MD-CardioNet: A Multi-Dimensional Deep Neural Network for Cardiovascular Disease Diagnosis from Electrocardiogram

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning for electrocardiogram (ECG) analysis is crucial for cardiovascular disease diagnosis.
• Existing methods often fail to integrate inter-channel and temporal information effectively.
• A need exists for comprehensive feature extraction from multi-lead ECG signals.

Purpose of the Study:

• To introduce MD-CardioNet, an efficient deep learning architecture for automated cardiovascular disease classification from ECGs.
• To address limitations in capturing temporal, spatial, and global feature contexts in current deep learning models.
• To develop a method that reduces computational complexity while maintaining high classification performance.

Main Methods:

• Utilized multidimensional (1D, 2D, 3D) convolutions for capturing temporal, spatial, and volumetric features from multi-lead ECG signals.
• Employed sequential feature extractors for time-dependent information and 2D convolutions for inter-channel feature extraction.
• Implemented a knowledge distillation framework to significantly reduce model parameters and computational load.

Main Results:

• Achieved 97.3% accuracy in classifying six distinct heartbeat rhythms on a large public ECG dataset (>10,000 patients).
• Demonstrated superior performance compared to several state-of-the-art approaches.
• Reduced trainable parameters by up to eight times through knowledge distillation, from 4,304,910 to 94,842.

Conclusions:

• MD-CardioNet offers a robust and accurate deep learning solution for cardiovascular disease classification using ECG signals.
• The architecture effectively integrates diverse features, overcoming limitations of previous 1D and 2D-based methods.
• The proposed method provides a computationally efficient and high-performing alternative for clinical applications.