Evaluating Temporal Dynamics in Breast Cancer Survival Predictions with Machine Learning and Cox Regression Analysis

Accurate prediction of breast cancer-specific survival is crucial for guiding personalized treatment decisions and improving patient outcomes. This study evaluated the performance of machine learning approaches (Random Survival Forest, RSF and Generalized Boosted Model, GBM) alongside traditional Cox proportional hazards models for predicting survival in 21,574 women diagnosed with stage I-IV breast cancer in New Zealand between 2000-2019. Performance comparisons using time-dependent Area Under the Curve and Brier score metrics demonstrated that RSF consistently outperformed both Cox regression variants and GBM across all time points. Distinct differences emerged in survival predictions between modelling approaches: RSF captured a sharper initial decline in survival for most tumour receptor subtypes and better differentiated the favourable prognosis of ER+/HER2- tumours compared to other subtypes. Notably, variable importance analysis revealed fundamentally different prognostic emphases between modelling approaches—disease stage dominated Cox model predictions while tumour receptor subtype most strongly influenced RSF predictions. These findings highlight how machine learning approaches can capture complex, nonlinear relationships between clinical variables and survival outcomes that may be missed by traditional statistical models. The complementary insights provided by different modelling approaches suggest potential value in their combined use for enhanced risk stratification and more tailored treatment planning in breast cancer management, particularly when accounting for tumour biological characteristics alongside conventional staging factors.