Intelligent Deep Models Based on Scalograms of Electrocardiogram Signals for Biometrics.

Area of Science:

• Biometrics
• Deep Learning
• Signal Processing

Background:

• Electrocardiogram (ECG) biometrics are susceptible to noise.
• Transforming ECG signals to a frequency domain using wavelets creates a 2D scalogram, enabling multiresolution analysis.
• This transformation complicates morphological analysis, potentially hindering simple classifiers.

Purpose of the Study:

• To investigate the efficacy of ECG scalograms as input for deep convolutional neural networks (CNNs).
• To compare the performance of different deep learning models (AlexNet, GoogLeNet, ResNet) for ECG-based biometrics.
• To evaluate the use of transfer learning with pre-trained deep models for efficient ECG classification.

Main Methods:

• Comparative analysis of deep learning models: AlexNet, GoogLeNet, and ResNet.
• Utilizing scalograms (absolute values of continuous wavelet transform coefficients) of ECG signals as input.
• Experimentation on two ECG databases: Physikalisch-Technische Bundesanstalt (PTB)-ECG and Chosun University (CU)-ECG.

Main Results:

• ResNet models achieved higher classification accuracy than AlexNet and GoogLeNet on both PTB-ECG and CU-ECG databases.
• ResNet performance was 0.73%-0.27% higher on PTB-ECG and 0.94%-0.12% higher on CU-ECG compared to the other models.
• Deep learning models, particularly ResNet, show promise for analyzing complex ECG scalogram data.

Conclusions:

• ECG scalograms are effective inputs for deep learning models in biometrics.
• ResNet architecture offers superior performance for ECG-based biometric classification compared to AlexNet and GoogLeNet.
• Deep learning, especially with transfer learning, provides a robust approach for noisy ECG signal analysis.