Arrhythmia detection with transfer learning architecture integrating the developed optimization algorithm and...

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence

Background:

• Electrocardiography (ECG) is crucial for diagnosing heart rhythm abnormalities but can be time-consuming and challenging, especially in pediatric patients due to unique ECG patterns.
• Existing methods for ECG analysis, while effective, often depend on clinician expertise and may lack generalizability across diverse datasets.

Purpose of the Study:

• To develop and validate an optimized deep learning model for accurate and stable classification of pediatric ECG data.
• To improve the generalizability of ECG analysis models for pediatric patients by integrating novel optimization and regularization techniques.

Main Methods:

• A custom pediatric ECG dataset was created, comprising 1318 abnormal and 1403 normal beats.
• The MobileNetv2 transfer learning architecture was enhanced with Proposed Optimization Algorithm V5 and Proposed Regularization Method V5.
• The model was evaluated on both a balanced 2-category dataset and a more complex 6-category dataset.

Main Results:

• The enhanced MobileNetv2 model achieved training and test accuracies of 0.9801 and 0.9509 on the 2-category dataset, outperforming the original architecture (0.9633 and 0.9399).
• On a 6-category dataset, the model achieved training and test accuracies of 0.9200% and 0.8975%, demonstrating acceptable performance and generalizability.
• The proposed optimization and regularization methods contributed to improved stability and accuracy in ECG classification.

Conclusions:

• The integrated MobileNetv2 model with Proposed Optimization Algorithm V5 and Proposed Regularization Method V5 offers a robust solution for pediatric ECG analysis.
• The findings suggest that the developed approach can be generalized to more complex and diverse ECG datasets, aiding in the accurate detection of cardiac arrhythmias.