ECG-Mamba vs Improved ECG-Mamba: Cardiac Abnormality Classification

Overview

1. Baseline ECG-Mamba - A straightforward Vision Mamba architecture for ECG classification
2. Improved ECG-Mamba - An enhanced version featuring:
   • Lead-Specific Multi-Branch Architecture
   • Mamba-Transformer Hybrid with Attention layers
   • Bi-Directional Scanning capabilities

Noise Handling Comparison

Non-Uniform-Mix vs Contrastive Learning + Masking

• Non-Uniform-Mix Augmentation: A novel augmentation technique specifically designed for ECG signals that applies non-uniform mixing across different leads and time segments
• Contrastive Learning + Masking: Traditional contrastive learning approaches combined with random masking strategies

Reference

Dataset

• 300 records sampled from the full PTB-XL database
• 12-lead ECG recordings
• Multiple cardiac abnormality classes

Dataset Citation

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P. C., Mark, R., ... & Stanley, H. E. (2000).
PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals.
Circulation [Online]. 101 (23), pp. e215–e220.

Features

Baseline ECG-Mamba

• Vision Mamba-based architecture
• Patch-based ECG signal processing
• Efficient sequence modeling with State Space Models (SSM)

Improved ECG-Mamba

• Multi-Branch Architecture: Separate processing branches for different ECG leads
• Hybrid Architecture: Combines Mamba SSM with Transformer attention mechanisms
• Bi-Directional Scanning: Processes ECG signals in both forward and backward directions
• Enhanced feature extraction and representation learning

Requirements

torch
numpy
pandas
wfdb
mamba-ssm
scikit-learn
matplotlib
seaborn
tqdm

Installation

Usage

• Data loading and preprocessing from PTB-XL
• Model architecture definitions
• Training procedures
• Evaluation and comparison metrics
• Visualization of results

Running the Notebook

1. Ensure you have GPU access for optimal performance
2. Install all required dependencies
3. Open ECG_Mamba_Improved_Comparison.ipynb
4. Run all cells sequentially

Model Architecture

ECG-Mamba (Baseline)

• Vision Mamba blocks for sequence modeling
• Patch embedding for ECG signals
• Classification head for cardiac abnormality detection

Improved ECG-Mamba

• Lead-specific processing branches
• Mamba-Transformer hybrid layers
• Bi-directional state space modeling
• Enhanced multi-scale feature fusion

Results

• Classification accuracy
• Confusion matrices
• Per-class performance metrics
• Training curves and convergence analysis
• Noise robustness evaluation: Comparing Non-Uniform-Mix augmentation vs Contrastive Learning + Masking under various noise conditions

Hardware Requirements

• Recommended: GPU with CUDA support
• Minimum RAM: 8GB
• Storage: ~2GB for dataset and model checkpoints

Project Structure

.
├── ECG_Mamba_Improved_Comparison.ipynb  # Main implementation notebook
├── README.md                             # This file
└── LICENSE                               # License information

Contributing

License

Acknowledgments

• PTB-XL dataset from PhysioNet
• Mamba SSM architecture developers
• IEEE JTEHM for publishing the original ECG-Mamba paper

Citation