Higher-Order Spectral Analysis Combined with a Convolution Neural Network for Atrial Fibrillation...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Medicine

Background:

• The global prevalence of atrial fibrillation (AFIB) is rising, posing significant public health challenges.
• Early and accurate detection of AFIB remains a critical unmet need, driving research into advanced prediction and management methods.
• Electrocardiogram (ECG) signals, while informative, are inherently non-stationary, non-linear, and non-Gaussian, complicating traditional analysis.

Purpose of the Study:

• To propose and evaluate a novel method for automatic atrial fibrillation (AFIB) detection using higher-order spectral analysis of ECG signals.
• To investigate the efficacy of bispectrum analysis, a higher-order spectral technique, in capturing essential features from complex ECG signals.
• To compare the performance of a proposed Convolutional Neural Network (CNN), AFIB-NET, against a pre-trained GoogLeNet model for AFIB classification.

Main Methods:

• Utilized higher-order spectra analysis, specifically the bispectrum, to extract features from non-stationary, non-linear ECG signals.
• Converted 2D bispectrum images into input data for two distinct CNN architectures: AFIB-NET and a modified GoogLeNet.
• Evaluated the performance using the MIT-BIH Atrial Fibrillation Database (AFDB), focusing on sensitivity, specificity, and Area Under the ROC Curve (AUC).

Main Results:

• The proposed AFIB-NET achieved a sensitivity of 95.3%, specificity of 93.7%, and an AUC of 98.3% for AFIB detection.
• The modified GoogLeNet model demonstrated a sensitivity of 96.7%, specificity of 82%, and an AUC of 96.7%.
• Preliminary findings indicate that bispectrum images are effective inputs for 2D CNNs in detecting AFIB rhythms.

Conclusions:

• Bispectrum analysis combined with 2D CNNs offers a promising approach for the accurate detection of atrial fibrillation.
• The proposed AFIB-NET model shows competitive performance, highlighting the potential of tailored CNN architectures for biomedical signal analysis.
• This methodology contributes to advancing automatic AFIB prediction and management strategies, addressing a key public health concern.