and the clinical applicability of the first three domains was also evaluated. Results were categorized into three risk levels: low, high, and unclear.

Statistical analysis

The software used for the meta-analysis included Excel, MetaDiSc 1.4, and STATA 14. The heterogeneity was assessed using the Spearman’s correlation coefficient, Cochran’s Q test and I² statistic. A random-effects model was used to pool the effect sizes of diagnostic accuracy indicators, fully accounting for the impact of heterogeneity among studies and ensuring the robustness of the results. Publication bias was qualitatively assessed by constructing a funnel plot. The significance level set at α = 0.05. A P-value of less than 0.05 indicated the presence of statistically significant publication bias.

For the VOCs analysis, extract all VOCs reported in the included studies, then classify these VOCs into chemical categories including alcohols, aldehydes, alkanes, alkenes, amines, aromatic compounds, esters, ethers, heterocyclic compounds, ketones, and sulfur-containing compounds. Use R software (version 4.4.0) to summarize the kinds of VOCs included in each chemical category. Visualization was performed with the ggplot2 package to generate a circular bar plot, illustrating the relative frequency and distribution of VOCs across the collected studies.

Results

Study selection and characteristics

After removing duplicate literature, a total of 469 articles were screened. Among these, 430 were excluded for failing to meet the inclusion criteria, and 39 studies were finally included in the systematic review (Table 1). These studies involved 2,325 COPD patients and 1,574 health controls from multiple 14 countries. The control group comprised healthy individuals without other lung diseases. The primary methods for collecting exhaled breath samples in the studies were Tedlar bags, Bio-VOCs and glass syringe. Collected breath samples could be directly detected by an eNose or RT-MS. Alternatively, they could be absorbed onto sorbent tubes or SPME fibers for subsequent indirect analysis via GC-MS. Among these studies, 11 reported information on VOCs detected in exhaled breath (Table 1). A comprehensive analysis revealed that a total of 89 VOCs were detected in the exhaled breath of COPD patients, which then were classified into chemical categories including alcohols, aldehydes, alkanes, alkenes, amines, aromatic compounds, esters, ethers, heterocyclic compounds, ketones, and sulfur-containing compounds (Table S1, Fig. 2). The repeatability of VOCs biomarkers for COPD diagnosis was low across different studies, and only the VOC of hexanal appeared in 3 studies.

Table 1.

Characteristics of included studies

Study	Year	Country	Detection method	VOCs reported	Collection method for exhaled breath samples	Breath portion	Volume	Data Analysis Methods
Allers et al.[18]	2016	Germany	GC-IMS/MCC, GC-APCI-MS	Y	Stainless steel tube, Sorbent Tubes	all/end	0.2mL, 50mL	Welch’s t-test
Avian et al. [19]	2022	China	e-nose	N	Integrated device	all	-	CNN
Basanta et al.[20]	2012	UK	GC-TOF-MS	Y	pressure-monitored adaptive sampler/Sorbent Tubes	end	3 L	LR, PCA, DFA
Basanta et al.[21]	2010	United Kingdom	GC-DMS	N	Resmed Mirage™ NIV full face mask, Sorbent Tubes	end	2.5 L	Kruskal-Wallis ANOVA, PLS-DA
Van Berkel et al.[22]	2009	Netherlands	GC-TOF-MS	Y	Tedlar bag, Sorbent Tubes	all	5 L	SVM, John Platt’s sequential minimal optimization algorithm, RF
Besa et al. [23]	2015	Germany	IMS/MCC	N	Teflon tube	end	10mL	Wilcoxon rank-sum test, Kolmogorov-Smirnov test, ANOVA, Kruskal-Wallis test, Mann-Whitney U test
Bessa et al.[24]	2011	Germany	IMS/MCC	N	Teflon tube	end	10mL	Wilcoxon rank-sum test, Box-and-Whisker plots
Binson et al. (1)[25]	2021	India	e-nose	N	Tedlar bag	all	1 L	SVM, K-NN, LDA, Naive Bayes
Binson et al. (2)[26]	2021	India	e-nose	N	Tedlar bag	all	IL	SVM, PCA
Binson et al. (3)[27]	2021	India	e-nose	N	Tedlar bag	all	1 L	XGBoost, AdaBoost, RF, KPCA
Binson et al. (4)[28]	2021	India	e-nose	N	Tedlar bag	all	1 L	SVM, XGBoost, combined with PCA, LDA, ICA, KPCA
Binson et al. (5)[29]	2021	India	e-nose	N	Tedlar bag	all	1 L	PCA, K-NN, SVM
Bregy et al.[30]	2018	Switzerland	SESI-HRMS	Y	Stainless steel sampling tube	all	-	Whitney-Mann U test, Pearson’s correlation coefficients
Callol-Sanchez et al.[31]	2017	Spain	TD-GC-MS	N	Bio-VOCs, Sorbent Tubes	end	1 L	Mann-Whitney test, Linear correlation, LR
Cazzola et al.[9]	2015	Italy	e-nose, GC-MS	Y	Tedlar bag	all	500mL	PLS-DA, Pearson correlation coefficient
Cristescu​et al. [32]	2011	Netherlands	PTR-MS	N	Tedlar bag	all	1 L	Bootstrapping, LR
de Vries et al. [33]	2023	Netherlands	e-nose	N	SpiroNose	all	-	PCA, LDA, Gradient Boosting Machine, Lasso
Dragonieri​et al. [34]	2009	Netherlands	e-nose	N	Tedlar bag	all	10 L	PCA, CDA
Fens et al. [35]	2009	Netherlands	e-nose	N	Tedlar bag	all	10 L	PCA, CDA
Gaida et al. [36]	2016	Germany	TD-GC-MS	Y	Sorbent Tubes	all	2.5 L	LDA
Hauschild et al. [37]	2012	Germany	IMS/MCC	N	BioScout	end	10mL	Decision Tree, Naive Bayes, Linear SVM, Neural Network, Radial SVM, RF
Incalzi et al. [38]	2012	Italy	e-nose	N	Tedlar bag	all	3 L	Spearman’s rho, PLS-DA
Kistenev et al.[39]	2015	Russia	LPAS	N	Bio-VOCs	all	150 mL	PCA, SIMCA
Kistenev et al. [40]	2017	Russia	LPAS SPME-GC-MS	N	Bio-VOCs, SPME	all	150 mL	PCA, SVM
Kistenev et al. [41]	2019	Russia	LPAS	N	Bio-VOCs	end	150mL	PCA, SVM
Krauss et al. [42]	2020	Germany	e-nose	N	Integrated device	-	-	Neural Network
Mahdavi et al. [43]	2024	Iran	e-nose	N	Tedlar bag	all	1 L	SVM, KNN, MLP, RF
Monedeiro et al.[44]	2021	Poland	GC-MS	Y	Tedlar bag/Needle Trap Device	all	50mL	PCA, Mann-Whitney test, RF, Multinomial LR
Phillips et al.[45]	2012	UK	GC-MS	N	Bio-VOCs, Sorbent Tubes	end	129mL	J48, JRIP, PART, SMO, FNN, RF, FRNN, VQNN,
Phillips et al. [46]	2014	UK	GC-MS	Y	Bio-VOCs, Sorbent Tubes	end	129mL	J48, JRIP, PART, paired t-tests
Pizzini et al. [47]	2018	Austria	TD-GC-ToF-MS	Y	glass syringes	end	400mL	One-way ANOVA, Post hoc test, RF
Rodríguez-Aguilar et al.[48]	2019	Mexico	ultrafast GC-enose	Y	Tedlar bag	all	1 L	PCA, CAP
Rodríguez-Aguilar et al.[49]	2021	Mexico	e-nose	N	Metallized plastic bag	all	1.4 L	PCA, CAP, One-way ANOVA, Dunnett’s post-hoc test, Fisher’s exact test​
Scarlata et al.[50]	2017	Italy	e-nose	N	Adsorbing cartridge	all	-	PLS-DA
Shafiek et al.[51]	2015	Spain	e-nose	N	Tedlar bag	all	5 L	LDA, Chi-square, Mann-Whitney, Kruskal-Wallis tests
Sibila et al. [52]	2014	Spain	e-nose	N	Tedlar bag	all	10 L	PCA, CDA, LR, ANOVA
Sinues et al. [53]	2014	Switzerland	SESI	N	Teflon tube	end	-	Mann-Whitney U test, PCA, KNN
Westhoff et al. [54]	2011	Germany	IMS/MCC	N	-	-	-	Mann-Whitney-Wilcoxon rank sum test, PCA, Decision Trees, Correlation analysis, Box-and-Whisker plots
Xiong et al. [55]	2025	China	TD-GC-MS	Y	Self-modified sampling devices/Sorbent Tubes	all	1.5 L	PCA, LASSO LR

GC-IMS/MCC GC-Ion Mobility Spectrometry/Multi-Capillary Column, GC-APCI-MS GC-Atmospheric Pressure Chemical Ionization- MS, GC-TOF-MS GC-Time-of-Flight Mass Spectrometry, GC-DMS GC-Differential Mobility Spectrometry, PTR-MS Proton-Transfer Reaction MS, TD-GC-MS Thermal Desorption-GC-MS, LPAS Laser Photoacoustic Spectroscopy, SPME-GC-MS Solid Phase Microextraction-GC-MS, TD-GC-ToF-MS Thermal Desorption-Gas Chromatography-Time-of-Flight MS, SESI Secondary Electrospray Ionization, SESI-HRMS Secondary Electrospray Ionization-High Resolution MS, CNN Convolutional Neural Network, LR Logistic Regression, PCA Principal Component Analysis, DFA Discriminant Function Analysis, PLS-DA Partial Least Squares Discriminant Analysis, SVM Support Vector Machine, SMO Sequential Minimal Optimization, RF Random Forest, K-NN k-Nearest Neighbors, LDA Linear Discriminant Analysis, KPCA Kernel Principal Component Analysis, ICA Independent Component Analysis, CDA Canonical Discriminant Analysis, SIMCA Soft Independent Modeling of Class Analogy, CAP Canonical Analysis of Principal Coordinates, J48 C4.5 Decision Tree, JRIP Repeated Incremental Pruning to Produce Error Reduction, PART Partial Decision Tree, FNN Fuzzy Nearest-Neighbour Classifier, FRNN fuzzy-rough set nearest-neighbour approach, VQNN Noise-tolerant fuzzy-rough set-based classifier

Fig. 2.

The circle-plot of exhaled breath VOCs in COPD patients. The length of the column represents the frequency of the kind of VOCs appearing in the existing literature

In addition, 18 of 39 studies involving 699 COPD patients and 563 healthy controls were included in the meta-analysis part; the remaining 21 of 39 studies were excluded due to a lack of sensitivity and specificity data. Among the meta-analysis included studies, the sensitivity of exhaled VOCs detection for COPD ranged from 67% to 100%, and the specificity ranged from 50% to 100% (Table 2). Based on the differences in detection methods employed by the included studies, subgroup analysis was conducted for three categories: GC - MS, eNose and others detection methods (Fig. 1). For the GC-MS subgroup, 8 studies with 362 COPD patients and 234 healthy controls were included, with the sensitivity ranging from 67% to 100% and the specificity from 50% to 97.7% (Table 2). For the eNose subgroup, 7 studies involving 204 COPD patients and 199 healthy controls were included, with the sensitivity ranging from 72.5% to 100% and the specificity from 71% to 100% (Table 2). For other detection methods (e.g., RT-MS and LAPS), the sample size was too small to conduct a meta-analysis. Thus, this review only compared the diagnostic efficacy of GC-MS and eNose in distinguishing COPD from healthy individuals.

Table 2.

Diagnostic performance of each study

Study	Mata-analysis or systematic review	Detection methods	Sensitivity (95% CI)	Specificity (95% CI)	COPD	Control	Stage of COPD (number of patients)	Other lung diseases
Avian et al. 2022[19]	Both	eNose, null	100	100	20	10	-	null
Binson et al. 2021 (2)[26]	Both	eNose, TGS	72.5	82.35	22	39	-	unknown
Binson et al. 2021 (4)[28]	Both	eNose, TGS	88.14	91.3	52	63	-	unknown
Cazzola et al. 2015[9]	Both	eNose, null	96	71	27	7	-	null
Mahdavi et al. 2024[43]	Both	eNose, TGS-2602	78.79	82.35	33	34	-	null
Rodríguez-Aguilar et al. 2019[48]	Both	eNose, null	95.8	90.9	23	33	GOLD I: 3, GOLD II: 8, GOLD III: 8, GOLD IV: 4	null
Sibila et al. 2014[52]	Both	eNose, Cyranose 320	81	86	27	13	GOLD I: 0, GOLD II: 5, GOLD III: 26, GOLD IV: 6	infection
Basanta et al. 2010[21]	Both	GC-MS	88	81	20	6	-	null
Basanta et al. 2012[20]	Both	GC-MS	85	50	39	32	GOLD I: 2, GOLD II: 19, GOLD III: 16, GOLD IV: 2	null
Berkel et al. 2009[22]	Both	GC-MS	100	81	16	16	-	null
Kistenev et al. 2017[40]	Both	GC-MS	68	60	12	11	-	null
Monedeiro et al.2021[44]	Both	GC-MS	67	97.7	12	20	-	null
Phillips et al. 2012[45]	Both	GC-MS	79	64	119	63	GOLD I: 7, GOLD II: 48, GOLD III: 38, GOLD IV: 22	null
Phillips et al. 2014[46]	Both	GC-MS	80	62	118	63	-	null
Xiong et al. 2025[55]	Both	GC-MS	77	87	26	23	-	null
Bregy et al. 2018[30]	Both	RT-MS	93	87	22	14	GOLD I: 4, GOLD II: 10, GOLD III: 7, GOLD IV: 1	null
Sinues et al. 2014[53]	Both	RT-MS	88	92	25	25	GOLD I: 2, GOLD II: 1, GOLD III: 2, GOLD IV: 10	unknown
Kistenev et al. 2019[41]	Both	LPAS	86	83	12	29	-	unknown
Binson et al. 2021 (1)[25]	Systematic review	eNose, TGS	-	-	48	90	-	null
Binson et al. 2021 (3)[27]	Systematic review	eNose, TGS	-	-	55	93	-	null
Binson et al. 2021 (5) [29]	Systematic review	eNose, TGS	-	-	38	72	-	null
de Vries et al. 2023[33]	Systematic review	eNose, null	-	-	682	unknown	GOLD I:59,GOLD II:327, GOLD III:202,GOLD IV:57	unknown
Dragonieri​ et al.2009[34]	Systematic review	eNose, Cyranose 320	-	-	10	10	-	null
Fens et al. 2009[35]	Systematic review	eNose, Cyranose 320	-	-	30	40	GOLD I: 0, GOLD II: 20, GOLD III: 10, GOLD IV: 0	null
Incalzi et al. 2012[38]	Systematic review	eNose, null	-	-	20	5	GOLD I: 5, GOLD II: 5, GOLD III: 5, GOLD IV: 5	null
Krauss et al. 2020[42]	Systematic review	eNose, Aeonose®	-	-	23	33	-	null
Rodríguez-Aguilar et al. 2021[49]	Systematic review	eNose, Cyranose 320	-	-	50	50	GOLD I: 8, GOLD II: 18, GOLD III: 9, GOLD IV: 14	null
Scarlata et al. 2017[50]	Systematic review	eNose, null	-	-	20	56	-	unknown
Shafiek et al. 2015[51]	Systematic review	eNose, Cyranose 320	-	-	143	30	-	pneumonia
Allers et al.2016[18]	Systematic review	GC- MS	-	-	79	73	-	null
Callol-Sanchez et al.2017[31]	Systematic review	GC- MS	-	87.6	40	89	GOLD I: 4, GOLD II: 13, GOLD III: 10, GOLD IV: 13	null
Gaida et al. 2016[36]	Systematic review	GC-MS	-	-	100	122	GOLD I: 31, GOLD II: 38, GOLD III: 16, GOLD IV: 4	null
Kistenev et al. 2015[39]	Systematic review	LPAS, GC-MS	-	-	12	11	-	null
Pizzini et al. 2018[47]	Systematic review	GC-MS	-	-	30	24	GOLD I: 0, GOLD II: 10, GOLD III: 7, GOLD IV: 13	null
Bessa et al. 2011[24]	Systematic review	RT-MS	-	-	13	33	GOLD I: 0, GOLD II: 0, GOLD III: 4, GOLD IV: 9	unknown
Besa et al. 2015[23]	Systematic review	RT-MS	-	-	45	51	GOLD I: 3, GOLD II: 5, GOLD III: 16, GOLD IV: 21	null
Cristescu et al.2011[32]	Systematic review	RT-MS	-	-	83	121	GOLD I: 53, GOLD II: 25, GOLD III: 5, GOLD IV: 0	unknown
Hauschild et al. 2012[37]	Systematic review	RT-MS	-	-	84	35	-	bronchial carcinoma
Westhoff et al. 2011[54]	Systematic review	RT-MS	-	-	95	35	-	bronchial carcinoma

TGS Taguchi gas sensors (TGS) manufactured by Figaro Engineering Inc

Studies quality analysis

The quality assessment results for the 39 studies included in the systematic review were shown in Figure S1. For the 18 studies included in the meta-analysis comparing COPD with healthy controls, the quality assessment results were presented in Figure S2. Additionally, the quality assessment results for the meta-analysis comparing the diagnostic efficacy between GC-MS and eNose in COPD were displayed in Fig. 3. In the patient selection domain, higher risks of bias and applicability issues were observed because almost all included studies were primarily case-control studies. Similarly, most studies had a high risk of bias in the index test domain, as they relied on clinically diagnosed COPD results as the reference standard, although applicability concerns here were minimal. In the reference standard domain, nearly all studies used the GOLD guidelines for COPD diagnosis, resulting in low risks of bias and high applicability. For the flow and timing domain, most studies had a low risk of bias; however, 3 GC-MS studies did not clearly state whether all patients underwent the same testing standards, leading to unclear bias risks.

Fig. 3.

Risk of bias and applicability concerns for the VOCs detection methods of GC-MS (A) and eNose (B) on the COPD diagnosis using the QUADAS-2 tool

Meta-analysis

This meta-analysis consists of three parts: first, a meta-analysis of VOCs for COPD diagnosis; second, a meta-analysis of the diagnostic performance of exhaled VOCs detected by GC-MS in COPD; and finally, a meta-analysis of eNose in COPD diagnosis. Meanwhile, the diagnostic performance of GC-MS and eNose was compared based on sensitivity, specificity and AUC values.

Firstly, a total of 18 studies were included in meta-analysis of exhaled VOCs for COPD diagnosis. The Spearman coefficient was −0.162 (P=0.521) and the summary receiver operating characteristic (sROC) curve showed no shoulder-arm pattern, indicating that threshold effects did not contribute to heterogeneity. The Cochran-Q test for the diagnostic odds ratio (DOR) was 45.38 (P=0.001), indicating the presence of heterogeneity due to non-threshold effects. The I² values for sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, and DOR were from 39.8% to 73.2%, indicating moderate to high heterogeneity. Thus, a random-effects model was used to combine these effect sizes. Consequently, the feasibility of exhaled VOCs detection for COPD was estimated. The meta-analysis results showed that the combined sensitivity of all included studies was 83% (95% CI: 80%−86%), and the combined specificity was 78% (95% CI: 75%−82%). The analysis of the sROC showed an AUC of 0.8995, indicating that the exhaled VOCs detection technology performed well in the COPD diagnosis (Fig. 4).

Fig. 4.

Clinical sensitivity (A), specificity (B), and sROC curve (C) the included studies which was used for discrimination between COPD patients and health control based on the exhaled breath VOCs

Second, the Spearman’s correlation coefficient for GC-MS and eNose in discriminating COPD patients from healthy control were 0.357 (P=0.385) and −0.429 (P =0.337), respectively, and neither sROC curves showed a shoulder-arm pattern. The Cochran-Q test for the DOR of GC-MS and eNose yielded values were 10.96 (P=0.140) and 11.47 (P=0.075), respectively. The I² values for both GC-MS and eNose indicated moderate to high heterogeneity. So a random-effects model was used to combine these effect sizes. For GC-MS, the pooled sensitivity was 80% (95% CI: 76%−84%), the specificity was 68% (95% CI: 62%−74%), and the AUC was 0.8291. For eNose-related studies, the pooled sensitivity was 87% (95% CI: 82%−92%), the specificity was 87% (95% CI: 82%−92%), and the AUC was 0.9281) (Fig. 5). Thus, eNose showed comparable sensitivity but higher specificity than GC-MS, suggesting that eNose may reduce misdiagnosis in clinical screening for COPD. This result may be influenced by heterogeneity due to variations in the disease stages of COPD cases or differences in the breath portion across studies. The publication bias test yielded P values of 0.08 for GC-MS and 0.75 for eNose, and the funnel plots were symmetrical, indicating no publication bias (Fig. 6).

Fig. 5.

Clinical sensitivities (A, D), specificities (B, E), and sROC curves (C, F) of different VOCs detection methods in the included studies. A-C for GC-MS and D-F for eNose

Fig. 6.

Publication bias analysis of GC-MS (A) and eNose (B)

Discussion

The bronchial provocation test may induce airway spasm or drug side effects. Meanwhile, PFT is time-consuming, requiring high equipment and labor costs, which hinders its use in primary medical institutions. In contrast, exhaled breath VOCs detection offers advantages such as non-invasiveness, easy of operation, highly repeatability, and good patient compliance, making it a promising method for COPD diagnosis.

Firstly, this review indicates that exhaled VOCs detection technology has potential in diagnosing COPD. The combined sensitivity, specificity, and AUC values demonstrate its good diagnostic accuracy for COPD. Secondly, both GC-MS and eNose show favorable performance in COPD diagnosis based on VOCs. Thirdly, eNose showed comparable sensitivity but higher specificity than GC-MS, suggesting that eNose may reduce misdiagnosis in clinical screening for COPD. This result may be influenced by heterogeneity due to variations in the disease stages of COPD cases or differences in the breath portion across studies. Meanwhile, heterogeneity may reflect the different disease subtypes, progression stages, or inflammatory phenotypes, exhibiting potential clues for analyzing the pathophysiological mechanisms of diseases.

In addition, the limited number of studies analyzing the correlation between COPD severity grades, subtypes, and exhaled VOCs, relevant data were insufficient to conduct in-depth statistical analysis. And the variability exists across included studies in terms of breath collection protocols, sample volumes, breath fractions, detection parameters, and analytical methods. Such inconsistencies have led to a lack of standardized VOCs biomarker combinations, posing substantial challenges for COPD diagnosis using exhaled breath VOCs. Furthermore, the application of exhaled breath VOCs testing for COPD diagnosis has not been validated in large-scale clinical studies. Therefore, extensive research remains necessary to facilitate the widespread clinical application of exhaled breath VOCs in COPD diagnosis.

Differential vocs’ mechanism

Although numerous VOCs have been reported as potential biomarkers on COPD diagnosis, they have not been validated in large-scale clinical cohorts or translated into clinical applications. And the sources and mechanisms of most VOCs are unclear, hindering their clinical translation. Moreover, like COPD, other lung diseases may have involved similar inflammatory or oxidative stress reactions, leading to comparable changes in VOCs.

Inflammatory reaction in COPD patients may induce abnormal fatty acid metabolism, leading to changes in the concentration of short-chain alkanes/alkenes or fatty aldehydes [56, 57]. For example, isoprene - a representative alkene - is associated with the cholesterol biosynthesis pathway [58], though its specific mechanism requires further investigation. Hexanal, which appears frequently in multiple studies, may be linked to lipid peroxidation caused by oxidative stress and inflammatory response in the lungs of COPD patients, but its specific mechanism also warrants in depth exploration [59]. Ketone substances such as acetone are mainly produced through β-oxidation of fatty acids in the body [60].

Benzene and its derivatives may originate from the external environment sources, such as air pollution and smoking. Due to the large surface area and abundant capillaries in the lungs, benzene and its derivatives can rapidly enter lung tissue through the alveolar epithelium and capillary walls, where they are metabolized by cytochrome P450 enzymes [61]. For instance, benzene is first oxidized to benzene epoxide, and then further metabolized into phenol, catechol, or hydroquinone [56]. The metabolic processes of toluene and xylene are similar to those of benzene [56, 62]. Some of these metabolites can be excreted via urine, while others - such as benzene and hydroquinone - may be excreted through respiration [63].

Disease severity

As COPD is a heterogeneous disease, the pathophysiological states such as degree of airway inflammation, lung function impairment, systemic oxidative stress levels, different severities as GOLD stages 1–4, and acute exacerbation/stable phase, vary significantly [64, 65]. These differences directly affect the composition of exhaled VOCs, leading to the different VOCs composition of COPD patients with different severities. For example, Pizzini et al. found that levels of VOCs such as n-butane, cyclohexanone, 2-pentanone, and 4-heptanone were elevated in the exhaled breath of patients with acute exacerbation of COPD (AECOPD) compared with those in stable stage COPD, with a sensitivity of 0.78 and specificity of 0.91 [47]. Moreover, indole and 2-pentanone were the main contributing VOCs for distinguishing the two groups, confirming the potential of exhaled VOCs in detecting and distinguishing between the acute exacerbation phase and stable phase of COPD. Sinues et al. also observed differences in exhaled VOCs among COPD patients with different GOLD stages, and the sensitivity and specificity for distinguishing between GOLD I/II and GOLD III/IV were 0.92 and 0.83, respectively [53]. Besides, Sibila et al. reported that VOCs can differentiate COPD patients with bacterial airway colonization from those without, offering a potential approach for the precise diagnosis and personalized treatment of COPD [52].

In this review, only 16 out of the 39 included studies reported the GOLD stage, and among these, 7 both reported GOLD stage and met the meta-analysis inclusion criteria. Thus, the subgroup analysis of COPD was not included in the current review.

Other lung diseases

Currently, the research on exhaled VOCs in the field of respiratory diseases has emerged as a pivotal direction for non-invasive diagnosis, demonstrating substantial potential for clinical application. Beyond its utility in COPD diagnosis, exhaled VOCs detection technology has also been applied to the field of asthma, non-small cell lung cancer (NSCLC), and respiratory infections. For instance, Fan et al. revealed that over 190 types of VOCs have been screened for lung cancer, with hexanal being the most frequently detected VOC [66]; Rufo et al. conducted a systematic review and meta-analysis, and showed that alkane compounds were the VOCs class most significantly associated with asthma-related inflammation [56]. In the field of infectious diseases, numerous studies were reported that they screened COVID-19 and its variants using exhaled breath VOCs [67, 68].

For distinguishing COPD from other diseases, Quaranta et al. analyzed exhaled VOCs using the Cyranose 320 in 98 COPD patients including 42 patients with bronchiectasis-COPD overlap syndrome [69]. Tian et al. identified 5 VOCs capable of distinguishing COPD from asthma, with an AUC of 0.92 ± 0.01 [70]. Dragonieri et al. conducted a differential analysis of exhaled VOCs between NSCLC and COPD patients [34]. van der Sar et al. analyzed exhaled VOCs using the SpiroNose including 50 COPD patients and 161 ILD patients, and the results showed that the AUC reached to 0.96 (95% CI: 0.90–1.00.90.00), with a specificity of 0.94 and a sensitivity of 0.89 [71]. The results of differences in VOCs between COPD and other lung diseases can serve as evidence for the existence of COPD-specific VOCs. However, these truly differential VOCs that can be used in clinical applications remain to be explored.

Sample collection methods

The collection segment (e.g., full-exhaled breath, end-tidal breath) and volume of exhaled breath samples can also affect the detection results. End-tidal breath, used in 13 studies, contains a higher proportion of alveolar gas, and its VOCs profile is less affected by the external environment, thus better reflecting pulmonary metabolic status [9]. However, there is currently no unified standard for the optimal collection volume, with volume in the studies ranging from 0.2 ml to 10 L. Tedlar bags and Bio-VOCs are the main two types of exhaled breath collection devices. Given that VOCs levels in exhaled breath are at the parts per million (ppm) level or lower, multiple studies have used pre-concentration devices to enrich breath samples into sorbent tubes or fibers prior to measurement, such as thermal desorption (TD), solid-phase microextraction (SPME), and needle trap devices (NTD) [9, 31, 44]. For eNose and RT-MS, breath samples can be collected using integrated devices; some samples are collected in Tedlar bags and then measured offline. GC-MS, when combined with pre-concentration devices, is commonly measured offline.

Detection methods

GC-MS is currently recognized as the gold standard for VOCs detection, as it can accurately identify and quantitatively analyze the compounds in exhaled breath [72]. However, due to its high cost and requirement for professional operation, GC-MS is more suitable for VOCs discovery in laboratory settings rather than clinical test [73]. GC×GC-MS has significantly improved the accuracy and sensitivity of VOCs detections in exhaled breath. This method has been used to distinguish between neutrophilic and eosinophilic asthma [74], but has not yet been reported in COPD research.

ENose analyzes exhaled VOCs through pattern recognition algorithms, classified by their core sensor technologies. Metal Oxide Semiconductor (MOS) eNoses rely on changes in the conductivity of gas-sensitive materials; conductive polymer eNoses depend on resistance variations of polymers; piezoelectric eNoses leverage frequency shifts caused by gas adsorption; optical eNoses rely on the absorption/scattering of light by gases; electrochemical eNoses depend on current generated from gas-electrode reactions; gas chromatography (GC)-integrated eNoses first separate gas components before detection; biological eNoses rely on specific binding of biological receptors; and hybrid eNoses integrate multiple sensors with AI to analyze multi-source signals, all for the qualitative and quantitative detection of gases [75–77]. But these methods cannot accurately identify specific VOCs [78].

Once VOCs are validated by GC-MS in COPD or other diseases, eNoses, RT-MS or other portable exhaled breath detection devices - due to their simplicity of operation, rapid detection, and low cost - can be customized for clinical application. Portable, compact devices can facilitate point-of-care testing (POCT), allowing for rapid VOCs detection in outpatient clinics, community healthcare centers, or even home monitoring scenarios, which is far more accessible than large-scale lab-based GC-MS systems. Thus, we propose that developing miniaturized clinical devices based on biomarkers identified by GC-MS is a viable approach to accelerate the clinical translation of VOCs in COPD.

Statistical methods

Current research on VOCs in COPD primarily focuses on identifying potential biomarkers, with two broad analytical approaches commonly employed. The first is a hypothesis-driven framework based on classical statistical tests - such as t-tests, Wilcoxon rank-sum tests and analysis of variance (ANOVA) - often combined with multivariate analysis techniques like partial least squares discriminant analysis (PLS-DA) or orthogonal PLS-DA (OPLS-DA). These methods offer a degree of interpretability and are suitable for detecting group-level differences; however, they are generally limited in their ability to capture complex, nonlinear interactions among VOCs [79]. The second approach involves feature selection through machine learning algorithms, including LASSO regression, Elastic Net and Random Forest. These techniques are well-suited to handle high-dimensional and multicollinear data, enabling the identification of more subtle biomarker patterns. Their performance, however, is highly dependent on sample size and data quality [80]. A key limitation of many current studies is their small sample size, which introduces a significant risk of overfitting and reduces model generalizability [81, 82].

In exhaled breath VOCs studies, the diagnosis model of disease needs to be validated. Selecting appropriate validation strategies is essential to demonstrate the reliability and generalizability of newly proposed biomarkers or diagnostic tests. In general, data validation can be divided into internal and external validation. Internal validation (such as k-fold cross-validation or random data splitting) is used to assess the stability of the proposed biomarkers within the same dataset [83]; however, due to the homogeneity of data sources, this approach may still lead to overfitting and an overestimation of performance [84]. External validation, by contrast, evaluates the true performance of biomarkers or diagnostic tests using samples collected at different locations, times, or from independent populations, and is a powerful approach to assess their reproducibility and clinical generalizability [85]. Among the studies included in our analysis, only two employed external validation, while the remaining studies relied on internal validation, which may have contributed to the risk of overfitting in the present study.

Moreover, existing studies often lack transparency and consistency in data preprocessing procedures, such as normalization, batch effect correction, imputation of missing values and VOCs filtering criteria. These steps are critical in breathomics; insufficient reporting of data processing may compromise the reproducibility and cross-study comparability of findings. To enhance the robustness and translational potential of diagnostic models, future studies should not only increase sample size but also adopt reasonable analytical workflows and integrate interpretable modeling strategies to enhance diagnostic feasibility and reliability.

Suggestions

Currently, studies on COPD diagnosis using exhaled VOCs exhibit inconsistencies in detection methods, sample collection, data statistics and patient selection, leading to greater variability compared with well-established PFT diagnostic methods [86]. Therefore, developing a unified standard operating procedure (SOP) for exhaled breath testing is crucial. Additionally, a “standard exhaled VOCs profile for healthy individuals” should be established, accounting for factors such as age, gender, and ethnicity, to serve as a reference for subsequent research and clinical applications [87].

Emerging VOCs detection technologies, such as GC×GC-MS, which enhance the accuracy and separation efficiency of VOCs, should be widely adopted. The lack of large-scale clinical cohorts has also hindered the validation of VOCs for COPD diagnosis; thus, combining exhaled VOCs detection with PFT diagnostic methods may offer a more precise means for the early diagnosis of COPD.

Furthermore, although 89 VOCs were discovered in discriminating between COPD and healthy control, only a few are consistently reported across studies. Exhaled VOCs are highly susceptible to multiple confounding factors, making it necessary to stratify patients in detail. Specifically, analyzing the relationships between COPD severity grades, smoking history, and comorbid lung diseases can help exclude interfering VOCs [27, 53]. Future research should integrate cytological experiments with clinical data to explore the underlying mechanisms by which key VOCs act as biomarkers for COPD, clarifying why and how they function in this role.

Limitations

This study demonstrates the feasibility of exhaled breath VOCs detection for COPD, and after comprehensive evaluation, indicates that eNose exhibits better diagnostic performance for COPD than GC-MS. However, several limitations remain: Firstly, the included studies employed diverse exhaled breath collection methods, breath fractions, detection techniques, and analytical approaches, making cross-study comparisons challenging. Secondly, the sample size of nearly all included studies was fewer than 100 participants, which may compromise the reliability of statistical results. Thirdly, a notable limitation is that many studies did not report the severity of COPD in patients, potentially hindering robust comparisons between GC-MS and eNose. Finally, this study focuses on exploring VOCs in COPD compared to healthy individuals, without including comparisons between COPD and other respiratory diseases such as asthma, lung cancer, interstitial lung disease, and bronchiectasis.

Conclusion

This study synthesizes existing research findings on exhaled breath detection for COPD diagnosis. The results indicate that exhaled breath detection - being a non-invasive and convenient method - yields promising outcomes for identifying exhaled VOCs in COPD diagnosis. After comprehensive evaluation of the included studies, eNose demonstrates better diagnostic performance for COPD than GC-MS. However, large-scale, multi-center studies are still needed to further compare the diagnostic feasibility of GC-MS and eNose. Additionally, establishing a “standardized” framework to unify breath sampling procedures, VOCs detection protocols, and statistical analysis methods will accelerate the clinical translation of exhaled breath-based diagnostics for COPD.

Abbreviations

VOCs

Volatile Organic Compounds

COPD

Chronic Obstructive Pulmonary Disease

eNose

Electronic Nose

GC-MS

Gas Chromatography-Mass Spectrometry

GC×GC

Two-Dimensional Gas Chromatography

RT-MS

Real-Time Mass Spectrometry

SROC

Summary Receiver Operating Characteristic Curve

DOR

Diagnostic Odds Ratio

TD

Thermal Desorption

SPME

Solid-Phase Microextraction

NTD

Needle Trap Devices

ANOVA

Analysis of Variance

PLS-DA

Partial Least Squares Discriminant Analysis

OPLS-DA

Orthogonal PLS-DA

AECOPD

Acute Exacerbation of COPD

PFT

Pulmonary Function Tests

TP

True Positives

FP

False Positives

TN

True Negatives

FN

False Negatives

Author contributions

This study was designed by YH and ZQ. YC (Yaoxi Chen) and YC (Yingying Chai) performed the literature search and data collection. YC (Yaoxi Chen) and ZJ conducted the meta-analysis. The manuscript was written by YC (Yingying Chai), YC (Yaoxi Chen), YH and ZQ drafted the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by National Natural Science Foundation of China (CN) (82370100); National Science and Technology Innovation 2030 Major Project for Prevention and Treatment of Cancer, Cardiovascular, Respiratory and Metabolic Diseases (2023ZD0506100, 2023ZD0506106); 1.3.5 Project for disciplines of excellence, West China Hospital, Sichuan University (ZYGD22009; RHM24202); 1.3.5 Project of State Key Laboratory Health and Multimorbidity, West China Hospital, Sichuan University (RHM24202).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.