Selection of effective features for ECG beat recognition based on nonlinear correlations.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electrocardiogram (ECG) beat classification is crucial for diagnosing cardiac conditions.
• Traditional feature selection methods may not effectively handle the complexity and redundancy in ECG data.
• Higher-order statistics (HOS) and wavelet transforms (DWT) combined with neural networks (FFBNN) offer potential for improved ECG analysis.

Purpose of the Study:

• To develop nonlinear correlation-based feature selectors for ECG beat classification.
• To identify the most effective and least redundant features for an HOS-DWT-FFBNN system.
• To enhance the accuracy and efficiency of ECG beat recognition.

Main Methods:

• Proposed three nonlinear correlation-based filters (NCBFs): NCBF1, NCBF2 (feature-feature and feature-class correlation), and SUFCO (feature-class correlation only).
• Compared NCBFs with Relief-F (nonlinear) and linear correlation-based filter (LCBF).
• Evaluated feature discrimination and redundancy quantitatively.

Main Results:

• NCBF1 and NCBF2 achieved higher accuracy (up to 96.34%) with fewer features (as few as eight) compared to SUFCO and Relief-F.
• Features selected by NCBF1/NCBF2 were demonstrably more discriminative and less redundant.
• The NCBF1/NCBF2 approach combined with HOS-DWT-FFBNN outperformed LCBF and other existing heartbeat classifiers, enabling discrimination of more beat types with fewer dimensions.

Conclusions:

• The proposed nonlinear correlation-based feature selection approach is effective for ECG beat recognition.
• The NCBF1/NCBF2 methods demonstrate superiority over other nonlinear and linear feature selection techniques.
• This approach offers a promising direction for developing robust and efficient ECG analysis systems.