Population-based cancer registries are an important data source for understanding cancer incidence, mortality, and survival (1). U.S. state-based cancer registries collect data related to cancer diagnosis, the first course planned treatment, and vital status; however, they do not routinely follow patients longitudinally to collect intermediate outcomes such as disease recurrence, progression or subsequent treatment. Given the increasing complexity of cancer care and improvements in patient survival after diagnosis, researchers and program planners are interested in data collection as well as methods to better understand intermediate outcomes, including the number and percent of patients living with recurrent or metastatic disease (2). This information would provide a more comprehensive knowledge of the disease course following initial diagnosis and the effectiveness of cancer treatment and better quantify the cancer burden (3).

A modeling method was recently developed to estimate the risks of breast cancer metastatic recurrence using survival data from cancer registries (4). To overcome the absence of data on the recurrence outcome and estimate survival from metastatic recurrence, the method combined registry survival data from patients initially diagnosed with metastatic disease with data from studies reporting survival from recurrent disease (5). For patients diagnosed with local or regional cancers, the model assumed that recurrence or progression to metastatic disease must occur before a cancer death and provided estimates on proportions recurring or progressing to metastatic disease. Estimates from this method showed that risk of recurrence or progression to metastatic breast cancer was higher for women diagnosed at older ages, earlier year of diagnosis, more advanced stage, and hormone receptor (HR) status negative tumors (4). For women aged 60–74 who were diagnosed in 2000–2013 with local and regional breast cancer, the estimated percent recurring within 5 years was 5.3% and 21.6%, respectively (4). While data on recurrence is not available at the population level, this approach holds potential for using cancer registry data to provide grouped population-based estimates on the percentages that will recur with metastasis by time since diagnosis. These estimates can also provide comparisons to evaluate efforts in the collection of recurrence data. Simulations were conducted and showed that the method was robust to a variety of assumptions. However, the method was not validated using external recurrence data collected from medical records (“gold standard” data) as no such data were available at the time.

The Centers for Disease Control and Prevention’s (CDC) National Program of Cancer Registries (NPCR) conducted a special project to support Patient-Centered Outcomes Research (PCOR)(6). Five state central cancer registries received support to collect additional information that included disease-free status, recurrence, progression, and respective dates for colon, rectum, and breast cancer cases diagnosed in 2011. The registries collected this information through active periodic review of medical records and active searching of information, beyond the standard cancer registry data collection procedures. Longitudinal follow-up for collecting these intermediate outcomes was completed by November 2017. The NPCR-PCOR project provided “gold standard” data on intermediate outcomes for up to five years from diagnosis and represents the most comprehensive population-based data on recurrence and progression in the U.S. to date.

This study aims to validate the modeling estimates with similar empirical estimates obtained using the NPCR-PCOR data on female breast, colon, and rectum cancer patients. Because the modeling approach assumes that either recurrence or progression to metastatic disease occurs before a cancer death, we combine information on progression and metastatic recurrence. We compare the percentages of patients who are recurrence-free by time since diagnosis and stage at diagnosis using the modeling approach to the abstracted NPCR-PCOR data. While similar estimates for breast cancer were previously available (4), they were not validated. Thus, the estimates in this paper are new and provide important information on the burden of recurrence or progression with metastatic disease for female breast, colon, and rectum cancer patients. These estimates represent population-based summaries of the risk of metastatic recurrence in the absence of other causes of death and reflect detection and treatment patterns as observed in the population. They can also be used to inform new studies on the collection of recurrence information in the US and elsewhere.

In our study we included 12,420 patients, 8,393 female breast, 2,880 colon, and 1,147 rectum cancers diagnosed in 2011 from NPCR-PCOR (Table 1). The proportion of female breast, colon, and rectum cancer patients who had a metastatic event (recurrence or progression) either reported or imputed were respectively, 792 (9%), 638 (22%), and 317 (28%).

Survival data from SEER was used to calculate the model estimates. There were some differences between the patients included in the SEER data vs. NPCR-PCOR data (Table 2). There was a higher percent of white cancer patients in the NPCR-PCOR data compared to SEER data. Stage distribution was similar for patients diagnosed with colon and rectum cancers. For female breast cancer patients, NPCR-PCOR had a slightly higher percent of patients diagnosed with stage I. NPCR-PCOR rectum cancer patients were slightly younger compared to patients in the SEER data. Sex distribution was similar in the two datasets.

All the modeled recurrence-free survival estimates are for the period of diagnosis 2004–2015 to be in better alignment with the 2011 NPCR-PCOR estimates (Table 3). When grouping patients with stages I-III, the NPCR-PCOR and modeled 5-year recurrence-free estimates are very similar being respectively, 88.6% vs 90.2% for female breast cancer, 75.3% vs 74.6% for colon cancer, and 68.5% vs 68.8% for rectum cancer. When controlling by stage, the 5-year recurrence-free NPCR-PCOR and modeled estimates are similar and differ less than 3.6 percentage points. The largest differences observed between the NPCR-PCOR and modeled 5-year recurrence-free estimates are 95.3% and 98.8% for stage I female breast cancer; 64.9% and 61.3% for stage III breast cancer and, 90.7% and 87.1% for stage I colon cancer. The model predicts that after 15 years from diagnosis, 84.1%, 63.4% and 57.0% patients diagnosed with stages I-III female breast, colon and rectum cancers respectively will be recurrence-free (Table 3).

While the modeled recurrence-free survival estimate using the SEER data are very similar to the empirical NPCR-PCOR estimates at 5 years from diagnosis, they differ considerably at years 1 to 3 from diagnosis getting closer at 4 years from diagnosis (Figure 3). The largest differences are observed at 1 year from diagnosis, especially for stage III. For patients diagnosed with stage III, the percent recurrence free at 1 year from diagnosis from NPCR-PCOR and modeled were respectively, 93% and 79% for breast cancer, 93% and 68% for colon cancer, and 91% and 74% for rectum cancer. However, the differences almost disappear at 5 years from diagnosis.

This study provides an external validation of a modeling method to estimate recurrence-free survival, using abstracted recurrence and progression information with respective dates from the NPCR-PCOR study as the gold standard. The NPCR-PCOR and modeling recurrence-free estimates are very similar at 5-years after diagnosis, differing less than 4 percentage points in survival. Much larger differences were observed in the 1 to 3 years from diagnosis, especially at 1 year for stage III disease, showing that the model method is not able to provide accurate recurrence-free estimates in years 1 to 3 after diagnosis. The alignment between empirical PCOR-NPCR and the modeling estimates provide validity of the modeled estimates at 4 to 5 years after diagnosis.

This article provides the first U.S. population-based estimates of the percent of cancer patients that are recurrence-free 5 years after diagnosis, additionally supported by the fact that NPCR-PCOR and modeled estimates are very similar. When considering all stages combined, the percent of cancer patients who are recurrence-free 5 years after diagnosis are highest for female breast cancer 89% (90%), followed by colon cancer 75% (75%) and rectum cancer 68% (69%) using the NPCR-PCOR data (modeled). Results from our study are also consistent with international studies that collect recurrence information for their national cancer registries. A population-based study (9) reported 5-year and 10-year recurrence-free survival estimates for patients diagnosed with colon and rectum cancers in Denmark, and the estimates were generally very similar to the modeled numbers. For example, they reported 83%, 78% and 60% recurrence-free after 10 years for patients diagnosed with stage I, II and III colon cancer, while the modeled estimates are 84%, 72% and 53%, respectively.

The NPCR-PCOR study represents a population-based benchmark resource on recurrence in the U.S. The study conducted extensive reviews of medical records to capture data on disease-free status, recurrence, or progression of disease on all 2011 incident study cancer cases diagnosed in 5 U.S. states to which we compared our model. The modeling approach assumes that cancer patients either progress or recur with metastasis before a cancer death. Thus, to compare the model recurrence-free survival we constructed an algorithm that combined the information on recurrence and progression and imputed events and respective dates where data were uncertain or missing.

The first results using the combined NPCR-PCOR data (6) differ from the results shown in this study. They reported recurrence-free survival from the surgery date among patients who had documented disease-free status. The modeled method uses data readily available from cancer registries and does not rely on disease-free status or date of definitive treatment information which may be missing or incomplete. To provide a fair comparison, this study used NPCR-PCOR recurrence-free survival estimates for all patients diagnosed with stage I-III cancer, irrespective of disease-free status, and the start time is the date of diagnosis. The estimates represent the proportion of patients that did not have metastasis at the time of diagnosis who later progressed or recurred with metastasis. Reporting recurrence-free survival from the date of end of initial treatment with confirmation that the patient is disease free is very relevant for clinical decision making and to inform the disease-free group of patients of their risk of recurrence However, collection of complete and accurate disease-free status and treatment dates is challenging. Recurrence-free survival from date of diagnosis including both progression and recurrence can be useful as a population-based cancer control measure for tracking progress against cancer over time or for comparisons across groups of patients.

The modeling method relied on assumptions. It assumed that recurrence or progression with metastatic disease occurs before a cancer death thus only allowing estimation of recurrence or progression combined. The method also assumed that survival from recurrence is known and similar to survival from de novo metastatic cancer (colon and rectum) For breast cancer we used mortality hazard ratio r= 1.35 from a study (5) that estimated a poorer survival from recurrence breast cancer compared to survival from de novo metastatic breast cancer. A study has shown that survival after recurrence for colorectal cancer patients improves with larger intervals between the initial diagnosis and recurrence (10), which may indicate a better recurrence survival and r<1.0 for colon cancer. The model and empirical estimates differed at 1–3 years from diagnosis. Sensitivity analysis (Supplementary Table 1) and simulation studies (4) showed that the model estimates are less accurate in the earlier years, since recurrence-free survival becomes more accurate as information is accumulated with time since diagnosis. It is also possible that the empirical estimates are more uncertain in short duration follow-up due to greater differences in the course of disease and care in this period. The model can be applied to other cancers, in particular those amenable to mixture cure modeling and those with data on survival from metastatic recurrence. In the absence of this information, survival from de novo distant disease could be used as an approximation and sensitivity analyses using different values of the mortality hazard ratio performed.

The NPCR-PCOR study demonstrated feasibility of capturing intermediate outcomes at the registry level. However, it required additional funding and staff resources beyond the scope of routine operations of central cancer registries. The study also identified need for standardized disease progression and recurrence definitions in population-based registries and provided useful data and experience to inform future work. For example, the NPCR-PCOR study demonstrated that identifying progression or disease-free status was more challenging than determining recurrence (6), a consideration for other studies.

There are many ongoing efforts on leveraging medical records (e.g., EHR, claims, pathology and imaging reports) to capture individual information on what happens after a cancer diagnosis. Individual data on intermediate outcomes such as recurrence are important for cancer research and to understand risks of recurrence and disparities in outcomes. Ideally, these efforts will also include information on what occurs after a recurrence, including post-recurrence treatment, quality of life and patient-reported outcomes for the purposes of cancer survivorship research and to improve the lives and outcome of patients. Until these efforts mature, the modeling approach and NPCR-PCOR study can be leveraged to inform current efforts to collect intermediate outcomes, and provide provisional population-based estimates of 5-year recurrence-free survival.