Classification of Imbalanced Electrocardiosignal Data using Convolutional Neural Network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Wearable heart monitors generate electrocardiogram (ECG) signals crucial for anomaly detection.
• ECG arrhythmia classification is challenging due to imbalanced datasets and time-series signal characteristics, particularly for minority classes.

Purpose of the Study:

• To present a novel method for imbalanced ECG arrhythmia classification.
• To address the difficulties in recognizing minority classes within imbalanced ECG datasets.

Main Methods:

• An improved data augmentation technique combining variational auto-encoder (VAE) and auxiliary classifier generative adversarial network (ACGAN) was implemented.
• Convolutional neural network (CNN) classifiers were utilized on the augmented dataset for automated arrhythmia recognition using 2D ECG images.

Main Results:

• The proposed method achieved 98.45% accuracy and 97.03% sensitivity on the MIT-BIH arrhythmia database.
• Sensitivities for two minority classes reached 95.83% and 97.37%, demonstrating effective recognition of underrepresented data.

Conclusions:

• Sensitivity of the minority class is a critical metric in imbalanced classification.
• The proposed method significantly enhances minority class sensitivity in ECG arrhythmia classification.
• Experimental results show superior performance compared to traditional augmentation and classification methods for imbalanced ECG data.