You are here

The IASLC Lung Cancer Staging Project: External Validation of the Revision of the TNM Stage Groupings in the Eighth Edition of the TNM Classification of Lung Cancer

Journal of Thoracic Oncology, July 2017, Volume 12, Issue 7, Pages 1109-1121

Commentary by Solange Peters
The AJCC/UICC stage classification of lung cancer is relatively unique among stage classifications of cancer sites. It is uniquely and rationally based on a large international database and an extensive analysis with multiple levels of internal validation.

Revisions to the TNM stage classifications for lung cancer, informed by the international database (N = 94,708) of the International Association for the Study of Lung Cancer (IASLC) Staging and Prognostic Factors Committee, need external validation. The objective of this manuscript is an external validation of the revisions by using the National Cancer Data Base (NCDB) of the American College of Surgeons.

The NCDB database is particularly useful for TNM classification validation because it is strong in an area in which the IASLC database is weak. Indeed, there is a relative underrepresentation of North America and nonsurgically managed patients in the IASLC database.

Cases presenting from 2000 through 2012 were drawn from the NCDB and reclassified according to the eighth edition stage classification. The T, N, and overall TNM classifications were evaluated according to clinical, pathologic, and “best” stage (N = 780,294).

Interesting points reported were a better survival for all stage categories, revealing possibly some better patient’s selection. Some of this difference can be explained by the higher proportion of surgical cases, and perhaps also by the higher proportion of adenocarcinoma cases in this NCDB database. Additionally, more recent years of analysis, and potentially variability in transversal quality of care, from surgery to death, can also explain this improved outcome across stages, with this database focusing on more performant centers than the whole database.

Despite the size and detailed nature of the NCDB database, there are some essential limitations. To me, the NCDB data does not allow to conclude anything about stage IV disease, knowing that if surgery is performed in that setting, this is only for highly selected patients (oligometastatic ?) or incidental stage IV diagnosis.   Validation of the M1a/b/c categories could therefore not be performed in this trial.

However, and this is the most important goal of this validation, the eighth edition TNM stage classification system demonstrated consistent ability to discriminate TNM categories using NCDB database, and confirming its adequacy for NSCLC stage grouping. 

Abstract

Introduction

Revisions to the TNM stage classifications for lung cancer, informed by the international database (N = 94,708) of the International Association for the Study of Lung Cancer (IASLC) Staging and Prognostic Factors Committee, need external validation. The objective was to externally validate the revisions by using the National Cancer Data Base (NCDB) of the American College of Surgeons.

Methods

Cases presenting from 2000 through 2012 were drawn from the NCDB and reclassified according to the eighth edition stage classification. Clinically and pathologically staged subsets of NSCLC were analyzed separately. The T, N, and overall TNM classifications were evaluated according to clinical, pathologic, and “best” stage (N = 780,294). Multivariate analyses were carried out to adjust for various confounding factors. A combined analysis of the NSCLC cases from both databases was performed to explore differences in overall survival prognosis between the two databases.

Results

The databases differed in terms of key factors related to data source. Survival was greater in the IASLC database for all stage categories. However, the eighth edition TNM stage classification system demonstrated consistent ability to discriminate TNM categories and stage groups for clinical and pathologic stage.

Conclusions

The IASLC revisions made for the eighth edition of lung cancer staging are validated by this analysis of the NCDB database by the ordering, statistical differences, and homogeneity within stage groups and by the consistency within analyses of specific cohorts.

Keywords: Lung cancer, Staging, National Cancer Database, AJCC, UICC, Validation.

Introduction

Proposals to revise the staging criteria for lung cancer were published in 2015 by the Staging and Prognostic Factors Committee of the International Association for the Study of Lung Cancer (IASLC). 1 2 3 4 5 The proposals take effect on January 1, 2017, in the staging guidelines for the Union for International Cancer Control (UICC) and on January 1, 2018, for the American Joint Committee on Cancer (AJCC). Changes to the T component included subclassification of T1 and T2 by size in 1-cm increments and reclassification of tumors larger than 5 cm as T3 and tumors larger than 7 cm as T4. Diaphragm invasion became a T4 descriptor. Lung atelectasis, whether partial or total, and all cases of main bronchus invasion regardless of the distance from the carina were classified as T2. Tumors with extrathoracic metastases were subdivided into M1b for a single distant metastasis and M1c for multiple distant metastatic lesions. No changes were made to the N component for the eighth edition. Following the proposed revisions to the T and M components, the TNM stage grouping scheme was revised accordingly. All of these revisions were informed by the analyses of an international database, with participants submitting data from 46 sites from 19 countries. 6

Internal validation of the proposed revisions has been described in a separate publication. 7 However, recommendations based on a single database, regardless of the size, geographic representation, and heterogeneity of data sources, should also be validated externally, through one or more separate databases. The National Cancer Data Base (NCDB) provides an excellent opportunity for external validation of the TNM staging recommendations. The NCDB is a large and inclusive North American database, with broad representation from all treatment modalities for lung cancer and the entire range of institution types, from community hospitals to university research institutions. The eighth edition TNM staging criteria were evaluated against this large and detailed database.

Methods

The NCDB is a hospital-based registry jointly produced by the Commission on Cancer of the American College of Surgeons (ACoS) and the American Cancer Society. The NCDB asserts that “the National Cancer Data Base (NCDB) is a joint project of the Commission on Cancer (CoC) of the American College of Surgeons and the American Cancer Society. CoC’s NCDB and the hospitals participating in the CoC NCDB are the source of the de-identified data used herein; they have not verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors.” The de-identified cases, which were originally entered by participating institutions with appropriate human subjects' approval, were received and analyzed in the Division of Thoracic Surgery at the Swedish Cancer Institute, Seattle, Washington, which is an ACoS member hospital. Institutional review board review of the analysis of the de-identified data set at the Swedish Cancer Institute was not required.

Cases entered into the registry with invasive NSCLC or SCLC presenting from January 2000 to December 2012 were used for analyses. Cases that were classified as a second primary tumor, or without sufficient anatomical information to be able to classify them according to eighth edition stage were excluded. To maintain consistency with the analyses of the IASLC database, cases with yp staging (stage based on pathologic evidence after neoadjuvant therapy [approximately 2% of NCDB cases]) were excluded from analyses of pathologic stage. Most of these yp cases were also effectively excluded from analyses of clinical stage as well, because the tumor descriptors were based on surgical findings after neoadjuvant therapy. Cases were also excluded if the TNM descriptors were in conflict with the recorded stage. Surgical cases from both databases were included regardless of margin status after surgery.

TNM components for the NCDB cases were entered by the participating registrars according to sixth or seventh edition AJCC criteria, depending on the period during which they were registered. Pathologic TNM stage was provided in surgical cases, and clinical stage was provided in both surgical and nonsurgical cases. Anatomic T descriptors were provided following the AJCC Collaborative Stage guidelines, with only the “highest” T descriptor given for each case. For each case, there is a field indicating the source of the recorded T descriptor information, either surgical/pathologic staging or pretreatment radiologic/clinical. Information about the NCDB and its data elements resides at the website of the ACoS. 8

Cases were reclassified according to the eighth edition stage groupings after translation of the T and M categories using the descriptors provided with each case. In cases originally entered under the fifth or sixth edition, particular care was taken to ensure that cases were reclassified correctly, including in the event of additional nodules. Cases were included in either the analyses of clinical stage or analyses of pathologic stage depending on the basis of the size and extent descriptors. Size and extent descriptors were based on pathologic staging in surgical cases and clinical staging in nonsurgical cases. A small subset of cases could be included in analyses of both clinical and pathologic stage. In these cases, both cTNM and pTNM stage categories were recorded; additional anatomical descriptors were not needed to reclassify for the eighth edition (e.g., seventh edition T4). Additionally, all cases were included together in analyses of “best” stage, which is pathologic stage where available and clinical otherwise. Because of the hierarchical reporting of a single T-descriptor per case when in fact there could be more than one, the placements of individual anatomic descriptors within the T-component categories were not evaluated. The revised M component of classification could not be validated because the number of distant metastatic lesions was not provided. All M1 cases were therefore classified simply as stage IV.

NSCLC and SCLC were analyzed separately. For NSCLC, survival curves according to eighth edition stage were drawn for both data sets. For reasons previously explained, 5 adjusted survival curves were drawn for the IASLC data using inverse probability weights applied to the survival calculations, based on the proportion of cases that were from registry databases (versus others) in each stage category. 9 This weighting procedure was not applicable to the NCDB data. The analysis data sets for the IASLC data were the same data sets that were used for the originally reported component analyses of T, N, and TNM. Stages IVA and IVB were combined to form a stage IV group in the IASLC data, consistent with the NCDB data.

Formal comparisons between adjacent stage groups were performed using Cox proportional hazards regression, adjusting for age, sex, and cell type for NSCLC in each data set. In the case of best overall stage, regression analyses were stratified by whether treatment included surgical resection of the primary tumor. For the IASLC data set, Cox regression analyses were also stratified by type of source data: registry versus other. R 2 statistics were calculated by using the method of O’Quigley and Xu. 10

Similar analyses were carried out for SCLC. However, in keeping with the original analyses of the IASLC database, survival curves in the SCLC database were not adjusted for registry submission. Because the overwhelming majority of SCLC cases are in the clinically staged subset, best stage was not analyzed for SCLC.

To explore potential differences in overall prognosis between the NCDB and IASLC NSCLC data sets, the NCDB and IASLC data sets were combined to analyze the entire group with a Cox model. The database of origin (NCDB versus IASLC) was included as a predictor variable, stratified by surgical treatment and adjusting for the baseline factors age, sex, and cell type as already described. All analyses, including the reclassification of NCDB cases to the eighth edition stage criteria, were done with SAS System for Windows software, version 9.4 (SAS Institute, Inc., Cary, NC).

Results

Cases could be classified according to the eighth edition in all but two circumstances. The first is T3 cases that may have had diaphragm invasion in the presence of another “higher” T3 descriptor. This would effectively mask the diaphragm invasion, thereby preventing the case from being correctly classified as T4. As a result, there may have been some cases in which diaphragm invasion was present but not classified as such and the case was inappropriately classified as T3. The second circumstance is M1 cases that could not be classified according to the newly proposed distinction between single (M1b) and multiple lesions within single or multiple sites (M1c). Whether disease was restricted to the thorax (M1a, stage IVA) and whether there were multiple sites of extrathoracic metastases (M1c) could be determined; however, whether there was a single extrathoracic lesion (M1b, stage IVA) or multiple extrathoracic lesions at a single or multiple sites (M1c, stage IVB) could not be determined. Because the distinction between M1b and M1c is the basis for distinguishing stage IVA from stage IVB, all stage IV cases were grouped together in the analyses.

NSCLC

In total for the NCDB, 1,419,185 NSCLC cases presenting in 2000–2012 were subjected to initial screening. Among these, 314,992 (22%) were eliminated at initial screening on account of missing information related to stage, histologic type, or vital status or on account of classification as noninvasive or a second primary tumor ( Supplementary Fig. 1 A ). From the remaining cases, 612,534 were determined to be eligible for analyses of clinical stage and 182,616 for pathologic stage. There was some overlap where specific descriptors were not needed for reclassification, with 14,856 cases available for both analyses. Analyses according to best stage (pathologic where available, otherwise clinical) could therefore be accomplished in 780,294 cases. Characteristics for this group and for the NSCLC data set from the IASLC Staging and Prognostic Factors Committee database, are shown in Table 1 . The NCDB had a lower percentage of adenocarcinomas compared with the IASLC database (50% versus 61% after exclusion of not otherwise specified cases), most likely reflecting the fact that the IASLC database includes many cases from Asia, which are known to have a higher proportion of adenocarcinomas ( Supplementary Fig. 2 ).

Data Set NSCLC SCLC
NCDB IASLC Database NCDB IASLC Database
Overall data set (best stage)        
Time frame 2000–2012 1999–2010 2000–2012 1999–2010
No. 780,294 39,441 157,910 5002
Median follow-up (mo) a 40 63 27 31
Surgical patients 25% 88% 3% 23%
Male patients 54% 63% 50% 68%
Clinically staged cases        
No. 612,534 17,477 155,293 4848
Surgical <5% 87% 1% 12%
No. also analyzed for pathologic stage 14,857 12,972 994 428
Also analyzed for pathologic stage 2% 74% 0.6% 8.8%
Pathologically staged cases        
No. 182,616 34,936 3611 582

a In patients alive at last contact.

IASLC, International Association for the Study of Lung Cancer; NCDB, National Cancer Data Base.

Table 1Patient Characteristics for the Lung Cancer Data Sets Derived from the National Cancer Data Base, 2000–2012, and the IASLC International Staging and Prognostic Factors Committee Database, 1999–2010

 

As with the IASLC database, 6 the NCDB contains all treatment modalities ( Fig. 1 A ) However, because cases with ypTNM staging after neoadjuvant therapy were excluded, modalities that included preoperative chemotherapy or radiotherapy are not included and not shown in the treatment distribution. Approximately 16% of patients with NSCLC in the analysis data set had an operation only and 25% received chemotherapy only. Eight percent had an indication of radiation therapy as the only mode of treatment and 22% had some combination of surgery, chemotherapy, and RT, whereas 28% were untreated. Treatment modalities were unknown in 2% of cases. The NCDB contained 25% surgical cases overall, as opposed to 88% in the IASLC database, in which surgical services made up a large proportion of the data contributors.

gr1
 

Figure 1
Treatment recieved in the National Cancer Data Base. Treatment received according to eighth edition stage groups (best stage) in the National Cancer Data Base. ( A ) NSCLC. ( B ) SCLC. Ch, chemotherapy; ChRT, combined chemotherapy and radiotherapy; No Tmt, no treatment; RT, radiotherapy; S, surgical resection; S-Ch, surgery followed by chemotherapy; S-ChRT, surgery followed by combined chemotherapy and radiotherapy; S-RT, surgery followed by radiotherapy; Unk, treatment unknown.

 

The median overall survival (OS) in the NCDB clinically staged subset ranged from 37.5 months in the stage IA1 category to 4.8 months in the stage IV category. In the clinically staged cases of the IASLC data set, the median survival was not reached in any stage I category and was 8.8 months in the stage IV category ( Fig. 2 A ). In previous analyses of the IASLC TNM database, a more granular view of clinical stage IV survival showed a median survival of 11.5 months in the stage IVA category and 6.0 months in the stage IVB categories. 5 An evaluation of the comparison between stage IIIC and IVA could not be accomplished with the NCDB because of the aforementioned inability to distinguish stage IVA from stage IVB. Formal comparisons between adjacent clinical stage groupings in the NCDB data set, adjusted for cell type, age, and sex, showed significant differences between each stage group with the exception of IA1 and IA2, as well as IIA and IIB, which had overlapping survival ( Table 2 ). The overlap between these two pairs of stage groupings in the clinically staged cases is a reflection mainly of overlap between corresponding T-category cohorts ( Supplementary Fig. 3 A ) and may be a result of imprecise tumor measurement in these clinically staged tumors. This overlap between T categories (between T1a and T1b and between T3 and T2b), which has relevance to the proposed revisions to the T component, was observed in the clinically staged group but not in the pathologically staged group ( Supplementary Fig. 3 B ). Accordingly, for pathologic stage ( Fig. 2 B ), there was separation between all adjacent stage I to IIIC groups ( p < 0.0001 for all comparisons [see Table 2 ]).

gr2
 

Figure 2
Stage groups for NSCLC. Overall survival in patients with NSCLC according to the eighth edition stage groups in the National Cancer Data Base (NCDB) and International Association for the Study of Lung Cancer (IASLC) data sets (IASLC survival curves are weighted by type of submitting database). ( A ) Clinically staged (cTNM) tumors. ( B ) Pathologically staged (pTNM) tumors. MST, median survival time (months).

NSCLC
Comparison
Clinical Stage Pathologic Stage
NCDB 2000–2012 Data Set IASLC 1999–2010 Data Set NCDB 2000–2012 Data Set IASLC 1999–2010 Data Set
HR p Value HR p Value HR p Value HR p Value
IA2 vs. IA1 1.04 0.3208 1.83 <0.0001 1.13 <0.0001 1.43 <0.0001
IA3 vs. IA2 1.18 <0.0001 1.41 <0.0001 1.15 <0.0001 1.32 <0.0001
IB vs. IA3 1.21 <0.0001 1.29 <0.0001 1.16 <0.0001 1.31 <0.0001
IIA vs. IB 1.13 <0.0001 1.29 0.0017 1.11 <0.0001 1.27 <0.0001
IIB vs. IIA 1.03 0.0703 1.32 0.0005 1.33 <0.0001 1.39 <0.0001
IIIA vs. IIB 1.21 <0.0001 1.52 <0.0001 1.32 <0.0001 1.65 <0.0001
IIIB vs. IIIA 1.23 <0.0001 1.36 <0.0001 1.55 <0.0001 1.68 <0.0001
IIIC vs. IIIB 1.15 <0.0001 1.34 <0.0001 1.47 <0.0001 1.83 <0.0001
IV vs. IIIC a 1.54 <0.0001 2.41 <0.0001 1.01 0.8837 b b
R 2 15.9% 68.0% 31.2% 46.1%

a This represents all stage IV cases versus stage IIIC cases because the NCDB cannot distinguish between stage IVA and IVB cases with available NCDB data fields.

b Pathologic stage IV cases were poorly represented in the IASLC 1999–2010 database.

Note: Hazard ratios for the IASLC data are also stratified by type of submitting database: registry versus other. Results are from a multiple-variable Cox model.
NCDB, National Cancer Data Base; IASLC, International Association for the Study of Lung Cancer; HR, hazard ratio; R 2 , a statistical measure of discrimination reflecting the percent of overall variance explained by the factors in a model.

Table 2Overall Survival Comparisons by Eighth Edition Clinical and Pathologic Stage Groups in the NCDB and IASLC Data Sets for NSCLC, Adjusted for Age (>70 y), Sex, and Adenocarcinoma Histologic Type

 

Although no changes were made to the N component for the eighth edition, clinical and pathologic N categories were examined in the NCDB as in the IASLC M0 cases ( Supplementary Fig. 4 ). The N categories are well separated, with the exception of clinical N3 compared with N2 in the NCDB data set. The statistical comparison of clinical stage N3 versus clinical stage N2, adjusted for age, sex, and cell type was statistically significant ( p < 0.0001) but the hazard ratio (HR) was somewhat small at 1.11 (data not shown). The survival estimates for pathologic N categories are very similar in both data sets, with similar separation between all of the N categories ( Supplementary Fig. 4 B ).

For the 780,294 cases analyzed for best stage in the NCDB data set ( Fig. 3 ), all stage groupings were ordered and distinct from one another, with a median OS ranging from 103 months for the stage IA1 group to 4.9 months for the stage IV group. The 2- and 5-year survival estimates exhibited similar patterns. The stage I groups had 5-year survival rates ranging from 69% to 53%. The 5-year survival rate for the stage IV group was 3% compared with 6% in the IASLC data set. Stage IVA and IVB were combined in the IASLC data set to be consistent with the NCDB database. All stage groups were statistically distinct from one another, with all p values less than 0.0001 ( Table 3 ).

gr3
 

Figure 3
Stage groups for NSCLC (best stage [American Joint Committee on Cancer best stage , means pathologic if available, otherwise clinical stage]). Overall survival in patients with NSCLC according to the eighth edition stage groups in the National Cancer Data Base (NCDB) and International Association for the Study of Lung Cancer (IASLC) data sets (IASLC survival curves are weighted by type of submitting database). MST, median survival time (months).

Comparison NCDB 2000–2012 Data Set IASLC 1999–2010 Data Set
HR p Value HR p Value
IA2 vs. IA1 1.12 <0.0001 1.50 <0.0001
IA3 vs. IA2 1.19 <0.0001 1.31 <0.0001
IB vs. IA3 1.17 <0.0001 1.31 <0.0001
IIA vs. IB 1.16 <0.0001 1.28 <0.0001
IIB vs. IIA 1.19 <0.0001 1.38 <0.0001
IIIA vs. IIB 1.27 <0.0001 1.68 <0.0001
IIIB vs. IIIA 1.17 <0.0001 1.38 <0.0001
IIIC vs. IIIB 1.11 <0.0001 1.18 <0.0001
IV vs. IIIC a 1.55 <0.0001 1.35 <0.0001
R 2 43.2% 62.8%

a This represents all stage IV cases versus IIIC cases because NCDB cannot distinguish between stages IVA and IVB with available NCDB data fields.

Note: Hazard ratios for the IASLC data are also stratified by type of submitting database: registry versus other. Results are from a multiple-variable Cox model.
IASLC, International Association for the Study of Lung Cancer; HR, hazard ratio; NCDB, National Cancer Database; R 2 , a statistical measure of discrimination reflecting the percent of overall variance explained by the factors in a model.

Table 3Overall Survival Comparisons by Eighth Edition Best Stage Group in the NCDB and IASLC Data Sets for NSCLC, Adjusted for Age (>70 y), Sex, and Adenocarcinoma Histologic Subtype and Stratified by Type of Primary Treatment: Surgical Versus Nonsurgical

 

Overall and within each stage, the survival prognosis was better in the IASLC data set. Some of this difference can be explained by the higher proportion of surgical cases, and perhaps also by the higher proportion of adenocarcinoma cases in this database. However, these factors do not seem to explain the difference entirely. In a combined Cox regression analysis of both databases together, adjusting for stage, histologic type, age, sex, and whether a resection was attempted on the primary tumor, the IASLC data set reflected a higher OS (HR = 1.22 in favor of the IASLC data set [ p < 0.0001]).

A simultaneous test for the significance of all database-by-stage interactions was performed in a combined NCDB-IASLC data set. This test was performed to assess whether the HRs (compared with baseline [stage IA1]) for any of the individual stage categories differed by database. The test was significant, with each individual interaction term being significant ( p < 0.0001) and HRs on the interaction terms all being less than 1 (data not shown). This indicates that each stage group in the NCDB data set shows a less pronounced (albeit still significant) difference from baseline compared with the IASLC data set, which could result from the two data sets being different in terms of population and range of OS outcomes, but it does not detract from the overall validation of discrimination between stage groupings in the eighth edition.

SCLC

The SCLC subset from the NCDB contained 215,395 cases after screening out of an initial 256,913 cases ( Supplementary Fig. 1 B ), and 40,556 cases (16%) were classified as a second primary and excluded. From the initially filtered group, there were 155,293 cases (60%) with complete and conflict-free clinical TNM stage and 3611 (1%) with a pathologic stage. Both clinical and pathologic stage could be analyzed in 994 cases. A comparison of patient characteristics (see Table 1 ) again shows that there is a higher proportion of surgical cases in the IASLC data set than in the NCDB data set (23% versus 1%) ( Fig. 1 B ).

The survival outcomes according to clinical stage in both the NCDB and IASLC data sets share common features, particularly in stages III and IV: stage IV is distinct from stage IIIC in terms of prognosis, there is less distinction between stage IIIB and IIIC, and stage IIIA is well separated from both IIB and IIIB. ( Fig. 4 A ). Statistically, this can be seen in the HRs for both the NCDB and the IASLC database ( Table 4 ). There were too few cases in the previous analyses of stage I and II in the IASLC database to achieve robust results, with fewer than 50 events per stage group. Although these early-stage cases are more numerous in the larger NCDB data set, the differences between the lower-stage groups still do not reach statistical significance, with small differences in survival. In the pathologically staged cases (see Table 4 and Fig. 4 B ), stage IIIC (T4N3) is very rare in both databases, and stage IIA is very poorly represented as well. With the exception of the poorly represented groups and the overlap among the stage IA groups, the eighth edition staging scheme appears to work well for pathologic stage, especially in the later-stage SCLC cases. The result of the trend test of no difference between stages against an ordered alternative was significant ( p < 0.0001) for both clinical and pathologic TNM stages, suggesting a trend in the appropriate direction for the stage groups.

gr4
 

Figure 4
Stage groups for SCLC. Overall survival in patients according to the eighth edition stage groups in the National Cancer Data Base (NCDB) and International Association for the Study of Lung Cancer (IASLC) data sets. ( A ) Clinically staged (cTNM) tumors. ( B ) Pathologically staged (pTNM) tumors. MST, median survival time (months).

SCLC
Comparison
Clinical Stage Pathologic Stage
NCDB 2000–2012 Data Set IASLC 1999–2010 Data Set NCDB 2000–2012 Data Set IASLC IASLC 1999–2010 Data Set
HR p Value HR p Value HR p Value HR p Value
IA2 vs. IA1 0.74 0.0051 2.30 0.1721 1.08 0.6235 1.91 0.1766
IA3 vs IA2 1.08 0.1933 0.75 0.3649 0.91 0.3586 0.79 0.3868
IB vs. IA3 1.08 0.1488 1.73 0.1255 1.28 0.0334 1.34 0.2765
IIA vs. IB 1.10 0.1398 1.37 0.5101 1.25 0.1540 0.87 0.7091
IIB vs. IIA 0.96 0.4307 0.86 0.7456 1.20 0.2016 1.63 0.2008
IIIA vs. IIB 1.35 <0.0001 2.14 0.0026 1.20 0.0169 1.91 0.0006
IIIB vs. IIIA 1.20 <0.0001 1.26 0.0102 1.24 0.0401 1.63 0.0529
IIIC vs. IIIB 1.10 <0.0001 1.13 0.1235 1.45 0.2182 a a
IV vs. IIIC 1.76 <0.0001 1.59 <0.0001 1.36 0.2866 a a
R 2 15.9% 18.9% 35.1% 29.4%

a Pathologic stage IV cases were poorly represented in the IASLC 1999–2010 database.

Note: Results are from a multiple-variable Cox model.
IASLC, International Association for the Study of Lung Cancer; HR, hazard ratio; NCDB, National Cancer Database; R 2 , a statistical measure of discrimination reflecting the percent of overall variance explained by the factors in a model.

Table 4Overall Survival Comparisons by Eighth Edition Clinical and Pathologic Stage Groups in the NCDB and IASLC Data Sets for SCLC Adjusted for Age (>70 y) and Sex

 

The T4 group is distinct from the other T categories in both clinically and pathologically staged SCLC ( Supplementary Fig. 5 and IASLC data not shown), which may partially account for the distinction between stages IIB and IIIA, at least in node-negative cases. The clinical N2 and N3 groups have a similar prognosis that is distinct from that of the N0 and N1 groups (see Supplementary Fig. 5 ). For SCLC, the NCDB best stage results mainly reflect clinical stage analyses, which accounts for most of the cases; therefore, survival by best stage is not shown.

Overall, in SCLC, early-stage cases in the IASLC data set show higher survival rates than the NCDB, which is consistent with what is seen in NSCLC. The same can be said for both clinically and pathologically staged cases, so that the proportion of surgical cases in the IASLC data set cannot account entirely for the difference. For the clinically staged III and IV cases, however, in which resection attempts would be very uncommon, the two data sets demonstrate very similar survival in the respective stage categories.

Discussion

The AJCC/UICC stage classification of lung cancer is relatively unique among stage classifications of cancer sites; driven by the initiative of the IASLC, it is based on a large international database and an extensive, sophisticated analysis with multiple levels of internal validation. 6 7 This is true for both the seventh and eighth edition TNM classifications, in contrast to classifications of many other disease sites and prior editions, in which the classification rests largely on empiric categories and consensus. Furthermore, the eighth edition lung cancer stage classification has undergone specific analyses, as should ideally be done for a classification system, 11 demonstrating multiple levels of generalizability (i.e., discrimination is maintained across time intervals, geographic regions, patients identified by different methods, a broad spectrum of disease, and follow-up duration). 7

Nevertheless, external validation in a sufficiently large independent cohort by using appropriate analyses of discriminatory ability is important. The primary purpose of the TNM classification is to provide a nomenclature for the anatomic extent of disease in a way that discriminates distinct patient groups. The actual outcomes of groups (calibration [e.g., 5-year survival]) in a data set may differ, reflecting the influence of the many factors besides disease extent that affect prognosis, such as the differences in the health care system, time period, treatment received, patient characteristics of the cohort (age, comorbidities), and so forth, that characterize the particular data set. 7 Neither the IASLC outcomes nor those of the NCDB should be viewed as an absolute benchmark for any individual group or region. For example, there is a marked improvement (∼25%–50%) in survival between the IASLC 1990–1999 and the 1999–2010 data sets used for the seventh and eighth edition analyses. 7 However, the discriminatory ability of the classification system must be maintained in different data sets if a classification system is to be valid.

The registry of the Surveillance, Epidemiology, and End Results program of the U.S. National Cancer Institute was considered for this validation exercise. In fact, this database was used previously for external validation of the seventh edition revision. However, because of the structure of the Extent of Disease codes through 2004 and only partial implementation of updated collaborative stage codes before 2010, there were too few cases that could be accurately reclassified according to the eighth edition. Therefore, we did not use the Surveillance, Epidemiology, and End Results registry.

The NCDB represents an excellent validation data set because of both its size and level of detail regarding the tumor characteristics. Use of this database allowed for detailed assessment of the T and N components and the stage groupings of the eighth edition TNM classification. However, there was insufficient information regarding number of individual metastatic lesions to assess the newly developed M component. As shown in this article, the ordering of T and N categories and stage groups is maintained in the NCDB with good discrimination (i.e., R 2 ). This is true across multiple subgroups (clinical and pathologic stage, best stage, NSCLC, and SCLC).

The NCDB database is particularly useful for validation because it is strong in areas in which the IASLC database is weak. There is a relative underrepresentation of North America and nonsurgically managed patients in the IASLC database. The NCDB on the other hand offers a completely North American cohort, as well as a representative majority of nonsurgically managed cases. The treatment sequences received by the nonsurgical patients are well described in the NCDB, with a higher level of detail than was available in most of the IASLC cases. The IASLC database includes several large data sources with patients receiving therapies that may be essentially the same within the data source but different from those of other large participating data submitters, which is an obvious cause for concern. The NCDB analysis strongly confirms the IASLC internal analysis, demonstrating that the eighth edition stage classification works well both in North America and in a primarily nonsurgical cohort. Concern regarding bias is ameliorated. In other words, the data show that the anatomic extent of disease as captured by the TNM categories does discriminate among patient groups even in a primarily nonsurgical setting. (This does not diminish the fact that other factors can also provide useful information to distinguish patient cohorts.)

As would be expected in a primarily nonsurgical data set, the NCDB outcomes for NSCLC are markedly worse relative to those of the primarily surgical IASLC data set (see Fig. 3 ), yet the discrimination is maintained (see Fig. 3 and Table 2 ). This is harder to assess in the SCLC data set. In general, an ordered progression of worsening survival in the NCDB data set is maintained, particularly within the broader stage I through IV categories. However, there is generally no statistical difference between groups within stage I and II, despite the larger patient numbers. Nevertheless, for the pathologically staged SCLC subset, better case representation in the NCDB data enabled the confirmation of appropriate ordering and good separation between stage groupings for pathologic stage in SCLC, which had not been previously accomplished.

The NCDB is also useful in providing information on how the patients within each stage group were treated during the period 2000–2012 (see Fig. 1 ). This allows the clinician to better interpret the survival outcomes seen in the NCDB analysis, and to make a better assessment of the degree to which the NCDB outcomes may estimate the outcome for a particular patient with a tumor of a certain stage group who will be treated in a certain way. However, it must be recognized that in this analysis the treatment groups are not fully represented, given the exclusion of all yp-staged cases that received neoadjuvant therapy. The basis for tumor size and extent was strictly pathologic in all surgical cases, which means that the yp-staged cases were also excluded from analyses of clinical stage and best stage.

Despite the size and detailed nature of the NCDB database, there are some limitations. First, the NCDB data did not permit validation of the M1b category, given that single-site metastasis in NCDB could involve either a single metastasis or multiple metastases. An analysis of single versus multiple sites in the NCDB could be undertaken as an approximation of the eighth edition M1b versus M1c categories, but this would not reflect true validation. Second, the ability to identify all cases within NCDB with diaphragmatic invasion was incomplete. Third, in the pathologically staged SCLC subset, stage IIA and stage IIIC were poorly represented. Sample size issues were also seen and previously reported in the IASLC SCLC data. 3 Fourth, as previously mentioned, and in contrast to the IASLC data sets, surgical cases could not be included in analyses of clinical stage because the basis for tumor descriptors was pathologic in surgical cases. Finally, the source of NCDB cases, registries at Commission on Cancer-accredited hospitals, may not be representative of the typical therapy received at other hospitals in the United States or in other countries. The more global geographic coverage and increasing inclusiveness of the IASLC effort is invaluable in this respect, and it is another factor in the complementary nature of these two databases.

This validation was carried out in accordance with proposed characteristics for a robust external validation. 7 Regarding the potential for case overlap, the NCDB cohort included less than 5% patients who were also included in the IASLC data set. We can be sure of this because of the small percentage (<5%) of cases in the IASLC database that were contributed from North American sites. We have reported follow-up information as well as the number of cases and events per subgroup analyzed. We have also reported on the ordering, statistical differences, and homogeneity within subgroups and examined and found the discrimination to be consistent within analyses of specific cohorts (i.e., NSCLC, SCLC, clinical, pathologic, and best stage). Multivariate analyses were carried out to adjust for various confounding factors.

This validation of the eighth edition stage classification brings the effort underlying the NCDB full circle. The NCDB was established to collect detailed data regarding patients with cancer to enhance the ability to understand the impact of many factors and changes over time. To accomplish this, an extensive infrastructure was developed to collect stage information, including trained tumor registrars, institutional certification, and extensive quality control. As with any major infrastructure, there is always tension between the cost of maintenance and appreciation of the value. However, the value of the NCDB cannot be overstated. There are very few data that permit such a robust validation of the eighth edition stage classifications—and this is a major testament to the value of the NCDB and the infrastructure and people it represents.

Conclusion

The NCDB provides an extensive resource for a robust external validation of the eighth edition UICC and AJCC stage classification for lung cancer. This analysis demonstrates that the discrimination of the eighth edition categories and stage groups validates very well for NSCLC and SCLC in clinical, pathologic, and best stage analyses. This validation also further underscores the geographic transportability of the eighth edition stage classification and the applicability to primarily nonsurgically treated cohorts.

Acknowledgments

This work was supported by the International Association for the Study of Lung Cancer .

Appendix

IASLC Staging and Prognostic Factors Committee

Peter Goldstraw, past chair, Royal Brompton Hospital and Imperial College, London, United Kingdom; Ramón Rami-Porta, chair, Hospital Universitari Mutua Terrassa, Terrassa, Spain; Hisao Asamura, chair elect, Keio University, Tokyo, Japan; David Ball, Peter MacCallum Cancer Centre, Melbourne, Australia; David G. Beer, University of Michigan, Ann Arbor, Michigan; Ricardo Beyruti, University of Sao Paulo, Brazil; Vanessa Bolejack, Cancer Research And Biostatistics, Seattle, Washington; Kari Chansky, Cancer Research And Biostatistics, Seattle, Washington; John Crowley, Cancer Research And Biostatistics, Seattle, Washington; Frank Detterbeck, Yale University, New Haven, Connecticut; Wilfried Ernst Erich Eberhardt, West German Cancer Centre, University Hospital, Ruhrlandklinik, University Duisburg-Essen, Essen, Germany; John Edwards, Northern General Hospital, Sheffield, United Kingdom; Françoise Galateau-Sallé, Centre Hospitalier Universitaire, Caen, France; Dorothy Giroux, Cancer Research And Biostatistics, Seattle, Washington; Fergus Gleeson, Churchill Hospital, Oxford, United Kingdom; Patti Groome, Queen’s Cancer Research Institute, Kingston, Ontario, Canada; James Huang, Memorial Sloan-Kettering Cancer Center, New York, New York; Catherine Kennedy, University of Sydney, Sydney, Australia; Jhingook Kim, Samsung Medical Center, Seoul, Republic of Korea; Young Tae Kim, Seoul National University, Seoul, Republic of Korea; Laura Kingsbury, Cancer Research And Biostatistics, Seattle, Washington; Haruhiko Kondo, Kyorin University Hospital, Tokyo, Japan; Mark Krasnik, Gentofte Hospital, Copenhagen, Denmark; Kaoru Kubota, Nippon Medical School Hospital, Tokyo, Japan; Antoon Lerut, University Hospitals, Leuven, Belgium; Gustavo Lyons, British Hospital, Buenos Aires, Argentina; Mirella Marino, Regina Elena National Cancer Institute, Rome, Italy; Edith M. Marom, M. D. Anderson Cancer Center, Houston, Texas; Jan van Meerbeeck, Antwerp University Hospital, Edegem (Antwerp), Belgium; Alan Mitchell, Cancer Research And Biostatistics, Seattle, Washington; Takashi Nakano, Hyogo College of Medicine, Hyogo, Japan; Andrew G. Nicholson, Royal Brompton and Harefield National Health Service Foundation Trust and Imperial College, London, United Kingdom; Anna Nowak, University of Western Australia, Perth, Australia; Michael Peake, Glenfield Hospital, Leicester, United Kingdom; Thomas Rice, Cleveland Clinic, Cleveland, Ohio; Kenneth Rosenzweig, Mount Sinai Hospital, New York, New York; Enrico Ruffini, University of Torino, Torino, Italy; Valerie Rusch, Memorial Sloan-Kettering Cancer Center, New York, New York; Nagahiro Saijo, National Cancer Center Hospital East, Chiba, Japan; Paul Van Schil, Antwerp University Hospital, Edegem (Antwerp), Belgium; Jean-Paul Sculier, Institut Jules Bordet, Brussels, Belgium; Lynn Shemanski, Cancer Research And Biostatistics, Seattle, Washington; Kelly Stratton, Cancer Research And Biostatistics, Seattle, Washington; Kenji Suzuki, Juntendo University, Tokyo, Japan; Yuji Tachimori, National Cancer Center, Tokyo, Japan; Charles F. Thomas Jr, Mayo Clinic, Rochester, Minnesota; William Travis, Memorial Sloan-Kettering Cancer Center, New York, New York; Ming S. Tsao, The Princess Margaret Cancer Centre, Toronto, Ontario, Canada; Andrew Turrisi, Sinai Grace Hospital, Detroit, Michigan; Johan Vansteenkiste, University Hospitals, Leuven, Belgium; Hirokazu Watanabe, National Cancer Center Hospital, Tokyo, Japan; Yi-Long Wu, Guangdong Provincial People's Hospital, Guangzhou, People’s Republic of China.

Advisory Board of the IASLC Mesothelioma Domain

Paul Baas, The Netherlands Cancer Institute, Amsterdam, The Netherlands; Jeremy Erasmus, M. D. Anderson Cancer Center, Houston, Texas; Seiki Hasegawa, Hyogo College of Medicine, Hyogo, Japan; Kouki Inai, Hiroshima University Postgraduate School, Hiroshima, Japan; Kemp Kernstine, City of Hope, Duarte, California; Hedy Kindler, The University of Chicago Medical Center, Chicago, Illinois; Lee Krug, Memorial Sloan-Kettering Cancer Center, New York, New York; Kristiaan Nackaerts, University Hospitals, Leuven, Belgium; Harvey Pass, New York University, New York, New York; David Rice, M. D. Anderson Cancer Center, Houston, Texas.

Advisory Board of the IASLC Thymic Malignancies Domain

Conrad Falkson, Queen’s University, Ontario, Canada; Pier Luigi Filosso, University of Torino, Italy; Giuseppe Giaccone, Georgetown University, Washington, District of Columbia; Kazuya Kondo, University of Tokushima, Tokushima, Japan; Marco Lucchi, University of Pisa, Pisa, Italy; Meinoshin Okumura, Osaka University, Osaka, Japan.

Advisory Board of the IASLC Esophageal Cancer Domain

Eugene Blackstone, Cleveland Clinic, Cleveland, Ohio.

Participating Institutions in the New IASLC Lung Cancer Staging Project

F. Abad Cavaco and E. Ansótegui Barrera, Hospital La Fe, Valencia, Spain; J. Abal Arca and I. Parente Lamelas, Complejo Hospitalario de Ourense, Ourense, Spain; A. Arnau Obrer and R. Guijarro Jorge, Hospital General Universitario de Valencia, Valencia, Spain; D. Ball, Peter MacCallum Cancer Centre, Melbourne, Australia; G. K. Bascom, Good Samaritan Hospital, Kearney, Nebraska; A. I. Blanco Orozco and M. A. González Castro, Hospital Virgen del Rocío, Sevilla, Spain; M. G. Blum, Penrose Cancer Center, Colorado Springs, Colorado; D. Chimondeguy, Hospital Universitario Austral, Argentina; V. Cvijanovic, Military Medical Academy, Belgrade, Serbia; S. Defranchi, Hospital Universitario-Fundacion Favaloro, Buenos Aires, Argentina; B. de Olaiz Navarro, Hospital de Getafe, Getafe, Spain; I. Escobar Campuzano and I. Macía Vidueira, Hospital de Bellvitge, L’Hospitalet de Llobregat, Spain; E. Fernández Araujo and F. Andreo García, Hospital Universitari Germans Trias i Pujol, Badalona, Spain; K. M. Fong, Prince Charles Hospital, Brisbane, Australia; G. Francisco Corral and S. Cerezo González, Hospital La Mancha Centro, Ciudad Real, Spain; J. Freixinet Gilart, Hospital Universitario Dr. Negrín, Las Palmas de Gran Canaria, Spain; L. García Arangüena, Hospital Sierrallana, Torrelavega, Spain; S. García Barajas, Hospital Infanta Cristina, Badajoz, Spain; P. Girard, L'Institut Mutualiste Montsouris, Paris, France; T. Goksel, Turkish Thoracic Society, Turkey; M. T. González Budiño, Hospital General Universitario de Oviedo, Oviedo, Spain; G. González Casaurrán, Hospital Gregorio Marañón, Madrid, Spain; J. A. Gullón Blanco, Hospital San Agustín, Avilés, Spain; J. Hernández Hernández, Hospital de Ávila, Avila, Spain; H. Hernández Rodríguez, Hospital Universitario de Tenerife, Santa Cruz de Tenerife, Spain; J. Herrero Collantes, Hospital Universitario Nuestra Señora de la Candelaria, Santa Cruz de Tenerife, Spain; M. Iglesias Heras, Hospital de Ávila, Ávila, Spain; J. M. Izquierdo Elena, Hospital Nuestra Señora de Aránzazu, Donostia, Spain; E. Jakobsen, Danish Lung Cancer Registry, Denmark; S. Kostas, Athens School of Medicine, Athens, Greece; P. León Atance and A. Núñez Ares, Complejo Hospitalario de Albacete, Albacete, Spain; M. Liao, Shanghai Lung Tumor Clinical Medical Center, Shanghai, People's Republic of China; M. Losanovscky, Clinica y Maternidad Suizo Argentina, Buenos Aires, Argentina; G. Lyons, Hospital Britanico de Buenos Aires, Buenos Aires, Argentina; R. Magaroles and L. De Esteban Júlvez, Hospital Joan XXIII, Tarragona, Spain; M. Mariñán Gorospe, Hospital de San Pedro de Logroño, Logroño, Spain; B. McCaughan and C. Kennedy, University of Sydney, Sydney, Australia; R. Melchor Íñiguez, Fundación Jiménez Díaz, Madrid, Spain; L. Miravet Sorribes, Hospital La Plana, Castellón, Spain; S. Naranjo Gozalo and C. Álvarez de Arriba, Hospital Universitario Marqués de Valdecilla, Santander, Spain; M. Núñez Delgado, Hospital de Meixoeiro, Vigo, Spain; J. Padilla Alarcón and J. C. Peñalver Cuesta, Instituto Valenciano de Oncología, Valencia, Spain; J. S. Park, Samsung Medical Center, Seoul, Republic of Korea; H. Pass, New York University Langone Medical Center and Cancer Center, New York, New York; M. J. Pavón Fernández, Hospital Severo Ochoa, Leganés, Spain; M. Rosenberg, Alexander Fleming Institute and Hospital de Rehabilitación Respiratoria, Buenos Aires, Argentina; E. Ruffini, University of Torino, Torino, Italy; V. Rusch, Memorial Sloan-Kettering Cancer Center, New York, New York; J. Sánchez de Cos Escuín, Hospital de Cáceres, Cáceres, Spain; A. Saura Vinuesa, Hospital de Sagunto, Sagunto, Spain; M. Serra Mitjans, Hospital Universitari Mutua Terrassa, Terrassa, Spain; T.E. Strand, Cancer Registry of Norway, Norway; D. Subotic, Clinical Centre of Serbia, Belgrade, Serbia; S. Swisher, M. D. Anderson Cancer Center, Houston, Texas; R. Terra, University of Sao Paulo Medical Center, Sao Paulo, Brazil; C. Thomas, Mayo Clinic Rochester, Rochester, Minnesota; K. Tournoy, University Hospital Ghent, Ghent, Belgium; P. Van Schil, Antwerp University Hospital, Edegem (Antwerp), Belgium; M. Velasquez, Fundacion Clinica Valle del Lili, Cali, Colombia; Y. L. Wu, Guangdong General Hospital, Guangzhou, People's Republic of China; K. Yokoi, Japanese Joint Committee for Lung Cancer Registry, Osaka, Japan.

Supplementary Data

Figure S1

Inclusion Diagram

Diagram of cases includeed and excluded in the National Cancer Database dataset. A) non-small cell lung cancer B) small cell lung cancer.

Figure S2

Histologic Types among Non-Small Cell Lung Cancer

Histologic type as a percentage of all analyzed non-small cell lung cancer cases in the NCDB and IASLC datasets.

BAC, bronchoalveolar carcinoma; IASLC, International Association for the Study of Lung Cancer Database; NCDB, National Cancer Database; NEC, neuroendocrine carcinoma (Carcinoid and larger cell neuroendocrine carcinoma); NOS, non-small cell lung cancer, not otherwide specified.

Figure S3

T Categories, Non-Small Cell Lung Cancer

Overall survival in patients with non-small cell lung cancers (N-any M0) according to the 8 th edition T categories in the NCDB and IASLC datasets. A) Clinically staged tumors B) pathologically staged tumors

IASLC, International Association for the Study of Lung Cancer Database; NCDB, National Cancer Database.

Figure S4

N Categories, Non-Small Cell Lung Cancer

Overall survival in patients with non-small cell lung cancers (T-any M0) according to the 8 th edition N categories in the NCDB and IASLC datasets. A) Clinically staged tumors B) pathologically staged tumors

IASLC, International Association for the Study of Lung Cancer Database; NCDB, National Cancer Database.

Figure S5

T and N Categories, Small Cell Lung Cancer

Overall survival in patients with small cell lung cancers (N-any M0) according to the 8 th edition T and N categories in the NCDB dataset. A) Clinically staged tumors B) pathologically staged tumors

NCDB, National Cancer Database.

 

References

  • 1H. Asamura, K. Chansky, J. Crowley. The IASLC Lung Cancer Staging project: proposals for the revision of the N descriptors in the forthcoming eighth edition of the TNM classification for lung cancer. J Thorac Oncol. 2015;10 :1675-1684
  • 2W. Eberhardt, J. Mitchell, J. Crowley. The IASLC Lung Cancer Staging Project: proposals for the revision of the M descriptors in the forthcoming (8th) edition of the TNM classification of lung cancer. J Thorac Oncol. 2015;10 :1515-1522
  • 3A. Nicholson, K. Chansky, J. Crowley. The IASLC Lung Cancer Staging Project: proposals for the revision of the clinical and pathologic staging of small cell lung cancer in the forthcoming eighth edition of the TMN classification for lung cancer. J Thorac Oncol. 2016;11 :300-311
  • 4R. Rami-Porta, V. Bolejack, J. Crowley. The IASLC Lung Cancer Staging Project: proposals for the revisions of the T descriptors in the forthcoming eighth edition of the TNM classification for lung cancer. J Thorac Oncol. 2015;10 :990-1003
  • 5P. Goldstraw, K. Chansky, J. Crowley. The IASLC Lung Cancer Staging Project: proposals for the revision of the TNM stage groupings in the forthcoming (eighth) edition of the TNM classification for lung cancer. J Thorac Oncol. 2015;11 :39-51
  • 6R. Rami-Porta, V. Bolejack, D.J. Giroux. The IASLC Lung Cancer Staging Project: the new database to inform the eighth edition of the TNM classification of lung cancer. J Thorac Oncol. 2014;9 :1618-1624
  • 7F. Detterbeck, K. Chansky, P. Groome. The IASLC Lung Cancer Staging Project: methodology and validation used in the development of proposals for revision of the stage classification of non-small cell lung cancer in the forthcoming (8th) edition of the TNM classification of lung cancer. J Thorac Oncol. 2016;11 :1433-1446
  • 8American College of Surgeons. Participant user files., . Accessed May 12, 2017.https://www.facs.org/quality-programs/cancer/ncdb/puf
  • 9S.R. Cole, M.A. Hernán. Adjusted survival curves with inverse probability weights. Comput Methods Program Biomed. 2004;75 :45-49
  • 10J. Crowley, D. Ankerst. Handbook of Statistics in Clinical Oncology. 2nd ed. 2006 Chapman and Hall/CRC Press: New York, NYJ. O’Quigley, R. Xu. Explained variation in proportional hazards regression. 347-363
  • 11A.C. Justice, K.E. Covinsky, J.A. Berlin. Assessing the generalizability of prognostic information. Ann Intern Med. 1999;130 :515-524