This retrospective cohort analysis included patients pathologically diagnosed with sRCC at our institution between December 2009 and September 2024. The study protocol was approved by the Ethics Committee of The Affiliated Hospital of Qingdao University (Approval No: QYFYWZLL30031) and conducted in accordance with the ethical principles of the Declaration of Helsinki (2013 revision). Informed consent was waived due to the retrospective nature of the study. Clinical data were independently and blindly collected by two researchers. Inclusion criteria were: (1) postoperative pathological confirmation of sRCC with complete clinical records; (2) abdominal CT scan performed within one month before surgery. Exclusion criteria were: (1) incomplete clinical, pathological, or follow-up data; (2) concurrent other malignancies or multi-organ dysfunction; (3) previous neoadjuvant therapy; (4) absence of DICOM-format CT images meeting quality standards; (5) death due to complications within 30 days after surgery; (6) active infection or recent use of anti-inflammatory/immunosuppressive drugs. The study flowchart is shown in Supplementary Figure 1 . Clinical variables included age, blood biochemical indicators, and pathological characteristics. Missing values (<5%) were handled using multiple imputation.

A standardized postoperative follow-up protocol was implemented: assessments every 3–4 months in the first year, every 6 months from years 2 to 5, and annually thereafter. Evaluations included clinical symptoms, laboratory tests (e.g., complete blood count and biochemistry), and imaging (CT or MRI). Follow-up concluded on April 1, 2025. The primary endpoint was overall survival (OS), defined as the duration from pathological diagnosis to death from any cause or the last confirmed follow-up.

Preoperative non-contrast CT images were obtained using multiple scanners: GE Optima CT620, LightSpeed CT750 HD, Optima CT670, Revolution CT (GE Healthcare, USA), and Siemens SOMATOM Sensation64 and Definition Flash (Siemens Healthineers, Germany). Scanning parameters were: tube current 240–320 mAs (automatically modulated), voltage 120 kVp, pitch 1.375, reconstruction matrix 512×512, and slice thickness 5 mm. All images were exported in DICOM format from the PACS for further processing.

Tumor segmentation and feature extraction were performed using a standardized protocol. One radiologist (7 years of abdominal imaging experience) and one urologist (15 years of urologic oncology experience), both blinded to pathology, manually delineated tumor boundaries on non-contrast CT images slice-by-layer using ITK-SNAP (v4.2.0) to generate 3D regions of interest (ROIs), carefully excluding adjacent renal parenchyma and perinephric fat. Discrepancies were resolved by a third urologist with 35 years of experience. Prior to feature extraction, all images underwent standardized preprocessing including resampling and gray-level discretization. Features were extracted in Python 3.7 using the pyradiomics toolbox, following the Image Biomarker Standardisation Initiative (IBSI) guidelines (21). Extracted features included first-order statistics, shape, gray-level co-occurrence matrix (GLCM), gray-level dependence matrix (GLDM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighboring gray-tone difference matrix (NGTDM), and wavelet-derived features. To evaluate segmentation reproducibility, two blinded urologists independently segmented ROIs on 30 randomly selected CT images. The first reader repeated the segmentation after one month for intra-observer consistency assessment. Features with an intraclass correlation coefficient (ICC) > 0.75 were retained for further analysis.

Body composition was quantified at baseline using CT axial images at the third lumbar (L3) level. SliceOmatic 5.0 (Tomovision, Canada) was used to measure cross-sectional areas (cm²) of total abdominal muscle (TAM), psoas muscle (PM), and paraspinal muscles (PS). Muscle tissue was defined using Hounsfield unit (HU) thresholds (−29 to 150 HU) (22), with manual correction for accuracy, as illustrated in Figure 1 . All analyses were performed by one radiologist with 7 years of experience. Height-adjusted indices (TAM/height², PM/height²) were derived (23, 24). Sarcopenia was defined using established criteria (23–25): height-adjusted TAM index <52.4 cm²/m² (men) or <38.5 cm²/m² (women); PM index <6.36 cm²/m² (men) or <3.92 cm²/m² (women); absolute PS area <31.97 cm² (men) or <28.95 cm² (women).

Radiomic features were Z-score normalized. A multi-stage selection strategy was applied: first, features with ICC > 0.75 were retained; second, low-variance features (variance threshold <0.1) were removed, and highly correlated features (|r| > 0.9) were reduced by retaining those with higher variance. Univariate Cox regression (p < 0.001) identified prognosis-related features, followed by Least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation (optimal λ selected via minimum error) for dimensionality reduction. Five machine learning algorithms (SuperPC, stepwise Cox, random survival forest [RSF], CoxBoost, and plsRcox) were used to build prognostic models. The model with the highest C-index was selected to generate a radiomics score (Rad-score) for subsequent analysis.

Three models were developed and compared. Univariate Cox regression (p < 0.05) identified potential prognostic variables, followed by LASSO regression (10-fold cross-validation, λ.min) for variable selection. Multivariable Cox regression with backward likelihood ratio (LR) method identified independent prognostic factors to build a clinical model centered on sarcopenia. A combined model was constructed by integrating the Rad-score and significant clinical variables via multivariable Cox regression. Model performance was compared among the radiomics, clinical, and combined models following TRIPOD guidelines.

Internal validation included repeated 10-fold cross-validation for C-index calculation and bootstrap resampling (1000 repetitions) for confidence intervals. Delong’s test and bootstrap methods (1000 repetitions) were used to compare C-indices between models. A nomogram based on the combined model was developed to predict survival probabilities. Time-dependent ROC curves assessed discrimination at 1, 2, and 3 years. Calibration curves (1000 bootstrap samples) evaluated agreement between predicted and observed outcomes. Decision curve analysis (DCA) quantified clinical utility by calculating net benefit across threshold probabilities. SHAP (Shapley Additive exPlanations) analysis interpreted feature contributions and enhanced model transparency (26, 27).

All analyses were performed using IBM SPSS Statistics 26.0 and R 4.4.3. Two-sided p-values < 0.05 were considered statistically significant. Categorical variables are presented as counts and percentages, compared using Pearson’s χ² or Fisher’s exact test. Continuous variables were tested for normality using Shapiro–Wilk test; normally distributed variables are expressed as mean ± standard deviation and compared with t-tests, while non-normal variables are reported as median (IQR) and compared with Mann–Whitney U test.

This retrospective cohort study strictly adhered to predefined inclusion and exclusion criteria, ultimately enrolling 121 patients with pathologically confirmed sRCC. All patients received standardized treatment and systematic follow-up. The median follow-up time for the entire cohort was 21 months (range: 1–183 months). By the end of follow-up, 60 deaths had been recorded. Detailed baseline characteristics, including demographic, clinical, and pathological parameters, are summarized in Table 1 .

A total of 854 quantitative radiomic features were extracted from the ROIs. After evaluating intra- and inter-observer consistency (ICC > 0.75), 707 features were retained for further analysis. Subsequent low-variance filtering (variance threshold < 0.1) and removal of highly correlated features (retaining those with higher variance in each correlated group) yielded 186 features. Univariate Cox regression identified 10 features significantly associated with prognosis (P < 0.001). LASSO regression was then applied for further dimensionality reduction, resulting in six highly predictive features for model construction ( Figures 2A, B ). To comprehensively evaluate predictive performance, five algorithmic strategies were systematically compared. The plsRcox model demonstrated optimal performance ( Figure 2C ), achieving a C-index of 0.696 via 10-fold cross-validation. Time-dependent ROC analysis showed that the model yielded AUC values of 0.706, 0.726, and 0.725 for predicting 1-, 2-, and 3-year OS, respectively ( Figure 2D ). SHAP analysis was used to interpret the plsRcox model. The global SHAP summary plot ( Figure 2E ) illustrated the direction and magnitude of contributions of the six key features, all of which acted as positive predictors. An individual prediction analysis ( Figure 2F ) deconstructed the prediction for a high-risk patient: the baseline prediction (E[f(x)] = −1.97×10^-17) represents the model’s output reference, while the individual prediction value (f(x) = 4.02) indicated elevated mortality risk.

Univariate Cox regression identified clinical features significantly associated with OS in sRCC patients ( Table 2 ). LASSO regression was used to screen prognostic variables, including PNR, PAR, HALP, Ki-67 index, tumor size, N stage, M stage, sarcopenia defined by TAM index, and sarcopenia defined by PS area ( Supplementary Figure 2 ). Subsequent multivariate Cox regression using the backward likelihood ratio method identified the following independent prognostic factors for OS: PNR (HR = 0.981, 95% CI: 0.971–0.991; P < 0.001), HALP (HR = 0.979, 95% CI: 0.963–0.995; P = 0.01), tumor size (HR = 1.074, 95% CI: 0.996–1.157; P = 0.064), N stage (HR = 2.434, 95% CI: 1.387–4.270; P = 0.002), and PS-defined sarcopenia (HR = 3.046, 95% CI: 1.119–8.289; P = 0.029). A clinical prognostic model based on these variables was constructed, and a clinical risk score was computed for each patient. Finally, a combined prognostic model was established by integrating the radiomics risk score (Rad-score) with the clinical model.

Using the selected clinical prognostic factors and the Rad-score, a nomogram was developed to predict 1-, 2-, and 3-year OS in sRCC patients ( Figure 3A ). The nomogram is applied as follows: (1) determine the points for each variable on the top point scale; (2) project each point vertically to the “Points” axis; (3) sum all points to obtain the total score; (4) determine the corresponding 1-, 2-, and 3-year survival probabilities on the bottom survival probability axis. SHAP analysis was further employed to interpret the combined model. The global SHAP beeswarm plot ( Figure 3B ) revealed that all four key predictive features exhibited positive contributions (SHAP values > 0), indicating significant associations with poor prognosis. Individual prediction visualization ( Figure 3C ) illustrated an example of a high-risk patient: the baseline prediction (E[f(x)] = 0) represents the model’s risk reference, while the individual prediction (f(x) = 4.14) was substantially higher, consistent with actual high-risk clinical outcomes.

The combined model demonstrated superior discriminative ability for predicting OS in sRCC patients compared to the sarcopenia-based clinical model and the radiomics model alone. Based on repeated cross-validation, the mean C-indices for the clinical, radiomics, and combined models in the training cohort were 0.746, 0.696, and 0.783, respectively. Pairwise comparisons using Delong’s test indicated a statistically significant difference between the combined model and the radiomics model (p = 0.002), while differences between the clinical and radiomics models (p = 0.081) and between the clinical and combined models (p = 0.216) were not statistically significant.

Time-dependent ROC analysis further validated the predictive accuracy of the models. The sarcopenia-based clinical model achieved AUC values of 0.814 (95% CI: 0.726–0.902), 0.749 (95% CI: 0.651–0.847), and 0.780 (95% CI: 0.684–0.876) for predicting 1-, 2-, and 3-year OS, respectively ( Figure 4A ). The corresponding AUC values for the combined model were 0.849 (95% CI: 0.773–0.926), 0.804 (95% CI: 0.725–0.883), and 0.819 (95% CI: 0.733–0.905) ( Figure 4B ). Calibration curves showed good agreement between predicted and observed survival probabilities for the combined model ( Figure 4C ). DCA indicated that the combined model offered high clinical utility across most threshold probabilities for 1-, 2-, and 3-year survival predictions, with net benefit exceeding those of the “treat-all” and “treat-none” strategies ( Figures 4D–F ), supporting its potential for clinical application.