A Neural Network-Based Prediction for Time-to-Metastasis in Breast Cancer of Later Occurrence

Abstract

Metastatic breast cancer (MBC) poses a significant global health concern, with increasing incidence rates, notably in regions like Nigeria. Despite advancements in diagnosis and prognosis, challenges persist in accurately predicting MBC, particularly when dealing with imbalanced datasets. This study introduces a neural network-based model for MBC prediction, integrating feature importance measures (FIM) such as the chi-square filter, Jaya algorithm wrapper, and gini-index random forest embedded, alongside data imbalance handling techniques including oversampling (ROS), under sampling (RUS), and synthetic minority oversampling technique (SMOTE). Initially, the three FIM methods were used on the original unbalanced dataset to independently select the top 10 features from a pool of 24 features in a 5-year MBC dataset. The selected features from each FIM method were then passed to the neural network classifier. Among these methods, the chi-square consistently demonstrated superior performance in accuracy, F1-score, and sensitivity metrics. Subsequently, RUS, ROS, and SMOTE were applied to balance the selected dataset subsets, including all features. Extensive experiments revealed that utilizing all 24 features with SMOTE consistently yielded superior performance across all metrics with significant margins, highlighting the importance of comprehensive FIM strategies and holistic data imbalance handling methods for enhancing BCM prediction.