Comparison of Random Survival Forest to traditional Cox regression in predicting cancer-specific survival of breast carcinosarcoma

Yalong ZhangYajuan Ren^*

• Longhua Hospital affiliated to Shanghai University of TCM, shanghai, China

This study compares the predictive performance of Random Survival Forest (RSF) and traditional Cox regression for cancer-specific survival (CSS) in breast carcinosarcoma (BS) patients.BS patients were identified from the SEER database and randomly divided into training (70%) and validation (30%) cohorts. Predictor variables were selected using univariable and multivariable Cox proportional hazards analyses. Nomogram, RSF, and Cox proportional hazards models were developed and evaluated using the area under the receiver operating characteristic curve (AUC), decision curve analysis, and calibration curves. Model interpretability was evaluated using time-varying feature importance, partial dependence plots, SurvSHAP(t), and SurvLIME analyses.variables were significantly associated with CSS: age, median household income, year of diagnosis, histological subtype, surgical treatment, radiation therapy, tumor grade, summary stage, and tumor size. The RSF model demonstrated superior discriminative ability, with higher AUC values at 1, 3, and 5 years compared to the Cox model. Decision curve analysis confirmed greater clinical utility for the RSF model. Summary stage, tumor size, age, tumor grade, and surgical treatment were identified as the most influential predictors.The RSF model outperformed Cox regression in predicting CSS for BS patients, offering superior accuracy and interpretable, personalized survival predictions.