Recurrence Plot-Based Approach for Cardiac Arrhythmia Classification Using Inception-ResNet-v2.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiac arrhythmia (CA) classification from electrocardiography (ECG) data is crucial for diagnosing heart conditions.
• Current deep learning (DL) methods, primarily 1D CNNs analyzing time-domain ECG signals, lack sufficient robustness and satisfactory performance.
• There is a need for advanced analytical techniques to improve the accuracy and reliability of CA detection.

Purpose of the Study:

• To develop a more robust and accurate method for classifying cardiac arrhythmias using a 2D deep learning approach.
• To introduce a novel technique combining recurrence plots (RP) with the Inception-ResNet-v2 network for ECG data analysis.
• To evaluate the proposed 2D RP-based method's performance against existing 1D ECG-based methods and traditional 2D transformation techniques.

Main Methods:

• Selected optimal ECG leads (lead II and aVR) from 12-lead data.
• Transformed 1D ECG segments into 2D texture images using the recurrence plot (RP) method.
• Utilized the Inception-ResNet-v2 network to classify nine types of cardiac arrhythmias using the generated 2D RP images as input.

Main Results:

• The 2D RP-based method achieved high average F1-scores across multiple datasets (e.g., 0.8862 on PTB_XL).
• The proposed method demonstrated excellent generalization ability on independent ECG databases.
• Outperformed existing 1D ECG-based works and traditional 2D transformation methods (time waveform, CWT) with an average F1-score of 0.844, using only two ECG leads.

Conclusions:

• The 2D recurrence plot combined with Inception-ResNet-v2 offers a highly effective and robust approach for cardiac arrhythmia classification.
• This method shows significant potential for clinical application, enabling accurate CA detection with a reduced number of ECG leads.
• The findings highlight the advantages of transforming 1D ECG signals into 2D representations for enhanced deep learning-based analysis.