Towards Automated Optimization of Residual Convolutional Neural Networks for Electrocardiogram Classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Electrocardiogram (ECG) signal interpretation is crucial for assessing heart function and diagnosing arrhythmias.
• Deep neural networks, particularly convolutional neural networks (CNNs), are effective for ECG processing but require extensive hyperparameter tuning.
• Manual hyperparameter optimization is labor-intensive and often suboptimal.

Purpose of the Study:

• To develop an automated hyperparameter optimization method for deep learning models in ECG analysis.
• To optimize a Residual 1D Convolutional Neural Network (R-1D-CNN) architecture for improved arrhythmia classification.
• To evaluate the proposed optimization approach on publicly available ECG datasets.

Main Methods:

• A two-level optimization strategy was employed: manual selection of key hyperparameters and automatic optimization using a proposed Bayesian Optimization (BO) algorithm.
• The architecture, termed BO-R-1D-CNN, integrates residual convolutional layers and multilayer perceptron layers for patient-specific ECG feature learning.
• The model was evaluated on the MIT-BIH database and a dataset of 10,000 ECG patients.

Main Results:

• The BO-R-1D-CNN achieved an exceptional 99.95% accuracy in discriminating five types of heartbeats on the MIT-BIH database.
• The optimized architecture demonstrated superior performance compared to other proposed models on the 10,000 ECG patients dataset.
• The proposed method significantly outperformed previous works on benchmark ECG datasets.

Conclusions:

• Automated hyperparameter tuning using Bayesian Optimization offers an effective solution for optimizing deep learning models in ECG analysis.
• The BO-R-1D-CNN architecture provides a highly accurate and efficient method for arrhythmia classification.
• This approach advances the potential of AI in non-invasive cardiac diagnostics.