Survival prediction in triple-negative breast cancer: a Cox model with fairness assessment using ISO/IEC TR...

Triple-negative breast cancer (TNBC) is an aggressive subtype with limited therapeutic options and poor survival outcomes. Prognostic models developed in Western cohorts rarely assess algorithmic fairness. This study aimed to develop and internally validate a clinically interpretable Cox survival model for TNBC using baseline diagnostic variables and to evaluate its fairness according to ISO/IEC TR 24027:2021 guidelines in a Middle East and North Africa (MENA) cohort.

A total of 138 TNBC patients were included after merging two institutional datasets and removing variables with > 25% missingness. Baseline features comprised age, tumor size, lymph node involvement, tumor grade, Ki-67, type of surgery, metastasis at diagnosis, chemotherapy, and radiotherapy. A Cox proportional hazards (CoxPH) model with six clinically established predictors was fitted to reduce overfitting. Model performance was assessed through five-fold stratified cross-validation using Harrell's concordance index (C-index), receiver operating characteristic area under the curve (AUROC), and calibration curves. Fairness was evaluated using demographic parity, equality of opportunity, predictive equality, and equalized odds metrics following ISO/IEC TR 24027:2021.

During follow-up, 34 patients (24.6%) died. Metastasis at diagnosis, high tumor grade, and radical mastectomy were significantly associated with mortality. The CoxPH model achieved a C-index of 0.80 [SE = 0.04; 95% confidence interval (CI): 0.72-0.87] and an AUROC of 0.81 (95% CI: 0.72-0.90). Calibration plots showed strong agreement between predicted and observed survival probabilities, with a modest overall bias of -8.8%. Fairness assessment revealed small but notable disparities in false-positive rates across age groups and surgical categories, while lymph node status and other variables showed no significant bias.

This study presents a robust and fairness-aware survival prediction model for TNBC using routinely available clinical features. The model demonstrates strong discrimination, good calibration, and quantifiable fairness across patient subgroups, offering a clinically interpretable and ethically aligned tool to support TNBC risk stratification and decision-making in the MENA region.