Deep learning-mediated prediction of concealed accessory pathway based on sinus rhythmic electrocardiograms

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Concealed accessory pathways (APs) can cause atrial ventricular reentrant tachycardia.
• These APs are often asymptomatic and undetectable during normal sinus rhythm, posing a diagnostic challenge.

Purpose of the Study:

• To develop and evaluate a deep learning model for detecting concealed APs from standard electrocardiogram (ECG) images.
• To assess the efficacy of various convolutional neural networks (CNNs) for this diagnostic task.

Main Methods:

• Collected normal sinus rhythm ECG images from patients with concealed APs and healthy controls.
• Trained and tested six popular CNNs (ResNet26, SE-ResNet50, MobileNetV3_large_100, DenseNet169, etc.) using both ImageNet pre-training and random initialization.
• Utilized separate training and testing datasets for model validation.

Main Results:

• Deep learning models achieved high diagnostic performance, with specific models reaching >87.0% sensitivity and >98.0% specificity.
• No significant differences in PR and QRS intervals were observed between groups, but QT and QTc intervals showed slight variations.
• Models trained from random initialization demonstrated comparable performance to those using ImageNet pre-training.

Conclusions:

• Deep learning presents an effective strategy for predicting concealed APs using routine ECGs.
• The findings support the potential of training deep learning models from random initialization for ECG-based diagnostic tasks.