A supervised machine-learning method for detecting steady-state visually evoked potentials for use in brain computer interfaces: A comparative assessment

Master Thesis

2017

Permanent link to this Item
Authors
Journal Title
Link to Journal
Journal ISSN
Volume Title
Publisher
Publisher

University of Cape Town

License
Series
Abstract
It is hypothesised that supervised machine learning on the estimated parameters output by a model for visually evoked potentials (VEPs), created by Kremlácek et al. (2002), could be used to classify steady-state visually evoked potentials (SSVEP) by frequency of stimulation. Classification of SSVEPs by stimulus frequency has application in SSVEP-based brain computer interfaces (BCI), where users are presented with flashing stimuli and user intent is decoded by identifying which stimulus the subject is attending to. We investigate the ability of the model of VEPs to fit the initial portions of SSVEPs, which are not yet in a steady state and contain characteristic features of VEPs superimposed with those of a steady state response. In this process the estimated parameters, as a function of the model for a given SSVEP response, were found. These estimated parameters were used to train several support vector machines (SVM) to classify the SSVEPs. Three initialisation conditions for the model are examined for their contribution to the goodness of fit and the subsequent classification accuracy, of the SVMs. It was found that the model was able to fit SSVEPs with a normalised root mean square error (NRMSE) of 27%, this performance did not match the expected NRMSE values of 13% reported by Kremlácek et al. (2002) for fits on VEPs. The fit data was assessed by the machine learning scheme and generated parameters which were classifiable by SVM above a random chance of 14% (Reang 9% to 28%). It was also shown that the selection of initial parameters had no distinct effect on the classification accuracy. Traditional classification approaches using spectral techniques such as Power Spectral Density Analysis (PSDA) and canonical correlation analysis (CCA) require a window period of data above 1 s to perform accurately enough for use in BCIs. The larger the window period of SSVEP data used the more the Information transfer rate (ITR) decreases. Undertaking a successful classification on only the initial 250 ms portions of SSVEP data would lead to an improved ITR and a BCI which is faster to use. Classification of each method was assessed at three SSVEP window periods (0.25, 0.5 and 1 s). Comparison of the three methods revealed that, on a whole CCA outperformed both the PSDA and SVM methods. While PSDA performance was in-line with that of the SVM method. All methods performed poorly at the window period of 0.25 s with an average accuracy converging on random chance - 14%. At the window period of 0.5 s the CCA only marginally outperformed the SVM method and at a time of 1 s the CCA method significantly (p<0.05) outperformed the SVM method. While the SVMs tended to improve with window period the results were not generally significant. It was found that certain SVMs (Representing a unique combination of subject, initial conditions and window period) achieved an accuracy as high as 30%. For a few instances the accuracy was comparable to the CCA method with a significance of 5%. While we were unable to predict which SVM would perform well for a given subject, it was demonstrated that with further refinement this novel method may produce results similar to or better than that of CCA.
Description

Reference:

Collections