Characterization and Robust Classification of EEG Signal from Image RSVP Events with Independent Time-Frequency Features

Jia Meng,Nima Bigdely Shamlo,Lenis Mauricio Meriño,Kay Robbins,Scott Makeig,Yufei Huang

doi:10.1371/journal.pone.0044464

Abstract

This paper considers the problem of automatic characterization and detection of target images in a rapid serial visual presentation (RSVP) task based on EEG data. A novel method that aims to identify single-trial event-related potentials (ERPs) in time-frequency is proposed, and a robust classifier with feature clustering is developed to better utilize the correlated ERP features. The method is applied to EEG recordings of a RSVP experiment with multiple sessions and subjects.The results show that the target image events are mainly characterized by 3 distinct patterns in the time-frequency domain, i.e., a theta band (4.3 Hz) power boosting 300–700 ms after the target image onset, an alpha band (12 Hz) power boosting 500–1000 ms after the stimulus onset, and a delta band (2 Hz) power boosting after 500 ms. The most discriminant time-frequency features are power boosting and are relatively consistent among multiple sessions and subjects.Since the original discriminant time-frequency features are highly correlated, we constructed the uncorrelated features using hierarchical clustering for better classification of target and non-target images. With feature clustering, performance (area under ROC) improved from 0.85 to 0.89 on within-session tests, and from 0.76 to 0.84 on cross-subject tests. The constructed uncorrelated features were more robust than the original discriminant features and corresponded to a number of local regions on the time-frequency plane. Availability: The data and code are available at: http://compgenomics.cbi.utsa.edu/rsvp/index.html

Highlights

A brain computer interface (BCI) system allows human subjects to use their brains to directly communicate with or control an external device [1]
One BCI system for image classification is realized with the rapid serial visual presentation (RSVP) of images [2,3,4]
The significance of a single discriminant pc,f,t(e) Vc,f,t is evaluated by its discriminant power defined as the area under receiver operating characteristic (Az score) of an linear discriminant analysis (LDA) classifier. For this case, the LDA assumes that the probability density functions (PDFs) of fpc,f,t(e)Dle~1g and fpc,f,t(e)Dle~0g are both normally distributed but with different mean and the same variance parameters, or and, respectively

Summary

Introduction

A brain computer interface (BCI) system allows human subjects to use their brains to directly communicate with or control an external device [1]. Characterization of event-specific signatures As a first step, a successful BCI system needs to extract the event-specific signatures that characterize the brain signals specific to the target (or non-target) images embedded in the EEG recordings. This is often achieved with a training process, in which the signatures are extracted from training data whose eventassociation are already known. Classification of unknown recordings A successful BCI system needs to effectively utilize the eventspecific signatures for classification of EEG recordings whose eventassociation is unknown This concerns building an efficient and robust classifier. After combining the correlated features using hierarchical clustering with the cluster medians as new features, the cluster based LDA classifier is capable of incorporating more information than traditional feature-based LDA methods and performs much better in terms of Az score (the area under the receiver operator characteristic curve)

Materials and Methods

Results

Conclusions and Discussion