High-performance next-generation screening technology system using teaching–learning-based optimization for cancer detection

Abstract

Early and accurate detection of breast cancer remains challenging due to high-dimensional clinical data, feature redundancy, and limitations of existing diagnostic decision systems. To address these issues, this study proposes a hybrid feature selection and classification framework that improves prediction performance while reducing overfitting. The method combines a filter-based ReliefF algorithm with a wrapper-based Teaching–Learning-Based Optimisation (TLBO), referred to as Re-wTLBO, to identify a compact and informative feature subset. Furthermore, an enhanced TLBO incorporating inertia weight is integrated with a Support Vector Machine (SVM) classifier, enabling simultaneous optimisation of feature selection and model parameters using SVM accuracy as the fitness function. This hybrid optimisation overcomes the shortcomings of conventional evolutionary wrappers and improves early-stage cancer prediction. An experimental evaluation of the Wisconsin Breast Cancer Dataset (WBCD) demonstrates that the proposed approach outperforms traditional wrapper-based methods, achieving an average accuracy of 99.30%, a sensitivity of 97.12%, a specificity of 98.01%, an F-measure of 98.23%, and an AUROC of 99.02%. These results confirm the effectiveness of the proposed framework for reliable and high-precision breast cancer diagnosis on the datasets and protocol examined.