Knowledge Graph-Based Feature Selection for Improving Breast Cancer Prognosis

In the last few years, breast cancer has become a major global health challenge since the patient numbers rise daily, which demands advanced diagnostic tools for early detection and accurate prediction. Typically, traditional breast cancer prediction models suffer from high dimensionality issues due to using all the features or inefficient feature selection techniques, resulting in increased computational complexity and reduced classification performance. To overcome the high dimensionality issue in breast cancer prediction, this paper presents a method combining the knowledge-graph-based features, state-of-the-art machine learning, and deep learning techniques. The approach uses a knowledge graph to extract semantically related features, thus eliminating feature redundancy and enhancing model performance. It can reduce the number of features from 30 to 9. Three classifiers, namely Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Bidirectional Long Short-term Memory (BiLSTM), are used for assessing the performance of the proposed feature selection algorithm. The experimental results indicate that the KG-BiLSTM approach produces better classification performance. It reaches an accuracy of $\mathbf{9 9. 7 \%}$, which is higher than that of any other method tested. The results show that using features derived from the knowledge graph outperforms other techniques in identifying relationships within the data and is used to diagnose breast cancer with increased accuracy and reliability.