Combining Low-dimensional Wavelet Features and Support Vector Machine for Arrhythmia Beat Classification.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Accurate electrocardiogram (ECG) beat recognition is crucial for diagnosing cardiac conditions.
• Effective feature extraction is a prerequisite for high-performance ECG classification.
• Existing methods may struggle with dimensionality and feature quality.

Purpose of the Study:

• To develop an effective method for low-dimensional ECG beat feature vector extraction.
• To apply these features for accurate classification of abnormal ECG beats.
• To evaluate the performance of the proposed method on a large ECG dataset.

Main Methods:

• Wavelet multi-resolution analysis for time-frequency domain feature extraction.
• Principal Component Analysis (PCA) for feature vector dimension reduction.
• One-versus-one Support Vector Machine (SVM) classifier with 10-fold cross-validation.

Main Results:

• The proposed method achieved high performance metrics: 99.09% sensitivity, 99.82% specificity, and 99.70% accuracy using a beat-based training scheme.
• Performance varied significantly between beat-based and record-based training schemes.
• The method was tested on 107,049 beats from the MIT-BIH Arrhythmia Database.

Conclusions:

• The developed method demonstrates high efficacy in abnormal ECG beat recognition and classification.
• Wavelet analysis combined with PCA provides effective low-dimensional feature representation for ECG data.
• The choice of training scheme significantly impacts classification performance in ECG analysis.