Optimal cutoff of pretreatment neutrophil-to-lymphocyte ratio in head and neck cancer patients: a meta-analysis and validation study

This meta-analysis was performed according to the recommendations of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2009 guidelines [24]. We systematically searched the MEDLINE, EMBASE and Cochrane library databases focusing on clinical studies published prior to April 30th, 2017. Potentially relevant studies were identified using the following key words: [neutrophil lymphocyte ratio], in combination with [head and neck], [cancer], and [prognosis] (Additional file 1: Table S1). In addition, reference lists of retrieved articles were screened manually to identify additional eligible studies. No language restriction was imposed. Eligibility of these studies was decided through comprehensive reviews and discussions with multiple researchers.

Inclusion criteria for these studies were as follows: (i) enrolled patients had a histological diagnosis of HNSCC, (ii) pre-treatment NLR was calculated from peripheral blood samples before any treatments, and (iii) information about clinical outcomes (disease recurrence, metastatic tumor progression and death from any cause) was available (disease-free survival DFS, progression-free survival PFS, overall survival OS). Studies were excluded if there was insufficient information to calculate the hazard ratio (HR) and 95% confidence interval (95%CI) of outcomes.

Two authors (JKC, HSJ) independently identified eligible articles and collected the following data: (i) publication information; the first author’s name, year of publication, country of study conducted, (ii) pre-treatment NLR data, (iii) clinical features: total patient numbers, TNM stages at diagnosis, disease outcomes and follow-up duration. Any disagreement was resolved by discussion. The Newcastle-Ottawa quality assessment scale for non-randomized studies was implemented to evaluate the quality of included studies (Additional file 1: Table S2) [25].

NLR status, a binary variable (high NLR and low NLR), was defined according to the study-specific NLR cutoff value. To determine the effect of NLR status on clinical outcomes, HRs were pooled using the random effect model [26, 27]. HR of more than 1 indicated worse outcome for the group having NLR above cutoffs compared to the group having NLR below cutoffs. The precision of estimates was quantified by 95%CI.

Heterogeneity was measured by Higgins and Green I² test [27]. Values of I² ranged between 0% (no heterogeneity) and 100% (maximal heterogeneity). Heterogeneity of the study was considered to be substantial at P < 0.1 and I² > 50%. We also evaluated potential publication bias with Egger’s regression test and funnel plot [26]. A sensitivity analysis with a trim-and-fill method was conducted. All above analyses were executed using R 3.3.2 (Vienna, Austria; http://​www.​r-project.​org/​) with a package of metafor. A two-sided P value of less than 0.05 was considered statistically significant.

Study-specific NLR cutoff values ranged from 1.9 to 6.0 (median = 3). Meta-regression analyses were then performed to determine whether the impact of NLR status on clinical outcomes was different according to each study-specific factor: the cutoff value of NLR, age, gender, tumor stage (I/II versus III/IV), tumor subsite (oral cavity, pharynx, larynx, others) and number of index tumors (single versus multiple). Results are presented as change of HR of OS and DFS.

To confirm the prognostic impact of NLR cutoff values estimated from previous analyses, we conducted a validation study using an independent cohort in our institution. All HNSCC patients were enrolled prospectively into our head and neck cancer registry and they provided written informed consents for use of their clinical and biological data under an Institutional Review Board approved protocol. From registered HNSCC patients, we included patients who had been treated for their HNSCC (newly diagnosed) between 2010 and 2014 (n = 540).(Additional file 2: Raw data for a validation study) The clinical characteristics of our cohort were comparable to the study patients enrolled in the meta-analyses (Additional file 1: Table S3). All patients followed the current standard treatment protocols (The National Comprehensive Cancer Network guideline, http://​www.​nccn.​org). Patients with secondary cancers, other pathologies, or palliative treatments were excluded. HR on outcomes (OS and DFS) of NLR status according to possible NLR cutoffs in the range of 2 to 6 by increment of 0.1 were calculated in multivariate Cox proportional hazard regression analyses. The present study focused on pretreatment NLR values and clinical outcomes was approved by our Institutional Review Board again before data collection.

We identified 38 potentially relevant articles through multiple database searches. Among them, 14 were further excluded, mainly due to inability to estimate HR and 95%CI of outcomes. The article by Charles et al. [14] presented results with oropharyngeal cancer and non-oropharyngeal cancer separately. We regarded their study results as two independent studies. Thus, 25 observational studies in 24 articles were included in our meta-analysis (Fig. 1). Characteristics of included studies are listed in Table 1. Among them, 24 reported HRs for OS outcomes, including 12 and 5 studies for HRs of DFS and PFS, respectively.

Twenty four eligible studies were analyzed in OS meta-analysis. The total HR of the random effect model was 1.96 [95%CI: 1.66–2.31] (P < 0.001) (Fig. 2). However, substantial heterogeneities across these studies were noted (I² = 48.29%, P = 0.0053). Regarding DFS meta-analysis, 12 studies were enrolled. An overall HR was 1.90 [95%CI: 1.41–2.54] (P < 0.0001), indicating that high NLR value above cutoff was a significant predictor for DFS in HNSCC patients. The heterogeneity of these studies was also significant (I² = 70.4%, P = 0.0002).

In terms of outcomes for PFS, there were only five studies describing PFS in their reports. An overall HR for PFS was 1.82 [95%CI: 1.43–2.33] (P < 0.0001), with low heterogeneity (I² = 0.0%, P = 0.92) (Additional file 3: Figure S1). Thus, high NLR value (above cutoffs) could predict disease progression in metastatic settings of HNSCC, as well as OS and DFS. Because the number of studies was small (n = 5) with an outcome of PFS, we mainly focused on OS and DFS in subsequent analyses.

Next, we evaluated publication bias by Egger’s regression test for funnel plot asymmetry. In OS meta-analysis for enrolled studies, the funnel plot showed a relatively asymmetric distribution with Egger’s P value of 0.0007. Similarly, there was a significant publication bias in DFS meta-analysis (P = 0.0017) (Fig. 2). However, a trim-and-fill method, in which we calculated the overall HR by making additional data set symmetrical to y-axis according to the midpoint of funnel plot, did not reverse results of the random effect model, confirming that there was no significant difference in the outcome (Table 2, Additional file 4: Figure S2).

Through the meta-analyses described above, we confirmed the prognostic significance of pretreatment NLR status for OS and DFS in HNSCC. However, cutoff values of NLR for dividing HNSCC patients into high and low NLR groups, ranged from 1.9 to 6.0 in enrolled studies. Thus, we next determined whether the effect of NLR status on clinical outcomes was different according to the cutoff value of NLR. To address this, we applied meta-regression for the association between the HRs of OS (or DFS) and study-specific NLR cutoff values (Fig. 3).

NLR cutoff values of 1.9 to 6.0 did not affect the extracted HRs of OS or DFS (P = 0.16 for OS; P = 0.18 for DFS). In short, absolute NLR cutoff values between 1.9 and 6.0 did not seem to matter, although groups below and above NLR cutoffs did show significant survival differences. Among other potential factors, only age variable had influenced the HRs of NLR status in OS and DFS (Additional file 1: Table S4). A prognostic significance of NLR status did not differ according to tumor subsites in the head and neck, in addition to gender, tumor stage, and multiplicity.

Next, we conducted an external validation study. For this, we used 540 registered HNSCC patients in our institute. In the independent cohort, we evaluated the prognostic impact of NLR status according to NLR cutoffs in the range between 2.0 and 6.0 by increment of 0.1. The HRs of OS were significant in any NLR cutoff values between 2.0 and 6.0 (to 7.8) (P < 0.001) after adjusting for age, gender, TNM stage and tumor site (a multivariate Cox hazard regression model), in line with results of the previous meta-regression analyses (Fig. 3). However, the HRs of DFS were not significant in NLR cutoffs from 1.5 to 8.5 in our cohort (P = 0.089).

In other words, any NLR cutoff between 2.0 and 6.0 produced significant discrimination for better and worse prognosis group in terms of OS. For a more practical application of NLR status on prognosis estimation in HNSCC patients, our data suggested a three-tier classification of NLR status (NLR: < 2.0, 2.0–6.0, ≥ 6.0), instead of binary NLR grouping based on a single NLR cutoff. Our results also confirmed a significant discrimination of OS among groups of below NLR 2.0 (reference), NLR 2.0 to 6.0 (HR = 1.80, [95% CI: 1.13–2.86]) and above NLR ≥ 6.0 (HR = 3.92, [95% CI: 1.89–8.12]) in our cohort by a multivariate Cox proportional hazard model (adjusting for age, gender, TNM stage and subsite) (Fig. 4).