Prognostic role of neutrophil-to-lymphocyte ratio in breast cancer: a systematic review and meta-analysis

This analysis was reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [5]. The search strategy developed by Templeton et al. [4] was used with the addition of “breast neoplasms” and synonymous breast cancer-specific terms. An electronic search of the following databases was performed: Medline (host: OVID), Medline in Process, Medline Epub Ahead of Print (host: OVID), EMBASE (host: OVID), and Cochrane Database of Systematic Reviews. All databases were searched from January 2013 to April 2016, supplementing the initial systematic review that searched databases until different time points in 2013. Citation lists of retrieved articles were screened manually to ensure sensitivity of the search strategy. The full search strategy is described in Table 3 in Appendix 1.

In order to reduce clinical heterogeneity, the following eligibility criteria were utilized: studies of adult women with breast cancer reporting on the prognostic impact of the peripheral blood NLR, where NLR was treated as a categorical variable; NLR collected prior to all treatment (surgery and/or systemic therapy); reporting of a multivariable HR for OS, and/or DFS or progression-free survival (PFS), and corresponding 95% CI and/or P value; available as a full-text publication; clinical trials, cohort studies, or case–control studies; and English-language publication. Case reports, conference proceedings, and letters to editors were excluded. Corresponding authors were contacted to clarify missing or ambiguous data. When multiple publications or data analyses were available from the same dataset and if clarification on potentially duplicate data could not be obtained, the study reporting the larger number of patients was retained and other studies were excluded. Studies only presenting data in graphic form without reporting a numerical value for HR were excluded. All titles identified by the search were evaluated, and all potentially relevant publications were retrieved in full. Two reviewers (JE and DD) independently reviewed full articles for eligibility based on inclusion criteria and data extraction, and disagreements were resolved by consensus. Three relevant articles identified in the previous systematic review were also included [4].

The following details were extracted from included studies using predesigned data abstraction forms: name of first author, year of publication, journal, number of patients included in analysis, median age, disease stage (nonmetastatic, metastatic, mixed (nonmetastatic and metastatic)), collection of data (prospective, retrospective), cutoff value used to define high NLR, number of patients with each breast cancer subtype, number of premenopausal and postmenopausal patients, and HRs and associated 95% CIs for OS, PFS, or DFS. Where more than one multivariable model was reported, HRs were extracted from models including the most participants.

Validity of included studies was assessed by two independent reviewers (J-LE and DD) using the Quality in Prognostic Studies (QUIPS) tool as described previously [6]. The QUIPS tool comprises 30 questions categorized into six domains (study participation, study attrition, prognostic factor measurement, outcome measurement, study confounding, and statistical analysis and reporting). Studies were rated according to each domain as being at low, moderate, or high risk of bias, based on the likelihood that they might alter the relationship between the prognostic factor and outcome.

Extracted data were pooled using RevMan 5.3 analysis software (Cochrane Collaboration, Copenhagen, Denmark). A meta-analysis was conducted for all included studies for each of the endpoints of interest if appropriate when clinical heterogeneity was minimal. The primary outcome of interest was OS, and intermediate endpoints such as PFS and/or DFS were secondary outcomes. Estimates for HRs were pooled and weighted by generic inverse variance, and were computed by fixed-effects or random-effects modeling. Heterogeneity was assessed using Cochran Q and I ² statistics. If significant heterogeneity was present (I ² > 50% or Cochran Q < 0.1), a random-effects model was used. Predefined subgroup analyses were conducted for disease stage (early, metastatic, mixed) using methods described by Deeks et al. [7] Meta-regression was performed to evaluate the effects of NLR cutoff, proportion of ER-positive patients, proportion of HER2-positive patients, proportion of triple-negative patients, median age, proportion of premenopausal patients, and proportion of patients with metastatic disease on the HR for OS and DFS. Meta-regression comprised a univariable linear regression weighted by individual study inverse variance and was performed using SPSS version 24 (IBM Corp, Armonk, NY, USA). A post-hoc meta-regression analysis testing the association between median duration of follow-up and the prognostic value of NLR was also performed. Multivariable meta-regression was not performed due to the small number of eligible studies leading to an undesirable risk of over-fitting. Publication bias was assessed by inspecting funnel plots visually. All statistical tests were two-sided, and statistical significance was defined as P < 0.05.

Thirteen studies comprising a total of 8015 patients reported adjusted HRs for OS. The median cutoff value for high NLR was 3.0 (range 2.0–5.0). Median follow-up was reported in 11 studies, and ranged from 1.8 to 7.2 years (mean 4.69 years) (Table 4 in Appendix 2). Overall, a NLR greater than the cutoff value was associated with worse OS (HR 2.56, 95% CI = 1.96–3.35; P < 0.001; see Fig. 2). There was statistically significant heterogeneity (Cochran Q = 0.009, I ² = 55%). This seems to be largely influenced by one study which showed a large effect size [8]. However, the association between NLR and OS was maintained in a sensitivity analysis omitting this study (HR 2.42, 95% CI = 1.89–3.09; P < 0.001; Cochran Q = 0.03, I ² = 48%), although statistically significant heterogeneity remained.

Exploratory analysis identified breast cancer stage as an important source of heterogeneity. Subgroup analysis showed that the association between NLR and OS was maintained in studies including only early-stage disease, as well as those comprised of patients with both early and metastatic disease (HR 2.98 vs 2.30 respectively; P for subgroup differences = 0.36). There was no statistical heterogeneity when the study driving heterogeneity in the main analysis [8] was omitted from the early stage subgroup (Cochran Q = 0.28, I ² = 20%). Additionally, the effect of NLR on OS was retained (HR 2.56, 95% CI = 1.82–3.60; P < 0.001). Statistical heterogeneity remained among studies with mixed early and metastatic disease (Cochran Q = 0.01, I ² = 69%).

Adjustment for age differences between arms was examined in individual studies. In one study, patients were significantly older in the arm with low NLR, and it was unclear whether the multivariable model was adjusted for age [9]. In two other studies, the median age in each arm was not reported, and age did not seem to be included in the multivariable model [10, 11]. In a sensitivity analysis excluding these three studies, high NLR remained a significant predictor for shorter OS (HR 2.55, 95% CI = 2.59–8.26; P < 0.001). Table 2 presents the results of the meta-regression analysis. We did not identify any classical clinicopathologic factors that were effect modifiers for influence of NLR on OS. Additionally, the median duration of follow-up did not affect the association between high NLR and OS.

There was evidence of publication bias, with fewer smaller studies reporting lower magnitude associations between NLR and OS (Fig. 3).

Nine studies comprising 4864 patients reported HRs for DFS. All studies included only patients with nonmetastatic disease. The median cutoff value for high NLR was 2.5 (range 1.9–4.0). Median length of follow-up was reported in eight studies, ranging from 1.8 to 7.2 years (mean 4.5 years) (Table 4 in Appendix 2). Overall, a NLR greater than the cutoff value was associated with worse DFS (HR 1.74, 95% CI = 1.47–2.07; P < 0.001; see Fig. 2). There was no evidence of statistically significant heterogeneity (Cochran Q = 0.14, I ² = 35%).

Adjustment for age differences between arms was examined in individual studies. Two studies had significant age differences between arms and no clear model adjustment for age, including one study where patients were significantly older in the arm with low NLR [9] and one study where the same group was significantly younger [12]. Another study did not report the median age in each arm and did not adjust for age in the multivariable model [10]. In a sensitivity analysis excluding these three studies, high NLR remained a significant predictor for shorter DFS (HR 1.69, 95% CI = 1.40–2.03; P < 0.001).

All studies reported the number of patients with HER2-positive disease, while seven of nine studies included data on ER status (Table 4 in Appendix 2). Meta-regression analysis is presented in Table 2. Results showed that ER and HER2 positivity were negative effect modifiers of the association between NLR and DFS, indicating that the NLR has a greater prognostic value in breast cancers that are ER-negative and/or HER2-negative. The proportion of patients with triple-negative or metastatic disease, median age, disease stage, histologic tumor grade, presence of nodal involvement, premenopausal status, median duration of follow-up, and NLR cutoff value did not affect the association between high NLR and DFS. There was evidence of publication bias, with fewer smaller studies reporting lower magnitude associations between NLR and DFS (Fig. 3).

The risk of bias in individual studies is summarized in Figure 4 in Appendix 3. Overall, risk of bias was low, particularly in the domains of study attrition, prognostic factor measurement, outcome measurement, and statistical analysis and reporting. There was a low–moderate risk of bias for the study participation domain due to lack of completeness in description of the baseline study sample in three studies [8, 13, 14]. Risk of bias was moderate with regards to study confounding, because four studies failed to adequately detail covariates included in adjusted models [8, 10, 12, 15].