Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Electrocardiogram (ECG) is a critical non-invasive tool for diagnosing cardiovascular diseases.
• Increasing ECG examinations and a shortage of cardiologists necessitate automated diagnostic solutions.
• Accurate, automated ECG analysis is a significant research focus in digital health.

Purpose of the Study:

• To develop and evaluate a deep neural network for automatic cardiac arrhythmia classification using 12-lead ECG data.
• To compare the deep learning model's performance against traditional machine learning methods.
• To identify the most informative ECG leads for arrhythmia detection.

Main Methods:

• Development of a deep neural network architecture for classifying arrhythmias from 12-lead ECG signals.
• Training and validation on a public 12-lead ECG dataset.
• Comparative analysis with four established machine learning algorithms utilizing expert-extracted features.
• Application of SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The proposed deep neural network achieved an average F1 score of 0.813 for arrhythmia classification.
• The deep learning model demonstrated superior performance compared to four traditional machine learning methods.
• Utilizing all 12 leads simultaneously yielded better results than single-lead ECG models.
• Leads I, aVR, and V5 were identified as the most effective leads for classification.

Conclusions:

• Deep neural networks offer a powerful approach for automated cardiac arrhythmia detection from 12-lead ECGs.
• The developed model provides a reliable and effective tool for cardiovascular disease diagnosis.
• Interpretable AI methods like SHAP enhance trust and understanding of deep learning models in clinical settings.