Imputing Missing Values in EEG with Multivariate Autoregressive Models

Area of Science:

• Biomedical Engineering
• Neuroscience
• Signal Processing

Background:

• Wearable electroencephalogram (EEG) devices facilitate real-world brain-computer interfaces and neuroscience studies.
• Practical use of wearable EEG leads to incomplete data due to artifacts and electrode issues, unlike controlled lab settings.
• Existing signal analysis methods require complete data, necessitating effective imputation techniques for missing values.

Purpose of the Study:

• To propose a novel EEG signal imputation method for handling missing data in wearable recordings.
• To develop an iterative estimation and simulation procedure based on multivariate autoregressive (MAR) modeling.
• To address the challenge of data incompleteness in practical wearable EEG applications.

Main Methods:

• Utilized multivariate autoregressive (MAR) modeling for EEG signal imputation.
• Employed an iterative estimation and simulation approach inspired by multiple imputation procedures.
• Evaluated the method using real EEG data with artificially introduced missing entries.

Main Results:

• The proposed MAR-based imputation method demonstrated superior performance compared to existing baseline interpolation techniques.
• Experimental results confirmed the effectiveness of the iterative scheme in reconstructing missing EEG data.
• The method successfully handled artificial missing entries in real EEG datasets.

Conclusions:

• The developed iterative EEG imputation method is effective and outperforms conventional approaches.
• The proposed technique provides a robust solution for incomplete data from wearable EEG devices.
• This iterative scheme offers a foundational approach for future advancements in wearable EEG signal processing.