Deep learning-based model detects atrial septal defects from electrocardiography: a cross-sectional multicenter...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Atrial septal defect (ASD) poses a risk for adverse cardiovascular outcomes.
• ASD is often underdiagnosed due to subtle clinical presentation.
• Early detection is crucial for timely intervention and risk mitigation.

Purpose of the Study:

• To develop and validate a novel screening strategy for early detection of ASD.
• To implement a convolutional neural network (CNN) for identifying ASD from 12-lead electrocardiography (ECG).
• To bridge the diagnostic gap for underdiagnosed ASD cases.

Main Methods:

• Collected 671,201 ECGs from 80,947 patients across 3 international hospitals.
• Trained a CNN model on ECG data, excluding cases with closed ASD.
• Validated the model's performance using derivation, validation, and external test datasets.

Main Results:

• The CNN model achieved an area under the receiver operating characteristic curve (AUROC) of 0.85-0.90 for ASD detection.
• Demonstrated excellent generalizability across diverse patient subgroups and institutions.
• Screening simulation showed a significant increase in sensitivity (80.6% to 93.7%) at 33.6% specificity compared to overt ECG abnormalities.

Conclusions:

• A CNN-based model effectively identifies ASD using 12-lead ECG.
• The model exhibits strong generalizability across different institutions and continents.
• This AI-driven approach offers a promising tool for enhanced ASD screening and diagnosis.