Feature Selection for Nonstationary Data: Application to Human Recognition Using Medical Biometrics

Area of Science:

• Biometrics
• Signal Processing
• Machine Learning

Background:

• Electrocardiogram (ECG) and transient evoked otoacoustic emissions (TEOAE) are robust physiological signals for biometrics.
• Time-dependent nature of these signals poses challenges for across-session recognition.
• Existing methods struggle with enrollment limitations (single session).

Purpose of the Study:

• To develop a novel feature selection method for across-session human recognition.
• To address the challenge of time-dependent physiological signals in biometrics.
• To improve the robustness of biometric systems using ECG and TEOAE.

Main Methods:

• A novel feature selection method utilizing an auxiliary dataset with multiple sessions.
• Selection of features exhibiting persistence across different sessions.
• Incorporation of local sample margins and an across-session measure.

Main Results:

• Comprehensive experiments evaluated ECG and TEOAE variability due to time lapse and body posture.
• The proposed method demonstrated superior performance compared to seven state-of-the-art feature selection algorithms.
• Outperformed six other established ECG and TEOAE biometric recognition approaches.

Conclusions:

• The proposed feature selection method effectively handles time-dependent physiological signals for biometrics.
• It offers a significant improvement for across-session human recognition scenarios.
• The method is robust and suitable for real-world biometric applications using ECG and TEOAE.