Abstract PS1-13-04: Fairness-constrained logistic regression achieves superior performance and racial equity in breast cancer survival prediction

Abstract Machine learning models for clinical prediction often exhibit racial bias, resulting in inferior predictive accuracy and potentially poorer outcomes for African American (AA) breast cancer patients compared to non-AA patients. Such bias can result in delayed diagnosis, inappropriate treatment recommendations, and reduced access to clinical trials, thereby compounding existing survival disparities. Standard optimization approaches prioritize overall performance, risking the amplification of these inequities. We hypothesized that fairness-constrained hyperparameter optimization could enhance overall predictive accuracy while improving equity in AA-specific breast cancer survival prediction. We conducted a controlled study comparing standard logistic regression with a fairness-constrained version for 5-year breast cancer survival prediction using TCGA data (N=1,000, 18.6% AA patients, 70.8% survival rate). The standard model optimized overall predictive accuracy, while the fairness-constrained model also aimed to reduce racial disparities in predictions. Both approaches used identical 10-fold stratified cross-validation, enabling direct comparison. Primary endpoints were overall AUC and AA-specific AUC. The secondary endpoint was AUC difference between non-AA and AA patients as a disparity metric. Statistical analysis used paired t-tests with bootstrap confidence intervals (n=1,000 resamples) and effect size calculation via Cohen's d. Fairness-constrained logistic regression significantly improved clinical performance and racial equity compared to standard optimization. Overall AUC increased from 0.634 (95% CI: 0.582-0.664) in the control group to 0.641 (95% CI: 0.592-0.669) with fairness constraints, yielding a +0.007 AUC improvement (95% CI: +0.002 to +0.013, p=0.018, Cohen's d=0.917). AA-specific performance demonstrated even greater enhancement, with AUC increasing from 0.610 (95% CI: 0.518-0.687) to 0.632 (95% CI: 0.528-0.713), representing a +0.021 AUC improvement (95% CI: +0.003 to +0.040, p=0.027, Cohen's d=0.832). Most importantly, racial disparity was reduced, with AUC difference between non-AA and AA patients improving from +0.029 (indicating disparity favoring non-AA patients) to -0.012 (slight advantage for AA patients), representing a 41-point equity improvement. Non-AA performance was maintained at 0.639 AUC (95% CI: 0.596-0.678) with no significant change (p=0.276), demonstrating that equity gains can be achieved without compromising care for the majority population. Incorporating fairness-constrained hyperparameter optimization into predictive models can simultaneously improve overall accuracy and reduce racial disparities in breast cancer survival prediction. These findings, based on analyses only recently completed, provide new and timely evidence that equitable machine learning is both feasible and clinically advantageous. By demonstrating that fairness constraints can improve African American-specific outcomes without compromising performance in the majority population, this work represents a potential paradigm shift toward more equitable, precision-guided oncology care. Citation Format: J. Zeineh, R. DeAngel, E. Grullon, T. J. Lawton, K. J. Bloom. Fairness-constrained logistic regression achieves superior performance and racial equity in breast cancer survival prediction abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS1-13-04.