Validation of a novel prognostic staging system for breast cancer incorporating age at diagnosis Authors: Helen M Johnson, Wenli Dong, Yu Shen, William Irish, Jan H Wong, Nasreen A Vohra, Nina Tamirisa Published: 2026 ABSTRACT AbstractPurposeAge at diagnosis is an independent prognostic variable for breast cancer-specific and overall survival (OS). We previously developed a novel prognostic staging system that incorporates age and demonstrated refined risk stratification compared with the current American Joint Committee on Cancer (AJCC) staging schema. We now aim to externally validate this staging system.MethodsThe National Cancer Database was used to identify adult females diagnosed with invasive breast cancer from 2010-2015. Women with prior malignancy, unknown vital status, or unknown AJCC clinical prognostic stage (CPS) variables were excluded. Serial multivariable Cox’s proportional hazards models evaluated associations between OS and CPS +/- age +/- effect modification (CPS by age interaction). Models were evaluated using concordance index (C-index). ResultsAmong 663,659 patients, the median age was 60 years (IQR 51-70) and the median follow-up was 91.3 months. At last follow-up, 133,273 patients (20.1%) were dead. The model that incorporated both age and effect modification had the highest C-index (0.7852, versus 0.6881 for the model excluding age and 0.7847 for the model without effect modification), indicative of the best predictive performance. Using this model to predict OS stratified by CPS showed differential survival across the age spectrum, consistent with the initial model. Within each stage group, women diagnosed at age 40 had the best survival, whereas women at the extremes of age had inferior survival. ConclusionThis analysis provides external validation of our novel prognostic staging system for breast cancer. Future editions of AJCC may consider incorporating age to achieve more accurate survival predictions. Keywords: Breast cancer, Age, Stage, Prognosis, Survival, Model Introduction Accumulating evidence supports age at diagnosis as an independent prognostic variable for breast cancer-specific survival and overall survival (OS) [1–7]. Both younger [8–11] and older [12–14] women experience higher mortality than those diagnosed with breast cancer during middle age, even after controlling for patient-related, tumor-related, and treatment factors. The relationship between age at breast cancer diagnosis and survival is therefore quadratic or “U-shaped”, with women aged 45–50 years experiencing superior outcomes compared to those at the extremes of age [15–17]. Despite the prognostic importance of age at diagnosis, the American Joint Committee on Cancer (AJCC) staging system for breast cancer does not take this factor into account. After recognizing that patients within the same anatomic stage group have widely variable prognoses which are impacted by tumor biology, the AJCC introduced the Clinical Prognostic Stage (CPS) in the 8th edition [18]. The CPS therefore includes not only anatomic variables [tumor size (T stage), nodal involvement (N stage), distant metastasis (M stage)] but also biological variables [tumor grade, estrogen receptor (ER) status, progesterone receptor (PR) status, and human epidermal growth factor receptor 2 (HER2) status] [19]. Patient-related factors such as age at diagnosis, however, were not considered. We previously developed a novel prognostic staging system that incorporates age at diagnosis and demonstrated more refined risk stratification compared with the current AJCC CPS schema [20]. We now aim to externally validate this novel staging system. Methods Cohort selection The National Cancer Database (NCDB) was used to identify adult (age 18–90 years) females diagnosed with breast cancer from 1/12,010 to 12/312,015. Women with prior history of malignancy, in-situ disease (T0N0M0), unknown AJCC CPS variables (T stage, N stage, M stage, grade, ER, PR, HER2), unknown vital status, or unknown survival follow-up time were excluded. Summary statistics were provided, including mean, standard deviation, median, and range for continuous variables. The frequency counts and percentages for categorical variables were summarized. Kaplan-Meier curves were produced by CPS group for OS. OS was defined according to the variables PUF_VITAL_STATUS and DX_LASTCONTACT_DEATH_MONTHS. The Log-rank test was used to test the difference in survival distributions among the groups. Model selection Multivariable Cox’s hazards models were used to evaluate the association of CPS and age with risk of overall mortality. Stage IA was used as the reference group. Serial models were constructed to evaluate the additive and multiplicative effects of age, taking into account the quadratic relationship between age and OS [15, 16]. Models were as follows: Model 1: CPS. Model 2: CPS + age + age2. Model 3: CPS + age + age2 + CPS x age + CPS x age2. Models were evaluated using Harrell’s concordance index (C-index) to assess model predictive accuracy, − 2 * Log likelihood and the Akaike information criterion (AIC) to assess in-sample model fit. Model validation The predicated survival probabilities at 12, 36, 48, and 60 months were calculated using the previously described model developed using Surveillance, Epidemiology, and End Results (SEER) data [20]. The model-derived survival probabilities were estimated using NCDB survival Model 3. Line graphs were created to visualize the difference between predicted and model-derived survival probabilities across various age groups and prognostic stage levels. All tests were two-sided. All analyses were conducted using the SAS (version 9.4; SAS Institute, Cary, NC) and S-plus (version 8.04, TIBCO Software Inc., Palto Alto, CA) statistical software. Results Patient characteristics and outcomes From a pool of 3,690,015 patients, 663,659 met study criteria and comprised the analytic cohort. The median age was 60 years (IQR 51–70). Most patients were White (82.7%), while 11.6% were Black, 4.8% other race, and 0.9% unknown race. 482,898 patients (72.8%) had hormone receptor-positive/HER2-negative disease, while 29,278 (4.4%) had hormone receptor-negative/HER2 + disease, 67,637 (10.2%) had triple-positive disease, and 83,846 (12.6%) had triple-negative disease. Distributions of AJCC CPS and the seven variables used to determine CPS are detailed in Table 1. Table 1Frequencies of tumor characteristics and american joint commission on cancer (AJCC) 8th edition clinical prognostic stageCharacteristicLevelN (%)Clinical T stagecT0539 (0.1%)cT1408,771 (61.6%)cT2189,209 (28.5%)cT338,108 (5.7%)cT427,032 (4.1%)Clinical N stagecN0533,008 (80.3%)cN1101,437 (15.3%)cN217,616 (2.7%)cN311,598 (1.7%)Clinical M stagecM0637,265 (96.0%)cM126,394 (4.0%)Tumor grade1148,613 (22.4%)2294,195 (44.3%)3220,851 (33.3%)Estrogen Receptor (ER)Negative119,848 (18.1%)Positive543,811 (81.9%)Progesterone Receptor (PR)Negative181,788 (27.4%)Positive481,871 (72.6%)HER2Negative566,744 (85.4%)Positive96,915 (14.6%)AJCC CPSIA338,260 (51.0%)IB125,634 (18.9%)IIA69,736 (10.5%)IIB51,295 (7.7%)IIIA12,806 (1.9%)IIIB28,217 (4.3%)IIIC11,317 (1.7%)IV26,394 (4.0%) Frequencies of tumor characteristics and american joint commission on cancer (AJCC) 8th edition clinical prognostic stage The median follow-up was 91.3 months (95% CI 91.2–91.4). At last follow-up, 133,273 patients (20.1%) were dead. Unadjusted survival curves by CPS stage are presented in Fig. 1. As expected, survival rates were best for stage IA and increasingly poorer for each successive CPS group. Unadjusted overall survival by American Joint Commission on Cancer (AJCC) 8th edition Clinical Prognostic Stage Model selection The model that incorporated CPS, age, and effect modification (CPS by age interaction) had the highest C-index and the lowest values for both AIC and − 2 * Log likelihood, indicative of the best accuracy and model fit (Table 2). Similar results were obtained previously using SEER data [20]. Table 2Measures of model performanceModels−2 * Log likelihoodAkaike Information CriterionC-Index1: CPS3,400,159 on 7 df3,400,1730.68812: CPS + age + age23,326,892 on 9 df3,326,9100.78473: CPS + age + age2 + (CS x age) + (CPS x age2)3,320,635 on 23 df3,320,6810.7852 Measures of model performance Survival predictions Model 3 was used to predict OS at 12, 36, 48, and 60 months, stratified by CPS. Within each CPS group, differential survival was seen across the age spectrum (Fig. 2), consistent with the initial model developed in SEER [20]. Within each stage group from IB-IIIC, women diagnosed at age 40 had the best survival, whereas women at the extremes of age had inferior survival (Table 3). For example, 3-year OS for patients with stage IIIA disease was 0.648 across all ages, highest (0.831) for patients diagnosed at age 40, lower (0.714) for age 20, and lowest (0.127) for age 90. Model-based predictions of overall survival at 12, 36, 48, and 60 months post-diagnosis across the age spectrum, stratified by American Joint Commission on Cancer (AJCC) 8th edition Clinical Prognostic Stage Table 3Model-based 3-year overall survival (OS) estimates by American Joint Commission on Cancer (AJCC) 8th edition clinical prognostic stage (CPS), overall and by selected agesAJCC CPS3y OS, Overall3y OS, Age 203y OS, Age 303y OS, Age 403y OS, Age 503y OS, Age 603y OS, Age 703y OS, Age 803y OS, Age 90IA0.8580.9700.9760.9760.9700.9510.9000.7520.372IB0.8060.9110.9390.9470.9410.9150.8470.6680.282IIA0.7510.8200.8850.9070.9040.8740.7930.5980.228IIB0.7150.8220.8700.8840.8740.8330.7340.5220.183IIIA0.6480.7140.8000.8310.8250.7790.6680.4420.127IIIB0.6070.6690.7510.7810.7740.7250.6140.4090.139IIIC0.4600.5480.6190.6420.6200.5500.4200.2310.056IV0.2860.4440.4450.4200.3690.2920.1970.1000.030 Model-based 3-year overall survival (OS) estimates by American Joint Commission on Cancer (AJCC) 8th edition clinical prognostic stage (CPS), overall and by selected ages Model validation For external validation, the estimated predictor was calculated for each NCDB patient using the coefficients from the previously developed prognostic model based on SEER data. Harrell’s concordance index (C-index) was then estimated in the NCDB cohort to assess predictive accuracy of the SEER-derived model. The estimated C-index was 0.784, indicating good prognostic performance and suggesting that the model correctly identifies the patient with poor survival outcome in approximately 78% of comparable patient pairs. Model-derived OS of patients in this NCDB cohort was compared to predicted OS using the previously-described model developed using SEER data [20]. Overall trends were similar (Fig. 3), consistent with external validation of the prior model. Predicted (prior model) versus model-derived (current model based on NCDB data) overall survival probabilities at (A) 12, (B) 36, (C) 48, and (D) 60 months post-diagnosis, stratified by American Joint Commission on Cancer (AJCC) 8th edition Clinical Prognostic Stage Discussion This analysis provides external validation of our novel prognostic staging system for breast cancer. Including age at diagnosis as a prognostic variable results in more refined risk stratification across all AJCC stage groups. Therefore, the AJCC may consider utilization of not only anatomic and biologic variables, but also age at diagnosis, to achieve more accurate survival predictions in future editions of the staging manual. The AJCC staging system was developed to standardize nomenclature and classification of patients with cancer into groups with similar prognoses. The manual is revised periodically to ensure robust prognostication and current applicability. The 8th edition’s shift from a purely anatomic classification system to one that incorporates tumor biology marked a major transition in our conceptualization of staging. AJCC described the introduction of the CPS as a “different paradigm for the use of ‘stage’ than used in the past”, emphasizing that the prognosis associated with a CPS group is based on anticipated response to standard treatment [19]. That is, the primary purpose of the CPS is accurate prognostication, and it is not intended to influence treatment decisions. To that end, age at diagnosis may serve as an additional variable to refine risk stratification and tailor prognostication when discussing CPS with patients. Given the complex relationship between age at diagnosis and breast cancer outcomes, incorporation of this prognostic variable into the AJCC CPS would require restructuring of the staging schema, which is currently presented as a table of categorical variables. In contrast to differentiated thyroid cancer, for which the effect of age on survival is binary [21], it is not possible to simply insert additional rows in the staging tables for breast cancer to account for age. Indeed, the AJCC anticipated that as our survival predictions become more sophisticated, future editions may require complex models in lieu of Table [19]. We have already developed a user-friendly online calculator (https://tinyurl.com/y3qeh8cl) based on our novel model, which can be adapted readily for mobile application and electronic medical record integration. It should be acknowledged that age and OS are inextricably linked. While age is independently associated with both OS and breast cancer-specific survival [15, 16], OS was selected as the primary endpoint of this study as this is the outcome on which the AJCC CPS groups are based. Although poorer OS may be expected for older women due to age-related competing mortality risks, this does not diminish the relevance of our novel staging system. Patients facing a new cancer diagnosis want to know their prognosis, and they seek realistic estimates rather than theoretical calculations that do not comprehensively reflect their authentic situation. There will inevitably be some variance in survival within a single CPS group, and this certainly may be impacted by co-morbidities, performance status, frailty, and other age-related factors. Future work should explore whether inclusion of geriatric indexes may further refine AJCC prognostication. Ongoing research seeks to clarify why younger women with breast cancer have higher morality than those diagnosed during middle age. While younger women have higher rates of germline pathogenic variants, breast cancer-specific survival is similar to noncarriers [22], so this does not explain the observed survival disparities. Parity and the time interval from last childbirth to breast cancer diagnosis are independently associated with survival [23–25]. Breast cancers diagnosed within 5–10 years postpartum have the poorest prognosis, potentially related to pro-tumorigenic effects of post-lactational involution [26]. In addition, breast tumors in women ≤ 40 have been shown to have unique genomic features and molecular profiles compared to those occurring in older women [27–29], with further distinctions dependent on parity and childbirth recency [30]. Further research is needed to determine whether age at diagnosis is a prognostic factor intrinsically or whether it is a surrogate marker for other factors that directly impact breast cancer outcomes, such as postpartum status. Others have proposed revisions to the AJCC staging system for breast cancer. Plichta and colleagues developed a novel classification system for patients with de novo metastatic breast cancer using NCDB data [31], which was validated in SEER [32], a population-based national database [33], and an institutional cancer registry [34]. Of note, they reported a significant association between age and OS and performed age-stratified analyses but elected not to include this prognostic variable in their final models, in keeping with the established precedent of using only tumor-related factors in the current schema for stages I-III. Limitations of this study include those intrinsic to retrospective analyses of large databases. As only US databases have been used to date, there is potential lack of generalizability to other cohorts. Validation in international tumor registries and population-based databases is important for ensuring broad applicability. In addition, SEER and NCDB lack detailed information on treatments, which indubitably impact survival outcomes. While validation in a richly-annotated institutional database or multicenter registry could lead to further refinement of our novel system, as the AJCC CPS is intended for prognostication of patients assumed to be receiving standard treatment, our system is highly relevant in its current form. Furthermore, as previously mentioned, it is possible that age is a surrogate marker for other factors that more directly impact survival. Nonetheless, it is important to include it in the staging schema until such time as a different primary driver of mortality is identified to ensure accurate prognostication for current patients. Finally, it should be noted that validated prognostic tools represent additional resources to estimate survival, and a few – such as PREDICT [35] – already include age at diagnosis. While these tools may be used alongside the AJCC CPS to support clinical discussions about prognosis, the frequency of their utilization is unknown. Refinement of the AJCC CPS, which is used routinely in oncologic clinical practice, may therefore impact a greater magnitude of patients. In summary, this novel, validated staging system may warrant close examination by the AJCC Breast Cancer Expert Panel for future editions of the staging manual. While including age may add a level of complexity to the current staging system, these data suggest that incorporating age at diagnosis alongside the current anatomic and biologic variables may result in more accurate survival predictions for patients with breast cancer. Acknowledgements These data were presented in part as a poster at the 2025 San Antonio Breast Cancer Symposium. Author contributions Conceptualization: All authors. Design: HMJ, WD, YS, WI, NT. Formal analysis: WD, YS. Interpretation of data: All authors. Writing – original draft: HMJ. Writing – review & editing: All authors. Project administration: NT. Funding This work was supported in part by the Cancer Center Support Grant, NCI Grant # P30 CA016672. Data availability The datasets generated and analyzed during the current study were derived from a deidentified file from the National Cancer Database: [https://www.facs.org/quality-programs/cancer-programs/national-cancer-database/]. The National Cancer Database is a joint program of the American College of Surgeons Commission on Cancer and the American Cancer Society, neither of which have verified the data nor are they responsible for the analysis and interpretation of this data. Declarations Ethics approval Due to the de-identified nature of the public-access user files in the National Cancer Database, the study was deemed exempt from institutional review board approval. Competing interests The authors have no relevant financial or non-financial interests to disclose. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and analyzed during the current study were derived from a deidentified file from the National Cancer Database: [https://www.facs.org/quality-programs/cancer-programs/national-cancer-database/]. The National Cancer Database is a joint program of the American College of Surgeons Commission on Cancer and the American Cancer Society, neither of which have verified the data nor are they responsible for the analysis and interpretation of this data.