Inflammation and nutrition-based biomarkers in the prognosis of oesophageal cancer: a systematic review and meta-analysis

Abstract

Background

Accumulating literature has shown the predictive values of inflammation and nutrition-based biomarkers in the prognosis of oesophageal cancer but with inconsistent findings.

Method

We performed a meta-analysis to systematically evaluate the predictive value of the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), C reactive protein-to-albumin ratio (CAR), systemic inflammation index (SII), prognostic nutritional index (PNI), Glasgow Prognostic Score (GPS) and modified Glasgow Prognostic Score (mGPS) in oesophageal cancer. The outcome indicators include the overall survival (OS), disease-free survival (DFS) and cancer-specific survival (CSS). We applied pooled HR, sensitivity, specificity, positive likelihood ratio, negative likelihood ratio, diagnostic odds ratio and area under the curve together with 95% CI to estimate the predictive accuracy.

Results

A total of 72 studies, including 22 260 patients, were included in the meta-analysis. Elevated NLR, PLR CAR, SII, GPS, mGPS and decreased LMR and PNI were associated with poor OS of oesophageal cancer. A high level of NLR, PLR and GPS was related to poor DFS. A high level of NLR and GPS was related to poor CSS. The summarised AUC of CAR (0.72, 95% CI: 0.68 to 0.75) and mGPS (0.75, 95% CI: 0.71 to 0.78) surpassed any other indicators.

Conclusions

Clinical indicators such as NLR, PLR, LMR, PNI, SII, CAR, GPS and mGPS have the moderate predictive ability in OS, DFS and CSS of oesophageal cancer. The pretreatment level of CAR and mGPS showed an outstanding prediction value in 5-year OS for oesophageal cancer.

Strengths and limitations of this study.

• We used internationally recognised critical assessment tools to assess the quality of individual studies.

• The pooled results were stable due to the large sample size.

• The prognostic performance of biomarkers for oesophageal cancer was systematically compared for the first time.

• Different cut-off values may result in heterogeneity and bias.

• Heterogeneity was not fully explained.

Backgrounds

Globally, oesophageal cancer is the seventh most common cancer and the sixth leading cause of cancer death.¹ In 2020, there were 570 000 new cases of oesophageal cancer and about 500 000 deaths worldwide.² Pathologically, squamous cell carcinoma (SCC) and adenocarcinoma are the major histological types. Oesophageal adenocarcinoma is mainly observed in industrialised countries, and nearly half of the cases occur in Northwest Europe and North America, while oesophageal squamous cell carcinoma (ESCC) is more common in China, Central Asia or South Africa. Despite substantial efforts in diagnosis, accurate staging and advanced treatments, the 5-year survival rate remains unfavourable with frequent metastasis and recurrence.³ The pathological tumor-node-metastasis (TNM) stage is the gold standard for predicting oncological outcomes after surgery.⁴ However, with the diversification of treatment methods and the complexity of prognostic factors, prognosis prediction tends to be unsatisfactory. Therefore, it is urgent to find better prognostic biomarkers to guide clinical treatment and appropriate follow-up.

Increasing evidence indicates that systemic inflammatory response and nutritional status are involved in tumour development and influence the clinical prognosis. Principal inflammation-based prognostic scores^5–7 include a neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), C reactive protein-to-albumin ratio (CAR), systemic inflammation index (SII), pretreatment albumin levels and lymphocyte to monocyte ratio. Typical nutrition-based prognostic scores^{8 9} are prognostic nutritional index (PNI) based on serum albumin and total lymphocyte count, Glasgow Prognostic Score (GPS) based on elevated C reactive protein (CRP) concentration and low levels of albumin and modified Glasgow Prognostic Score (mGPS), a modified version of GPS. Recently, accumulating literature has shown the prognostic values of these inflammation and nutrition-based prognostic markers in oesophageal cancer, but with inconsistent findings. Hence, it is meaningful to distinguish an accurate prognosis index for patients with oesophageal cancer to guide individualised therapy and precision service.

In the current study, we performed a systematic review of relevant literature. We applied the meta-analysis to explore the accuracy of inflammation and nutrition-based prognostic scores for patients with oesophageal cancer.

Materials and methods

Literature search

We followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement. Two students (YJ and DX) independently searched PubMed, Web of Science and Cochrane Library Databases for eligible articles from the inception of databases to March 2020. Additionally, references in the eligible publications were also reviewed for potential studies. The language was limited to English. The search terms are listed in online additional file 1. The detailed search procedure is illustrated in figure 1.

Figure 1.

Flow diagram of the search process. AUC, area under the curve

The definition of outcome

Overall survival (OS) was defined as the time from the beginning of treatment to the death due to all causes or last follow-up. Cancer-specific survival (CSS) was defined as the time from the beginning of therapy to the time of cancer-related death. Disease-free survival (DFS) was defined as the time from the start of treatment to the cancer recurrence or the occurrence of the second primary malignancy.¹⁰

Selection criteria

Articles were included if they met the following criteria: (1) patients were histopathologically confirmed to be primary oesophageal cancer; (2) prognostic indicators were measured before esophagectomy, chemotherapy or radiotherapy; (3) hazard ratios (HRs) with 95% CIs were reported in multivariate analysis. Studies were excluded if they were: (1) reviews, case reports, letters or conference abstracts; (2) studies with insufficient data; or (3) duplicate publications.

Data extraction

For each study, the following information was extracted by two students (YJ and DX): the name of the first author, year of publication, country, study design, pathological type, number of patients, age, sex, end-point, follow-up time, cut-off selection, therapy, tumour stage, cut-off values and HRs (95% CIs). We further collected the data of true-positive (TP), false-positive (FP), true-negative (TN) and false-negative (FN) for 5-year OS directly provided in the paper or calculated by comparable data (the number of people in the high-risk and low-risk groups according to the cut-off values and the corresponding number of deaths and survivors). If only the area under the curve (AUC) was reported, we contacted the corresponding author for original data. If we could not get a response, we only included this study in the first part of the analysis.

Quality assessment

Two reviewers (HS and BQ) independently assessed the methodological quality of the studies using the Quality Assessment of Diagnostic Accuracy Studies 2 tool.¹¹ Each item was judged as ‘yes’, ‘no’ or ‘unclear’. Any signalling question answered ‘yes’ indicated a low risk of bias, while ‘no’ showed a high risk of bias. If the answer was uncertain, the domain was judged as having an uncertain risk of bias.

Statistical analysis

The risk of bias was analysed and plotted using Review Manager V.5.3 (London, UK). The meta-analysis was performed using STATA V.15.0 (Texas, USA). The strength of NLR, PLR, LMR, PNI, SII, CAR, GPS, mGPS in association with OS, CSS and DFS was measured by the combined HRs and their 95% CIs. Cochran’s Q test and Higgins I² statistics were undertaken to assess the heterogeneity of studies. If p≥0.10 in the Q test or I²<50%, we used the fixed-effect model; otherwise, we used the random-effect model. Publication bias was assessed by Begg and Egger test. The sensitivity analysis was utilised by omitting individual study one-by-one to evaluate the robustness of the results. All p values were two-tailed, and a p value <0.05 was considered statistically significant.

The pooled sensitivity, specificity, AUC and the corresponding 95% CI were calculated by TP, FP, FN and TN using a bivariate regression model. The threshold effects were calculated by testing the Spearman correlation using Meta-DiSc (Madrid, Spain).¹² If I²≥50% and p value ≤0.05, the heterogeneity was significant due to the non-threshold effect, and then we used the meta-regression analysis to find the source of heterogeneity. The pooled positive likelihood ratio (P-LR), negative likelihood ratio (N-LR) and diagnostic odds ratio (DOR) were also calculated to understand the performance of the prognostic index better. Deek’s funnel plot was used to detect publication bias. To evaluate the difference of AUC between biomarkers, we checked the overlap of 95% CIs. If not, we used the following z-test ((X₁−X₂)/(SE₁²+SE₂²)^1/2), where X₁ and X₂ represented the indicators, and SE₁ and SE₂ were the corresponding standard errors.¹³ It was considered significantly different if the p value obtained from the z-test was less than P′ (0.05/n, n was the number of comparisons). The comparison for sensitivity, specificity, P-LR, N-LR, or DOR was also performed.

Results

Literature selection and study characteristics

The initial search identified 662 potentially relevant records. After removing duplicates and papers that did not meet the inclusion criteria, 72 studies with 22 260 subjects remained for the systematic review (online additional file 2). A flowchart demonstrating the process of study selection is illustrated in figure 1. Most studies were carried out in Asia (42 in China; 23 in Japan). Before treatment, the blood cell counts used to calculate NLR, PLR, LMR and CAR were obtained. The baseline characteristics and treatment methods are presented in online additional file 2.

Risk-of-bias and quality assessments

Figure 2 illustrates the risk assessment of bias. A high risk of selection bias was observed in all studies. Nearly one-third of the studies had an unclear bias in study attrition. One study had an unclear bias for detection bias, and two studies had the risk of bias in measuring prognostic factors and outcomes, respectively. Six studies were judged as having an unclear performance bias.

Figure 2.

Risk of bias and applicability concerns.

Prognostic indicators in OS, DFS and CSS of esophageal cancer

As shown in figure 3 (A–H), factors significantly contributing to a short OS were a high level of NLR (HR: 1.43, 95% CI: 1.30 to 1.58, p<0.001; I²=61.7%, p_het<0.001), PLR (HR: 1.26, 95% CI: 1.18 to 1.35, p<0.001; I²=29.8%, p_het=0.108), CAR (HR: 1.84, 95% CI: 1.60 to 2.10, p<0.001; I²=41.8%, p_het=0.079), SII (HR: 1.46, 95% CI: 1.30 to 1.65, p<0.001; I²=41.0%, p_het=0.118), GPS (HR: 2.35, 95% CI: 1.99 to 2.76, p<0.001; I²=36.5%, p_het=0.078) or mGPS (HR: 1.69, 95% CI: 1.49 to 1.92, p<0.001; I²=48.4%, p_het=0.022), and low level of LMR (HR: 1.37, 95% CI: 1.14 to 1.65, p=0.001; I²=84.9%, p_het<0.001) and PNI (HR: 1.51, 95% CI: 1.36 to 1.68, p<0.001; I²=45.8%, p_het=0.048).

Figure 3.

Forest plot of HR for 5- year OS, DFS and CSS in patients with oesophageal cancer. (A) NLR-OS; (B) PLR-OS; (C) LMR-OS; (D) CAR-OS; (E) SII-OS; (F) PNI-OS; (G) GPS-OS; (H) mGPS-OS; (I) NLR-CSS; (J) GPS-CSS; (K) NLR-DFS; (L) PLR-DFS; (M) LMR-DFS; (N) GPS-DFS. CAR, C reactive protein-to-albumin ratio; CSS, cancer-specific survival; DFS, disease-free survival; GPS, Glasgow Prognostic Score; LMR, lymphocyte-to-monocyte ratio; mGPS, modified Glasgow Prognostic Score; NLR, neutrophil-to-lymphocyte ratio; OS, overall survival; PLR, platelet-to-lymphocyte ratio; PNI, prognostic nutritional index; SII, systemic inflammation index.

Patients with an elevated NLR (HR: 1.21, 95% CI: 1.04 to 1.41, p=0.011; I²=43.4%, p_het=0.089) and GPS (HR: 1.64, 95% CI: 1.33 to 1.94, p<0.001; I²=45.5%, p_het=0.119) had a worse CSS (figure 3I–J).

NLR (HR: 1.39, 95% CI: 1.10 to 1.75, p=0.005; I²=60.9%, p_het=0.018), PLR (HR: 1.30, 95% CI: 1.12 to 1.51, p<0.001; I²=33.0%, p_het=0.202) and GPS (HR: 2.44, 95% CI: 1.28 to 4.66, p<0.007; I²=57.5%, p_het=0.052) were negatively correlated with DFS. No significant association was found for LMR (HR: 1.08, 95% CI: 0.85 to 1.38, p=0.522; I²=79.8%, p_het<0.001) (figure 3K–N).

Subgroup analysis and meta-regression

Subgroup analysis and meta-regression were further conducted according to the cut-off value, sample size, follow-up time, sex, age, clinical stage and region (online additional file 3). The heterogeneity of OS studies was relatively low except LMR (I²=84.9%) and NLR (I²=61.7%). The pooled HR was significantly different between studies with more or less than 280 patients, indicating that the sample size may be the source of heterogeneity for LMR. Similarly, we found the source of heterogeneity for other indicators. The follow-up time may be the source of heterogeneity for PLR (p=0.004) and GPS (p=0.027). The sample size may be the source of heterogeneity for SII (p=0.047) and mGPS (p=0.014). The sex ratio may be the source of heterogeneity for CAR (p=0.045). In DFS analysis, we found that cut-off value and region may be the source of high heterogeneity of LMR (p=0.034) and NLR (p=0.018), respectively.

Publication bias

Begg and Egger’s tests were applied to estimate the publication bias. As shown in online additional file 3, no significant publication bias was observed.

Sensitivity analysis

We performed a sensitivity analysis by excluding one study each time. As shown in online additional 4, the results were not substantially changed, demonstrating the reliability and stability of the current meta-analysis.

Pooled sensitivity, specificity, DOR, and AUC of indicators

We further extracted TP, FP, FN and TN from each study (online additional file 2) to calculate the pooled accuracy of each indicator for a 5-year OS. There were 11 studies for NLR, 11 studies for PLR, 7 studies for LMR, 6 studies for CAR, 6 studies for SII, 7 studies for PNI, 6 studies for GPS and 5 studies for mGPS.

Threshold effect

The Spearman correlation coefficient (p value) for NLR, PLR, LMR, PNI, SII, CAR, GPS and mGPS was 0.56 (0.07), 0.59 (0.06), 0.57 (0.18), 0.75 (0.05), 0.77 (0.07), 0.20 (0.70), 0.77 (0.07) and −0.10 (0.87), respectively, indicating no significant threshold effect.

Forest plot and subgroup analysis

Forest plots of sensitivity and specificity are shown in figure 4. SII had the highest pooled sensitivity (0.61, 95% CI: 0.48 to 0.73), while CAR (0.84, 95% CI: 0.71 to 0.91) had the highest pooled specificity. The I² of the sensitivity and specificity of these prognostic indicators were relatively high (around 70%–90%) except GPS (sensitivity: 43.76%; specificity: 6.96%). We further conducted a subgroup analysis and meta-regression (online additional file 5). For SII, the sensitivity of studies with a cut-off level ≥410 ng/mL (0.47, 95% CI: 0.37 to 0.57) was significantly lower than the studies with a cut-off level <410 ng/mL (0.73, 95% CI: 0.66 to 0.81), while studies with a cut-off level ≥410 ng/mL (0.76, 95% CI: 0.72 to 0.81) had a significantly higher specificity than studies with a cut-off level <410 ng/mL (0.42, 95% CI: 0.38 to 0.47). Therefore the cut-off value may be the source of heterogeneity in both sensitivity and specificity of SII. Similarly, we found that sample size may be the source of sensitivity for mGPS (p<0.001), PLR (p=0.02), GPS (p=0.03), CAR (p=0.04) and LMR (p=0.04), and the source of heterogeneity in the specificity of NLR (p=0.03) and GPS (p<0.001). Additionally, the study area may be the source of heterogeneity in the specificity of mGPS (p=0.01). Also, age and clinical stage may be the source of heterogeneity in specificity for PLR (p<0.001) and PNI (p=0.01), respectively. However, we failed to find the source of heterogeneity for the sensitivity of NLR or PNI and the specificity of CAR or LMR.

Figure 4.

Forest plot of sensitivity and specificity. (A) NLR; (B) PLR; (C) LMR; (D) CAR; (E) SII; (F) PNI; (G) GPS; (H) mGPS. CAR, C reactive protein-to-albumin ratio; GPS, Glasgow Prognostic Score; LMR, lymphocyte-to-monocyte ratio; mGPS, modified Glasgow Prognostic Score; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; Q, Cochran Q statistic; PNI, prognostic nutritional index; SII, systemic inflammation index.

Comparison of AUC

Figure 5 shows the summarised receiver-operating characteristic curves of eight indicators. We found that the scope of pooled AUC of CAR (0.72, 95% CI: 0.68 to 0.75) and mGPS (0.75, 95% CI: 0.71 to 0.78) surpassed other indicators except GPS (0.67, 95% CI: 0.63 to 0.71). We further compared CAR, mGPS and GPS by z test. The pooled AUC of CAR or mGPS was larger than GPS (p=0.033; p=0.002), but there was no significant difference between CAR and mGPS (online additional file 6).

Figure 5.

SROC curves of 5-year overall survival. (A) NLR; (B) PLR; (C) LMR; (D) CAR; (E) SII; (F) PNI; (G) GPS; (H) mGPS; AUC, area under curve. CAR, C reactive protein-to-albumin ratio; GPS, Glasgow Prognostic Score; LMR, lymphocyte-to-monocyte ratio; mGPS, modified Glasgow Prognostic Score; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; PNI, prognostic nutritional index; SII, systemic inflammation index. SROC, summary receiver-operating characteristic.

Publication bias and sensitivity analysis

Only PNI (p=0.03) and mGPS (p=0.02) had a significant publication bias (online additional file 7). The sensitivity analysis of combined DOR showed a robust finding (online additional file 4 and figure 2).

Discussion

In this meta-analysis, we summarised data from 72 studies and estimated the predictive ability of inflammation and nutrition-based indicators in oesophageal cancer. In general, these indicators showed an excellent ability to predict the OS, DFS and CSS of patients with oesophageal cancer. The pretreatment level of CAR and mGPS showed an outstanding prediction value for 5-year OS than other indicators.

Inflammation plays an essential role in the development and progression of various malignant tumours.¹⁴ In addition, nutritional status is closely related to carcinogenesis, cancer growth, tumour progression and tumour prognosis.¹⁵ The peripheral blood cell analysis is a good choice for establishing a prognostic prediction model based on inflammatory and nutrition-related indicators due to its convenience, repeatability and low cost.¹⁶ Previous studies have systematically reviewed the role of some inflammation and nutrition-based indicators in the prognosis of oesophageal cancer, most of which focused on ESCC. Yang et al¹⁷ investigated the relationship between NLR and oesophageal cancer by summarising six studies involving 1633 patients. Sun and Zhang⁵ reviewed 26 studies to explore the NLR, PLR and LMR in the OS, CSS and DFS in ESCC. Li et al¹⁸ reviewed nine observational studies and showed that a low PNI score was significantly correlated with a poor OS of oesophageal cancer and recurrence-free survival of ESCC. Liu et al collected eight observational studies and showed that high CAR was related to a worse OS.¹⁹ Although previous meta-analyses have reported the prognostic value of these indicators, this is the first study to comprehensively estimate the popular inflammatory and nutrition-related markers in OS, DFS and CSS of oesophageal cancer. Moreover, this is the first systematic review to summarise the sensitivity and specificity and compare the AUC of these predictors in the 5-year OS of oesophageal cancer.

In this review, we observed that the AUC of CAR and mGPS was significantly higher than NLR, PLR, SII, PNI, LMR and GPS, indicating their outstanding predictive value in oesophageal cancer. CAR and mGPS are calculated based on the level of CRP and albumin. CRP is a kind of acute reactive protein synthesised by liver cells or cancer cells, which can produce an attractive environment for tumour growth, induce DNA damage, promote angiogenesis and favour neoplastic spread and metastasis, revealing levels of inflammation in the body.^{20 21} Albumin reflects the malnutrition status of the host, triggers malignant transformation and tumour progression or even causes cachexia.²² Combining the two indicators can reveal a patient’s inflammatory status and nutritional status, which can effectively predict prognosis. These may explain the prominent prognostic role of CAR and mGPS. Additionally, some prospective studies have revealed the better predictive power of CAR and mGPS in other types of cancer. For example, it was reported that the CAR had a better predictive performance for hepatocellular carcinoma and colorectal cancer than NLR, PLR or CRP alone.²³ Other studies demonstrated that mGPS was an independent marker of poor prognosis for patients with SCC and superior to NLR, PLR and PNI.⁹

Although the TNM staging system is well known as a predictive clinical parameter in terms of guiding treatment and clinical prognosis, the survival outcomes for patients with oesophageal cancer with the same TNM stage still vary widely.⁴ In addition, many patients with oesophageal cancer cannot undergo surgery to obtain pathological identification for various reasons. Thus their prognosis cannot be obtained. Moreover, multiple factors influence the prognosis of patients with oesophageal cancer, such as neoadjuvant therapy, psychological factors and behaviour and eating habits, which will change the postoperative pathological stage of the tumour and thus affect the evaluation of disease progression.^24–26 Therefore, pathological diagnosis is not sufficient to accurately predict the prognosis of patients with oesophageal cancer. More readily available objective indicators with high specificity and sensitivity are needed to predict the prognosis of patients with cancer. Our results of this meta-analysis will help clinicians and patients to select appropriate indicators for prognosis prediction. In this way, patients can be classified, and appropriate treatment strategies and postoperative management methods can be selected, providing policymakers with ideas. Taken mGPS as an example, patients with oesophageal cancer with a score of 2 may have a high risk of prognosis, which may provide an effective way for clinicians to select high-risk patients with worse prognosis or severe adverse events before treatment and further timely adjust individualised treatment regimens and enhance postoperative rehabilitation. In addition, policymakers should develop policies to strengthen community guidance and management of such high-risk postoperative patients.

Malnutrition is closely related to carcinogenesis, cancer growth, tumour progression and tumour prognosis.²⁷ The Global Leadership Initiative on Malnutrition (GLIM) standards integrate current best evidence and expert opinion on malnutrition to promote the prevention, identification and treatment of malnutrition in patients with cancer.²⁸ Inflammation is one of the aetiological criteria in GLIM classification, and studies have demonstrated that the changes of GPS score, CRP and albumin are highly consistent with the GLIM criteria in identifying malnutrition in patients.²⁹ Similar to this study, our findings confirm the value of mGPS and CAR in predicting the prognosis of oesophageal cancer. GLIM criteria are re-evaluated every 3–5 years based on new research. Our results may provide a basis for the optimisation of GLIM criteria. Additionally, previous studies have reported that the combination of PNI and GLIM criteria has significant advantages in predicting the incidence and survival rate of perioperative malnutrition.³⁰ Our results show that the prognostic indicators we studied have high specificity but unsatisfactory sensitivity. More well-designed studies are needed to develop joint indicators to improve the sensitivity and specificity of prediction.

Some limitations should be acknowledged. First, some heterogeneity was not fully explained. This may be due to the fact that some factors that may affect survival were not included, such as living behaviour and eating habits, comorbidities, neoadjuvant therapy and psychological factors.^{10 24} Second, the cut-off value of indicators varied between studies, affecting the pooled analysis results and induce unavoidable potential heterogeneity and bias. Therefore, a standard and uniform cut-off value need to be defined. Third, publication bias was detected in studies on PNI and mGPS. Papers that failed to get published due to negative or null results could not be identified in our literature search and thus were not included in the meta-analysis. This may overestimate the prognostic effect of PNI and mGPS. Therefore, more well-designed prospective studies with large samples are needed to verify our findings.

Conclusion

NLR, PLR, LMR, PNI, SII, CAR, GPS and mGPS are commonly used as clinical indicators to predict OS, DFS and CSS of oesophageal cancer, but with unsatisfactory sensitivity. Pretreatment CAR and mGPS showed outstanding prognostic values in 5-year OS for patients with oesophageal cancer. Future extensive prospective studies with rigorously designed methodologies are warranted to confirm our results.

Data availability statement

Data are available in a public, open access repository. All data generated or analyzed during this study are included in this published article.

Ethics statements

Patient consent for publication

Not required.

Ethics approval

All data were downloaded from the public database and followed the data access policies. This study was exempted from ethical review by the ethics committee of Nanjing Medical University. This study did not involve individual information, so there was no requirement for informed consent.