Cost-Sensitive Learning for Anomaly Detection in Imbalanced ECG Data Using Convolutional Neural Networks.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning for arrhythmia detection is crucial for diagnosing cardiac abnormalities.
• Extracting deep features from electrocardiograms (ECGs) is challenging due to inter-patient variability in morphology.
• Imbalanced datasets in ECG analysis lead to overfitting on minority classes, hindering model generalization.

Purpose of the Study:

• To develop a balanced deep feature representation for ECG beats, capturing both short-term and long-term morphological characteristics.
• To address the challenge of imbalanced data in ECG arrhythmia detection.
• To improve the accuracy and generalization of deep learning models for cardiac abnormality diagnosis.

Main Methods:

• Designed a temporal transition module with convolutional layers of varying kernel sizes for efficient ECG feature extraction.
• Proposed a novel, cost-sensitive loss function to mitigate imbalanced data issues by dynamically assigning weights to classes.
• The loss function ensures balanced deep representation by adjusting class weights based on batch distribution and model performance.

Main Results:

• Achieved high classification accuracy: 99.81% for intra-patient and 96.36% for inter-patient heartbeat classification.
• Demonstrated effective mitigation of imbalanced data issues through the proposed cost-sensitive loss function.
• The approach learned a balanced ECG beat representation, outperforming existing studies.

Conclusions:

• The proposed deep learning method effectively learns balanced ECG beat representations by addressing data imbalance.
• The cost-sensitive loss function is key to improving classification performance in arrhythmia detection.
• This approach offers a promising advancement for accurate and generalized cardiac abnormality diagnosis using ECGs.