Classification of ECG beats using deep belief network and active learning.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Accurate classification of cardiac irregularities from electrocardiogram (ECG) signals is crucial for patient diagnosis.
• Existing methods may require extensive labeled data or lack efficiency in real-time applications.

Purpose of the Study:

• To develop and validate a semi-supervised deep learning approach for classifying cardiac irregularities using ECG beats.
• To improve the accuracy and sensitivity of arrhythmia detection while minimizing the need for expert labeling.

Main Methods:

• A three-phase model integrating Gaussian-Bernoulli deep belief networks for feature learning and Support Vector Machine (SVM) for classification.
• Incorporation of an active learning query generator to interactively label ECG beats with expert input.
• Adherence to Association for the Advancement of Medical Instrumentation (AAMI) standards for ECG classification.

Main Results:

• Achieved high accuracy rates: 99.5% for Supraventricular Ectopic Beats (SVEB) and 99.4% for Ventricular Ectopic Beats (VEB) on the MIT-BIH Arrhythmia Database.
• Attained 97.5% accuracy for SVEB and 98.6% for VEB on the MIT-BIH Supra-ventricular Arrhythmia Database (SVDB).
• Demonstrated significant accuracy improvements with minimal expert queries and fast online training capabilities.

Conclusions:

• The proposed semi-supervised deep learning and active learning model offers an efficient and accurate method for ECG signal classification.
• This approach effectively identifies cardiac irregularities, showing promise for clinical applications in arrhythmia diagnosis.
• The model's ability to learn from limited expert feedback enhances its practicality and scalability.