Deciphering Heartbeat Signatures: A Vision Transformer Approach to Explainable Atrial Fibrillation Detection from ECG Signals | JoVE Visualize

Area of Science:

Cardiology
Artificial Intelligence
Biomedical Engineering

Background:

Wearable single-lead electrocardiogram (ECG) devices combined with artificial intelligence (AI) show promise for early heart disease detection.
Current AI models for ECG analysis often lack interpretability, hindering clinical adoption due to their 'black-box' nature.
Identifying key ECG signal features is crucial for enhancing the reliability and clinical utility of AI diagnostic tools.

Purpose of the Study:

To develop and compare AI models, specifically a vision transformer and a ResNet, for identifying atrial fibrillation using single-lead ECG data.
To enhance the interpretability of AI models by identifying critical ECG signal features contributing to accurate diagnoses.
To classify atrial fibrillation, sinus bradycardia, and normal sinus rhythm heartbeats.

Main Methods:

Developed a vision transformer model and a residual network (ResNet) model for ECG analysis.
Applied both models to the Chapman-Shaoxing dataset for classification tasks.
Utilized the models to identify key ECG signal regions and features influencing classification outcomes.

Main Results:

Both the vision transformer and ResNet models successfully classified atrial fibrillation, sinus bradycardia, and normal sinus rhythm.
The models identified specific ECG regions and features, such as P-waves, T-waves, heartbeat duration, and signal amplitude, as critical for distinguishing between different heart rhythms.
The study demonstrated the potential for AI to provide interpretable insights into ECG diagnostics.

Conclusions:

AI models, including vision transformers and ResNets, can effectively analyze single-lead ECG data for arrhythmia detection.
Interpretable AI approaches can highlight the importance of specific ECG waveform components (P-waves, T-waves) and signal characteristics in diagnosing heart conditions.
This research contributes to the development of more reliable and clinically applicable AI-driven tools for remote patient monitoring and cardiovascular disease diagnosis.