[Automatic classification method of arrhythmia based on discriminative deep belief networks]

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Traditional arrhythmia classification relies on manual ECG feature selection, leading to subjectivity and complexity.
• This impacts the overall accuracy of identifying cardiac arrhythmias.

Purpose of the Study:

• To propose a new method for automatic arrhythmia classification using discriminative deep belief networks (DDBNs).
• To enhance feature extraction and classification accuracy in ECG signal analysis.

Main Methods:

• Automatic extraction of morphological and RR interval features using generative restricted Boltzmann machines (GRBM).
• Implementation of discriminative restricted Boltzmann machines (DRBM) for feature learning and classification.
• Conversion of DDBNs to deep neural networks (DNNs) with a Softmax regression layer for fine-tuning via backpropagation.

Main Results:

• Achieved 99.84% ± 0.04% classification accuracy on consistent data sources from the MIT-BIH Arrhythmia Database.
• Reached 99.31% ± 0.23% accuracy on inconsistent data sources, enhanced by active learning (AL).

Conclusions:

• The DDBN method demonstrates significant effectiveness in automatic ECG feature extraction and arrhythmia classification.
• Presents a novel deep learning solution for automated ECG analysis and arrhythmia detection.