Heartbeat classification using feature selection driven by database generalization criteria.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Accurate heartbeat classification is crucial for diagnosing cardiac conditions.
• Existing electrocardiogram (ECG) classifiers often struggle with generalization across different datasets.
• Developing robust ECG analysis tools is essential for remote patient monitoring and clinical decision support.

Purpose of the Study:

• To develop and validate a simple heartbeat classifier utilizing ECG feature models.
• To enhance the generalization capability of the classifier across diverse patient populations and recording conditions.
• To identify a parsimonious set of features that yield high classification performance.

Main Methods:

• Feature extraction from RR series, ECG samples, and wavelet transform at multiple scales and leads.
• Utilized a floating feature selection algorithm to identify optimal feature subsets.
• Trained and validated the classifier on publicly available databases: MIT-BIH Arrhythmia, MIT-BIH Supraventricular Arrhythmia, and INCART.
• Adhered to Association for the Advancement of Medical Instrumentation (AAMI) standards for labeling and reporting.

Main Results:

• The best model comprised eight features, achieving 93% global accuracy on a disjoint test set of the MIT-BIH Arrhythmia database.
• Achieved high sensitivity (S) and positive predictive value (P(+)) for normal (S: 95%, P(+): 98%) and ventricular beats (S: 81%, P(+): 87%).
• Demonstrated comparable performance on the INCART database, indicating strong generalization capability.

Conclusions:

• The developed eight-feature ECG heartbeat classifier exhibits superior performance and generalization compared to state-of-the-art methods.
• The model's simplicity and effectiveness suggest its potential for practical clinical application.
• The findings support the use of selected ECG features and robust validation strategies for reliable cardiac rhythm analysis.