Lesion distribution and prognosis of renal cell carcinoma bone metastasis: A novel evaluation model based on random survival forests

Abstract The prognostic value of lesion distribution in renal cell carcinoma bone metastasis (RCC‐BM) is unclear. This study aimed to quantify the association between BM distribution and prognosis in RCC‐BM patients and to employ a predictive model based on the random survival forests (RSF) algorithm. At first BM diagnosis, 122 patients were stratified by Memorial Sloan‐Kettering Cancer Center (MSKCC)/Motzer risk score and classified into locoregional (21.3%), stochastic (56.6%), and extensive (22.1%) groups based on bone lesion distribution. Spinal and pelvic involvement was observed in 39.3% and 35.2% of patients. Univariate, logistic regression, and Kaplan–Meier survival analyses indicated that locoregional spread, spinal involvement (odds ratio [OR] 3.30; 95% confidence interval [CI] 1.20–9.09), and advanced age (OR 1.04; 95% CI 1.00–1.08; p < .05) correlated with higher risk stratifications, while pelvic metastasis was linked to shorter median overall survival (32 vs. 49 months; p < .05). The RSF model was trained in 70% and validated in 30% of the series, incorporating spatial lesion involvement (pelvic, spinal, and upper extremities involvement), MSKCC/Motzer score, and age as principal contributing variables. Time‐dependent area under the curve (AUC) values achieved in single‐split validation for 1‐ and 3‐year survival were 0.90 and 0.87. Consistent performance was observed across 100 repeated splits, with median AUCs of 0.89, 0.86, and 0.89 for 1‐, 3‐, and 5‐year survival, respectively. A cut‐off value of 15.03 effectively separated high‐ and low‐risk groups ( p < .05). RSF demonstrated superior accuracy over Cox regression (median AUC 0.89 vs. 0.59 for 1‐year survival). Overall, integrating bone lesion patterns into RSF modeling facilitates personalized prognosis and supports more precise care in RCC‐BM.