Performance of a convolutional neural network derived from an ECG database in recognizing myocardial infarction

Area of Science:

• Medical Technology
• Artificial Intelligence
• Cardiology

Background:

• Medical technology is rapidly advancing, with artificial intelligence (AI) showing significant potential in image analysis.
• Electrocardiogram (ECG) diagnosis involves analyzing 2D waveform images, similar to other image analysis tasks.

Purpose of the Study:

• To investigate the capability of an AI, specifically a convolutional neural network (CNN), to accurately recognize myocardial infarction (MI) from ECG images.
• To compare the MI recognition performance of the developed CNN model against that of human physicians.

Main Methods:

• A 6-layer CNN model was developed and trained using the PTB ECG database, which includes 148 myocardial infarction (MI) cases.
• The CNN model and 10 physicians were tested on a separate set of ECGs to evaluate their MI recognition capabilities using F1 score and accuracy metrics.
• The impact of removing Goldberger-leads and ECG image compression on the CNN's recognition performance was also assessed.

Main Results:

• The CNN model achieved significantly higher F1 scores (83±4%) and accuracy (81±4%) compared to physicians (F1: 70±7%, accuracy: 67±7%, P<0.0001).
• Performance degradation was not significant even when Goldberger-leads were removed or ECG images were compressed to quarter resolution.

Conclusions:

• Deep learning with a simple CNN can achieve physician-level capability in recognizing myocardial infarction (MI) on ECG images.
• Further research is warranted to explore the broader application of AI in ECG image assessment.