This multicenter retrospective study integrated clinical data from the SEER database of the National Cancer Institute and Qingdao University Affiliated Hospital. The SEER database covers approximately 30% of the US population and provides comprehensive data, including clinicopathological characteristics, demographic information, and survival outcomes. Using SEER*Stat software (version 8.4.5), we identified patients diagnosed with pRCC (ICD-O-3 code 8260/3) between 2004 and 2015, resulting in an initial cohort of 15,040 cases. The inclusion criteria were: (1) histopathologically confirmed pRCC; (2) a single primary tumor; and (3) treatment with either partial or radical nephrectomy. Exclusion criteria included: (1) missing data for key variables (age, sex, race, pathological grade, TNM stage, tumor size, tumor laterality, surgical approach, follow-up time, marital status, or survival outcomes); (2) pRCC not being the first primary malignancy; and (3) data sourced solely from autopsy reports or death certificates. Ultimately, 5,690 eligible patients were included. The SEER cohort was randomly split in a 7:3 ratio into a training set (n=4,001) for nomogram development and an internal validation set (n=1,689). This ratio was chosen to ensure model stability while utilizing an independent set to assess generalization capability and prevent overfitting.

To validate model generalizability, an additional 172 pRCC patients treated at Qingdao University Affiliated Hospital between February 2013 and December 2021 were enrolled as an external validation cohort. The rationale for this multi-center validation strategy was twofold: 1) to leverage the large sample size of the SEER database for robust model construction, and 2) to evaluate the model’s applicability in a real-world clinical setting using an independent external cohort. The inclusion criteria for the external validation set were consistent with those for the training set, with the last follow-up date in December 2024. The study protocol was approved by the Institutional Review Board of Qingdao University Affiliated Hospital. Informed consent was waived due to the retrospective nature of the study. The study flowchart is presented in Figure 1 .

The following variables were collected: 1) baseline demographic characteristics (age at diagnosis, race, and marital status); 2) tumor characteristics (maximum tumor diameter, histological grade, TNM stage); 3) treatment information (surgical procedure); 4) GRANT score; and 5) survival outcomes (vital status and survival time). The GRANT score was calculated based on four clinicopathological parameters, each assigned a binary score (0 or 1 point) as follows: 1) Tumor grade: Fuhrman nuclear grade 1–2 scored 0 points; grade 3–4 scored 1 point. 2) Age: ≤60 years scored 0 points; >60 years scored 1 point. 3) Primary tumor extent (pT stage): pT1–T3a scored 0 points; pT3b–T4 scored 1 point. 4) Lymph node status: pN0 or pNx (no pathologically confirmed lymph node metastasis) scored 0 points; pN1 (pathologically confirmed lymph node metastasis) scored 1 point. It is noteworthy that for lymph node status assessment, pN0 (pathologically confirmed negative nodes) and pNx (nodes not assessed or status unknown) were combined into a single category. This approach was justified based on: 1) the substantial proportion of pNx cases in the SEER database, thereby enhancing the model’s applicability; 2) consistency with previous studies (17); and 3) recognition of the high false-positive rate associated with radiographic lymph node assessment, whereas this study primarily relied on pathological staging.

TNM staging was performed according to the AJCC 6th edition criteria. The total GRANT score ranged from 0 to 4 points. Patients were stratified into three risk groups for prognostic assessment: low-risk (0–1 points), intermediate-risk (2 points), and high-risk (3–4 points) (see Supplementary Table 1 ). The primary endpoints were OS and CSS. OS was defined as the time from diagnosis to death from any cause, and CSS was defined as the time from diagnosis to death specifically attributed to pRCC.

Continuous variables are presented as mean ± standard deviation or median (interquartile range), while categorical variables are expressed as frequencies (percentages). Kaplan-Meier survival curves were plotted to assess survival differences across GRANT score groups, with inter-group comparisons performed using the log-rank test. A multi-stage variable selection strategy was employed to identify independent prognostic factors. First, univariate Cox regression analysis was conducted for all candidate variables (including age, sex, race, marital status, tumor grade, T stage, N stage, M stage, tumor size, and GRANT score), with statistical significance set at P < 0.05. Variables showing significance in the univariate analysis were subsequently subjected to Least Absolute Shrinkage and Selection Operator (LASSO) regression for dimensionality reduction and feature selection. The optimal penalty parameter (λ) for LASSO was determined via 10-fold cross-validation, retaining predictors with non-zero coefficients. These selected variables were then incorporated into a multivariable Cox proportional hazards regression model using forward LR selection. Finally, a prognostic nomogram was constructed based on the statistically significant variables in the final multivariable model to predict 1-, 3-, and 5-year CSS and OS in pRCC patients.

To evaluate the predictive performance of the GRANT score and the nomogram, both internal and external validations were performed. The concordance index (C-index) and its 95% confidence interval were used to assess the overall discriminative ability of the models. Time-dependent receiver operating characteristic (ROC) curve analysis was used to calculate the area under the curve (AUC) at 1, 3, and 5 years, evaluating the models’ discriminative power at specific time points. Calibration curves were plotted to analyze the agreement between predicted probabilities and observed outcomes. Internal validation of discrimination and calibration was performed using the bootstrap method with 1000 resamples. Furthermore, based on the risk scores derived from the nomogram, optimal cut-off values were determined using X-tile software (version 3.6.1) to categorize patients into low-, intermediate-, and high-risk groups.

All statistical analyses were performed using SPSS 26.0 and R language (version 4.4.3). A two-sided P value < 0.05 was considered statistically significant.

A total of 5,690 patients with pRCC meeting the inclusion criteria were included in this study, with data sourced from the SEER database. These patients were randomly assigned to a training set (n=4,001) and an internal validation set (n=1,689). Additionally, 151 pRCC patients from our institution were enrolled as an external validation cohort. As shown in Table 1 , significant differences in baseline characteristics were observed among the cohorts (χ² test, specific P-values are detailed in Table 1 ). In the SEER cohort, male patients accounted for 73.41% (4,177/5,690), and White individuals constituted 69.67% (3,964/5,690). Regarding pathological stage and grade, 51.88% (2,952/5,690) of patients had Fuhrman grade 2, 72.58% (4,130/5,690) were classified as T1 stage (AJCC 6th edition), and 76.19% (4,335/5,690) were in GRANT grade 1. In the external validation cohort, male patients accounted for 70.20% (106/151), 43.71% (66/151) had Fuhrman grade 2, 38.41% (58/151) were T1 stage, and 71.52% (108/151) were in GRANT grade 1. Other detailed clinicopathological characteristics are presented in Table 1 .

Stratified analysis of OS and CSS was performed for pRCC patients according to their GRANT scores. Kaplan-Meier survival curves demonstrated significant differences in both OS and CSS among the different GRANT score groups in both the training and validation cohorts (log-rank test, all P < 0.05; Figure 2 ). To further evaluate the predictive performance of the GRANT score, the C-index and time-dependent AUC at 1, 3, and 5 years were calculated in the training cohort. For OS, the C-index was 0.621 in the training cohort, with 1-, 3-, and 5-year AUCs of 0.696, 0.675, and 0.657, respectively. In the internal validation cohort, the C-index for OS was 0.621, with corresponding AUCs of 0.697, 0.671, and 0.662. The external validation cohort showed a C-index of 0.661 for OS, with AUCs of 0.676, 0.696, and 0.675, respectively. The corresponding results for CSS are detailed in Supplementary Table 2 .

To identify independent prognostic factors for OS and CSS, univariate Cox regression, followed by LASSO regression, and finally, multivariable Cox proportional hazards regression with forward stepwise selection were performed sequentially in the training set. Univariate analysis revealed that marital status, TNM stage (AJCC 6th edition), maximum tumor diameter, Fuhrman grade, and GRANT score were significantly associated with OS (all P < 0.05). TNM stage, maximum tumor diameter, Fuhrman grade, and GRANT score were significantly associated with CSS (all P < 0.05). Subsequently, variables with statistical significance from the univariate analysis were included in the LASSO regression for further screening ( Supplementary Figure 1 ), and the final retained variables were incorporated into the multivariable Cox regression model.

Multivariable analysis identified seven independent prognostic factors for OS, including marital status, TNM stage, maximum tumor diameter, Fuhrman grade, and GRANT score ( Table 2 ). Six factors were significantly associated with CSS, namely TNM stage, maximum tumor diameter, Fuhrman grade, and GRANT score ( Table 3 ). Notably, the widely used TNM staging system was confirmed as an independent prognostic factor for both OS and CSS in this study, consistent with previous findings. Furthermore, multivariable analysis validated the GRANT score as an independent predictor for both OS and CSS. Additionally, marital status was identified as a potential protective factor for OS.

Based on the results of the multivariable Cox regression analysis, we constructed nomograms to predict prognosis in pRCC patients. The analysis indicated that surgical approach at diagnosis, marital status, N stage, M stage, maximum tumor diameter, Fuhrman grade, and GRANT score were independent prognostic factors for 1-, 3-, and 5-year OS ( Figure 3A ). Concurrently, surgical approach, TNM stage, maximum tumor diameter, Fuhrman grade, and GRANT score were confirmed as independent prognostic factors for 1-, 3-, and 5-year CSS ( Figure 3B ). The nomogram is used as follows: first, for each predictor variable, locate its value on the top scale and draw a vertical line downward to the Points axis to determine the assigned score. Second, sum the scores obtained for all variables to get the Total Points. Finally, locate the total score on the respective bottom scales for 1-, 3-, and 5-year survival probability to obtain the predicted survival probability for the patient.

Time-dependent ROC curve analysis was employed to systematically evaluate the predictive performance of the nomogram for survival rates calculated on a monthly basis. The C-index and the AUC were used as quantitative indicators of the model’s discriminative ability. These metrics theoretically range from 0.5 (no discrimination) to 1.0 (perfect discrimination), with higher values indicating superior discriminative power.

The results demonstrated that the nomogram model exhibited good predictive accuracy across the training, internal validation, and external validation sets. For OS prediction, the C-index was 0.711 (95% CI: 0.697-0.724) in the training set, with 1-, 3-, and 5-year AUC values of 0.811 (95% CI: 0.780-0.844), 0.791 (95% CI: 0.769-0.811), and 0.764 (95% CI: 0.744-0.783), respectively ( Figure 4A ). In the internal validation set, the C-index was 0.720 (95% CI: 0.700-0.741), with corresponding AUCs of 0.835 (95% CI: 0.783-0.879), 0.800 (95% CI: 0.766-0.833), and 0.771 (95% CI: 0.742-0.803) ( Figure 4B ). The external validation set yielded a C-index of 0.740 (95% CI: 0.665-0.814) for OS, with AUCs of 0.750 (95% CI: 0.608-0.881), 0.770 (95% CI: 0.678-0.858), and 0.771 (95% CI: 0.678-0.858) ( Figure 4C ).

For CSS prediction, the C-index was 0.860 (95% CI: 0.843-0.876) in the training set, with 1-, 3-, and 5-year AUC values of 0.911 (95% CI: 0.881-0.937), 0.915 (95% CI: 0.896-0.933), and 0.895 (95% CI: 0.874-0.916), respectively ( Figure 4D ). In the internal validation set, the C-index was 0.873 (95% CI: 0.849-0.898), with AUCs of 0.901 (95% CI: 0.852-0.943), 0.905 (95% CI: 0.870-0.938), and 0.907 (95% CI: 0.875-0.935) ( Figure 4E ). The external validation set showed a C-index of 0.826 (95% CI: 0.733-0.900) for CSS, with AUCs of 0.847 (95% CI: 0.681-0.964), 0.843 (95% CI: 0.714-0.936), and 0.842 (95% CI: 0.742-0.933) ( Figure 4F ). These results indicate that the predictive model developed in this study possesses high discriminative accuracy across different datasets.

Furthermore, calibration curve analysis was used to assess the accuracy of the predicted 1-, 3-, and 5-year survival probabilities in the training, internal validation, and external validation sets. Under ideal calibration, predicted values should align perfectly with the 45-degree reference line. As shown in Figures 5A–C , for OS prediction, the calibration curves demonstrated good agreement between the nomogram’s predictions and the actual observations, confirming satisfactory calibration capability. Similarly, the calibration curves for CSS prediction also showed good consistency ( Figures 5D–F ), further validating the model’s calibration performance.

Based on the nomogram model, we calculated a risk score for each patient in the training cohort and performed risk stratification accordingly. Using X-tile software (version 3.6.1), optimal cut-off values were determined to categorize patients into three subgroups: low-risk, intermediate-risk, and high-risk (18). For the OS nomogram, the stratification criteria were: low-risk (score ≤ 2.51), intermediate-risk (2.51 < score ≤ 6.04), and high-risk (score > 6.04) ( Supplementary Figures 2A, B ). For the CSS nomogram, the criteria were: low-risk (score ≤ 7.34), intermediate-risk (7.34 < score ≤ 27.23), and high-risk (score > 27.23) ( Supplementary Figures 2C, D ). Kaplan-Meier survival analysis revealed statistically significant differences in OS among the different risk subgroups in the training cohort ( Figure 6A ). Similarly, the risk stratification for CSS also showed statistically significant differences ( Figure 6B ).