A Spectral-longitudinal Model for Detection of Heart Attack from 12-lead Electrocardiogram Waveforms

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Cardiovascular diseases (CVDs) cause millions of deaths yearly, with myocardial infarction (MI) being the most common.
• Data-driven methods for CVD prediction from ECG exist, but integrating multiple ECG leads remains underexplored.

Purpose of the Study:

• To propose an end-to-end trainable spectrallongitudinal model for predicting heart attack (MI).
• To leverage data-level fusion of multiple ECG leads for improved diagnostic accuracy.

Main Methods:

• Developed a joint spectrallongitudinal model combining spectral analysis and recurrent neural networks.
• Transformed time-series ECG data into stacked spectrograms to capture frequency-time characteristics.
• Utilized recurrent networks to model temporal dependencies within ECG waveforms.

Main Results:

• The spectrallongitudinal model demonstrated superior performance in predicting MI compared to baseline methods.
• Data-level fusion of multiple ECG leads significantly enhanced diagnostic capabilities.
• The model effectively integrated spectral and temporal information for accurate heart attack prediction.

Conclusions:

• The proposed spectrallongitudinal model offers a powerful new approach for diagnosing myocardial infarction.
• Integrating multiple ECG leads through data fusion enhances CVD prediction accuracy.
• This method shows promise for improving early detection and management of heart attacks.