Meta-analysis of Circulating Tumor Cells as a Prognostic Marker in Lung Cancer

Lung cancer was the most common cancer as well as the leading cause of cancer death. Approximately 1.6 million new cases of lung cancer will be diagnosed and 1.4 million deaths will occur from lung cancer during 2008 (Jemal et al., 2011). The presence of circulating tumor cells (CTCs) in the blood was first reported by T. R. Ashworth more than a century ago (Ashworth, 1869). The level of detected CTCs was widely used in the diagnosis of breast (Cristofanilli, 2006), colorectal (Cohen et al., 2008), lung (Krebs et al., 2011) and prostate cancers (Helo et al., 2009). The detection of CTCs have been recently developed to reflect the progression and survival of the disease. Many studies reached in a positive conclusion towards the role of CTCs in prognostic prediction of lung cancer. However, some other study stood with the opposite attitude (Chen et al., 2007). Thus, it still remained a question whether CTCs can warn for disease progression and survival earlier and less invasively than conventional methods currently available. The aim of this study is to comprehensively and quantitatively summarize the evidence for the use of CTCs


Introduction
Lung cancer was the most common cancer as well as the leading cause of cancer death. Approximately 1.6 million new cases of lung cancer will be diagnosed and 1.4 million deaths will occur from lung cancer during 2008 (Jemal et al., 2011).
The presence of circulating tumor cells (CTCs) in the blood was first reported by T. R. Ashworth more than a century ago (Ashworth, 1869). The level of detected CTCs was widely used in the diagnosis of breast (Cristofanilli, 2006), colorectal (Cohen et al., 2008), lung  and prostate cancers (Helo et al., 2009). The detection of CTCs have been recently developed to reflect the progression and survival of the disease. Many studies reached in a positive conclusion towards the role of CTCs in prognostic prediction of lung cancer. However, some other study stood with the opposite attitude (Chen et al., 2007). Thus, it still remained a question whether CTCs can warn for disease progression and survival earlier and less invasively than conventional methods currently available.
The aim of this study is to comprehensively and quantitatively summarize the evidence for the use of CTCs Xue-Lei Ma 1& , Zhi-Lan Xiao 1& , Lei Liu 1& *, Xiao-Xiao Liu 1 , Wen Nie 1 , Ping Li 1 , Nian-Yong Chen 1 , Yu-Quan Wei 1 to predict the clinical results of lung cancer patients.

Search strategy
Medline and EMBASE were searched for the last time on Feb 26, 2012. The search strategy included the following keywords variably combined by ''CTCs'', ''circulating tumor cells'' and ''lung cancer''.

Study inclusion/exclusion criteria
Studies were considered eligible if they met all of the following inclusion criteria, (i) discussed patients with lung cancer, (ii) measured the appearance of CTCs in peripheral blood, and (iii) investigated the association between CTCs' appearance rate and survival outcome (overall survival, OS or progression free survival, PFS). Studies wereexcluded based on any of the following criteria, (i) were review articles or letters (ii) analyzed in varioustumors but with no specific results of lung cancer, (iii) lacked keyinformation for analysis with methods developed by Parmar et al. (1998), Williamson et al. (2002), and Tierney et al. (2007).

Data Extraction
Articles were reviewed independently by two investigators (Ma XL and Xiao ZL) for article inclusion and exclusion. Disagreements were resolved by consensus. Data were extracted from eligible studies by two investigators (Ma XL and Liu L) independently. The primary data were p-value, the Kaplan-Meier survival curves or HR and 95% confidence interval (CI) of survival outcomes. Additional data obtained from the studies included first author, publication year, study size, patients age and sexuality, TNM stage, histological classification, methods to detect CTCs, positive CTCs definition, the attitude conclusion and other clinical characteristics.

Statistical Methods
The logHR and SE (logHR) (SE) were used for aggregation of the survival results, but these statistical variables were not given explicitly in most studies. We calculated the necessary statistics on the basis of available numerical data with methods developed by Parmar, Williamson, and Tierney. We performed meta-analysis in OS and PFS, the subgroup research were given when the article number ≥ 2. Calculation was accomplished by the software designed by Matthew Sydes and Jayne Tierney with these methods (Medical Research Council Clinical Trials Unit, London, UK) (Tierney et al., 2007).
We also examine the correlation between CTCs appearance and the clinical variables including TNM stage, the depth of invasion, lymph node status, distant metastasis, sexuality and smoking status. According to clinical characteristics, Stage I and Stage II were combined and Stage III and Stage IV were combined; pT1 and pT2 were combined and pT3 and pT4 were combined. Odds ratio (OR) was used as the measure index to describe the correlation.
Forrest plots were used to estimate the effect of CTCs appearance on survival outcome and the correlation between CTCs appearance and the clinical variables. Heterogeneity was defined as p < 0.10 or I 2 > 50% (Higgins et al., 2003). When homogeneity was fine (p ≤ 0.10, I 2 ≤ 50%), a fixed effect model was used for secondary analysis. If not, a random effect model was used. An observed HR>1 indicated worse outcome for the positive group relative to the negative group and would be considered statistically significant if the 95% CI did not overlap 1.The Begg's rank correlation also was applied to assess the potential publication bias, p > 0.05 was considered that there was no potential publication bias (Begg, 1994). All above calculations were performed using RevMan5.1 (Cochrane collaboration, Oxford, UK) Publication biases were evaluated using the Begg's funnel plot by STATA 11.0 (STATA Corporation, College Station, TX).

Eligible Studies
The initial search yielded 1457 articles. We did another electronic search with the same key words using online EMBASE, which was unable to retrieve additional pertinent references. In all yielded publications including potential ones in reviews, reviewers identified 69 potential studies for full-text review. 42 studies were excluded for follow reasons: they did not mention survival outcomes     (Kubuschok et al., 1999), or used exactly identical cases in Kurusu' study ) and Yamashita's study (Yamashita et al., 2002). We finally used the information from both of the two articles and named it Kurusu Y in our list. Okumura's study (Okumura et al., 2009) referred the survival outcome of OS, but we can't calculate the HR (95% CI). Thus, we only extracted the patients' clinical characteristics in this article. Finally, we enrolled 12 Yamashita et al., 2000;Sher et al., 2005;Chen et al., 2007;Liu et al., 2008;Hou et al., 2009;Yoon et al., 2011;Hou et al., 2012;Nieva et al., 2012) articles containing survival outcomes and patients' clinical characteristics and 15 articles (Castaldo et al., 1997;Peck et al., 1998;Li et al., 2005;Hayes et al., 2006;Sheu et al., 2006;Huang et al., 2007;Sawabata et al., 2007;Guo et al., 2009;Okumura et al., 2009;Tanaka et al., 2009;Wu et al., 2009;Farace et al., 2011;Devriese et al., 2012;Wendel et al., 2012) containing only patients' clinical characteristics in our analysis (Figure 1). These studies were published between the year of 1997 and 2012. The total number of patients included was 2615, ranging from 9 to 250 patients per study (median, 78). HRs on OS, and PFS could be extracted for 11 and 5 studies respectively. Patients' clinical characteristics were listed in Table 1 and an overview of the study design variables were listed in Table 2.

Correlation between CTCs appearance and survival outcome (OS and PFS)
Overall Analyses: The meta-analysis of all studies on OS showed significant prognostic effects on CTCs detected in samples collected before and after treatment. The HR (95% CI) of 9 studies Yamashita et al., 2000;Sher et al., 2005;Chen et Table 3). Subgroup analysis: As several studies collected samples at various time points, we separately summarized them according to the time points in subgroup analyses stratified by either of patients' clinical characteristics we analyzed. When there was more one study focusing on a subgroup, we conducted a meta-analysis and listed the result in Table 3; otherwise, we listed the result of the original study without analysis.
As shown by the subgroup analysis stratified by method used to identify CTCs in peripheral blood, we found OS prediction effect of CTCs in the analyses of studies applied RT-PCR (shown in Table 2) using samples collected before treatment (HR = 3.04 [1.71, 5.42], n=390, I 2 = 65%, p =

Assessment of publication bias
As shown in Figure 4, Begg's test was used to examine publication bias. No significant publication biases were found in results of HRs for OS both using samples collected before and after treatment (P = 0.118 and P = 0.221 respectively). As for PFS, we obtained similar   results (P = 0.086 and P = 0.296 when using samples before and after treatment respectively).

Discussion
As we know, it was the first time that a comprehensive and detailed meta-analysis revealed the prognostic role of CTCs for lung cancer. CTCs expression was confirmed with a poor survival outcome according to the evidencebased medicine in our study.
Our results revealed CTCs' prognostic value in lung cancer (Table 3), which was in agreement with the recent meta-analysis in colorectal cancer (Rahbari et al., 2010), breast cancer (Zhao et al., 2011), melanoma (Mocellin et al., 2006) and prostate cancer (Wang et al., 2011). As referred in Hayes (Hayes et al., 2001), a prognostic factor with RR > 2 is considered as useful practical value. Fortunately, all the pooled HRs were above 2.0 in our study. These results indicated that detected CTCs appearance in peripheral blood of lung cancer patients could predict their prognosis practically.
Comparing the results yielded in studies using samples collected before and after treatments, we could find out that the HRs for survival outcome were significantly higher in post-treatment group ( These results indicated that the post-treatment detection of CTCs was more persuasive than that at baseline, which recommended us detecting CTCs after treatment rather than before to predict patients' survival. Furthermore, four studies (Yamashita et al., 2000;Chen et al., 2007;Hou et al., 2012) examined CTCs on the respectively identical populations both before and after treatment CTCs support our finding with higher HRs after treatment.
In SCLC subgroup analysis using random mode, we noticed that 2 included studies had significant results (1.43 [1.09, 1.89] and 3.56 [2.10, 6.04]), but they reached a conclusion of negative (HR 2.19 [0.90,5.34]). This could be explained by an HR compensation on confidence interval on the smaller side when a random model was applied, which leads to an overlap with 1 (Hedges & Vevea, 1998). This puzzle could be solved when much more studies were conducted to confirm clinical value of the CTCs tested in SCLC. For there were not always sufficient subgroup studies, when grouping studies by different detecting methods, the HRs could be only obtained in OS prediction by pro-and post-treatment CTCs detected by RT-PCR and PFS prediction by post-treatment CTCs detected by CellSearch. Thus, we could not reach in a conclusion which method was more accurate in detection of CTCs of prognostic value. However, Hofman's study  showed that HR value was higher using CellSearch than that of ISET in clinical research consisted of 208 patients. Future study could pay attention to this question to optimize the detection method.
In the correlation study of CTCs appearance with patients' clinical characteristics, the ORs revealed that pro-treatment CTCs appearance was correlated with TNM staging, lymph node status and distant metastasis. No significant or weak correlation had been observed with the depth of invasion, sexuality, histological differentiation and smoking status. Experimental studies had proven CTCs was correlated to distant metastasis former (Kim et al., 2009). Hou JM and colleagues summarized that CTCs is a factor that promotes metastasis as well (Hou et al., 2011). Coupled with a gradually increase OR of TNM staging through treatment, the detection of post-treatment CTCs had a potential ability in earlier, less invasive and more reliable discovery of disease progression in the follow up. Similarly, Tanaka et al. (2009) demonstrated that CTCs as a diagnostic marker in lung cancer,showed good sensitivity and specificity in distinguishing clinical stage. Lymph node status and happened distant metastasis were associated with pro-treatment CTCs but not during or after. This might be explained by that these clinical factors were obtained before treatment, whereas CTCs detection during or after treatment might be affected by the treatment.
Besides, the limitation still existed in the present detection method. As referred in Pantel K's study (Pantel and Alix-Panabieres, 2010), CTCs positive rate detected by identification of EpCAM in patients with happened distant metastasis were lower than that in non-metastasis patients. They hypothesized that it was the epithelialmesenchymal transition (EMT) that led to a decline in the EpCAM expression. Thus, CTCs of an EMT phenotype could be missed by current detection methods.Intriguingly, we found that the positive rate of CTCs after treatment was smaller than that before treatment in all the studies referred (Yamashita et al., 2000;Chen et al., 2007;Hou et al., 2012). This might be explained by platelet's role in promoting EMT with the influence of surgery which leads to local platelet accumulation (Labelle et al., 2011).
Significant heterogeneity was found in the metaanalysis of the prognostic role of CTCs collected before treatment (69%, 0.001).When we divided studies into subgroups of NSCLC and SCLC, the heterogeneity could not be eliminated (53%, 0.05). To exclude technique biases, subgroup analyses were performed for the most frequently used methods, RT-PCR, CellSearch and ISET (Pantel and Alix-Panabieres, 2010). This suggested that the techniques were unlikely to be a source of biases. Therefore, histological classification and detection methods were not major sources of heterogeneity. This could be explained by different cut-off values and different composition of NSCLC in each study. The meta-analysis performed in subgroup of post-treatment had revealed a fine homogeneity in both OS and PFS.
A potential source of biases was related to the HRs and 95% CI extrapolation. Once the key information was not provided by the authors, we calculated them from the data available in the article. Once there was no sufficient information for calculation, we extracted them from the survival curves. Multivariate survival analysis reported in the article was included in the our analysis; if these data were not available, we extracted univariate data instead. These results should be confirmed by an adequately designed prospective study. Furthermore, there was also some tiny bias derived from the software we used, designed by Matthew Sydes and Jayne Tierney. This was because this software retained only percentile when calculated the logHR and SE. However, when we verified the data again by STATA 11.0, only minimal bias was observed. The publication biases were additional problem for the meta-analysis. Fortunately, the Begg's test showed no significant publication bias (p > 0.05).
In conclusion, the meta-analysis suggested that the both pro-and post-treatment CTCs appearance in peripheral blood were associated with poor prognosis in lung cancer patients. It was of more significance using CTCs to predict survival after treatment. In addition, the detection of post-treatment CTCs had a potential ability in earlier, less invasive and more reliable discovery of disease progression in the follow up. These results should be confirmed by adequately multi-center designed prospective studies in future.