Blood neutrophil counts are associated with exacerbation frequency and mortality in COPD

Seven thousand two hundred twenty people with a COPD diagnosis were identified within the TARDIS (Tayside Allergy and Respiratory Disease Information System) database. Details of TARDIS have been previously described and are detailed online [10]. In brief, since 2008 patients age > 40 years, with a post-bronchodilator FEV₁/FVC ratio < 0.7 (70%) and a clinical diagnosis of COPD were invited to participate in a structured primary care cohort. Patients were reviewed annually for spirometry, recording of symptoms and exacerbation history. Patient’s TARDIS data was linked to regional prescribing databases, and databases that record hospitalizations and deaths nationally. The study was approved by the local research ethics committee approval number 13/ES/0030.

The index date for each participant was set as the date of the first stable BNC recorded after they had been in the TARDIS study for 12 months. A stable BNC was one considered representative of times when that person’s COPD was stable and unaffected by exacerbations or short-tern medical treatment. BNCs made between 1 week before and 4 weeks after the start of a COPD exacerbation, defined as either receipt of a short course of prednisolone (defined as an acute prescription for at least 30 mg/day for at least 5 days) or a respiratory hospitalization, were excluded as potentially being elevated by the immediate effects of the exacerbation or associated corticosteroid treatment. The online supplementary material gives details of the derivation of each variable used.

Extremely low and extremely high BNCs can result from various diseases and have the potential to dominate analyses that dichotomize or treat effects as linear. Therefore, the BNCs were split into four categories: low (<2000cells/μL); normal (2000–6000cells/μL); elevated (6000–15,000cells/μL); and extreme (> 15,000cells/μL). The upper bound of the normal range for BNC has often been taken to be 8000cells/μL in clinical practice so the models were refitted using that cut-off point in a sensitivity analysis. Models were also fitted, including all individuals, with BNC as a linear term as a sensitivity analysis. The primary interest was how people with elevated BNC differed from those in the normal range.

Two hundred seventy-three patients in the TARDIS cohort had been previously recruited into prospective longitudinal studies as previously described [10]. Participants were clinically assessed after giving written informed consent (Relevant medical history, Spirometry, St Georges Respiratory Questionnaire (SGRQ), GOLD 2017 scores, COPD Assessment Test (CAT), and Medical Research Council (MRC) Dyspnoea Score), and sputum and blood obtained when clinically stable at study enrolment. Full blood counts and C-reactive protein were measured at every visit. DNA was extracted from the induced sputum obtained at visits and 16S rRNA microbiome sequencing performed as detailed in the online supplement. All sequences generated are available in the NCBI Sequence Read Archive under the Bioproject accession numbers PRJNA316126 and PRJNA539959.

BNCs were measured as a biomarker in the ECLIPSE cohort, a longitudinal COPD cohort study that has been previously described [11]. The original study had a duration of 3 years. Patients who provided consent for additional follow-up to 8 years were included and the relationship between baseline BNCs and survival was examined after adjustment for confounders using Cox proportional hazards regression.

Data were summarized by BNC group to give simple comparisons. Regression models were then used to look at how other potential explanatory variables might change the results. Cox proportional hazard models of all-cause mortality were fitted, along with survival models of mortality recorded as being due to ICD10 code J44 (COPD) and ICD10 codes J00–99 (all respiratory mortality). The models of disease specific mortality treated other causes of mortality as competing risks. Generalized linear models (GLMs), with log links and negative binomial errors, were used to look at the frequency of severe exacerbations (defined as hospitalization with the main condition recorded as ICD10 code J44), and all acute exacerbations (defined as respiratory hospitalization or receiving a course of prednisolone), in the 12 months from the index date.

To examine changes in stable FEV₁, mixed models were fitted. These contained an intercept and slope for each BNC group as fixed effects, along with an intercept and slope for each individual as random effects. Because the mixed model assumes linear changes in FEV₁, and there was high mortality in this cohort, a second set of mixed models was fitted using only measurements taken within 3 years of the index dates.

Unadjusted models were fitted, using the four BNC levels to predict the outcomes, comparing the other groups to the normal BNC group. The models were then adjusted for age, sex and smoking status and refitted. To examine the stability of the results, 35 further versions of each model were fitted, each adjusting for another plausible confounding variable as listed in e-Table 1.

The association between index blood eosinophil counts and mortality was weak and only evident in patients with blood eosinophil counts <100cells/μL (Fig. 2a). No significant difference in all-cause mortality, mortality due to COPD or change in FEV₁ was observed between the different blood eosinophil count groups. The total exacerbation rates were also indistinguishable, though the rate of severe exacerbations was lower among those with elevated index blood eosinophils than in all other groups (e-Fig. 1 and e-Table 2).

BNC were significantly higher during exacerbations (P < 0.001; Fig. 2b), though this difference was small compared to the within-individual variability (median coefficient of variation 0.36; IQR 0.26–0.48; e-Figure 2). Despite this variability, a pattern of mortality risk being higher for those with BNC outside the normal range, was evident. Unsurprisingly, the association of mortality with index BNC (Fig. 2c) was less strong than that for the final BNC (Fig. 2d). These patterns were similar for all-cause mortality and that recorded as due to COPD. The patterns in the raw data (Fig. 2c & d) remained as statistically significant differences (P < 0.001) in survival between the BNC groups even after adjustment for age, sex and smoking history, with 10 year survival for those with an elevated BNC being about 2/3 of that for those with normal BNC at index (Fig. 3a and b). Very similar results were obtained for unadjusted models, and those considering all respiratory (ICD code J) deaths (e-Figure 3). The adjusted model with BNC as a linear term estimated the hazard ratio for all-cause mortality at 1.067 (95%CI: 1.060–1.073 per 1000cells/μL), and that for COPD mortality at 1.091 (95%CI: 1.082–1.100). These estimates, while still significant, are lower than the between group differences (e-Table 3) because there is high mortality among the group with low index BNC. Additional models did not identify any important confounding effects (e-Figures 4 and 5).

The elevated BNC group had more exacerbations than the normal BNC group in both the year before index (mean 2.3 vs 1.3 exacerbations/year, P < 0.001) and the year following index (1.9 vs 1.5, P < 0.001). The incident rate ratios for all exacerbations or severe exacerbations alone in the 12 months following index date were determined, with the normal BNC group as reference; two models are shown (Fig. 3c), one unadjusted and the second adjusted for age, sex and smoking history. The adjusted models indicated that, for patients with an elevated BNC, there was a 40% increase in total exacerbations, and a 170% increase in severe exacerbations in the year from index relative to the rates for those with normal BNC (P < 0.001 in both cases). The adjusted model with BNC as a linear term, estimated that each increase of 1000cells/μL in BNC was associated with the total exacerbation rate increasing by a factor of 1.05 (95%CI: 1.04–1.06). For severe exacerbations, the figure was 1.17 (95%CI: 1.14–1.20). In the year from index date the much smaller group of patients with an extreme BNC had an increase in all and severe exacerbations of 96 and 450% (P < 0.001 in both cases; e-Figure 6).

Additional GLMs did not identify any important confounding effects (e-Figures 7 and 8). When looking at all exacerbations, the neutrophil effect remained significant in every case, with only the total number of exacerbations in the year up to index date attenuating the magnitude of the effect. (e-Figure 7). Restricting the analysis to the younger half of the cohort did not change the results (e-Figure 9). Changing the upper limit of the normal range from 6000 to 8000cells/μL halved the size of the elevated BNC group, and widened confidence intervals, without changing the overall results or conclusions (e-Figure 10).

The overall rate of change in FEV₁, across the whole cohort was − 0.038 (95%CI: − 0.034, − 0.041) L/yr or 38 mL/yr. The rate within the elevated BNC group was indistinguishable from the normal BNC group (Fig. 3d). While estimated rate of decline for the elevated BNC group was slightly faster (difference − 0.004; 95%CI: − 0.010, 0.003), the difference was not statistically significant, and was reversed when consideration was restricted to only the first 3 years after the index date (difference between groups 0.002; 95%CI: − 0.012, 0.015). However, these confidence intervals are too wide to entirely rule out the existence of a clinically meaningful difference.

Two hundred seventy-three patients provided at least one sputum sample, when clinically stable, of which 246 samples passed sequencing quality control to be included in the analysis. Utilizing the same cut offs as the main TARDIS cohort, 1 patient had a low BNC, 159 were normal, 79 elevated and 7 extreme. For the purpose of the microbiome analysis, low and normal patients were grouped as <6000cells/μL, elevated and extreme as >6000cells/μL. At the phylum level, the microbiome was dominated by Proteobacteria and Firmicutes, corresponding to dominant genera of Haemophilus (n = 59 patients) and Streptococcus (n = 78) at the genus level (e-Figure 11). Additional dominant genera observed included: Veillonella (n = 55), Neisseria (n = 12), Pseudomonas (n = 11), Moraxella (n = 8) and Prevotella (n = 7). The BNC count of the samples was significantly associated with increased % Proteobacteria (P = 0.027) and reduced Shannon Wiener Diversity Index (P = 0.0017) (Fig. 4 a and b). Beta diversity was assessed using the Bray Curtis Dissimilarity Index and Weighted UNIFRAC (Fig. 4c). Between patients, the microbiome composition was shown to be significantly different by PERMANOVA when comparing patients with BNCs <6000cells/μL to patients with BNCs >6000cells/μL (P = 0.014), which remained statistically significant (P = 0.021) when adjusted for age, sex and smoking status (Fig. 4d).