Robust ECG Classification using Mamba and Self-Supervised Representation Learning.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Electrocardiograms (ECG) are vital for diagnosing heart conditions but are susceptible to noise, complicating automated analysis.
• Current machine learning methods for ECG analysis often require large datasets, leading to poor generalization and noise sensitivity.
• Addressing data scarcity and noise in biosignal analysis is crucial for reliable AI-driven diagnostics.

Purpose of the Study:

• To develop a self-supervised learning framework using a Mamba-based model for ECG denoising.
• To leverage the denoised ECG representations for improved downstream classification tasks.
• To evaluate the model's robustness to noise and compare its performance and computational efficiency against transformer-based methods.

Main Methods:

• Self-supervised pre-training of a Mamba-based neural network on ECG data.
• Incorporating noise augmentation during the pre-training phase.
• Evaluating the model's performance on ECG classification tasks using the learned latent representations.
• Comparative analysis with transformer-based approaches regarding performance and computational cost.

Main Results:

• Self-supervised pre-training significantly enhances the Mamba model's robustness to noise in ECG signals.
• The denoising approach leads to improved classification performance, particularly in noisy environments.
• Optimal noise parameters for pre-training were identified, balancing denoising effectiveness and model generalization.
• The proposed Mamba-based method demonstrated superior performance with reduced computational cost compared to transformer models.

Conclusions:

• Self-supervised pre-training offers an effective strategy for denoising ECG data and improving AI-based diagnostic accuracy.
• Mamba-based models show promise for robust and efficient biosignal processing in challenging, noisy conditions.
• This approach addresses data limitations and enhances the reliability of machine learning in clinical electrocardiography.