Renal cell carcinoma (RCC) is one of the most common urologic cancers worldwide, and predicting which patients will experience recurrence or reduced survival after surgery remains a major clinical challenge. Traditional prognostic tools rely on a limited set of clinical variables and often lack the precision needed to guide individualized treatment decisions.
Machine learning (ML) offers a promising alternative by identifying complex, nonlinear relationships among many variables simultaneously. In this study, researchers applied nine ML algorithms to the Korean Renal Cell Carcinoma (KORCC) database, which contains data from 10,068 patients across 8 institutions, making it one of the largest RCC cohorts assembled for this purpose.
The primary goals were to predict two key outcomes: disease recurrence and overall survival following surgical treatment. By benchmarking multiple algorithms against each other, the team aimed to identify which approach offers the best performance and to determine which clinical features matter most for each prediction task.
The KORCC database served as the primary training set, with external validation performed using 3,169 patients from Seoul National University Bundang Hospital (SNUBH). All patients had undergone surgical resection for RCC, and detailed clinical, pathological, and laboratory variables were collected.
Feature selection was performed using elastic net regularization, a technique that combines L1 (lasso) and L2 (ridge) penalties to identify the most informative variables while reducing overfitting. This process selected 9 variables for the recurrence prediction model and 13 variables for the survival prediction model.
A major challenge in this dataset was class imbalance: patients who experienced recurrence or death were far fewer than those who did not. To address this, the team applied SMOTE (Synthetic Minority Over-sampling Technique), which generates synthetic examples of the minority class to create a more balanced training distribution and prevent models from defaulting to the majority class.
The study evaluated nine ML algorithms spanning a range of complexity and approach: logistic regression (LR), support vector machine (SVM), decision tree (DT), random forest (RF), naive Bayes (NB), XGBoost, NGBoost, LightGBM, and CatBoost. Each was trained and tested using the same feature sets to allow fair comparison.
Gradient boosting methods (XGBoost, LightGBM, CatBoost, NGBoost) are ensemble techniques that build decision trees sequentially, each correcting the errors of the previous one. Random forest is a parallel ensemble of decision trees trained on random subsets of data. These ensemble methods generally outperform simpler models on tabular clinical data.
Model performance was assessed using accuracy, F1-score, area under the receiver operating characteristic curve (AUC-ROC), and calibration metrics. The external validation cohort from SNUBH allowed testing of how well models generalize to patients from a different institution, which is critical before clinical deployment.
Across both prediction tasks, the ML models achieved strong performance metrics. Accuracy ranged from 0.77 to 0.94, and F1-scores ranged from 0.77 to 0.97, with ensemble methods generally outperforming simpler classifiers. The best-performing models for recurrence prediction were gradient boosting variants.
For overall survival prediction, models again achieved high performance, with XGBoost and LightGBM consistently ranking near the top. The use of SMOTE meaningfully improved recall for the minority class (recurrence or death events), ensuring the models were not simply predicting that every patient would survive without recurrence.
External validation on the SNUBH cohort confirmed that performance was maintained outside the original training institutions, lending credibility to the generalizability of the selected models. This is a particularly important finding given the multi-institutional origin of the KORCC database.
The elastic net feature selection revealed distinct sets of variables most predictive of recurrence versus survival. For recurrence, pathologic T stage, tumor size, and histologic grade were among the strongest predictors, consistent with established clinical knowledge. Lymphovascular invasion also emerged as an important feature.
For survival prediction, the model required a broader set of 13 variables, suggesting that overall survival is influenced by a wider range of factors including patient comorbidities and systemic markers, not just tumor-specific features. This reflects the complexity of survival as an endpoint that captures both cancer-related and non-cancer-related mortality.
The agreement between ML-selected features and clinically recognized prognostic factors lends biological credibility to the models and supports the idea that these algorithms are capturing meaningful signal rather than random noise in the data.
The ability to accurately predict recurrence and survival after RCC surgery has direct implications for postoperative surveillance planning. Patients identified as high-risk could be monitored more intensively or enrolled in adjuvant therapy trials, while those at low risk might be spared the burden of frequent imaging and follow-up visits.
The KORCC database represents an unusually rich resource due to its size and multi-institutional design, which reduces the risk of institutional bias. The external validation at SNUBH further strengthens the case that these models could be adopted in real-world Korean clinical settings with confidence.
Future work should explore incorporating genomic and molecular data, which could improve predictive accuracy beyond what is achievable with clinical and pathological variables alone. Integration of these models into electronic health record systems would also be a critical step toward practical clinical deployment.