A Denoising and Fourier Transformation-Based Spectrograms in ECG Classification Using Convolutional Neural Network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Electrocardiogram (ECG) signals are vital for assessing heart conditions and diagnosing cardiac diseases.
• Traditional ECG interpretation is complex, time-consuming, and requires specialized expertise due to large data volumes.
• Advancements in neural networks show promise for efficient biomedical signal interpretation, including ECG and electroencephalogram (EEG).

Purpose of the Study:

• To investigate the efficacy of using spectrograms, derived from ECG signals, for improved cardiac disease classification.
• To develop a simplified Convolutional Neural Network (CNN) architecture for accurate and computationally efficient ECG analysis.
• To compare the diagnostic performance of spectrogram-based ECG analysis against traditional raw signal analysis.

Main Methods:

• ECG data from the PTB-XL dataset were processed into both raw signal and spectrogram formats.
• Spectrograms were generated using Short-Time Fourier Transformation (STFT) and data reduction via frequency filtration.
• A simplified CNN model was trained and evaluated on both raw signal and spectrogram datasets for binary classification.

Main Results:

• The proposed approach utilizing spectrograms achieved a highest accuracy of 99.06% in ECG classification.
• Spectrogram-based analysis demonstrated superior performance compared to raw ECG signal analysis.
• The method reduced memory usage and computational power requirements by employing a simpler CNN architecture.

Conclusions:

• Spectrograms offer a more effective representation of ECG data for cardiac disease classification than raw signals.
• The developed CNN approach provides a computationally efficient and highly accurate method for ECG interpretation.
• This technique holds potential for improving the accessibility and speed of cardiac diagnostics.