Electrocardiogram Biometrics Using Transformer's Self-Attention Mechanism for Sequence Pair Feature Extractor and...

Area of Science:

• Biometrics
• Artificial Intelligence
• Signal Processing

Background:

• Existing electrocardiogram (ECG) biometrics struggle with performance degradation due to ECG changes post-enrollment.
• Feature extraction methods often fail to correlate ECG data acquired at different times, limiting biometric accuracy.

Purpose of the Study:

• To propose a novel sequence pair feature extractor for ECG biometrics inspired by BERT's sentence pair task.
• To enhance ECG identification by dynamically representing ECG pairs and utilizing self-attention for inter-identity relationships.

Main Methods:

• Developed a sequence pair feature extractor using a transformer's self-attention mechanism.
• Trained a single model on 10 ECG databases and tested it on six additional databases without retraining.
• Evaluated model performance considering the time separation between ECG enrollment and classification.

Main Results:

• Achieved high identification accuracy (96.20%-100.0%) across multiple databases with short time separation.
• Demonstrated significant improvements over state-of-the-art methods with long time separation (92.70% on ECGIDDB, 64.16% on PTBDB).

Conclusions:

• The proposed sequence pair feature extractor effectively handles ECG changes over time, outperforming existing methods.
• The self-attention mechanism enhances inter-identity relationship modeling for robust ECG identification.