Clear cell renal cell carcinoma (ccRCC) is the most common subtype of kidney cancer, and approximately 30% of patients develop metastatic disease. Bone is one of the most frequent sites of metastasis, occurring in up to 35% of metastatic RCC patients, and is associated with significant morbidity including pathological fractures, spinal cord compression, and severe pain.
Despite the clinical importance of bone metastasis in RCC, robust prognostic tools for predicting survival in this specific population are lacking. Clinicians currently rely on general metastatic RCC prognostic scores, which may not adequately capture the unique biology and clinical trajectory of patients with bone involvement.
This study aimed to develop and validate machine learning models capable of predicting 1-year and 3-year overall survival (OS) specifically in ccRCC patients with bone metastasis, using a large population-based dataset supplemented by real-world clinical data from China.
The primary dataset was drawn from the SEER (Surveillance, Epidemiology, and End Results) database, a large US population-based cancer registry. After applying inclusion and exclusion criteria, 1,490 ccRCC patients with bone metastasis were identified. This cohort was split into training and internal validation sets.
For external validation, 42 patients with ccRCC and bone metastasis treated at hospitals in China were enrolled prospectively. This external cohort allowed testing of whether models trained on a US population could generalize to a different patient population and healthcare system, an important test of real-world applicability.
Four machine learning algorithms were compared: XGBoost (XGB), logistic regression (LR), random forest (RF), and naive Bayes (NB). Each model was trained to predict binary outcomes: survival versus death at 1 year and at 3 years following diagnosis of bone metastasis.
Candidate variables included demographic factors (age, sex, marital status), tumor characteristics (histologic grade, T stage, N stage, tumor size), and metastatic burden indicators (presence of brain, liver, and lung metastases in addition to bone). Treatment-related variables such as whether surgery was performed were also included.
Multivariate analysis identified the independent predictors that were incorporated into the final models. These included age at diagnosis, marital status, histologic grade, T stage, N stage, tumor size, presence of brain metastasis, liver metastasis, lung metastasis, and receipt of surgical treatment.
The inclusion of extraskeletal metastatic sites (brain, liver, lung) as predictors reflects the importance of overall metastatic burden in determining survival outcomes. Patients with bone-only metastasis have meaningfully different prognoses than those with concurrent visceral metastases.
Among the four algorithms, XGBoost achieved the best performance for predicting 1-year overall survival. In the training set, XGBoost attained an AUC of 0.891, demonstrating strong discrimination between patients who would survive one year and those who would not.
In internal validation, the AUC dropped to 0.711, and in the external Chinese cohort, the AUC was 0.812. The external validation AUC was notably higher than the internal validation AUC, which may reflect the smaller size and more homogeneous composition of the external cohort, or differences in patient selection between registries and clinical practice.
For 3-year survival prediction, model performance varied more across algorithms. Random forest and XGBoost again tended to outperform logistic regression and naive Bayes, consistent with findings in other survival prediction studies that demonstrate the advantage of ensemble methods over linear models for complex clinical outcomes.
Marital status emerged as a statistically significant predictor of survival, a finding that, while initially surprising, is consistent with a broader literature showing that social support influences cancer outcomes. Married patients may have better access to care, greater treatment adherence, and stronger emotional support networks.
Receipt of surgical treatment was also an independent predictor of improved survival, even in the setting of metastatic disease. This likely reflects patient selection, as patients fit enough to undergo surgery tend to have better performance status and less aggressive disease, rather than a direct therapeutic benefit of surgery in all bone metastasis patients.
The identification of modifiable factors such as social support and treatment access alongside tumor biology variables enriches the prognostic model and points to potential targets for intervention beyond oncologic therapy alone.
Based on the identified predictors, the study also developed nomograms, which are graphical tools that translate complex multivariate statistical models into a simple point-based scoring system. Clinicians can use these nomograms at the bedside to estimate individual patient survival probabilities without needing access to specialized software.
The practical utility of these tools lies in treatment planning and patient counseling. For patients predicted to have poor 1-year survival, aggressive local interventions such as surgery or radiation may be de-emphasized in favor of systemic therapy or palliative care. For those with predicted better outcomes, more active treatment may be warranted.
Limitations include the retrospective nature of the SEER data, potential coding inconsistencies in registry data, and the small size of the external validation cohort. Prospective validation in larger independent cohorts would be needed before clinical implementation of these models.