Cardiologist-level interpretable knowledge-fused deep neural network for automatic arrhythmia diagnosis.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Long-term Electrocardiogram (ECG) monitoring is vital for diagnosing arrhythmias, yet challenging in remote areas.
• Digital ECG and AI offer solutions for non-hospital-based arrhythmia diagnosis.
• AI can assist clinicians in diagnosing arrhythmias, improving accessibility.

Purpose of the Study:

• To develop and evaluate a deep learning-based AI model for multi-label arrhythmia diagnosis using a large-scale Chinese ECG dataset.
• To compare the AI model's diagnostic performance against experienced cardiologists and other benchmark models.
• To assess the interpretability and potential clinical utility of the AI diagnostic system.

Main Methods:

• A large-scale Chinese ECG dataset (272,753 patients) was compiled and labeled by expert cardiologists.
• A deep learning, multi-label, interpretable diagnostic model was developed for ECG recordings.
• Model performance was evaluated using Accuracy, F1 score, and AUC-ROC, compared against cardiologists and six other models.

Main Results:

• The AI model achieved an F1 score of 83.51%, mean accuracy of 93.74%, and AUC ROC of 0.977 for 6 common arrhythmias.
• Performance on a hidden dataset surpassed that of expert cardiologists.
• The model demonstrated interpretability, highlighting diagnostic regions in ECGs.

Conclusions:

• The AI diagnosis system exhibits superior performance compared to human clinicians for ECG-based arrhythmia detection.
• The system can aid clinicians in rapidly identifying abnormal ECG regions, boosting diagnostic efficiency and accuracy in China.
• This AI approach shows promise for improving out-of-hospital ECG diagnosis and advancing telemedicine capabilities.