cT1b renal cell carcinoma refers to kidney cancers measuring between 4 and 7 centimeters that are confirmed by imaging to be confined to the kidney. This is considered an intermediate stage, neither the smallest localized tumors nor the more advanced disease, and it represents a group where survival outcomes vary considerably between patients.
Standard survival prediction for kidney cancer relies on the American Joint Committee on Cancer (AJCC) TNM staging system, which classifies tumors based on size, lymph node involvement, and distant spread. While useful, TNM staging does not incorporate many factors known to affect how individual patients fare, such as age, overall health, tumor grade, or whether surgery was performed.
This study used a Random Survival Forest (RSF), a type of machine learning algorithm, to build a more personalized survival prediction model for cT1b RCC patients. The model was trained on data from over 22,000 patients and compared directly against conventional TNM-based predictions.
The key question the researchers aimed to answer was whether a machine learning model incorporating multiple patient characteristics could better predict which cT1b RCC patients would survive 5 and 10 years after diagnosis than the staging system currently used in clinical practice.
The study drew on the Surveillance, Epidemiology, and End Results (SEER) database, a nationally representative cancer registry maintained by the United States National Cancer Institute. SEER contains information on cancer diagnoses, treatments, and survival outcomes for millions of patients across the United States, dating back to 1973.
Using SEER allowed the researchers to study a very large, diverse population of cT1b RCC patients diagnosed between 2010 and 2019. After applying eligibility criteria to ensure data quality, the final study cohort included 22,426 patients. This large sample size provides statistical power to identify meaningful predictors and build a robust prediction model.
The primary outcome measured was overall survival (OS), defined as the time from diagnosis until death from any cause. A secondary outcome was cancer-specific survival (CSS), which counts only deaths directly attributable to kidney cancer. Comparing these two outcomes helps distinguish cancer-related mortality from deaths caused by other health conditions.
Variables available in the SEER database and included in the analysis encompassed patient demographics such as age, sex, race, and marital status, as well as tumor characteristics including histological type, grade, and laterality (which kidney was affected), and treatment information including type of surgery performed.
A Random Survival Forest is an extension of the random forest machine learning approach adapted for survival data, where the outcome includes both whether an event (death) occurred and when it occurred. Unlike traditional statistical methods, RSF does not make assumptions about the mathematical form of the relationship between predictors and survival.
The RSF works by building hundreds of individual decision trees, each trained on a random subset of the patient data and a random subset of the predictor variables. Each tree independently predicts survival based on different combinations of features. The final prediction is produced by averaging the results from all trees, a process that reduces the influence of any single noisy or unrepresentative tree.
To prevent the model from memorizing the training data rather than learning general patterns, the researchers used out-of-bag validation, where each tree is evaluated on the portion of data it was not trained on. The dataset was also split into a 70% training cohort and a 30% testing cohort to provide an independent evaluation of model performance.
Model performance was measured using the area under the receiver operating characteristic curve (AUC) at 5 and 10 years, as well as the concordance index (C-index), which measures how well the model ranks patients by survival time. These are standard benchmarks for evaluating survival prediction models in oncology.
SHAP (SHapley Additive exPlanations) analysis was applied to interpret the model and identify which variables had the greatest influence on predictions. SHAP assigns each feature a contribution score for each individual patient, making it possible to understand both the overall importance of each variable and how it affects survival in specific patients.
The RSF model achieved an AUC of 0.746 for 5-year overall survival and 0.742 for 10-year overall survival in the testing cohort. These values indicate good discriminatory ability, meaning the model reliably distinguishes patients who will survive from those who will not significantly better than chance.
When compared directly to the AJCC TNM staging system, the RSF model showed meaningfully higher AUC values. The TNM system, which classifies all cT1b patients into the same stage category, has very limited ability to discriminate survival outcomes within this group. The RSF model, by incorporating additional clinical variables, substantially improves on this baseline.
Calibration analysis, which tests whether predicted survival probabilities match actual observed survival rates, confirmed that the RSF model was well-calibrated. Patients predicted to have a 70% chance of 5-year survival did indeed survive at approximately that rate in the testing data, an important property for any tool intended to inform clinical decisions.
The model also performed well for cancer-specific survival, with AUC values comparable to those for overall survival. This suggests that the model captures factors specifically relevant to kidney cancer mortality rather than simply reflecting general health status.
SHAP analysis identified age at diagnosis as the single most influential predictor of overall survival in cT1b RCC patients. Older patients faced substantially higher mortality risk, driven in large part by competing non-cancer causes of death. This reflects the reality that many patients with smaller kidney tumors may ultimately die from cardiovascular disease or other age-related conditions rather than from their cancer.
Tumor size within the 4-to-7 centimeter range also emerged as a significant predictor, with larger tumors associated with worse outcomes even within this narrow size category. Tumor grade was another top predictor, confirming that the aggressiveness of cancer cells independently affects survival beyond what tumor size alone can capture.
Interestingly, marital status appeared among the top predictors. Married patients had better survival outcomes than unmarried patients. This finding, consistent with prior cancer research, likely reflects the social support, earlier symptom recognition, and improved treatment adherence that accompany partnership and close social relationships.
The type of surgery performed was also predictive, with patients who underwent partial nephrectomy (removing only the tumor and a margin of healthy tissue) showing better survival than those who had their entire kidney removed. This likely reflects both patient selection factors and the known benefits of preserving kidney function for long-term health.
For patients with cT1b kidney cancer, the decision between different surgical approaches, the intensity of follow-up surveillance, and the consideration of additional treatments after surgery depend heavily on individual risk. A more accurate survival prediction tool gives physicians and patients better information for making these decisions together.
The RSF model's ability to incorporate multiple variables simultaneously means it can identify high-risk patients within a stage category that conventional staging would treat uniformly. An older patient with a high-grade tumor who is unmarried may need more intensive monitoring and earlier intervention than a younger, healthy patient with a lower-grade tumor of the same size.
The finding that marital status affects survival highlights that cancer outcomes are influenced by social and behavioral factors alongside biological ones. Healthcare teams should consider social support assessment as part of comprehensive cancer care, potentially connecting at-risk patients with social workers, support groups, or other resources that can improve adherence and outcomes.
A practical nomogram tool was developed from the model to make individualized survival predictions accessible in clinical practice without requiring specialized software. Physicians can enter a patient's characteristics and receive a predicted probability of 5- and 10-year survival, which can then be incorporated into shared decision-making conversations.
This study demonstrates that a Random Survival Forest model trained on large real-world data can meaningfully improve survival prediction for intermediate-stage kidney cancer compared to the standard staging approach currently used in clinical practice.
The integration of SHAP analysis ensures that the model is not simply a black box but provides interpretable, clinically meaningful insights into which patient characteristics drive survival outcomes. This transparency is essential for building trust among clinicians and for identifying targets for intervention.
Limitations of the study include its retrospective nature and reliance on a single national database, which may not fully represent all patient populations. Future work should validate the model in prospective cohorts from multiple countries and explore whether adding molecular tumor characteristics, such as genetic mutations or biomarker profiles, further improves prediction accuracy.
For patients and families navigating an intermediate-stage kidney cancer diagnosis, this research represents progress toward more personalized, evidence-based guidance. Understanding individual risk more precisely enables more informed conversations about treatment options, follow-up schedules, and long-term health planning.