De-Redundant Biomedical Feature Extraction Based On Improved MultiSURF

Biomedical data contain valuable information which contributes to the prediction of clinical disease risk. However, the presence of redundant and irrelevant features will diminish prediction accuracy and augment model complexity. According to our research, MultiSURF is a famous feature selection method but falls short in eliminating redundant features. To address this constraint, this paper introduced an improved MultiSURF method known as EN-MultiSURF. The proposed method integrated the strengths of Elastic Net and MultiSURF, which is capable of detecting similarity between features as well as eliminating redundant and irrelevant features. The proposed method was compared with Relief, MultiSURF, Elastic Net, Random Forest and Decision Tree on one binary along with four multiclass numerical datasets. The experimental results substantiated that our method surpassed other methods and exhibited a higher degree of consistency with medical prior knowledge, which established itself as a valuable tool for biomedical data mining.