Discriminative training of subject-dependent Gaussian mixture models for portable SSVEP identification

Abstract

Abstract Various factors and challenges have impeded Brain Computer Interface (BCI) systems from moving outside laboratories towards real applications. Such impediments might be, reliability, information transfer rates, non-stationarity of Electroencephalogram (EEG), and subject variation. Moreover, due to the portable design requirements of our BCI systems, which consists of a wireless EEG device to record the data and an Android tablet that serves as visual stimuli, the system׳s performance is affected by the imprecise Steady-State Visual Evoked Potential (SSVEP) paradigm generation and the intermittent Android system calls. As such, we follow our previous effort, in which we introduced a signal processing solution that entailed partitioning the score spaces of Canonical Correlation Analysis (CCA) and Power Spectral Density Analysis (PSDA), extracting useful information from each partition, and augmenting them to generate a robust representation of the biased SSVEP recordings. Then, we incorporate subject specific information via Gaussian Mixture Model (GMM) training and adaptation utilizing Maximum A Posteriori (MAP) model adaptation. However, despite the robust performance of MAP, it is essentially a generative method which may not be very effective in capturing discriminative subject-specific information. As such, in this effort we introduce the Minimum Segment Classification Error (MSCE) algorithm for discriminative re-estimation of the GMM parameters in a time efficient manner by directly relating the parameters to the performance measure. Moreover, We conduct a comparison between the re-estimated subject-dependent GMMs utilizing our proposed discriminative MSCE algorithm and the standard GMM-MAP baseline model from our previous effort to demonstrate the significance of our proposed discriminative re-estimation algorithm. Additionally, we evaluate the performance of our proposed MSCE algorithm and compare it to a well-established SSVEP-identification method that utilizes Multivariate Linear Regression (MLR). Our 10-Fold Cross-Validation results demonstrated that while MLR achieved an overall 86% SSVEP identification accuracy, the GMM-MAP baseline model yielded an overall high identification accuracy of 92.9% and that accuracy is further improved to 97.41% when utilizing our proposed MSCE algorithm.