Comparing Random Survival Forests and Cox Regression for Nonresponders to Neoadjuvant Chemotherapy Among Patients With Breast Cancer: Multicenter Retrospective Cohort Study.

Abstract

Background: Breast cancer is one of the most common malignancies among women worldwide. Patients who do not achieve a pathological complete response (pCR) or a clinical complete response (cCR) post-neoadjuvant chemotherapy (NAC) typically have a worse prognosis compared to those who do achieve these responses.

Objective: This study aimed to develop and validate a random survival forest (RSF) model to predict survival risk in patients with breast cancer who do not achieve a pCR or cCR post-NAC.

Methods: We analyzed patients with no pCR/cCR post-NAC treated at the First Affiliated Hospital of Chongqing Medical University from January 2019 to 2023, with external validation in Duke University and Surveillance, Epidemiology, and End Results (SEER) cohorts. RSF and Cox regression models were compared using the time-dependent area under the curve (AUC), the concordance index (C-index), and risk stratification.

Results: The study cohort included 306 patients with breast cancer, with most aged 40-60 years (204/306, 66.7%). The majority had invasive ductal carcinoma (290/306, 94.8%), with estrogen receptor (ER)+ (182/306, 59.5%), progesterone receptor (PR)- (179/306, 58.5%), and human epidermal growth factor receptor 2 (HER2)+ (94/306, 30.7%) profiles. Most patients presented with T2 (185/306, 60.5%), N1 (142/306, 46.4%), and M0 (295/306, 96.4%) staging (TNM meaning "tumor, node, metastasis"), with 17.6% (54/306) experiencing disease progression during a median follow-up of 25.9 months (IQR 17.2-36.3). External validation using Duke (N=94) and SEER (N=2760) cohorts confirmed consistent patterns in age (40-60 years: 59/94, 63%, vs 1480/2760, 53.6%), HER2+ rates (26/94, 28%, vs 935/2760, 33.9%), and invasive ductal carcinoma prevalence (89/94, 95%, vs 2506/2760, 90.8%). In the internal cohort, the RSF achieved significantly higher time-dependent AUCs compared to Cox regression at 1-year (0.811 vs 0.763), 3-year (0.834 vs 0.783), and 5-year (0.810 vs 0.771) intervals (overall C-index: 0.803, 95% CI 0.747-0.859, vs 0.736, 95% CI 0.673-0.799). External validation confirmed robust generalizability: the Duke cohort showed 1-, 3-, and 5-year AUCs of 0.912, 0.803, and 0.776, respectively, while the SEER cohort maintained consistent performance with AUCs of 0.771, 0.729, and 0.702, respectively. Risk stratification using the RSF identified 25.8% (79/306) high-risk patients and a significantly reduced survival time (P<.001). Notably, the RSF maintained improved net benefits across decision thresholds in decision curve analysis (DCA); similar results were observed in external studies. The RSF model also showed promising performance across different molecular subtypes in all datasets. Based on the RSF predicted scores, patients were stratified into high- and low-risk groups, with notably poorer survival outcomes observed in the high-risk group compared to the low-risk group.

Conclusions: The RSF model, based solely on clinicopathological variables, provides a promising tool for identifying high-risk patients with breast cancer post-NAC. This approach may facilitate personalized treatment strategies and improve patient management in clinical practice.