Optimal Feature Selection for Artifact Classification in EEG Time Series
Identifying artifacts or non-brain electrical signals in EEG time series is often a necessary but time-consuming preprocessing step, as many EEG analysis techniques require that the data be artifact free. Because of this, reliable and accurate techniques for automated artifact detection are desirable in practice. Previous research has shown that coefficients obtained from autoregressive (AR) models can be used as feature vectors to classify among several different artifact conditions found in EEG. However, a statistical method for identifying significant AR features has not been presented. In this work we propose a method for determining the optimal AR features that is based on penalized multinomial regression. Our results indicate that the size of the feature vector can be greatly reduced with minimal loss to classification accuracy. The features selected by this algorithm localize to specific channels and suggests a possible BCI implementation with increased computational efficiency than with using all available channels. We also show that the significant AR features produced by this approach correlate to known brain physiological properties.
KeywordsAutoregressive (AR) model Artifacts Electroencephalography classification feature selection multinomial regression penalized regression machine learning
Unable to display preview. Download preview PDF.
- 11.Friedman, J., Hastie, T., Tibshirani, R.: Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software 33, 1 (2010)Google Scholar
- 13.U.S. Department of Defense, Office of the Secretary of Defense: Code of federal regulations, protection of human subjects. Government Printing Office. 32 CFR 19 (1999)Google Scholar
- 14.U.S. Department of the Army: Use of volunteers as subjects of research. Government Printing Office. AR 70-25 (1990) Google Scholar