Prognostic impact of tumor length in esophageal Cancer: a systematic review and Meta-analysis

In clinical studies, it has been observed that esophageal cancer (EC) patient prognosis can be very different even for those patients with tumors of the same TNM stage. Tumor length has been analysed as a possible independent prognostic factor in many studies, but no unanimous conclusion has been reached. Therefore, this review used a meta-analysis to evaluate the association between tumor length and prognosis in EC patients. A systematic search for relevant articles was performed in PubMed, Web of Science, and Embase. Hazard ratios (HRs) and 95% confidence intervals (CIs) were used as effective measures to estimate the correlation between tumor length and prognosis, including overall survival, disease-free survival, progression-free survival, disease-specific survival, and cancer-specific survival. STATA 15.0 software was used to perform the meta-analysis and the data synthesis. Finally, 41 articles with 28,973 patients were included in our study. The comprehensive statistical results showed that long tumors are an independent prognostic parameter associated with poor overall survival (OS) (HR = 1.30; 95% CI: 1.21–1.40, p < .001) and disease-free survival (DFS) (HR = 1.38; 95% CI: 1.18–1.61, p < .001) in EC patients. Subgroup analyses also suggested a significant correlation between long tumors and poor OS. Sensitivity analysis and publication bias evaluation confirmed the reliability and stability of the results. Similar results were obtained in the analyses of progression-free survival (PFS), disease-specific survival (DSS), and cancer-specific survival (CSS). The results of this meta-analysis showed that long tumors were related to poor OS, DFS, PFS, DSS and CSS in EC patients. Tumor length might be an important predictor of prognosis in EC patients, and it can be used as an independent staging index. Further well-designed and large-scale prospective clinical studies are needed to confirm these findings.


Background
Esophageal cancer (EC) is one of the most common and fatal cancers of the digestive system in the world. In the United States, more than 18,000 people are diagnosed with EC every year, and more than 15,000 people die of EC. Most patients die within 1 year of diagnosis [1]. The main methods to improve the prognosis of EC are surgery, chemotherapy and radiotherapy [2,3]. Although there are established radiotherapy and chemotherapy regimens, surgical resection is still the only possible cure for this disease. Complete surgical resection and radical lymph node dissection can provide accurate pTNM staging, which has important guiding significance for predicting prognosis and determining treatment options after surgery. An effective and reasonable EC staging system is important for the selection of treatment options and the prediction of long-term survival [4].
The current classification criteria for EC patients include histological grading, primary tumor (T), lymph node metastasis (N), and metastasis (M). Tumor length was not included as a prognostic indicator [5]. In recent years, many studies have suggested that tumor length is an independent prognostic factor. Gaur et al. [6] and Feng et al. [7] analysed 296 EC patients who underwent surgical resection and found that tumor length was a risk factor for long-term survival in both SCC and AC patients. Shimada et al. [8] and Song et al. [9] retrospectively reviewed 575 patients with EC who underwent curative resection and found that tumor length had a similar negative effect on patients with different pTNM stages. Our early studies showed that tumor length was an important prognostic factor for OS and cancerspecific survival (CSS) [10,11]. However, the prognostic role of tumor length in EC is still controversial. The results of Khan et al. [12] indicated that tumor length was not correlated with prognosis in N0 EC patients. Therefore, the purpose of this study was to evaluate the prognostic value of tumor length for the survival of EC patients by performing a meta-analysis.

Data sources and search strategy
All data included in this study are publicly available in PubMed, Embase, and Web of Science. Our metaanalysis was based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2009 guidelines [13]. This systematic review and meta-analysis has been registered on PROSPERO with registration number: CRD CRD42018106851. The PubMed, Embase, and Web of Science databases were searched to identify relevant articles for analysis from 1994 until December 2019. All keywords were searched through MeSH. The full text of all included articles must be available, and the reference lists of each identified publication were reviewed for potential studies to avoid omission.

Inclusion and exclusion criteria
Articles were included if they met the following criteria: 1. Patients with EC were histopathologically confirmed, including SCC and AC; 2. Studies explored the relationship between the tumor length and prognosis of EC; 3. The hazard ratios (HRs) and 95% CIs between tumor length and survival outcomes were reported.
Articles were excluded if they met the following criteria 1. They were abstracts, case reports, reviews, letters, or editorials; 2. Studies were not published in English; 3. Studies had overlapping or repeating data; 4. Studies did not present the cut-off value for tumor length.

Data extraction and quality assessment
Two researchers (ZW and YJ) independently reviewed articles that met the inclusion criteria. Any disagreements were resolved by discussion or arbitrator (WX). The HRs and 95% CIs were extracted and pooled according to different prognoses (OS or DFS/PFS/DSS/ CSS) to analyse the relationship between tumor length and prognosis. All HRs were extracted from the multivariate analysis. The data were obtained directly from the included articles or from simple indirect data calculations [14]. In addition, other parameters were extracted in a uniform format, including author, year of publication, patient source, total number of patients, patient age, follow-up time, treatment strategy, LN metastasis, histology, tumor location, TNM stage, cut-off value of tumor length and survival data. The quality of the included studies was assessed through the Newcastle-Ottawa Quality Assessment Scale (NOS), which consists of 3 parts: selection (0-4 points), comparability (0-2 points), and outcome assessment (0-3 points). The maximum score was 9 points, and studies with no less than 6 points were defined as high-quality studies [15].

Statistical analysis
All statistical analyses were performed using STATA software version 15.1 (STATA, College Station, TX). The pooled HRs and 95% CIs were used to assess the prognostic role of tumor length in EC patients. A pooled HR > 1 indicated a worse prognosis in EC patients with long tumors. P < .05 for the Q test or I 2 > 50% for the I 2 test indicated significant heterogeneity in the literature, and the random-effects model (DerSimonian-Laird method) was adopted [16]; otherwise, the fixed-effects model (Mantel-Haenszel method) was used [17]. Subgroup analyses were performed based on patient source, histology, treatment, sample size, median age, cut-off value, percentage of patients with LN metastasis, percentage of patients with TNM stage III/IV classification and HR analysis method to explore the potential sources of heterogeneity and assess the stability of our aggregated results. Sensitivity analyses were performed by sequentially removing individual studies to test the robustness of the pooled results of these studies. Publication bias was assessed by Begg's funnel plot and Egger's linear regression test [18,19]. The trim-and-fill method was used to evaluate the impact of publication bias on the pooled results in the presence of significant publication bias. P < .05 was considered statistically significant.

The characteristics of the included studies
Based on the search strategy mentioned above, a total of 1110 articles were searched from PubMed, Embase and Web of Science. Finally, 41 studies published between 1994 and 2019 with a total of 28,973 patients were included in our meta-analysis according to the inclusion and exclusion criteria [6][7][8][9][10]. The specific literature selection process is shown in Fig. 1. Table 1 summarizes the characteristics of all the studies included in this meta-analysis. All of the included studies were retrospective. Among these studies, 23 studies were from China [7,9,10,[30][31][32][33][34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49], 7 studies were from the USA [6,[50][51][52][53][54][55], 7 studies were from Japan [8,[24][25][26][27][28][29], and 1 study each was from Italy [22], Australia [20], the Netherlands [21], and Turkey [23]; the participants in 39 studies were patients with SCC. Three of the 41 studies included patients who received chemoradiotherapy only [26,48,53], and the remaining 38 studies included patients who underwent surgical resection with or without chemoradiotherapy. Among the 41 studies, 35 studies reported the relationship between the tumor length of EC and OS, 10 studies reported disease-free survival (DFS), 2 reported progression-free survival (PFS) [48,53], 1 reported cancer-specific survival (CSS) [55] and 1 reported disease-specific survival (DSS) [20]. The cut-off value (cm) applied in all studies was between 1.5 and 6. Eighteen studies used a tumor length cut-off value of ≥5, and 23 studies used a tumor length cut-off value of < 5. All the tumor length values in the Surg and Surg 2 groups were measured from postoperative pathological specimens, and the data from the chemoradiotherapy (CRT) group and Surg 3 group were measured by endoscopy or computed tomography (CT) before treatment. The Surg 1 group did not clearly describe the timing of measurement according to the different treatments. Because of the irregular shapes of the tumors, the actual size of them is difficult to measure. Most studies only analyzed the relationship between tumor length and patient prognosis, but did not discuss the effect of width. So, we default that the length of the tumor is proportional to its size, and combined the results of included studies on this basis. Two studies reported only relative risk (RR) data [41,42], and we used RR to replace HR when pooling the data.

Tumor length and OS in EC
A total of 35 studies evaluated the relationship between tumor length and OS in EC, and because of the significant heterogeneity among the included studies (I 2 = 66.8%; ph < .001), the pooled HR and 95% CI were calculated by a random-effects model. The pooled HR of 1.301 (95% CI: 1.210-1.399) suggests that long tumors are associated with poor OS (Fig. 2).

Tumor length and DFS in EC
Ten studies provided the HR and 95% CI of tumor length in association with DFS. The pooled data showed that long tumors were associated with poor DFS (HR = 1.378; 95% CI: 1.179-1.609, p < .001). Due to obvious heterogeneity (I 2 = 76.7%, ph < .001), the random-effects model was used to calculate the pooled HR and its 95% CI (Fig. 3).

Tumor length and PFS in EC
No significant heterogeneity was found among the studies that evaluated PFS, so a fixed-effects model was used to calculate the pooled HR and 95% CI (HR = 1.161; 95% CI: 1.008-1.337, p < .005). The data suggest that long tumors were associated with shorter PFS.

Tumor length and CSS/DSS in EC
Two studies reported preoperative tumor length and CSS/DSS data for EC. There was no significant heterogeneity (I 2 = 0.0%, ph = .551). The pooled HR and 95% CI (HR = 1.856; 95% CI: 1.173-2.937, p < .001) indicated that long tumors were also related to poor CSS/DSS.

Subgroup analysis
To analyse the impact of heterogeneity on the results of this study, we performed subgroup analyses on the extracted data. Subgroup analyses by patient source (China and others), histology (SCC, AC and mixed), treatment (Surg 2 and others), median age (< 60 and ≥ 60), cut-off value (< 5 and ≥ 5), sample size (< 200 and ≥ 200), percentage of patients with LN metastasis (< 50% and ≥ 50%), percentage of patients with TNM stage III/ IV classification (< 50% and ≥ 50%) and HR analysis method (MV and UV) were performed to explore the potential sources of heterogeneity for the pooled OS results. Because the cut-off values of tumor length were different among the included studies, we performed subgroup analysis according to different cut-off values.
Because 40% of articles chose 5 cm as the cut-off value for tumor length, we used this cut-off value as the boundary. For cut-off value ≥5, the pooled HR was 1.259 (95% CI: 1.096-1.446, I 2 = 59.4%, ph = .002). For cut-off value < 5, the pooled HR was 1.322 (95% CI: 1.210-1.443, I 2 = 67.9%, ph < .001). This suggests that we may be able to define 5 cm as a standard cut-off value and recommend it to other researchers to reduce heterogeneity between different studies. The results of subgroup analyses showed that the different classification methods had no obvious influence on HR (Table 2). In almost all subgroups, long tumors were significantly related to

Sensitivity analyses
After omitting each individual study for the sensitivity analysis, the results showed that when excluding any study, the pooled HRs did not change substantially, indicating that our results were stable (Figs. 4, 5).

Publication bias
Publication bias was evaluated by Begg's funnel plot and Egger's linear regression test. Considering the limited number of studies on the relationship between tumor length and DFS, PFS, DSS or CSS, we analysed the publication bias of tumor length and OS (Pr > |z| =0.127 for Begg's test and p < .01 for Egger's test) (Fig. 6). Therefore, we conducted a further trim-and-fill analysis (Fig. 7). The adjusted result (HR = 1.195, 95% CI: 1.111-1.284) was similar to our previous pooled result. This

Discussion
EC is one of the most lethal malignant tumors in the world, with a 5-year survival rate of less than approximately 20% [1]. With the development of clinical treatment technology, many comprehensive treatment methods, including surgery, chemotherapy and radiotherapy, have been used in EC patients. Combined with the extensive development of minimally invasive McKeown esophagectomy (MIE-McKeown), we can obtain accurate pTNM staging information for each EC patient and use the system to guide the choice of further treatment and the prediction of long-term survival [4]. These assessments are based on the latest American Joint Committee on Cancer staging system for EC [5]. However, it is not difficult to find a significant difference in the prognosis of EC patients with similar TNM stages. Song et al. [9] found that even for EC patients with similar TNM stages, tumor length is still an independent factor for prognosis. In previous studies, many efforts have been made to find new clinicopathological factors to predict the prognosis of EC patients [56,57]. Therefore, we conducted this meta-analysis to assess the prognostic significance of tumor length in EC patients. Tumor size has been used in staging assessment systems for lung and thyroid cancers [58,59]. Many studies have been conducted on the relationship between tumor length and prognosis in other tumors [60,61]. Our team published the protocol of this metaanalysis in September 2018 [62]. Later, Yang et al. [63] performed a meta-analysis of the relationship between prognosis and tumor length in patients undergoing radical resection of EC. On the basis of expanding the scope of the literature included, we excluded some literature with repeated data and included literature on the prognosis and tumor length of EC patients with non-surgical treatment. At the same time, the sample size and pathological types were expanded, and the guiding value of tumor length for the choice of treatment in different stages was discussed. In fact, in the American Joint Committee on Cancer (AJCC) TNM staging system in 1983, tumor length was used as an indicator to evaluate T staging [64,65]. Although it was replaced in the 1987 version by depth of invasion of the esophageal wall, tumor length is still considered by many researchers to be a factor related to the extent of the peripheral invasion of the tumor. Our overall pooled results demonstrated that long tumors had a negative effect on OS, DFS, PFS, CSS and DSS in EC patients. Moreover, our subgroup analysis and sensitivity analysis results showed that our pooled HRs were stable and reliable. However, in the US category, the pooled HR was 1.457 (95% CI: 0.851-2.495, I 2 = 84.10%, ph < .001) [6,[50][51][52][53][54][55]. This result suggests that the tumor length of EC has different effects on OS in different races and regions. In the adenocarcinoma group and the mixed group, the pooled HRs were 1.616 (95% CI: 0.950-2.748, I 2 = 67.90%, ph < .05) [6,22,52] and 1.128 (95% CI: 0.895-1.421, I 2 = 78.50%, ph < .01) [21,44,50,54,55]. This finding suggests that the prognostic significance of tumor length may be limited for other non-squamous carcinomas of We found that two of the three articles in which patients were only treated with CRT suggested that there was no significant correlation between tumor length and prognosis [26,48]. In studies of CRT only, the included patients had a later TNM stage; for advanced patients, the systemic progression of the disease is rapid, and OS is   [8, 25-27, 33, 50]. Based on the above findings, we believe that tumor length may be of greater value in the prognosis of SCC and early EC. We may regard tumor length as an independent factor to evaluate prognosis and as a further subdivision standard only for early EC patients so that more individualized treatment measures that patients can benefit from can be selected. Yamamoto et al. [29] considered that whether the length of the tumor is less than 5 cm and the circumferential spread is less than two-thirds can be used as the criteria for endoscopic resection of mucosal carcinoma; moreover, whether neoadjuvant therapy can cause some patients to be reconsidered for included based on this standard also reflects the importance of tumor length as  There were several limitations in this meta-analysis that need to be addressed. First, our study only included research published in English, which may cause publication bias in this study. Second, the number of articles describing tumor length and DSS/PFS/DSS/CSS was insufficient, and the reference value was limited. Third, there were no corresponding data to analyse the relationship between tumor length and prognosis in different tumor locations. More importantly, there were many confounding factors, such as different resources of patients, pathological types, treatment strategies, durations of follow-up, sample sizes, proportions of lymph node metastasis, TNM stages and cut-off values of tumor length, which will increase the heterogeneity of this study. Although we conducted subgroup analyses and sensitivity analyses on the extracted data, these analyses could not fully explain the heterogeneity. Significant heterogeneity may affect the reference value of the metaanalysis results. We also assessed Begg's funnel plot and performed Egger's test to evaluate publication bias, and the results indicated that 12 articles on the relationship between tumor length and OS in EC have not been published. Though we adjusted the pooled result with the trim-and-fill method, the existence of significant publication bias suggests that we may have deficiencies in the selection and inclusion of literature. All of these limitations may hinder the application of the results in clinical work. More studies and larger sample size meta-analyses are needed to correct these limitations and improve the accuracy of the results.

Conclusion
The tumor length of EC is widely measured in clinical work, but it is not often used as an indicator to evaluate prognosis and select treatment. Our results show that long tumors are related to poor survival outcomes in EC and can be considered an effective prognostic factor. This index can be used for the risk stratification of patients and is a promising prognostic index for clinical decision-making regarding EC treatment, especially for early-stage patients. However, due to the limitations listed above and the lack of more detailed case-control studies, more well-designed and large-scale studies are needed to confirm this conclusion in the future.

Availability of data and materials
The datasets that support the conclusions of this study are openly available in PubMed, Embase, and Web of Science.