Chronic obstructive pulmonary disease (COPD) is an inflammatory airway disease characterised by poorly reversible airway obstruction. In fact, COPD can be viewed as an umbrella term that encompasses a range of pulmonary and systemic manifestations. COPD severity is graded by forced expiratory volume in 1 second (FEV₁) 1, but this grading does not recognise the range of pathophysiological abnormalities that may be present in this heterogeneous condition. There is currently much interest in improving the phenotypic description of COPD by the use of biomarkers that allow distinct subgroups of patients with different prognosis or response to therapy to be identified 2.

Induced sputum is a safe and non-invasive method for studying biomarkers of airway inflammation in COPD patients, neutrophil 3 and eosinophil 4 numbers being the most valuable measures at present. An alternative biomarker is nitric oxide (NO), which is synthesized from L-arginine by nitric oxide synthase (NOS) enzymes and can be measured in exhaled breath (Fe_NO). Fe_NO (fractional exhaled nitric oxide) has not become widely used as a biomarker in COPD patients as it is reduced by current cigarette smoking 5 and can therefore mainly be used in ex-smokers 67 and subjects with unstable disease 8.

Biomarkers of airway inflammation, such as induced sputum and Fe_NO, clearly have the potential to define subgroups of COPD patients with different characteristics. In order to use these biomarkers to enhance phenotype description, it would be important to know other clinical characteristics associated with these biomarkers. For example, patients with COPD have increased levels of systemic inflammation 910, with CRP levels being associated with increased mortality 11, possibly through cardiovascular disease 10. Such associations between airway and systemic inflammation may point towards specific pathophysiological mechanisms that contribute to disease characteristics.

Multivariate modelling overcomes has been used to test pre-determined hypotheses concerning the relationships between biomarkers and other measurements in COPD 1213. An alternative strategy is to use methods that generate hypotheses rather than test pre-determined hypotheses. Exploratory factor analysis 1415 is a hypothesis generating method that identifies groups of associated parameters into factors that are responsible for disease heterogeneity. This approach has previously been used in COPD to demonstrate dissociation between airway inflammation and pulmonary function 16. Principal components analysis (PCA) is the commonest form of factor analysis and reduces a large number of variables to a much smaller number of components, explaining the variability within the data set. These components represent latent processes which cannot be directly measured. In the context of COPD, components may represent the pathophysiological processes responsible for disease heterogeneity.

We report the use of PCA to explore the heterogeneity in markers of airway and systemic inflammation and pulmonary function in a cohort of subjects with COPD. The primary aim of this study was to identify components representative of the different pathophysiological processes and hence generate hypothesis concerning COPD phenotype description. We also used traditional multivariate modelling to test the predetermined hypothesis that the non-invasive airway biomarkers studied were associated with other disease parameters.

Figure 1 shows that of the 127 patients, 10 patients could not perform Fe_NO adequately, 21 patients did not have blood taken for analysis, and 92 produced adequate sputum for analysis. There was no difference in pulmonary function or blood biomarker measurements between the patients who could and could not perform these measurements. All patients were included in the analysis, with 70 patients having a complete dataset with all variables. The post bronchodilator FEV₁ range for these COPD patients was 17.9 to 79.6%. 98 of the 127 patients had moderate COPD (GOLD stage II), while 22 had severe disease (GOLD stage III) and 4 very severe disease (GOLD stage IV).

The primary aim of this study was to generate hypotheses about COPD phenotype description and disease mechanisms by exploring the variability in markers of inflammation and lung function using PCA. This analysis suggests that COPD is a truly multi-dimensional disease. PCA identified four main components, each explaining similar amounts of the variance (between 10 and 20%). The first two components represented neutrophilic and eosinophilic inflammation, explaining 20.2% and 18.2% of the variance respectively. Lung function parameters formed a separate component, comprising measures of airflow obstruction and reversibility. CRP also formed a separate component. Some hypotheses about disease mechanisms can be generated from this analysis; component 1 suggests that the profile of neutrophilic airway inflammation is associated with systemic inflammation, and component 2 suggests that patients with sputum eosinophilia, which is associated with increased corticosteroid responsiveness 4 also have increased Fe_NO levels. Importantly, PCA indicates that these are distinct components of disease that could be used for patient phenotyping 19. Correlations between the components were weak despite the use of a Promax rotation. To validate the PCA components, we performed multivariate modelling, which confirmed our PCA findings.

The main limitation of any PCA is the selection of variables included. This analysis has focused on a selection of well studied markers of airway 34 and systemic inflammation 132021 as well as pulmonary function, by which COPD is classically defined 1. Other important biomarkers of COPD pathophysiology, explaining further disease heterogeneity may not have been included. Nevertheless, our study shows the potential utility of PCA, and further studies using other biomarkers of inflammation or clinical measurements would be of interest.

The results of PCA are critically dependent on the selection of subjects. If particular subgroups of patients are included or excluded from the study, the sources of variation in the dataset will be affected. A common issue in studies of airway sampling in COPD patients, either by induced sputum or Fe_NO, is that not all patients can complete each measurement 16. We used a well validated approach to this issue, excluding cases pairwise where data was missing 14, so that all of the 92 patients with induced sputum data and 117 patients with Fe_NO data could be analysed where possible e.g. all of these data could be analysed against pulmonary function. We did not exclude patients who could not perform certain analysis, as this may have introduced a bias into the dataset; e.g. certain patient phenotypes may produce less sputum than others, and by excluding such patients any such a phenotype would be poorly represented in the data set.

Factor analysis/PCA has been rarely used in COPD 1622232425. The sample for the currents study compares very favourably with these studies, which have often enrolled less than 100 subjects 22232425. Indeed, even if we accounted for the incomplete measurements in the current study, the sample size of patients with a "complete dataset" (n = 70), is still larger than the enrolled sample size of some of these studies 1622232425.

Recently, factor analysis has been used by Lapperre et al 16 in 114 COPD patients to generate hypothesis about disease description using many of the same parameters as the current study, but importantly systemic inflammation biomarkers were not investigated. A four factor solution was reported, with the character of the components being somewhat different to our findings. Firstly, a factor representing asthma like parameters (i.e. reversibility, bronchial hyper-reactivity and atopy) was identified. This may be due to some extent to differences in subject selection, as in our cohort subjects with asthma may have been more rigorously excluded, as atopic subjects were not recruited. Secondly, Lapperre et al reported a component including sputum percentage neutrophils and eosinophils, and that Fe_NO was not in the same component as sputum eosinophils, which differs from our results. Important methodological issues should be considered; (a) Lappere et al used 2 different Fe_NO analysers, but it is known that data from different analysers generates significantly different NO levels 26. This could explain why Fe_NO was in a distinct component, as associations with other parameters were not possible as the absolute Fe_NO values were actually mostly dependent on the type of analyser used rather than any patient characteristics (b) Laperre et al included both sputum eosinophil and neutrophil % in the factor analysis, which together were found to form a distinct component. However, there is a mathematical relationship between these parameters (as one increases, the other decreases). The impact on PCA analysis is that the mathematical relationship between sputum percentages will cause these parameters to be associated within the same component, and may cause associations with other parameters, such as Fe_NO, to be overlooked. We used sputum total neutrophil cell count and sputum eosinophil percentage in the PCA to avoid this issue. Reassuringly, our PCA findings concerning Fe_NO and eosinophils were confirmed by multivariate modelling showing a significant association between these parameters.

The positive relationship between Fe_NO and sputum eosinophils has been observed in a smaller COPD group 27 but not by Siva et al 28 in 83 COPD patients. Again, measurement methodology may give an explanation for the lack of positive findings, as Siva et al used a flow rate of 250 ml/s, which is well known to give very low Fe_NO readings, particularly in COPD patients, and so may not be able to discriminate between patients.

Component 1 suggests that sputum neutrophils and the neutrophil chemoattractant IL-8 describe a distinct component of disease that is associated with systemic inflammation, measured by plasma TNFα levels. It is perhaps surprising that the other systemic inflammation biomarker that we measured, CRP, was not associated with airway neutrophils. CRP is a known marker of cardiovascular disease risk 29 and levels in COPD are associated with mortality 11, leading to the hypothesis that CRP levels are indicative of cardiovascular disease risk in COPD patients. A possible explanation for our findings is that neutrophilic airway inflammation is associated with some systemic inflammation pathways, such as TNFα which is known to be involved in muscle inflammation 20, but not CRP which is an indicator of cardiovascular disease.

Inhaled corticosteroid use was associated with CRP levels; this may be due to more severe patients with higher CRP levels being prescribed inhaled corticosteroids. Inhaled corticosteroid use was not associated with pulmonary function (component 3); this may be viewed as a surprising finding as corticosteroids are used for patients with lower FEV₁ values who have exacerbations. The reason for the lack of an association in the current study was probably that the range of FEV₁ values was relatively narrow, as most patients had moderate COPD (GOLD stage 2), and that the inclusion of greater numbers of severe/very severe patients would be needed to assess this relationship further.

It has been reported that sputum neutrophil counts 30 and IL-8 31 levels are related to severity of airflow obstruction and subsequent decline in FEV₁ in COPD. Airway tissue immunohistochemistry studies clearly show that mucosal inflammation is associated with lower FEV₁ 32. However, in agreement with the previous study by Laperre et al we found dissociation between airway inflammation and pulmonary function. This suggests that luminal inflammation, sampled by induced sputum, is not associated with FEV₁. A strength of the current study in coming to this conclusion is the sample size used, and 2 independent statistical analysis techniques. However, it is possible that other population groups including more patients with very severe COPD may generate different results.

Our study population was composed of mainly GOLD stage 2 "moderate" COPD patients, although 26 severe/very severe patients were also recruited. This mix of patients reflects our strategy of recruiting from primary care. It would be of interest to repeat the current study using more severe patients, perhaps recruited from hospital clinics, to observe whether the same or different results are obtained.

We found that females with COPD had lower levels of Fe_NO than males, which has been shown to be true in healthy controls 33. Similarly, the associations of Fe_NO with age and with smoking that we observed have also previously been reported 34.

In summary, this study provides insights into the dimensions of COPD that can be described using non-invasive biomarkers of airway inflammation and pulmonary function. Independent components representing different types of airway inflammation, lung function and systemic inflammation have been identified providing novel concepts with regards to COPD pathophysiology. We hope that our results will provide an impetus to further explore the usefulness of these components in guiding clinical practice by assessing their ability to predict important features of COPD such as prognosis, systemic manifestations and treatment responses.

The authors declare that they have no competing interests.

KR conceived, designed and coordinated the study as well as performing a lot of the measurements in exhaled NO, spirometry, plethysmography, sputum induction and processing, peripheral blood sampling and statistical analysis. DS and JV helped to conceive the study and participated in its design. JS participated in the statistical analysis. UK performed sputum induction, sputum supernatant cytokine analysis and assays for CRP and TNF-alpha. ZB performed measurements in spirometry, plethysmography and peripheral blood sampling.

All authors read and approved the final manuscript.