Hybrid and advanced convolution techniques for highly accurate, reliable and automated ECG classification

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Cardiac arrhythmias are a major global cause of death.
• Automated analysis of electrocardiogram (ECG) data is crucial for clinical diagnosis.
• Existing methods may lack accuracy or efficiency for real-world applications.

Purpose of the Study:

• To develop a hybrid 1D CNN architecture for accurate automated ECG heartbeat classification.
• To optimize the model for low-power edge deployment on hardware accelerators.
• To evaluate the model's performance and generalization capabilities.

Main Methods:

• A hybrid 1D CNN incorporating Dynamic Kernel Switching (DKS), Self-Attention, and Dilated Convolutions.
• Evaluation of three quantization paths (Dynamic PTQ, Full-Integer PTQ, QAT) for edge deployment.
• Rigorous testing using the MIT-BIH Arrhythmia Database with stratified cross-validation.

Main Results:

• Achieved 98.99% test accuracy and a macro F1-score of 0.9484 on the MIT-BIH dataset.
• Demonstrated robust performance with 98.93% accuracy via five-fold cross-validation.
• Model size reduced to ≈0.32 MB with Dynamic PTQ, suitable for edge devices.

Conclusions:

• The proposed hybrid 1D CNN offers a highly accurate and efficient solution for automated ECG arrhythmia classification.
• The model's low-power characteristics and small footprint enable practical edge deployment.
• This advancement supports improved cardiovascular monitoring and diagnosis through AI.