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 » Results and Discussion

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


Summary of data

A 32-channel EEG was recorded during self-paced index finger extension for four untrained subjects. The acquired data was pre-processed, and 100 epochs of one second before onset of movement and 0.5 second after were extracted. The data was divided into 80% training and 20% validation data. The discrete wavelet transform was applied he EEG signal and five coefficients for each channel were obtained, resulting in 160 features in total. Classification of the resulting right/left finger movement epochs were attempted using multiple linear regression (MLR) and non-linear artificial neural networks (ANN) with different degrees of evolutionary optimization, including filter and wrapper feature selection.

Prediction accuracy

Due to the random nature of evolutionary algorithms and ANN training, each method was run 10 times and the best validation result was obtained. This process was repeated five times, and an average was formed and reported here. In order to compare the different approaches under similar conditions, all methods were restricted to select only 10 out of the 160 features. Random classification results in a score of 50%. The subject mean classification performances appear proportional to classifier complexity (Fig. 1). The non-linear methods performed consistently better than the linear approaches (subject mean for linear random, filter and wrapper: 58.75%, 63.25% and 73.50%, respectively; for non-linear random, filter and wrapper: 67%, 69.75.%, 75%, respectively; p = 0.016 using Friedman's non-parametric two-way ANOVA to test for the difference between linear vs. non-linear on subject levels while adjusting for possible method effects [28]). The random feature selection achieved a significantly lower score (linear: 58.75%, non-linear: 67%) than the high-performing wrapper classifiers (linear: 73.50%, non-linear: 75%; Friedman, p

Feature subset selection

The constitution of the final feature subsets differs between algorithm runs. However, taking all selected feature subsets into account, a selection frequency ranking is obtained. When the ranking is plotted per EEG channel, it is clear that some electrodes discriminate finger movements better than others (Fig. 2). There is large variation in spatial preference between individuals, yet one pattern is discernible: either of FC1, C3 or Cz is highly selected in all subjects. Plotting the frequency of selection against the wavelet coefficients, on the other hand, did not reveal any time or scale preference.

rating: 0.00 from 0 votes | updated on: 25 Nov 2007 | views: 12036 |

Rate article: