Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Computerized electrocardiogram (ECG) interpretation is vital for clinical workflows.
• Deep learning (DL) offers potential to enhance automated ECG analysis accuracy and scalability.
• A comprehensive evaluation of end-to-end DL for diverse ECG diagnostic classes was previously unreported.

Purpose of the Study:

• To develop and evaluate a deep neural network (DNN) for classifying 12 distinct cardiac rhythm classes from single-lead ECGs.
• To compare the diagnostic performance of the DNN against board-certified cardiologists.

Main Methods:

• A DNN was developed to analyze 91,232 single-lead ECGs from 53,549 patients.
• The DNN was validated on an independent test dataset annotated by expert cardiologists.
• Performance metrics included area under the receiver operating characteristic curve (ROC) and F1 score.

Main Results:

• The DNN achieved an average ROC of 0.97.
• The DNN's average F1 score (0.837) surpassed that of average cardiologists (0.780).
• At equivalent specificity, the DNN demonstrated superior sensitivity across all rhythm classes compared to cardiologists.

Conclusions:

• An end-to-end deep learning approach can classify a wide range of arrhythmias from single-lead ECGs with performance comparable to cardiologists.
• This AI-driven method may reduce misdiagnoses in computerized ECG interpretations.
• It could improve efficiency by triaging urgent cases, optimizing expert human interpretation.