The New American Joint Committee on Cancer/International Union Against Cancer Staging System for Adenocarcinoma of the Stomach: Increased Complexity without Clear Improvement in Predictive Accuracy

Between July 1985 and December 2009, a total of 2,589 patients with an adenocarcinoma of the stomach or GEJ underwent a gastrectomy at the Memorial Sloan-Kettering Cancer Center (MSKCC) and were entered in a prospectively maintained database. Patients with tumors of the GEJ (Siewert I–III, n = 669), and patients with a noncurative (R1 or R2) resection or with M1 disease (n = 358) were excluded. Because the data set focused on curative resections, all patients with M1 disease were excluded; all would have been staged as having stage IV disease in the 6th and 7th edition staging systems. Three patients with T0 N+ disease in their final pathology could not have a stage group assigned and were excluded, leaving 1,559 patients for analysis.

Most patients underwent a D2 lymph node dissection. Preoperative and postoperative therapy were administered according to the ongoing clinical trials and the standard of care at MSKCC during the study period. Adjuvant chemotherapy was provided infrequently from 1985 to 1999. From 2000 to 2009, perioperative chemotherapy became more common for advanced-stage tumors; postoperative chemoradiation was also administered between 2000 and 2007. Follow-up was generally conducted according to published National Comprehensive Cancer Network guidelines.6

Survival data were updated when available until March 2010. This study was approved by the institutional review board of MSKCC. This data set was used in part to help guide changes to the AJCC 7th edition.

In the Dutch Gastric Cancer Trial (DGCT, 1989–1993), 1,078 patients with adenocarcinoma of the stomach were randomized for D1 or D2 lymphadenectomy.7‐9 None of the patients had a tumor of the GEJ, while patients with metastatic disease (n = 367) and patients who underwent a noncurative resection (n = 74) were excluded, leaving 637 patients who underwent an R0 resection for this study. No adjuvant therapy was provided to these patients in the curative setting. Follow-up was conducted every 6 months. Recurrent disease was generally confirmed with radiology, endoscopy, and/or histology. Survival data were updated when available until November 2007.

Because the Union for International Cancer Control (UICC) and AJCC use the same staging definitions, for purposes of clarity, the UICC/AJCC staging system is referred to here as the AJCC staging system. Tumor, nodal, and metastasis stage and stage grouping are all based on final postoperative pathology. All staging parameters (T, N, M) and stage groupings of the 6th and 7th edition staging system were calculated on the basis of depth of invasion through the gastric wall, the number of positive lymph nodes, and the presence or absence of distant metastases. No patients were excluded as a result of incomplete staging data.

Survival probabilities were estimated with the Kaplan-Meier method, and differences in survival curves were assessed by the log-rank test. The end point in this study was disease-specific survival, which was recorded from the date of surgery until the date of death of disease; dead from other causes and alive at last date of follow-up were recorded as censored events.

The concordance index between survival and stage for the two staging systems was calculated by a previously described methodology.10 Concordance for a staging system can range from 0 to 100 %, with 100 % representing absolute concordance, 50 % indicating no association (no better than flipping a coin), and 0 % perfect discordance. The concordance index for a staging system was calculated by analyzing all possible pairs of two patients in the data set. A pair of two patients is concordant if the patient with the higher stage has the shorter survival. Concordant pairs are assigned a value of 1, and discordant pairs are assigned a value of 0. The concordance of the staging system is the sum of the values of all the individual pairs divided by the total number of pairs in the data set. For pairs where the shorter survival time was censored, the stage-specific Kaplan-Meier estimate of survival was used. Pairs in which both patients were in the same stage group were assigned a value of 0.5. Therefore, the maximum concordance of the staging system could never be 100 %. The maximum potential concordance in our data set for the 6th edition was 0.818 and 0.853 for the 7th edition. Confidence intervals and P-values for the difference in concordance indices of the two staging systems were calculated by bootstrap resampling.

To validate the results provided by concordance analysis, the Brier score was used to evaluate the expected error of the predictions in both staging systems. For every patient, the Brier score measures the difference between the survival probability predicted by the staging system and the observed survival. Kaplan–Meier estimates were used for censored observations. The average squared deviation for all patients gives the Brier score, in which a lower score represents a better predictive accuracy.

All 2,196 patients in this analysis underwent a radical (R0) resection for an adenocarcinoma of the stomach between July 1985 and December 2009, either at MSKCC (n = 1,559) or at one of the hospitals participating in the DGCT (n = 637). Patient characteristics are summarized in Table 3. Median follow-up was 98 months.

Figure 1 depicts the distribution of T and N classification of the 6th edition and 7th edition staging systems for all 2,196 patients. The redefined N1, N2, and N3 classifications were more evenly distributed. Among 2,196 patients, 674 (31 %) were assigned a higher N classification in the AJCC 7th edition. In the 7th edition staging system, the N3 category is divided into N3a (7–15 positive nodes) and N3b (16 or more positive nodes). This recognized the unique independent prognostic significance of an increasingly higher number of positive nodes, even at the high end.

Differences in stage distribution between the two systems are listed in Table 4. For stage IB to IV, T, N, and M as well as stage group definitions were altered, leading to a change in stage group for 1,302 (59 %) of 2,196 patients. In total, 748 patients (34 %) moved to a higher-stage group, and 186 patients (9 %) moved to a lower-stage group. A total of 368 patients (17 %) with a stage II tumor in the 6th edition staging system were distributed between stage IIA and IIB in the 7th edition staging system. Of note, stage grouping did not make use of the N3a/N3b classification.

Five-year survival estimates for both staging systems are shown in Table 5 and Fig. 2. In the 6th edition staging system (Fig. 2a), Kaplan–Meier survival estimates significantly differed for stage IA–IB, IB–II, II–IIIA, and IIIA–IIIB, but not for stage 0–IA (P = 0.64) and IIIB–IV (P = 0.60).

In the new staging system, stage group 0 and IA remain unchanged. Differences between the 7th edition stage groups were significant for stage IA–IB, IIA–IIB, IIB–IIIA, and IIIB–IIIC but not for stage 0–IA (P = 0.64), IB–IIA (P = 0.09), and IIIA–IIIB (P = 0.15, Fig. 2b).

Figure 3a shows patients from the 6th edition stage II, which is subdivided into stage IIA, IIB and IIIA in the 7th edition staging system. Differences between the curves were all significant. In Fig. 3b, the subdivision of 6th edition stage IIIA into 7th edition stage IIIA and IIIB is shown; no significant differences between the two new stage groups were detected (P = 0.26).

Overall, in the AJCC 6th edition, two out of six consecutive steps between stage groups were not significantly discriminant, while in the AJCC 7th edition, three out of seven consecutive steps were not significantly discriminant, indicating that the discriminant value between stage groups has decreased between the 6th and 7th edition staging system.

The concordance index of T staging did not change significantly from the 6th to the 7th edition (P = 0.36) (Table 6). The concordance index of N staging showed an increase from 0.659 to 0.665 (P = 0.03). Despite the change in definition for almost every stage group and the increased number of stage groups, the concordance estimate for the 7th edition was 0.697, which was significantly inferior to that of the 6th edition staging system (0.711, P < 0.01). Brier score for T, N, and overall stage groupings showed no significant improvement from the 6th to the 7th edition.

Establishing homogeneity within stage groups requires grouping of TNM combinations that have similar survival estimates (Table 2). For homogeneity testing, results are highly dependent on the size of the data set. Ahn et al.11 showed improved homogeneity of two homogeneous stage groups in the 7th edition compared to one homogeneous stage group in the 6th edition, using a data set of nearly 10,000 patients. In the current study, numbers are smaller, and therefore significant homogeneity within stage groups is hard to detect (results not shown).

Heterogeneity between stage groups can be assessed by comparing stage-specific survival estimates for significant differences. Whether differences between stage groups are significant is highly dependent on the size of the data set. Small differences in survival estimates between stage groups are more likely to be statistically significant in a large data set. In the current study, stage-specific heterogeneity has decreased in the 7th edition when compared to the 6th edition. Although AJCC 6th edition stage II contained a highly heterogeneous population (Fig. 3a), and distributing these patients between stages IIA, IIB, and IIIA in the 7th edition has created three groups with a significantly different prognosis, the distribution of 6th edition stage IIIA patients into AJCC 7th edition stages IIIA and IIIB has created two stage groups with almost identical stage-specific survival (Fig. 3b). Wang et al. showed decreased heterogeneity between stage groups in the 7th edition as well.12

Performance of a staging system can also be assessed on the individual patient level by comparing survival of patients with different stages. Several ways of comparing staging systems on an individual-patient level have been proposed, but there is no standard method.21 Commonly used methods include explained variation (or Brier score), the area under the receiver–operator characteristic curve, the concordance index, and a summary measure of separation. We decided to use the concordance index and Brier score to measure the prognostic accuracy of the staging systems because they analyze different, complementary measures. Concordance index is a measure of whether ranking of patients by staging is consistent with the ranking of their outcome. Its advantages include interpretation (because it is a probability), robustness (because it is based on ranks, it is not sensitive to small changes in the data), and availability of appropriate statistical methods for estimation. It also incorporates a built-in penalty for staging systems with a higher number of categories, so that with equally performing staging systems, the system with more categories will have a lower concordance probability. It does not penalize possible shifts (miscalibrations) between predicted and observed survival. Therefore, we also used the Brier score because it looks at the actual difference (in months) between predicted and observed survival, taking possible shifts into account.

In the current data set, concordance analysis showed no difference for T category, an improvement for N category, and a decline for stage grouping. Brier scores consistently showed no significant improvement from the 6th to the 7th edition. Therefore, it can be concluded that for individual patient outcome, no improvements were detected from the 6th to the 7th edition staging systems.

Only one of the previously published studies compared the two staging systems on an individual-patient level. It found increased predictive accuracy for the 7th edition staging system.12 A disadvantage of the method employed in that study is that the metric used for comparison, the Akaike information criterion, measures how well the staging system fits to the used data set without assessing the actual prognostic accuracy.

The larger number of stage group categories for the 7th edition of the staging system means that the system has become more complex. Increasing the number of categories of the staging system is not unique to gastric cancer.4 With the increasing availability of pathologic and molecular data, there is a trend toward incorporating more and more information into newer staging systems. Although these new categories might better reflect the natural history and prognosis of these diseases, there is a limit to the improvement of prognostic accuracy achievable with a categorical anatomic-based staging system like the TNM classification.22,23 At the same time, the goal of creating an intuitive, easy-to-use staging system disappears, and in daily clinical practice, cancer staging consists of using complex tables, if it is used at all.

Meanwhile, tools for individual patient prognostication have been developed that significantly outperform the TNM classification in prognostic accuracy. For gastric cancer, a nomogram has been developed based on a single US institution’s database.24,25 This nomogram has been validated in several international patient cohorts.26‐28 The question is whether the TNM classification should aspire to the same goal of highly accurate individual patient prognostication as these nomograms. Prognostication is only one of the five goals of the TNM classification; all the other goals are directed toward a simple, intuitive international language: to aid the clinician in planning and evaluating treatment, to facilitate the exchange of information, and to contribute to research.1

In summary, the 7th edition of the AJCC staging system for gastric cancer has resulted in improved predictive accuracy for the N classification but decreased heterogeneity among stage groups. The increased complexity of the 7th edition staging system is not accompanied by an improvement in prognostic accuracy of stage grouping. Staging represents a compromise in accounting for the most reproducible and prognostically relevant factors to aim at a simple, intuitive, useful, common language to describe the natural history of a tumor. It should not be confused with more complex multivariable prognostication models, which may be useful in defining groups of patients at homogenous risk of recurrence, regardless of anatomic TNM characteristics.