An explainable clinical-radiomics machine learning model for preoperative prediction of WHO/ISUP nuclear grade in clear cell renal cell carcinoma.

backgroundAccurate preoperative assessment of the WHO/ISUP nuclear grade of clear cell renal cell carcinoma (ccRCC) is critical for guiding individualized surgical strategies and prognostic evaluation. However, current grading still depends on postoperative pathology. This study aimed to develop and validate a noninvasive, machine learning-based model integrating clinical and radiomic features for predicting WHO/ISUP pathological grade in ccRCC.

methodsA total of 415 patients with pathologically confirmed ccRCC were retrospectively enrolled, including 320 for model development and 95 for external validation. Clinical variables significantly associated with tumor grade in univariate analysis were further analyzed using multivariate logistic regression and machine learning algorithms. Radiomic features were extracted from preoperative CT images and screened using three strategies: least absolute shrinkage and selection operator (LASSO), principal component analysis (PCA), and maximum relevance minimum redundancy (mRMR). Twelve algorithms-including Random Forest (RF), XGBoost, LightGBM, and Support Vector Machine (SVM)-were trained and validated. Model performance was evaluated by area under the curve (AUC), accuracy, sensitivity, specificity, and negative predictive value (NPV). Model interpretability was assessed using SHAP (SHapley Additive Explanations) analysis.

resultsSix independent predictors of WHO/ISUP grade were identified: body mass index (BMI), maximum tumor diameter, perirenal infiltration, serum albumin (SA), osmolality (OSM), and neuron-specific enolase (NSE). Among clinical-only models, RF achieved the best performance (AUC = 0.852, 0.734, and 0889 in training, test, and external validation cohorts, respectively). Incorporating radiomic features further improved prediction accuracy. In the combined LASSO-RF model, the external validation cohort achieved an AUC of 0.912 (95% CI: 0.854-0.970), accuracy of 0.853, sensitivity of 0917, and specificity of 0.831. SHAP analysis revealed that OSM, SA, and NSE were the most influential features, while radiomic texture features such as log-sigma-1-mm-3D first-order maximum also contributed significantly.

conclusionThe LASSO-RF model integrating clinical and radiomic features demonstrated excellent accuracy and generalizability for predicting WHO/ISUP grade in ccRCC. This noninvasive approach enables individualized preoperative grading and risk stratification, providing valuable support for nephron-sparing surgical decision-making and precision oncology in renal cancer management.