Abstract 4913: PREDICT Breast v4: Advancing equity in breast cancer prognosis for diverse populations

Abstract Breast cancer, the most common cancer among women globally, remains a significant public health challenge. Prognosis, influenced by factors such as tumor characteristics, patient demographics, and treatment choices, is a major concern for patients. Tools like the PREDICT Breast v3 model (https://breast.v3.predict.cam/) help weigh treatment benefits against short- and long-term risks, enabling more personalised prognostic predictions. However, its performance disparities among non-Hispanic Black and non-Hispanic Asian patients in the US highlight the need for more equitable and inclusive tools. Therefore, this study aims to develop a breast cancer prognostic model that improves predictive accuracy across diverse populations and reduces disparities in prognostic outcomes. Here, we analysed data from 273, 918 breast cancer patients in the Surveillance, Epidemiology, and End Results (SEER) database, each with clinical records and socioeconomic status (SES) information. The data were split 7:3 for model training and validation, with 10-fold cross-validation applied for parameter tuning. We first evaluated both conventional methods (PREDICT v3) and advanced machine learning models, including survival XGBoost, survival random forest, DeepHit, and DeepSurv, using calibration and discrimination as evaluation metrics. After comparing the models, we selected the DeepHit neural network as the best-performing approach due to its lowest calibration error and highest discrimination. We then examined whether incorporating SES improved performance by comparing SES-adjusted and unadjusted DeepHit models. Our results showed that the SES-adjusted DeepHit model achieved strong calibration (within 1.3% of observed deaths) and reduced racial disparities compared to PREDICT v3, which exhibited calibration errors of 30% for non-Hispanic Asians and 20% for non-Hispanic Blacks. These findings emphasise the importance of incorporating SES factors to enhance model accuracy and equity, underscoring the critical need for fairness in prognostic prediction. This work advances the application of machine learning in cancer prognosis and provides a foundation for addressing health disparities in clinical prediction tools. Citation Format: Yi-Wen Hsiao, Gordon C. Wishart, Paul D.P. Pharoah, Pei-Chen Peng. PREDICT Breast v4: Advancing equity in breast cancer prognosis for diverse populations [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4913.