Fikri Goksu

Adapting subject specific motor imagery EEG patterns in space–time–frequency for a brain computer interface

Abstract In this paper we propose a new technique that adaptively extracts subject specific motor imagery related EEG patterns in the space–time–frequency plane for single trial classification. The proposed approach requires no prior knowledge of reactive frequency bands, their temporal behavior or cortical locations. For a given electrode array, it finds all these parameters by constructing electrode adaptive time–frequency segmentations that are optimized for discrimination. This is accomplished first by segmenting the EEG along the time axis with Local Cosine Packets. Next the most discriminant frequency subbands are selected in each time segment with a frequency axis clustering algorithm to achieve time and frequency band adaptation individually. Finally the subject adapted features are sorted according to their discrimination power to reduce dimensionality and the top subset is used for final classification. We provide experimental results for 5 subjects of the BCI competition 2005 dataset IVa to show the superior performance of the proposed method. In particular, we demonstrate that by using a linear support vector machine as a classifier, the classification accuracy of the proposed algorithm varied between 90.5% and 99.7% and the average classification accuracy was 96%.

Sparse common spatial patterns in brain computer interface applications

The Common Spatial Pattern (CSP) method is a powerful technique for feature extraction from multichannel neural activity and widely used in brain computer interface (BCI) applications. By linearly combining signals from all channels, it maximizes variance for one condition while minimizing for the other. However, the method overfits the data in presence of dense recordings and limited amount of training data. To overcome this problem we construct a sparse CSP (sCSP) method such that only subset of channels contributes to feature extraction. The sparsity is achieved by a greedy search based generalized eigenvalue decomposition approach with low computational complexity. Our contributions in this study are extension of the greedy search based solution to have multiple sparse filters and its application in a BCI problem. We show that sCSP outperforms traditional CSP in the classification challenge of the multichannel ECoG data set of BCI competition 2005. Furthermore, it achieves nearly similar performance as infeasible exhaustive search and better than that of obtained by L1 norm based sparse solution.

Selection of spectro-temporal patterns in multichannel MEG with support vector machines for schizophrenia classification

We present a new framework for the diagnosis of schizophrenia based on the spectro-temporal patterns selected by a support vector machine from multichannel magnetoencephalogram (MEG) recordings in a verbal working memory task. In the experimental paradigm, five letters appearing sequentially on a screen were memorized by subjects. The letters constituted a word in one condition and a pronounceable nonword in the other. Power changes were extracted as features in frequency subbands of 248 channel MEG data to form a rich feature dictionary. A support vector machine has been used to select a small subset of features with recursive feature elimination technique (SVM-RFE) and the reduced subset was used for classification. We note that the discrimination between patients and controls in the word condition was higher than in the non-word condition (91.8% vs 83.8%). Furthermore, in the word condition, the most discriminant patterns were extracted in delta (1–4 Hz), theta (4–8Hz) and alpha (12–16 Hz) frequency bands. We note that these features were located around the left frontal, left temporal and occipital areas, respectively. Our results indicate that the proposed approach can quantify discriminative neural patterns associated to a functional task in spatial, spectral and temporal domain. Moreover these features provide interpretable information to the medical expert about physiological basis of the illness and can be effectively used as a biometric marker to recognize schizophrenia in clinical practice.

ECoG Based Brain Computer Interface with Subset Selection

We describe an adaptive approach for the classification of multichannel neural recordings for a brain computer interface. A dual-tree undecimated wavelet packet transform generates a structured redundant feature dictionary with different time-frequency resolutions computed on multichannel neural recordings. Rather than evaluating the individual discrimination performance of each electrode or candidate feature, the proposed approach implements a wrapper strategy combined with pruning to select a subset of features from the structured dictionary by evaluating the classification performance of their combination. The pruning stage and wrapper combination enables the algorithm to select a subset of the most informative features coming from different cortical areas and/or time frequency locations with faster speeds, while guaranteeing high generalization capability and lower error rates. We show experimental classification results on the ECoG data set of BCI competition 2005. The proposed approach achieved a classification accuracy of 93% by using only three features. This is a marked improvement over other reported approaches that use all electrodes or require manual selection of sensor subsets and feature indices and at best achieve slightly lower classification accuracies.

Classification of EEG with structural feature dictionaries in a brain computer interface

We present a new method for the classification of EEG in a brain computer interface by adapting subject specific features in spectral, temporal and spatial domain. For this particular purpose we extend our previous work on ECoG classification based on structural feature dictionary and apply it to extract the spectro-temporal patterns of multichannel EEG recordings related to a motor imagery task. The construction of the feature dictionary based on undecimated wavelet packet transform is extended to block FFT. We evaluate several subset selection algorithms to select a smell number of features for final classification. We tested our proposed approach on five subjects of BCI Competition 2005 dataset- IVa. By adapting the wavelet filter for each subject, the algorithm achieved an average classification accuracy of 91.4% The classification results and characteristic of selected features indicate that the proposed algorithm can jointly adapt to EEG patterns in spectm-spatio-temporal domain and provide classification accuracies as good as existing methods used in the literature.

Greedy solutions for the construction of sparse spatial and spatio-spectral filters in brain computer interface applications

In the original formulation of common spatial pattern (CSP), all recording channels are combined when extracting the variance as input features for a brain computer interface (BCI). This results in overfitting and robustness problems of the constructed system. Here, we introduce a sparse CSP method in which only a subset of all available channels is linearly combined when extracting the features, resulting in improved generalization in classification. We propose a greedy search based generalized eigenvalue decomposition approach for identifying multiple sparse eigenvectors to compute the spatial projections. We evaluate the performance of the proposed sparse CSP method in binary classification problems using electrocorticogram (ECoG) and electroencephalogram (EEG) datasets of brain computer interface competition 2005. We show that the results obtained by sparse CSP outperform those obtained by traditional (non-sparse) CSP. When averaged over five subjects in the EEG dataset, the classification error is 12.3% with average sparseness level of 11.6 compared to 18.4% error obtained by the traditional CSP with 118 channels. The classification error is 10% with sparseness level of 7 compared to that of 13% obtained by the traditional CSP using 64 channels in the ECoG dataset. Furthermore, we explored the effectiveness of the proposed sparse methods for extracting sparse common spatio-spectral patterns (CSSP).

Sparse common spatial patterns with recursive weight elimination

The past decade has shown the importance of adapting spatial patterns of neural activity while decoding it in a Brain Machine Interface (BMI) framework. The common spatial patterns (CSP) algorithm tackles this problem as feature extractor in binary BMI setups in which a number of spatial projections are computed while maximizing the variance of one class and minimizing of the other. Recent advances in data acquisition systems and sensor design now make recording the neural activity of the brain with dense electrode grids a possibility. However, high density recordings also pose new challenges such as overfitting to data as the number of recording channels increases dramatically compared to the number of training trials. In this study, we tackle this problem by constructing a sparse CSP algorithm through recursive weight elimination (CSP RWE), in which the spatial projections are computed using a subset of the recording channels. The sparse projections are expected to yield increased robustness and eliminate overfitting. We show promising results towards the classification of multichannel Electrocorticogram (ECoG) and Electroencephalogram (EEG) datasets with CSP RWE for a BMI.

Denoising of multiscale/multiresolution structural feature dictionaries for rapid training of a brain computer interface

Multichannel neural activities such as EEG or ECoG in a brain computer interface can be classified with subset selection algorithms running on large feature dictionaries describing subject specific features in spectral, temporal and spatial domain. While providing high accuracies in classification, the subset selection techniques are associated with long training times due to the large feature set constructed from multichannel neural recordings. In this paper we study a novel denoising technique for reducing the dimensionality of the feature space which decreases the computational complexity of the subset selection step radically without causing any degradation in the final classification accuracy. The denoising procedure was based on the comparison of the energy in a particular time segment and in a given scale/level to the energy of the raw data. By setting denoising threshold a priori the algorithm removes those nodes which fail to capture the energy in the raw data in a given scale. We provide experimental studies towards the classification of motor imagery related multichannel ECoG recordings for a brain computer interface. The denoising procedure was able to reach the same classification accuracy without denoising and a computational complexity around 5 times smaller. We also note that in some cases the denoised procedure performed better classification.