Hybrid feature selection for dimensionality reduction while preserving predictive performance in breast cancer diagnosis

Globally, Breast Cancer (BC) is the most common type of cancer among women, which is growing at a high rate. Recently, AI/ML-assisted algorithms have most frequently utilised to accurately, efficiently, and early detection of diseases in healthcare. Though many existing techniques rely on high dimensional features and analysis of performance-feature trade-offs are rarely addressed. This work proposes a novel Hybrid Machine Learning (HML) framework to address three key contributions. Firstly, a unified framework is proposed which integrates filter and wrapper-based Feature Selection (FS) approaches to enhance their feature relevance and subset optimization. Secondly, it presents a comprehensive comparative analysis of diverse hybrid FS models to identify optimal solutions for efficient and high-performance models. Third key contribution of this work demonstrates a quantitative analysis which clearly presents that optimized FS can significantly reduce dimensionality while preserves or gives comparable model predictive performance. Experimental results reveal that the proposed RF hybrid models give comparable performance with significant features reduced from 30 to 10. In the same manner, DT with hybrid models give comparable performance by using only 6–8 significant features i.e., 73 to 80% feature reduction. It is demonstrated that RF based approaches achieves accuracy nearly 98% whereas DT based approaches obtain accuracy nearly 95% in a single split test. Moreover, the reliability of results is assessed by using 10-fold cross-validation with 95% confidence interval and Wilcoxon statistical test. Feature importance analysis is demonstrated to strengthen interpretability as significant features contributes more into model decisions and accuracy. Consequently, the proposed hybrid framework achieves an optimal balance between performance, interpretability, and computational efficiency thereby offers a scalable and practical solution for BC diagnostics. However, this work is limited for subgroup specific performance and needs validation on large, independent, and diverge datasets.