Enhancing fairness in breast cancer recurrence prediction through temporal machine learning models

Breast cancer recurrence prediction is a significant challenge in oncology. Advanced methodologies are required to improve prediction accuracy and clinical decision-making. This study presents a novel approach to breast cancer recurrence prediction by integrating machine learning techniques and a hybrid data mining methodology incorporating a temporal dimension into dataset derivation. Our research is based on the Jordan Breast Cancer Dataset (JBRCA), which includes over 44,000 cases spanning 15 years collected from the King Hussein Cancer Center’s registry database in Amman, Jordan. The proposed methodology encompasses data understanding, preparation, and model development stages. We use a thorough data preparation process involving multicollinearity feature selection, feature scaling, and strategic sampling to address dataset challenges. Moreover, we introduce a temporal-derived dataset strategy, dividing the data into four distinct time intervals to capture evolving characteristics and optimize model relevance. We employ diverse base classifiers and ensemble methods to enhance predictive performance in model development. We use evaluation metrics such as accuracy, recall, specificity, G-mean, and ROC-AUC to assess model efficacy across temporal intervals. Our experimental findings reveal significant impacts on classifier performance with temporal dataset derivation, with notable strengths observed in specific classifiers and temporal intervals. For instance, the Naive Bayes model demonstrates efficacy in identifying recurrence cases, while logistic regression exhibits robust performance in ROC-AUC and G-mean metrics. Our study contributes to breast cancer recurrence prediction by introducing a novel methodology that addresses dataset challenges and leverages temporal insights for enhanced predictive accuracy. The findings have a direct impact on clinical practice, providing valuable tools for early detection and improved therapy planning. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.