Assessing a modified-AJCC TNM staging system in the New South Wales Cancer Registry, Australia

Under the NSW Public Health Act 2010, there is a mandatory requirement to report notifiable cancer cases to NSWCR by facilities that diagnose, manage, or treat cancer patients – these include public and private hospitals, public and private pathology laboratories, private and public day procedure centres, cancer treatment facilities, and residential aged care facilities [12]. Death data are notified from the NSW Registry of Births, Deaths and Marriages. Coded death data are supplied by the Australian Bureau of Statistics.

Notification data which pertains to a unique cancer type for the patient are consolidated to an incident cancer case through a process that entails computer-embedded business rules (implemented in the registry system) and manual coding.

RD-stage consists of numerical stages I-IV. The aim of the RD-staging project was to provide population-based data for five high incidence tumour groups (prostate, colorectal, breast, lung, and melanoma) diagnosed in 2011. The NSWCR only staged eligible invasive cancers with a morphology code ending in /3, in-situ tumours were excluded (see Additional file 1: Table S2 for eligible tumour morphology and topography codes). Cases deemed ineligible included: (i) sarcomas and lymphomas of the breast, colon, rectum, lung, and prostate; (ii) carcinoid tumours of the colon and rectum; and (iii) transitional cell carcinomas of the prostate. Business rules to derive RD-stage using T, N, and M values were provided to the NSWCR by VicCR. For eligible cases where T, N, and M values were not already recorded, values were assigned using all routinely available notification sources in NSWCR within 120 days of diagnosis – these included scanned pathology reports, electronic hospital notifications (using DoS data and ICD10 metastatic codes), and clinical information (if available). Values were assigned by registry coders supervised by a pathologist using the AJCC 7th Edition [4]. For prostate cases, business rules also incorporated PSA and Gleason scores. Where RD-stage could not be derived due to incomplete information provided by T, N, and M, RD-stage was recorded as stage missing or stage not-applicable.

Fig. 1 provides further information for how stage data was obtained.

To determine for which of the five cancers TNM stage data (both AJCC-SG and RD-stage) provided more complete staging than the current conventional DoS summary stage, stage data were extracted from NSWCR (pre- and post- RD-staging project) and compared across (i) tumour groups, (ii) stage systems, and (iii) time periods.

We compared alignments between surrogate stage variables (DoS/RD-stage) with AJCC-SG. Expected mappings were developed from inspecting (i) RD-staging business rules provided by VicCR, (ii) NSWCR business rules for deriving AJCC-SG, (iii) NSWCR business rules for consolidating notification data for DoS and, (iv) IARC documentation. To confirm our expected mappings, we compared expected and observed cross-tabulated frequency distributions using a subset of NSWCR data. All staging information available for eligible RD-staging cases were extracted for analysis. For each tumour group, overall agreement measures (concordance and kappa) were calculated. Sensitivity and specificity were calculated for each stage grouping, similar to Kwan and colleagues [14]. To ensure staging data were comparable, all DoS 7 (invasion of adjacent organs and regional lymph nodes involved) were re-assigned to DoS 3 (regional lymph nodes) (see Additional file 1: Table S1). Pathologically and clinically staged AJCC-SG were consolidated (prioritising pathologically staged values where possible), and AJCC-SGs were collapsed from alpha-numerical to numerical classifications (stage I, II, III or IV).

In epidemiological analyses, stage data are most commonly used as a prognostic factor to estimate outcomes, particularly mortality and survival. We compared survival models developed from each of the three stage variables (RD-stage, AJCC-SG, or DoS). For lung and colorectal cancer cases we fitted Cox proportional hazards regression models to compare the relative contribution of RD-stage, AJCC-SG, or DoS as an explanatory variable to survival time. Patient demographic information (age at diagnosis and sex) are readily available in the NSWCR, and were also considered as explanatory variables. Mortality follow-up data, including cancer-specific causes of death, were available within the NSWCR up until the end of 2014. Prostate, breast and melanoma cancers generally have high 5-year survival rates (ranging from 90.6 to 95% based on Australian cancer data, 2010–2014) compared to lung and colorectal cancers (lung cancer 5-year survival in men and women in 2010–2014 are 14.5 and 19.6% respectively, whereas colorectal survival rates in men and women are 69.0 and 70.0%) [15]. Only lung and colorectal cancer were chosen for the 4-year survival analyses as we expected higher numbers of deaths to occur in these patients compared to prostate, breast and melanoma. We checked the proportional hazards assumption using Schoenfield residuals, and transformed variables as appropriate. Stage and age variables were found to violate the proportional hazards assumption. Stage variables were stratified. Transformation of age at diagnosis to a categorical variable (using age group categories described in Benitez-Majano and colleagues [3]) minimised violation of the proportional hazards assumption.

We additionally fitted logistic regression models to compare the relative contribution of the association between RD-stage, AJCC-SG, or DoS and all-cause mortality after a 1-year period from date of diagnosis. We examined Akaike information criterion (AIC) and Akaike weights to compare across each set of models. All calculations and visualisations were performed in R statistical software version 3.4.1.

Prior to undertaking the RD-staging project (June 2017), AJCC-SG data were available for less than half of eligible cases (see Table 1). In June 2018, as a result of the staging project, AJCC-SG and RD-stage were available for over 90% of eligible cases. RD-staging improved completeness for TNM derivations (AJCC-SG/RD-stage) across all stage groups, with melanoma showing the greatest change in completeness followed by prostate, breast, colorectal, and lung cancer. Prostate cancer had the lowest staging completeness for DoS pre- and post- RD-staging, followed by lung cancer.

Mappings were developed across all three staging systems based on examination of relevant stage system documentation (details provided in Table 2 and Additional file 1: Tables S3, S4, S5, S6 and S7). Mappings (see Fig. 2) showed RD-stage to AJCC-SG alignments were mostly linear (RD-stage I = AJCC-SG I, RD-stage II = AJCC-SG II, etc.), whereas comparability of DoS and AJCC-SG was less well-defined with the exception of the colorectal tumour group. These findings were also reflected in our analysis of NSWCR case data where higher agreement scores were more evident in comparisons of AJCC-SG/RD-stage compared to AJCC-SG/DoS (see Fig. 3 and Additional file 1: Table S8).

Compared to DoS, RD-staging provided greatest improvements in accuracy (in terms of alignment to AJCC-SG) for prostate and breast cancers followed by lung cancer and melanoma. For colorectal cancer, we found all three classification systems showed near-linear alignment with scores of > 80% for all accuracy scores.

From examining AIC values (see Additional file 1: Tables S9 and S10), models using AJCC-SG consistently showed the highest fit and models using DoS consistently showed the poorest fit. Akaike weights indicated models using AJCC-SG best explained survival outcomes. In 4-year multivariable Cox proportional hazards models (see Additional file 1: Table S10), across all model sets, we observed highly similar hazard ratios (HRs) among RD-stage, AJCC-SG, and DoS models, suggesting that DoS and RD-stage are adequate alternatives to AJCC-SG for estimating survival and mortality outcomes for colorectal and lung cancers. From examining an additional set of 4-year multivariable Cox proportional hazards models where stage groups were not stratified, we observed models using DoS underestimated HRs for higher stage groups. Similar trends were found in 1-year all-cause mortality logistic regression models (see Additional file 1: Table S9) and univariable 4-year Cox proportional hazards models (see Additional file 1: Table S11).

Overall, we found RD-staging (compared to DoS) provided greater stage data completeness and accuracy (alignment to AJCC-TNM) for prostate cancer cases. RD-staging provided stage data for 98% of prostate cases compared to only 72% for DoS. Previous NSWCR studies have similarly shown low DoS stage data completeness for prostate cancer [16, 17]. RD-stage data was also much more aligned to AJCC-TNM with concordance/kappa scores of 80%/64%, compared to only 68%/35% for DoS. Based on clear improvements in stage data completeness and accuracy compared to DoS, prostate cancer would be a clear candidate for RD-staging in the NSWCR.

It is important to note the caveats that apply to both RD-staging and NSWCR’s AJCC staging systems. It is expected that the NSWCR will have a higher number of AJCC-SG I prostate cases that are actually AJCC-SG II given PSA and Gleason score are not factored into the business rule algorithm to calculate AJCC-SG. This was also reflected in the RD-stage where these non-anatomic variable were unavailable – in our study sample of 7223 prostate cases we found only 31% (N = 2210) had a valid Gleason score and only 23% (N = 1626) had a valid PSA score. Overall, this means that both AJCC-SG and RD-stage will potentially underestimate the incidence of stage group II prostate cancers as both NSWCR and VICCR algorithms simplify them to stage group I. With these points considered, within the NSWCR, RD-staging (rather than AJCC-SG) will provide stage group classifications that would more closely align to current (7th and 8th) AJCC editions for prostate cancer.

For colorectal cases, we found summary stage from DoS an adequate surrogate staging system: compared to DoS, RD-staging did not improve stage data completeness, and RD-stage only provided a small amount of improvement in accuracy (with concordance/kappa scores of 99%/99 and 88%/83% for RD-stage and DoS respectively). Both DoS and RD-stage 4-year multivariable Cox proportional hazards survival models showed highly similar hazard ratios (HR) to more clinically relevant AJCC-SG models when the stage variable was stratified, which suggest both stage variables are suitable alternatives to AJCC-SG for survival modelling. However, we observed DoS consistently underestimated odds ratios (ORs) and HRs in 1-year all-cause mortality logistic regression models, univariable 4-year Cox models, and multivariable 4-year Cox models.

For melanoma, we found concordance/kappa scores were higher for RD-stage (97%/91%) compared to DoS (83%/44%). However, compared to DoS, RD-staging provided only minimal improvements in terms of stage data completeness (DoS was available for 98% of melanoma cases compared to 99% for RD-stage).

For breast cancer, RD-staging showed near-perfect alignment to more clinically relevant AJCC-SG (with very high concordance/kappa scores of 100%/100% for RD-stage and fairly low scores of 56%/38% for DoS). However, compared to DoS, RD-staging provided minimal improvements in terms of stage data completeness (DoS was available for 95% of breast cancer cases compared to 97% for RD-stage).

We found RD-staging provided small improvements in stage data completeness (DoS was available for 85% of lung cancer cases compared to 87% for RD-stage). RD-staging however showed improvements in alignment to AJCC-SG (concordance/kappa scores increased moderately from 86%/74% for DoS to 96%/93% for RD-stage). As seen with colorectal cancer, lung cancer multivariable 4-year Cox proportional hazards survival models showed similar HRs among RD-stage, AJCC-SG, and DoS models, however this was only seen when the stage variable was stratified.

Both RD-stage and AJCC-SG data were derived from T, N and M values which were sourced from manual review of pathology reports by NSWCR coders (as part of the RD-staging project), and/or manual review of hospital in-patient notification and other clinical information sources by Cancer Information Managers (CIMs) (as part of routine NSWCR data collection). The RD-staging project, conducted in 2017, involved manual collection of T, N and M values from pathology reports of melanoma, breast, prostate, colorectal and lung cancer cases diagnosed in 2011. This exercise not only provided the RD-stage data, but also resulted in a substantial increase in AJCC-SG data coverage. While a formal comparison of procedure and workload for RD-staging and AJCC-TNM staging cannot be performed, we can provide comments around (i) routine data collections and (ii) data collections performed specifically for the RD-staging project.

Routine TNM data collections in the NSWCR are performed by CIMs and involve transcribing data from reports held in either data extracts from cancer treatment centres, or reports held at point of care in the NSWCR. Complete population coverage is not possible as CIMs generally collect data from public (as opposed to private) hospitals and treatment centres. When there are data inconsistencies or when data is missing, CIMs review clinical documents from cancer treatment centres and all inpatient hospital notifications sourced from hospitals. This can take years to get through full review due to the volume of inpatient notifications generated. The proportion of missing data is variable, however generally data completeness is poor across the board. It is also worth noting that even when the CIMs manually review 100% of patients in a period, recovering and providing TNM values for 100% of those patients is not possible primarily due to data governance (e.g. private consult notes cannot be provided and public treatment referral letters miss key information), and also due to TNM not being essential to some treatment decisions in some treatment modalities and/or protocols.

Collection of DoS is conducted routinely within the NSWCR and is part of coding a cancer case. DoS collection adheres to published IARC categories [13] and is comparatively straightforward for the tumours staged in this study. Generally, there is higher stage data completeness for DoS compared to collection of T, N and M data.

The RD-staging project involved extensive training of NSWCR coders to recognise and assign T, N and M based on review of available pathology reports in the NSWCR. Where T, N and M values were not able to be transcribed – information in reports were reviewed and interpreted by coders to assign T, N and M. We estimated NSWCR coders completed manual TNM staging of 16,007 cases within 61 working days. The time spent on the RD-staging project however impacted on routine coding procedures – for other PBCRs where additional resourcing is not available, collection of stage data may not be worthwhile.

Stage is currently not considered an essential variable for reporting by the International Association of Cancer Registries (IACR). However, with growth in capacity for PBCRs to store and manage clinical data, collection of stage data is becoming more feasible [4]. Furthermore, there is increasing interest in measuring global cancer survival outcomes [2, 18, 19]. Growing interest in using stage data for clinically-oriented population studies has also created a need for collection of TNM stage data.

While AJCC-TNM data are not mandatory for collection, a 2013 comparative analysis of international PBCRs found AJCC-TNM stage data were collected from PBCRs in 10 of 12 jurisdictions [2]. In England, increasing stage data completeness has been a national priority in recent years and resources have been specifically allocated to improve data collection processes [20]. In the United States, AJCC-TNM stage data have been collected since 2004 under a national Collaborative Stage Data Collection System which has recently been expanded to incorporate information on related biomarkers and prognostic factors [4]. Other PBCRs have conducted and published evaluations of completeness and accuracy of AJCC-TNM stage data within their respective registries [21‐23]. Australian PBCRs are considered well-resourced, high quality PBCRs [2, 24] and accordingly, should aim to meet high standards in cancer reporting, including provision of complete and accurate AJCC-TNM stage data.

A limitation of RD-staging is that other countries are not familiar with RD-stage and have no access to TNM information necessary for RD stage. In 2018 the Union for International Cancer Control (UICC) released Essential TNM a process for collecting stage data in PBCRs in low and middle income countries where there are insufficient resources to derive complete TNM data. [25] Essential TNM is aligned with the UICC staging system, not AJCC –differences between the two systems have previously been documented [26]. While a comprehensive formal mapping of Essential TNM to AJCC TNM and DoS was outside the scope of this study, we provide some brief comments based on the Essential TNM User Guide. [27] Generally, Essential TNM aligns more closely with DoS: DoS 1 (and DoS 6) would equate to L1/L2, DoS 2 would equate to A1/A2, DoS 3 and 7 map to R+, and DoS 4 map to M+. Examining the staging of prostate cancer in more detail – Essential TNM, like DoS, defaults N+ tumours to Stage III, which we found to map across AJCC SGs III and IV. T4 N0 M0 also maps to AJCC SG IV but aligns with DoS 2 and would align with Essential TNM TA (locally advanced). Given the simplification of T staging and the assumption of Stage III disease for node-positive prostate cancer, DoS and Essential TNM are likely to align in under-staging AJCC-SG IV cancers as well as resulting in a higher number of unknown stage cases for biopsy-only cases. It would be reasonable to consider DoS as a staging system for PBCRs in low and middle income countries given there is documentation available for most tumour groups (not just breast, cervix, colon and prostate cancer) [13].

Our comparisons of survival models show DoS in the context of a PBCR remains useful for epidemiological studies as traditionally intended and used. In this paper we provide comprehensive DoS to AJCC-TNM mappings based on the 7th edition AJCC which will be useful for researchers interested in consolidating stage data across the different stage classification systems. DoS can potentially be used in conjunction with TNM-derived data through mapping algorithms, as explored in previous studies [2, 14]. [28]

At its meeting in November 2018, the Australasian Association of Cancer Registries (AACR) discussed the value and feasibility of prospectively collecting and providing national stage data. In light of the findings of our analysis and those provided by a similar analysis undertaken by the South Australia Cancer Registry, there was a preliminary agreement for Australasian PBCRs to consider prospective collection of stage data, where possible, for melanoma, breast, and colon cancers with a diagnosis date of 2017 onwards. Lung cancers are considered difficult to accurately stage based on information available to PBCRs, and comprehensive AJCC-TNM stage data are already collected by the state-based Prostate Clinical Cancer Registries.

In light of the move toward Structured Reporting of Cancer nationally and internationally, the The Royal College of Pathologists of Australasia (RCPA) has issued a Position Statement [29] advising its Fellows to implement AJCC Staging (8th edition). In general, Australian pathologists have historically used AJCC staging in practice and NSWCR implemented Business Rules for AJCC accordingly.

We acknowledge that the findings drawn from this study may not be the same across other cancer registries, or for other diagnosis years or tumour groups. Our analyses only used data from the NSWCR for a subset of melanoma, prostate, colorectal, breast, and lung cancer cases diagnosed in 2011 that were eligible for RD-staging. Other Australian cancer registries will have different stage data collection practices – the value of RD-staging within their respective registry may be determined by different factors. Additionally, survival analyses conducted in this study only examined outcomes at 1 and 4 years after diagnosis, whereas in practice, models typically examine survival at 5 or 10 years after diagnosis.