Performance of a Convolutional Neural Network and Explainability Technique for 12-Lead Electrocardiogram...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Machine Learning for Healthcare

Background:

• Automated preliminary electrocardiogram (ECG) interpretation is crucial for millions of clinicians.
• A lack of critical comparisons exists between machine learning (ML)-based automated analysis and current clinical standards of care.

Purpose of the Study:

• To train and apply an explainability technique to a convolutional neural network (CNN).
• To utilize readily available 12-lead ECG data for CNN training.
• To achieve high performance against clinical standards of care for automated ECG interpretation.

Main Methods:

• A cross-sectional study using 992,748 ECGs from 365,009 adult patients (2003-2018).
• A CNN was trained to predict 38 diagnostic classes from 12-lead ECG data.
• The Linear Interpretable Model-Agnostic Explanations (LIME) technique was used for CNN explainability.

Main Results:

• The CNN achieved an area under the receiver operating characteristic curve (AUC) of at least 0.960 for 84.2% of diagnostic classes.
• CNN demonstrated higher frequency-weighted F1 scores than cardiologists and MUSE automated analysis across all 5 diagnostic categories.
• LIME highlighted physiologically relevant ECG segments contributing to CNN diagnoses.

Conclusions:

• Readily available ECG data can train CNN algorithms to match cardiologist performance and exceed MUSE automated analysis for most diagnoses.
• The LIME technique provides interpretable insights into CNN diagnostic reasoning for ECGs.
• This approach offers a promising advancement in automated ECG interpretation.