Enhancing convolutional neural network predictions of electrocardiograms with left ventricular dysfunction using a...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning (DL) in electrocardiogram (ECG) analysis shows potential for diagnosing cardiovascular abnormalities.
• Traditional ECG DL models often use full waveforms, leading to feature learning redundancies and prediction inaccuracies.

Purpose of the Study:

• To introduce a novel sub-waveform representation for ECG DL, focusing on a data-centric approach.
• To enhance prediction accuracy and reliability in cardiovascular diagnostics using ECG data.

Main Methods:

• Developed a sub-waveform representation leveraging ECG rhythmic patterns.
• Applied the representation to a large cohort of 92,446 patients for left ventricular dysfunction identification.
• Utilized convolutional neural networks (CNNs) with the proposed ECG representation.

Main Results:

• The sub-waveform representation significantly improved performance metrics compared to full-waveform analysis.
• Observed a 2% increase in area under the receiver operating characteristic curve and a 10% increase in area under the precision-recall curve.
• Demonstrated enhanced explainability, interpretability, and fairness, with reduced prediction uncertainties.

Conclusions:

• The sub-waveform ECG representation offers improved DL modeling for cardiovascular AI.
• This approach provides better control over data granularity, advancing AI in cardiology.
• Expectation of improved DL performance for future cardiovascular artificial intelligence technologies.