Comparing performance of iterative and non-iterative algorithms on various feature schemes for arrhythmia analysis

Area of Science:

• Cardiology
• Machine Learning
• Signal Processing

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing cardiac conditions.
• Machine learning algorithms offer potential for automated ECG interpretation.
• Evaluating feature sets and algorithms is key to improving diagnostic accuracy.

Purpose of the Study:

• To assess the performance of classic machine learning algorithms (SVM, LS-SVM, RF, NB) for ECG classification.
• To determine the effectiveness of various feature schemes in ECG analysis.
• To compare iterative and non-iterative algorithms in classifying normal, Atrial Fibrillation (AF), and ST change ECG recordings.

Main Methods:

• ECG recordings were filtered using a 0.1 Hz - 12 Hz bandpass filter.
• 80 features (time, frequency, time-frequency, PCA) were extracted into six feature schemes.
• Four algorithms (RF, SVM, LS-SVM, NB) were applied to binary (normal/AF) and tri-classification (normal/AF/ST change) tasks.

Main Results:

• Time domain, frequency domain, and PCA features provided reliable combinations for RF and SVM.
• Random Forest (RF) achieved the highest F1-scores: 0.8908 for binary and 0.7535 for tri-classification.
• RF outperformed SVM, LS-SVM, and NB in both classification tasks.

Conclusions:

• Random Forest (RF) demonstrates superior performance for ECG classification tasks.
• Feature combinations including time, frequency, and PCA domains are effective.
• Machine learning, particularly RF, holds significant promise for automated cardiac condition detection from ECGs.