Variable step dynamic threshold local binary pattern for classification of atrial fibrillation.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Atrial fibrillation (AF) detection from ECG signals is crucial for cardiovascular health.
• Existing feature extraction methods can be computationally intensive or lack detail.

Purpose of the Study:

• To propose novel, computationally efficient feature extraction methods for AF classification.
• To improve the representation of ECG signal morphology for enhanced classification.

Main Methods:

• Introduced a dynamic threshold for Local Binary Pattern (LBP) code generation for detailed morphological pattern representation.
• Incorporated a variable step value into the LBP algorithm to handle high-frequency ECG signals without downsampling.
• Utilized simple arithmetic operations (addition, division, comparison) for feature extraction, avoiding complex processes like filtering or wavelet transforms.

Main Results:

• Achieved 99.11% sensitivity and 99.29% specificity for AF classification using the MIT-BIH Atrial Fibrillation Database.
• Demonstrated strong performance on the MIT-BIH Arrhythmia Database with 99.38% sensitivity and 98.97% specificity.
• Maintained high accuracy across datasets with varying sampling frequencies.

Conclusions:

• The proposed variable step dynamic threshold LBP method offers a highly accurate and computationally efficient solution for AF detection.
• This method achieves state-of-the-art results in AF classification while being robust to different signal sampling frequencies.