Integrating Distance Correlation and Adaptive Weighting with RBF Kernel Transformations: A Novel Feature Selection...

Area of Science:

• Biomedical Informatics
• Machine Learning
• Cardiology

Background:

• Accurate feature selection is vital for machine learning in medical diagnosis.
• Conventional methods struggle with complex non-linear relationships in biomedical data.

Purpose of the Study:

• To introduce an advanced feature selection approach integrating distance correlation and adaptive weighting for enhanced cardiac arrhythmia detection.
• To improve the accuracy of machine learning models in classifying cardiac arrhythmias by addressing limitations of existing feature selection techniques.

Main Methods:

• Proposed a novel feature selection method ranking features by distance correlation.
• Applied inverse penalty weighting to manage feature correlations and RBF kernel transformation with LASSO refinement.
• Evaluated fifteen feature selection techniques on an electrocardiographic database using 4-fold cross-validation and a neural network classifier.

Main Results:

• The proposed method significantly outperformed all alternative techniques, including conventional approaches.
• Demonstrated superior performance in capturing non-linear dependencies and mitigating multicollinearity and overfitting.
• Successfully leveraged synergistic kernel-based interaction modeling with sparse selection for robust feature identification.

Conclusions:

• The combined approach of statistical dependence measures, adaptive regularization, and non-linear transformations offers a robust framework for feature selection.
• This method enhances cardiac arrhythmia classification and has potential applications in broader medical informatics.
• The study highlights the importance of advanced feature selection for accurate medical diagnosis using machine learning.