Automatic Triage of 12-Lead ECGs Using Deep Convolutional Neural Networks

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Medical Diagnostics

Background:

• Conventional computerized electrocardiogram (ECG) interpretation struggles to achieve physician-level accuracy for cardiac abnormalities.
• Automated ECG analysis is crucial for timely diagnosis and treatment of cardiac conditions.

Purpose of the Study:

• To develop and validate a deep neural network (DNN) for comprehensive automated ECG triage in clinical practice.
• To assess the DNN's ability to classify ECGs into four distinct triage categories.

Main Methods:

• A 37-layer convolutional residual DNN was developed using 336,835 free-text physician-annotated 12-lead ECGs.
• The DNN was trained on a large dataset and validated on 984 independent ECGs annotated by cardiologists.
• The algorithm classified ECGs into normal, abnormal not acute, subacute, and acute categories.

Main Results:

• The DNN demonstrated excellent overall discrimination with a concordance statistic of 0.93.
• A polytomous discriminatory index of 0.83 indicated strong performance in classifying ECGs.
• The model achieved high accuracy in differentiating between the four triage categories.

Conclusions:

• An end-to-end DNN can be accurately trained on unstructured ECG annotations for consistent triage.
• This deep learning-based ECG interpretation has the potential to enhance time to treatment.
• Further validation and fine-tuning could significantly decrease healthcare burden and improve patient outcomes.