Data Imputation in EEG Signals for Brainprint Identification

Abstract

Electroencephalograms (EEG) signals have very low signal-to-noise ratio, thus can be easily affected and changes over milliseconds. Normally, trials with excessive body movements or other types of artefacts with amplitude more than 100 µV should be discarded to reduce the noise stains. Scrapping the affected features is not advisable. Therefore, missing values imputation is essential to avoid incomplete data that may deteriorate the computational modelling performance. Hence, this paper proposes a similarity matching method to replace the missing values in the EEG trials. The main idea of the missing values imputation is based on the similarity measure between the trials. The trials with the highest similarity is taken to replace the missing values for the related EEG channels. Statistical evaluation and classification evaluation are used to evaluate the reliability of the proposed similarity matching method. The mean, variance and standard deviation are used for statistical evaluation. For the classification evaluation, the dataset is classified for brainprint identification by using the Incremental Fuzzy-Rough Nearest Neighbour (IncFRNN). The statistical evaluation proved that the proposed similarity matching imputation method is promising when the missing values are not come from the same channels. The classification results achieved the excellent performance with 98.19% in accuracy and 0.998 in AUC.