Predicting overall and breast cancer-specific survival: machine learning models and prognostic nomograms from a population-based registry

Abstract Background Breast cancer remains a leading cause of cancer-related mortality worldwide, with pronounced regional inequalities in middle-income countries such as Brazil. In Mato Grosso, structural disparities and limited access to specialized care pose additional challenges for accurate prognostic assessment. This study aimed to develop and internally validate Machine Learning (ML) models and prognostic nomograms to predict overall survival (OS) and breast cancer-specific survival (CSS), as well as to stratify patient risk in Mato Grosso, Brazil. Methods A retrospective population-based cohort study was performed using data from the Population-Based Cancer Registry (RCBP) (2001-2018), deterministically linked to the Mortality Information System (SIM) (2001-2022). Five survival prediction models were developed and internally validated, including four ML survival models, Random Survival Forest (RSF), Gradient Boosting Survival Analysis (GBS), Survival Support Vector Machine (SSVM), and ComponentWise Gradient Boosting (CWGB), and the conventional Cox proportional hazards (CPH) model. The dataset was split into training and test sets (80:20, stratified). Model performance was assessed using the concordance index (C-index), time-dependent area under the curve (AUCt), and integrated Brier score (IBS), with bootstrap confidence intervals. The highest-performing model was interpreted using SHAP values and permutation-based feature importance. Cox-based nomograms were then constructed and internally validated using 1,000 bootstrap resamples, followed by risk stratification and decision curve analysis to evaluate clinical utility. Results GBS demonstrated the highest overall performance for both OS and breast CSS, with strong discrimination and good calibration. For OS, the C-index reached 0.781 (95% CI 0.776-0.786), with an AUCt of 0.843 and an IBS of 0.129; for CSS, the C-index reached 0.790 (0.784-0.796), with an AUCt of 0.852 and an IBS of 0.120. CPH showed comparable discrimination but weaker calibration overall. The nomograms maintained stable discrimination at 1-, 3-, 5-, and 10-year time points and demonstrated good agreement between predicted and observed outcomes. Both ML models and Cox-derived nomograms achieved significant risk stratification (log-rank P< 0.001). SHAP analyses identified diagnostic method, tumor morphology, marital status, education level, and age as key predictors. Conclusions In this study, ML models, particularly GBS, and Cox-based nomograms demonstrated individualized prognostic capability for OS and breast CSS. These approaches may support effective risk stratification, facilitating the identification of patients at higher mortality risk and providing a scalable framework for risk-adapted management in real-world public health settings.