Prognostic Performance of Three Lymph Node Staging Systems in Patients After D2 Lymphadenectomy for Gastric Cancer

Gastric cancer is the fifth most common malignancy and the third leading cause of cancer-related death worldwide.¹ At present, advanced tumor invasion and lymph node (LN) involvement are the best indicators of poor prognosis in patients with resectable gastric cancer.^2,3

Several LN staging systems have been proposed for predicting prognosis in patients with gastric cancer. The numerical node staging system (pN), which was based on the number of metastatic lymph nodes (MLN), is the method most commonly used for nodal staging in clinical practice.⁴ However, this method does not account for the number of total harvested nodes (THNs) at the time of surgery, which might affect patient outcomes.⁵ The LN ratio (LNR), defined as the ratio between MLN and THN,^6–8 and the log odds of metastatic lymph nodes (LODDS), defined as the log of the ratio between the number of MLNs and the number of negative lymph nodes, were established to avoid stage migration and has been shown to be more reliable than the pN system for prognostic stratification.^9–13 Despite these findings, it remains inconclusive as to whether LNR and LODDS can replace pN staging in terms of accuracy and they are not yet widely accepted.

To address these issues, we compared the performances pN, LNR and LODDS staging and incorporated the optimal system into a nomogram with other significant prognostic indicators for predicting survival in patients with resectable gastric cancer.

Methods

Results

Table 1 summarizes patients' characteristics. The study cohort consisted of 423 patients. Among them 265 (62.6%) patients had LN metastasis; the retrieval of >15 LNs had been achieved in 226 (53.4%) of patients. The median follow-up period was 18.3 months (IQR: 9.5–27.2 months).

Table 1 Clinicopathologic characteristics of the study population

View Fullscreen

The value of LNR increased as the number of MLN increased, suggesting a close correlation between these two parameters (r = 0.879). However, the relationship was not linear and the value of LNR declined as THN >15 (Fig. 1A). A similar relationship was observed between LODDS and MLN (r = 0.888; Fig. 1B). In contrast, the correlation between LODDS and LNR remained approximately linear (r = 0.963) with the exceptions of the extremes (Fig. 1C)

View Figure

• Download as Powerpoint
• Download Figure

Multivariate cox regression analyses revealed that LNR had superior prognostic value (C-index: 0.712) compared to both pN (C-index: 0.695) and LODDS (C-index: 0.704). Furthermore, LNR remained the best performing system in patients with ≤15 or >20 LNs retrieved (C index: 0.647 and 0.725, respectively) when stratified by THN. However, in patients with 16 to 20 retrieved LNs, LODDS had the best prognostic performance (C index: 0.656; Table 2).

Table 2 Prognostic performances of pN, LNR and LODDS based on C-index

View Fullscreen

To provide an intuitive and quantitative method to better stratify patients with different prognosis, a nomogram incorporating LNR and other significant prognostic factors (age, CA 19-9 and CA 125) was established. The nomogram was able to predict each patient's prognosis individually: the higher the total score, the poorer the prognosis. For example, a 60 year-old patient (20 points) with an LNR of 0.2 (10 points), a serum CA 19-9 level of 300 (10 points) and a serum CA 125 level of 60 (22 points) would have a total of 62 points, yielding an estimated 2-year OS of 38%.

The unadjusted C-index for OS prediction was 0.712 and the bootstrap-corrected C-index was 0.704, which revealed minimal evidence of model overfit. The calibration plot showed close agreement between the actual 2-year OS values in the study population and the predicted OS according to the nomogram (Fig. 2B). Furthermore, the patients were divided evenly into 3 subgroups based on the tertiles of the total points and the outcomes within the subgroups were compared using the Kaplan-Meier curves. The results showed that each group represented a distinct prognosis (Fig. 2C).

View Figure

• Download as Powerpoint
• Download Figure

Discussion

Accurate staging of LN status is essential for predicting prognosis and planning treatments in patients with gastric cancer. In 2010, Sun et al reported that LODDS was superior to both the UICC pN and ratio-based LNR staging systems for prognostic assessment in patients after R0 resection.¹² However, subsequent reports have proved contradictory: whereas some studies supported their findings,^9–13 others did not find a superiority of LODDS over pN or LNR in discriminating prognoses in these patients.^18–20 The inconsistencies between the various studies may be due to differences in the statistical methods for analysis. All three LN staging systems are closely correlated, therefore studies using stepwise regression to determine the highest performing staging system have been inappropriate.^10,12,18 In addition, there is still no consensus on the ideal cutoff values for the three LN staging systems, and the choice of cutoff values between different studies may be affected by the size of the study cohort and the clinicopathologic characteristics of the patients.

Several reports have suggested that the prognostic value of LN staging can be improved by treating LN scores as continuous variables rather than categorical variables.^9,13 Therefore, we adopted these approaches to assess pN, LNR, and LODDS performances in the present study and the results demonstrated that LNR showed the highest prognostic accuracy in patients with R0 resected gastric cancer. Although the underlying mechanisms remain unknown, a possible explanation is that LNR may better mimic the actual mathematic interaction between MLNs and THNs on survival among this cohort. This finding was verified by bootstrap-validation.

Investigations to determine the prognostic accuracy of LN staging systems according to the number of THN have been conflicting: The studies by Sun et al¹² and Wang et al¹³ who found that LODDS constantly outperformed other LN systems irrespective of the number of THN. In contrast, Xu et al¹⁹ found that LODDS only gave the best performance if THN <10; LNR was better when THN >15; however, when Kaplan-Meier curves were used instead of ROC curves, pN showed the best discriminatory accuracy for THN >15. To date, Spoverato et al⁹ are the only investigators known to have incorporated pN, LNR and LODDS as continuous variables into a multivariate analysis to adjust for clinicopathologic factors. Their results indicated that LNR was the most reliable when the number of harvested LNs was ≤15 but that LODDS achieved the same C-index when THN >15.⁹ The present study applied the same methodology and obtained similar findings. The following factors may account for the inconsistencies: differences in statistical methods can lead to contradictory findings, even within the same study; for example, the study by Xu et al¹⁹ discussed above. The distribution of clinicopathologic characteristics can vary when stratified by THN; for example, higher THN values have been associated with younger patients, increasing stage and higher hospital volumes.^21,22 Therefore, further studies need to be conducted using methodologically sound approaches, such as matched case-control logistic regression, to investigate the prognostic performances of different LN staging systems within different THN subgroups.²³

Kim et al²⁴ constructed a nomogram incorporating patient age, gender, pT, tumor site and LNR to predict OS and showed it had better discriminatory power than both the UICC staging system and the nomogram established by the Memorial Sloan Kettering Cancer Centre.²⁵ The current nomogram achieved a similar level of discrimination to that of Kim et al.²⁴ Our cohort was selected over a narrower admission period (December 2012–July 2014) and the treatment regimens had been relatively standardized; however, only short-term survival (2-year OS) was assessed. Kim et al showed that pT was a strong predictor of OS, whereas in the present study pT failed to predict OS independently, possibly due to the small number of patients with T1/2 disease. An advantage of the nomogram in the present study over that of Kim et al. was that it incorporated preoperative serum values of CA 19-9 and CA 125. Previous studies have demonstrated that these factors were useful prognostic markers in patients who had undergone curative resection.^26,27 Further studies are still required to validate the value of incorporating CA 19-9 and CA 125 in prognostic nomograms.

In summary, this study has provided strong evidence that LNR was superior to other LN based staging systems in predicting survival outcome in patients who had undergone D2 resection. In addition, LNR was developed into a nomogram with age, and preoperative serum CA 19-9 and CA 125 to refine the prediction of OS in these patients. The nomogram was internally validated, and it may become a useful tool in prognosis, follow-up and treatment programming if externally validated.