Hybrid Deep Learning and Discrete Wavelet Transform-Based ECG Biometric Recognition for Arrhythmic Patients and...

Area of Science:

• Biometrics
• Signal Processing
• Machine Learning

Background:

• Electrocardiogram (ECG) signals are an emerging biometric modality due to their intrinsic and liveness detection properties.
• Challenges include low recognition performance with large, diverse datasets and short ECG signal intervals.

Purpose of the Study:

• To propose a novel method for feature-level fusion of discrete wavelet transform (DWT) and a 1D convolutional recurrent neural network (1D-CRNN) for enhanced ECG-based biometric recognition.
• To address the challenge of low recognition performance in large population datasets with short ECG signal intervals.

Main Methods:

• ECG signal preprocessing: removal of powerline interference, low-pass filtering (1.5 Hz cutoff), and baseline drift removal.
• Feature extraction: conventional using Coiflets' 5 DWT on PQRST-segmented signals and deep learning-based using a 1D-CRNN (2 LSTM, 3 convolutional layers).
• Feature-level fusion of conventional and deep learning features.

Main Results:

• Biometric recognition accuracies achieved: 80.64% (ECG-ID), 98.81% (MIT-BIH), and 99.62% (NSR-DB).
• An accuracy of 98.24% was achieved when combining all datasets.
• The proposed fusion method outperformed conventional and deep learning-based feature extraction alone, as well as transfer learning approaches (VGG-19, ResNet-152, Inception-v3) on short ECG segments.

Conclusions:

• The proposed feature-level fusion method significantly enhances ECG biometric recognition accuracy.
• This approach is effective even with short ECG signal segments and diverse datasets, offering a robust solution for forensic, surveillance, and security applications.