Pole-Zero REM Modeling with Application in EEG Artifact Removal.

Area of Science:

• Signal Processing
• System Identification
• Biomedical Engineering

Background:

• Pole-zero modeling is essential for system identification but challenging with white noise.
• Selecting the correct model order is critical to avoid overfitting or underfitting data.
• Existing methods like AIC and BIC may not optimally balance model complexity and data fit in noisy conditions.

Purpose of the Study:

• To propose a new method, RE Minimization (REM), for optimal pole-zero model order selection in the presence of white noise.
• To address the limitations of conventional methods in preventing overfitting and underfitting.
• To demonstrate the efficacy of REM in both synthetic and real-world data applications.

Main Methods:

• Developed a novel order selection technique based on minimizing the upper bound of the error between noisy and noiseless data (RE).
• Utilized conventional least squares estimation for model parameter calculation.
• Validated the proposed REM method against established criteria like AIC and BIC using synthetic datasets.

Main Results:

• REM successfully avoided the overparametrization issues often seen with AIC and the underparametrization common with BIC.
• Simulation results demonstrated REM's superiority over AIC and BIC in selecting appropriate model orders for noisy data.
• The method proved effective in a practical application, efficiently removing EOG artifacts from EEG data by accurately modeling the underlying EEG signal.

Conclusions:

• RE Minimization (REM) offers a robust and accurate approach for pole-zero model order selection in noisy environments.
• REM provides a significant improvement over existing methods, ensuring better data representation and preventing common modeling errors.
• The successful application in EEG artifact removal highlights REM's potential for practical biomedical signal processing tasks.