Online ECG Biometrics for Streaming Data with Prototypes Learning and Memory Enhancement.

Area of Science:

• Biometrics
• Signal Processing
• Machine Learning

Background:

• Electrocardiography (ECG) is a promising biometric for identification.
• Existing ECG biometric methods primarily use batch processing, which is inefficient for streaming data.
• Batch processing requires complete data retraining when new data arrives, leading to high computational costs.

Purpose of the Study:

• To propose a novel online method for ECG biometrics that incrementally learns from streaming data.
• To address the limitations of batch processing in ECG biometric systems.
• To develop an efficient and effective solution for real-time ECG biometric recognition.

Main Methods:

• An online learning approach for ECG biometrics that updates models incrementally with new data.
• Introduction of bidirectional regression and prototype learning modules to enhance feature representation.
• Incorporation of a memory enhancement module to combat catastrophic forgetting in online learning.
• Development of an efficient online optimization algorithm for loss minimization.

Main Results:

• The proposed online method effectively learns from streaming ECG data without retraining on old data.
• Bidirectional regression and prototype learning modules improve the discriminative power of learned representations.
• The memory enhancement module mitigates catastrophic forgetting, maintaining performance over time.
• Extensive experiments on two datasets validate the method's effectiveness and efficiency.

Conclusions:

• The developed online ECG biometric method offers an efficient alternative to traditional batch processing.
• The novel modules and optimization algorithm contribute to robust and adaptive biometric recognition.
• This approach is well-suited for real-world applications involving continuous data streams.