Enhancing ECG Classification in Cardiac Diagnostics: A Novel Approach Using Adaptive Focal Cross-Entropy Loss...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Heart disease is a leading global cause of mortality.
• Electrocardiograms (ECGs) are crucial for diagnosing heart conditions but traditional analysis faces challenges like time consumption, errors, and class imbalance.
• Existing artificial intelligence (AI) methods for ECG classification have limitations including poor interpretability and high computational costs.

Purpose of the Study:

• To introduce a novel deep learning algorithm, DRA-ECG (Depthwise Separable Residual Attention for ECG analysis).
• To develop a customized Adaptive Focal Cross-Entropy (AFCE) loss function to address class imbalance and enhance model performance.
• To improve the accuracy and reliability of cardiac condition classification using ECG data.

Main Methods:

• Utilized Continuous Wavelet Transform (CWT) to convert 1D ECG signals into 2D scalograms for enhanced feature representation.
• Incorporated edge feature detection as a preprocessing step to denoise and refine features in the 2D scalograms.
• Developed and integrated the DRA-ECG deep learning model with the customized AFCE loss function for classification.

Main Results:

• The proposed DRA-ECG model achieved a high accuracy of 98.17%.
• The model demonstrated strong performance with a recall of 95.78%, F1-score of 95.82%, and precision of 95.89%.
• The combination of CWT, edge detection, DRA-ECG, and AFCE effectively addressed class imbalance and improved feature representation.

Conclusions:

• The novel DRA-ECG algorithm and AFCE loss function offer a significant advancement in automated cardiac condition classification from ECGs.
• This AI-driven approach enhances diagnostic accuracy and efficiency, potentially reducing mortality associated with heart disease.
• The methodology provides a robust solution for handling challenges in ECG analysis, including class imbalance and feature representation.