Explainable artificial intelligence to detect atrial fibrillation using electrocardiogram.

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Medical Diagnostics

Background:

• Early detection of atrial fibrillation (AF) is crucial for preventing mortality.
• Current deep learning models (DLMs) for AF detection lack clinical interpretability.
• Explainable AI is needed to bridge the gap between DLMs and clinical practice.

Purpose of the Study:

• To develop an explainable deep learning model (DLM) for detecting atrial fibrillation (AF) using electrocardiograms (ECGs).
• To validate the performance and interpretability of the developed DLM across diverse ECG formats.
• To enhance the transparency of DLMs for potential clinical application in AF detection.

Main Methods:

• A retrospective study utilizing the Sejong ECG dataset (128,399 ECGs) for model development and internal validation.
• Development of a DLM with feature modules for decision interpretability.
• External validation using large datasets: PTB-XL (21,837 ECGs), Chapman (10,605 ECGs), and PhysioNet (8,528 ECGs).

Main Results:

• The explainable DLM achieved high diagnostic performance with Area Under the Curve (AUC) values of 0.997-0.999 for 12-lead ECGs.
• Excellent AUCs (0.990-0.999) were observed for 6-lead and single-lead ECGs.
• High AUCs (0.961-0.993) confirmed the explainability of features like rhythm irregularity and absence of P-wave.

Conclusions:

• The developed explainable DLM effectively detects AF from various ECG formats.
• The model provides interpretable insights into its detection mechanisms.
• This approach demonstrates the potential of explainable AI to enhance the clinical adoption of DLMs for ECG analysis.