Convolutional Neural Network for Individual Identification Using Phase Space Reconstruction of Electrocardiogram.

Area of Science:

• Biometrics
• Machine Learning
• Signal Processing

Background:

• Electrocardiogram (ECG) biometrics offer unique individual identification based on cardiac electrical activity.
• Convolutional neural networks (CNNs) enhance ECG biometrics by extracting discriminative features.
• Phase space reconstruction (PSR) transforms ECG signals into feature maps without precise R-peak alignment.

Purpose of the Study:

• To develop a PSR-based CNN for ECG biometric authentication.
• To investigate the impact of time delay and grid partition parameters on identification accuracy.

Main Methods:

• Utilized PSR with varying time delays and grid partitions on ECG data from 115 subjects (PTB Diagnostic ECG Database).
• Developed and evaluated a CNN model incorporating PSR features.
• Compared performance across different parameter settings and network sizes.

Main Results:

• Optimal identification accuracy was achieved with time delays between 20-28 ms, facilitating better expansion of ECG wave components (P, QRS, T).
• High-density grid partitions in PSR resulted in finer phase-space trajectories and improved accuracy.
• A smaller CNN using low-density PSR (32x32) achieved accuracy comparable to larger networks with high-density PSR (256x256), reducing network size and training time.

Conclusions:

• Parameter tuning (time delay, grid density) in PSR is crucial for enhancing ECG biometric performance.
• Efficient ECG biometric systems can be developed using optimized PSR-CNN approaches, balancing accuracy and computational resources.