SEER*Stat software (version 8.4.0.1) was used to extract data from patients with sarcomatoid mRCC who were diagnosed between 2010 and 2015. For this study, we signed a data agreement and used the SEER database under the username 17749-Nov2020. Because SEER is a publicly accessible database containing deidentified data, an ethical review is not needed. Sarcomatoid mRCC was defined as kidney parenchyma with sarcomatoid features and distant metastases (American Joint Committee on Cancer 7th M1). The histology recoded broad groupings: 8,140–8,389 adenomas and adenocarcinomas were included according to the 2016 WHO histopathological classification of renal cell tumors.

For external validation (n = 55), a retrospective cohort of patients with sarcomatoid mRCC who underwent CRN between January 2010 and December 2020 was extracted from the medical records of three medical centers in China: Nanjing University Jinling Hospital, Affiliated Hospital of Medical School, Nanjing University; Changhai Hospital, Naval Medical University; and Qilu Hospital of Shandong University. Ethical approval was obtained from the ethics committee of each institution. In cases in which individual patient consent was not required, the chairperson of the ethics committee waived the need for patient consent.

Patients aged ≥18 years and diagnosed with metastatic mRCC and sarcomatoid features were included. Patients with tumors without sarcomatoid features, unknown TNM stage, unknown survival data, nonunilateral tumor or unknown tumor laterality, not RCC, or unknown metastatic status (including bone, brain, liver, and lung) were excluded.

CRN was defined as radical nephrectomy (RX Summ-Surg Prim Site (1998+) surgery code: 50). The screening process of the study population and the establishment of the predictive model is depicted in a flowchart ( Figure 1 ).

According to whether the patients underwent CRN, the study population from the SEER database was divided into the CRN group and the non-CRN group. Multivariate logistic regression models using the stepwise method were used to predict CRN candidates to prove the necessity of propensity score matching (PSM). Multivariate Cox regression by the stepwise method was employed to determine the independent risk factors that would be included in the nomogram.

To minimize selection bias and baseline imbalance, PSM was used to match the non-CRN group to the CRN group. By logistic regression, a PS was generated using variables that could potentially influence treatment outcomes, such as age, sex, laterality, T-stage, N-stage, bone metastasis, brain metastasis, liver metastasis, lung metastasis, systematic therapy, metastatic site, and lymph nodes. The CRN and non-CRN groups were 1:1 matched using the nearest PS on the logit scale with a caliper of 0.08 without replacement. The balance of covariables in both groups before and after matching was evaluated by standardized differences, and a standardized difference of <10% was considered to indicate sufficient balance.

The Chi-square test was used to determine the significance of the difference in categorical variables. Cox proportional hazard regression was used to determine independent prognostic factors. Hazard ratios (HRs) were calculated with a 95% confidence interval (CI). Statistical analysis was performed with R version 4.2.1 (http://www.r-project.org/). A two-sided p-value of <0.05 was considered statistically significant.

The median overall survival (OS) in the non-CRN group was 6 months. After PSM, patients with an OS of >6 months in the CRN group were classified as the “CRN benefit group.” Conversely, those with an OS of <6 months in the CRN group were classified as the “no CRN benefit group.”

To identify patients with sarcomatoid mRCC who may benefit from CRN, a prediction model was established using multivariate logistic regression analysis. The matched CRN group (n = 215) was randomly divided into two groups for training (n = 150) and internal validation (n = 65) with a ratio of 7:3. The logistic regression model comprised the independent predictor variables from multivariate Cox regression analysis, including age, T-stage, systemic therapy, metastatic site, and lymph nodes. This prediction model was built on the training set and displayed as a nomogram. The probability of patients with sarcomatoid mRCC benefiting from CRN was calculated by summing the scores for each selected variable. The area under the receiver operating characteristic (AUC) curve was used to determine the prediction efficiency (sensitivity and specificity) in the training, internal validation, and external validation cohorts.

To compare the predicted and observed outcomes, a calibration plot and decision curve analysis were performed, with a p-value of >0.05 indicating a good model fit. A classification system based on the nomogram was established to predict the probability that a patient with sarcomatoid mRCC may benefit from CRN. A predicted probability of >0.5 indicated a good chance of CRN benefit and hence longer survival; conversely, a predicted probability of ≤0.5 indicated that the patient may not benefit from CRN. The Kaplan–Meier method was performed to estimate OS and cancer-specific survival (CSS) in CRN benefit and no CRN benefit groups. This method was performed to verify that the model can identify patients who may benefit from CRN and subsequently evaluate the clinical utility of the model.

A total of 900 sarcomatoid mRCC cases were included in this study based on the SEER database using the following SEER variables: “primary site—labeled: C64.9—kidney, NOS,” and “Histology recoded broad groupings: 8,140–8,389 adenomas and adenocarcinomas.” Multivariate logistic regression (stepwise method) of all patients showed that the independent predictors of CRN recipients were T-stage, bone metastasis, liver metastasis, systematic therapy, and lymph nodes ( Table 1 ).

Before matching, significant differences in T-stage, bone metastasis, liver metastasis, systematic therapy, and lymph nodes were found between the CRN and non-CRN groups, which were well balanced after matching (all p > 0.05), and 215 matched pairs of patients with metastatic sarcomatoid renal cell carcinoma with or without CRN were generated and enrolled in the subsequent analysis ( Table 2 ). The baseline characteristics of the three cohorts (training, internal validation, and external validation cohorts) are shown in Table 3 .

In the nomogram, the independent prognostic predictive variables from the multivariate Cox regression (stepwise method) for OS included age, T-stage, systematic therapy, metastatic site, and lymph nodes ( Figure 2 ).

Either before or after PSM, Kaplan–Meier analysis revealed that the CRN group was associated with significantly better OS and CSS than the non-CRN group ( Figure 3 ). Before PSM, the median OS and CSS were both 11 months in the CRN group and both 5 months in the non-CRN group. After PSM, the CRN group showed longer median OS (8 vs. 6 months, HR = 0.767, p = 0.0085) and longer median CSS (8 vs. 7 months, HR = 0.801, p = 0.033) than the non-CRN group.

After PSM, patients in the CRN group (n = 215) whose OS exceeded the median OS of the non-CRN group were assigned to the “CRN benefit group” (OS of >6 months, n = 124, 57.7%). The remaining patients were assigned to the “no CRN benefit group” (survival time of ≤6 months, n = 91, 42.3%).

According to the multivariate Cox regression analysis of the CRN group (n = 608) before PSM, clinical variables that could be accessed preoperatively were included for constructing the nomogram, including age, T-stage, systematic therapy, metastatic site, and lymph nodes. Based on multivariate logistic regression analysis, a prediction model was established as a nomogram to predict the probability that a patient with sarcomatoid mRCC may benefit from CRN ( Figure 4 ).

The prediction nomogram showed a good discrimination capacity for identifying patients with mRCC who may benefit from CRN in the training cohort (AUC = 0.766, 95% CI: 0.687–0.845), internal validation cohort (AUC = 0.796, 95% CI: 0.684–0.908), and external validation cohort (AUC = 0.911, 95% CI: 0.831–0.991) ( Figure 5 ). The outcomes of the actual calibration curve for training, internal validation, and external validation cohorts were all in perfect agreement with those of the nomogram ( Figure 6 ).

All corresponding scores of each variable were superimposed in the nomogram to calculate the CRN benefit probability. According to the total score from the nomogram, candidates with a predicted probability >0.5 at the cutoff point were stratified as the CRN benefit group. Otherwise, they are stratified as the no CRN benefit group.

Decision curve analysis showed the clinical value of the nomogram ( Figure 7 ). The Kaplan–Meier analysis showed that the OS of the CRN benefit group, no CRN benefit group, and non-CRN group was accurately distinguished in both the internal and external validation sets, suggesting good clinical value of the nomogram ( Figure 8 ).