An Effective LSTM Recurrent Network to Detect Arrhythmia on Imbalanced ECG Dataset

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Cardiovascular disease (CVD) mortality remains high, necessitating improved diagnostic tools.
• Electrocardiogram (ECG) beat analysis is crucial for computer-aided arrhythmia diagnosis systems.
• ECG datasets often exhibit significant class imbalance, complicating accurate classification.

Purpose of the Study:

• To propose an effective deep learning model for ECG beat classification.
• To address the challenge of imbalanced ECG beat data in arrhythmia diagnosis.
• To enhance the accuracy and objectivity of ECG signal interpretation.

Main Methods:

• Utilized a long short-term memory (LSTM) recurrent neural network to capture temporal features in ECG signals.
• Incorporated focal loss (FL) to mitigate the impact of class imbalance by downweighting common normal ECG beats.
• Validated the proposed model using the comprehensive MIT-BIH arrhythmia database.

Main Results:

• The LSTM network with FL demonstrated a reliable solution for imbalanced ECG beat classification.
• The model proved robust and insensitive to variations in ECG signal quality.
• Achieved high accuracy in distinguishing between different types of ECG beats.

Conclusions:

• The proposed LSTM network with focal loss effectively handles imbalanced ECG datasets for arrhythmia diagnosis.
• This method offers a reliable and objective approach to ECG signal analysis.
• The model is suitable for deployment in telemedicine to support cardiologists in diagnosing ECG signals.