Join for Free!
122503 members

table of contents table of contents

The aim of this study was to design and compare methods for …

Home » Biology Articles » Bioengineering » Evolutionary optimization of classifiers and features for single-trial EEG Discrimination » Discussion

- Evolutionary optimization of classifiers and features for single-trial EEG Discrimination

This study has demonstrated that individual classifier tailoring and feature subset selection significantly improves single-trial limb laterality discrimination, and that the optimal EEG channels differ much between subjects.

It should be noted that this study has focused on comparing classifiers rather than maximizing prediction accuracies. The number of features as well as the maximum generations allowed in the evolutionary algorithm were heavily limited due to time and computer restrictions. Similarly, the number of included movement epochs was reduced to only 100, a factor that significantly decreases the prediction accuracy. The non-linear classifiers performed better than the linear approaches, agreeing with previous studies [29,30]. Interestingly, the improvement between linear and non-linear classifiers is 14.04%, 10.28% and 2.04%, respectively, for the random, filter and wrapper approaches. This observation suggests that as the association between classifier training and feature selection increases, the non-linearity of the classifier becomes less important. In the evolutionary approach, the tailored feature subset is allowed to express either non-linearities or linearities in the data – whatever suits the given classifier optimally. More data is, however, required to establish this theory statistically.

The spatial preference found in this study partially agrees with previous research, which focuses on a few central electrodes for finger movement classification [4-8]. Channels FC1, C3 or Cz, highly selected in all subjects, are located close to the left-hand side motor cortex. Interestingly, C4, the right-hand side equivalent, is not ranked high in any subject, including the left-handed subject 3. Also, in three out of four subjects, P7 or P4 in the parietal regions, with no major established connection to motor areas, are ranked highly. However, the results can in part be explained by investigating the geometrical orientation of the electric fields (dipoles) generated by the activated neurons. For example, projecting the event-related potential for subject 3 on a human head model reveals that the signal source at peak EEG activity before the movement, -39 ms, has a tangential orientation. These results are in accordance with the physiological orientation of the pyramidal neurons responsible for finger movement. The neurons are located in the finger area of the motor cortex, which in turn is located within the central sulcus [9,31]. These pyramidal neurons are aligned along the surface of the cortex, thus generating a tangential dipole [1]. During the course of movement, different areas with different dipole orientations are activated, and it is therefore difficult to manually predict what combination of electrode locations will provide most useful information for laterality discrimination.

Different areas of the cortex are likely to be of varying importance at different times throughout a movement epoch, and each wavelet coefficient corresponds to a given scale – translatable into frequency – and point in time. In this study, however, no single wavelet coefficient was significantly more frequently selected than any other, indicating either that there was no time or frequency preference or that these wavelet coefficients capture the dynamics during finger extension cycle poorly. Not restricting the feature pool to a given number of wavelet coefficients based on an average, as was done here, could potentially resolve this issue.

Ideally, the feature pool would consist of several different types of signal parameters other than wavelet coefficients, such as Fourier frequencies, power spectral densities and autoregressive coefficients. Other state-of-the-art classifiers, such as support vector machines, can also be incorporated into the algorithm.

rating: 0.00 from 0 votes | updated on: 25 Nov 2007 | views: 11873 |

Rate article: