Explainable Artificial Intelligence Model for Diagnosis of Atrial Fibrillation Using Holter Electrocardiogram...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Atrial fibrillation (AF) is a significant cardiac arrhythmia.
• Existing machine learning models for AF diagnosis lack explainability for physicians.
• Need for transparent AI tools in clinical decision-making for arrhythmias.

Purpose of the Study:

• To develop and validate an explainable AI (XAI) model for accurate atrial fibrillation diagnosis.
• To enhance physician understanding of AI-driven diagnostic results using XAI.
• To leverage convolutional neural networks (CNN) and gradient-weighted class activation mapping (Grad-CAM) for AF detection.

Main Methods:

• Utilized a dataset of 57,273 Holter ECG waveform slots (30 seconds each) from January 2016 to October 2019.
• Developed a CNN-based AI model incorporating the Grad-CAM technique for explainability.
• Validated the model's performance using standard metrics and cardiologist annotations.

Main Results:

• Achieved high diagnostic performance: sensitivity 97.1%, specificity 94.5%, accuracy 95.3%.
• Area Under the Curve (AUC) for AF detection was 0.988, indicating strong discriminative power.
• XAI analysis confirmed that 94.5% of AI-identified regions of interest corresponded to cardiologists' identified characteristic AF sites.

Conclusions:

• The developed CNN-based XAI model accurately diagnoses atrial fibrillation from Holter ECG data.
• The model provides interpretable diagnostic insights, facilitating physician trust and adoption.
• This study represents a significant advancement towards practical XAI-based diagnostic tools for AF.