An End-to-End Cardiac Arrhythmia Recognition Method with an Effective DenseNet Model on Imbalanced Datasets Using ECG...

Area of Science:

• Medical Science
• Artificial Intelligence
• Cardiology

Background:

• Electrocardiography (ECG) is crucial for assessing heart rhythm and conditions.
• Automatic arrhythmia diagnosis can significantly reduce physician workload and enhance diagnostic accuracy.
• Existing methods face challenges with imbalanced datasets common in real-world ECG data.

Purpose of the Study:

• To develop and evaluate an end-to-end deep learning model for automatic ECG arrhythmia recognition.
• To address the challenge of class imbalance in ECG datasets.
• To improve the effectiveness and efficiency of arrhythmia diagnosis.

Main Methods:

• An automatic end-to-end 2D Convolutional Neural Network (CNN) deep learning method utilizing a DenseNet model.
• Training and evaluation on large, imbalanced datasets: MIT-BIH arrhythmia and INCART (97,720 and 141,404 beat images).
• Stratified 5-fold cross-validation strategy and classification based on AAMI standards (N, V, S, F types).

Main Results:

• The proposed model achieved high performance, outperforming state-of-the-art methods.
• Accuracy: 99.80% (MIT-BIH) and 99.63% (INCART).
• Precision: 98.34% (MIT-BIH) and 98.94% (INCART).
• F1-score: 98.91% (both datasets).
• Satisfactory results were also obtained using transfer learning on smaller, imbalanced datasets.

Conclusions:

• The proposed 2D CNN DenseNet model is a robust and generalized solution for arrhythmia recognition.
• It effectively handles class-imbalanced ECG datasets, showing promise for real-world clinical applications.
• The model offers a significant advancement in automated cardiac rhythm analysis.