A hybrid method for heartbeat classification via convolutional neural networks, multilayer perceptrons and focal loss

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Heart arrhythmia is a significant cardiovascular disease requiring advanced diagnostic methods.
• Traditional methods struggle with inter-patient variations in heartbeat morphology.
• The need for accurate and automated heartbeat classification is growing.

Purpose of the Study:

• To propose a hybrid deep learning model for accurate heartbeat classification.
• To address challenges posed by imbalanced heartbeat classes and inter-patient variations.
• To enhance the performance of automated cardiac monitoring systems.

Main Methods:

• A hybrid approach combining convolutional neural networks (CNNs) for morphological feature extraction and multilayer perceptrons (MLPs) for classification.
• Integration of RR interval features with morphological features to capture dynamic heartbeat information.
• Utilization of focal loss to mitigate performance degradation due to imbalanced heartbeat classes.

Main Results:

• The hybrid method achieved high performance metrics on the MIT-BIH arrhythmia database.
• Achieved an accuracy of 96.27%, sensitivity of 68.55%, positive predictive value of 64.68%, and F1-score of 66.09%.
• Demonstrated significant performance improvements compared to existing heartbeat classification methods.

Conclusions:

• The proposed hybrid deep learning method is effective and simple for heartbeat classification.
• This approach holds potential for real-time, personal automatic heartbeat classification in remote medical monitoring.
• Open-source code is available for further research and application.