The Methodology of Long-term EEG Signal Processing

This research is aimed primarily at developing a new robust complex methodology for automated processing and evaluation of long-term EEG signals. The proposed solution should deal with an unknown multichannel EEG signal and be as automated as possible.

Our intention is to arrive at a solution in which for each type of measured signal the whole chain of processing methods will be defined, starting from preprocessing, via feature extraction (feature selection is done during the design of the processing chain) to classification. All the main processing steps that we use are shown in the following figure.


The data processing structure

Fig. 1.: The data processing structure. a) Input signals. b) Signal preprocessing methods. The application of one signal segmentation algorithm is obligatory. Other preprocessing methods are optional. c) Feature extraction and selection methods. Different features can be computed for different datasets, but at least one feature must always be used. Whether feature normalization will be applied depends on the classification methods that are selected. d) Supervised and/or unsupervised methods can be used. e) The parameters of the applied methods can be optimized and various visualization techniques can be used.

Signal Preprocessing

The aim of signal preprocessing is to remove noise and prepare the raw signal for further processing. The application of a certain method depends on the kind of data that is being processed, how noisy it is, and what techniques will be used in the subsequent processing stages. The most important signal preprocessing techniques that are used are signal filtering and resampling, baseline wandering removal, constant and adaptive signal segmentation, detection of artifacts, and various averaging techniques.


The data processing structure

Fig. 2.: The results of adaptive and constant segmentation. Real clinical neonatal PSG data was used. Adaptive segmentation was applied to EEG signals (electrodes FP1, FP2, T3, C3, C4, T4, O1, O2), an EOG signal and an EMG signal. Constant segmentation was used for ECG and the respiration signal (PNG).

Feature Extraction and Selection

Each segment of an EEG signal may be represented by computed features (usually more than one) or by a combination of computed features. This part of the EEG processing process is crucial, because it provides the ability to distinguish between different classes. It thus directly affects the accuracy of the final classification. The set of features that is used may include statistical features, frequency features computed for typical and extended EEG bands, features obtained by interval or period analysis, entropy-based features, features extracted after application of the Wavelet Transform (WT), etc. In addition, we have used information extracted from the other polygraphic channels. There are for example EMG, ECG, EOG, and PNG signals.

With feature extraction from EEG several hundreds of features can be acquired. This may be burdensome for further processing. The dimensionality of the feature space can be reduced by selecting subsets of features. There are various strategies and criteria for searching useful subsets of relevant features from the initial set of features. Feature selection is important because it decreases the number of features that have to be measured and processed. In addition to the improved computational speed in lower dimensional feature spaces, there may also be an increase in the accuracy of the classification algorithms. In other words, feature selection is considered to be successful if the dimensionality of the data is reduced and the classification accuracy improves or remains the same.


The data processing structure

Fig. 3.: Illustration of feature extraction results for two comatose recordings: a) the physician's evaluation (comatose stages C1 to C7), b) feature 'Fz-Cz: MIN MAX NUMBER', c) feature 'Cz-T4: FFT REL GAMMA1 (30-35 Hz)'.

Classification and Visualization

Classification is one of the main aspects of EEG signal processing. Classification involves assigning a class to an unknown object. In the case of EEG signal processing, the objects are segments described by vectors of features. Both supervised and unsupervised classification methods have been used for obtaining the final results of the analysis. Note that the final classification can only be as good as the extracted features.

The additional aim of this project is to propose and implement a clear and concise visualization that may be useful for evaluating long-term EEG records. This includes displaying classification results in a very compressed form and/or displaying information that is not visible in raw EEG data and may be related to internal or external clinical factors (e.g. brain disorder, or a specific neurological state). Visualization proved to be very important for clinical practice, because physicians are used to working primarily with visual information.


The data processing structure

Fig. 4.: Example of hierarchical clustering results - the found clusters. An epileptic EEG record was used (19 EEG channels according to the 10/20 system.


PSGLab

PSGLab is Matlab toolbox for processing of polysomnographic (PSG) data. PSGLab implements signal preprocessing, feature extraction, classification, cluster analysis and data visualization methods.

Detailed information about PSGLab: http://bio.felk.cvut.cz/psglab/

Cooperation

Used clinical data are obtained form the cooperating medical institution - Na Bulovce University Hospital in Prague, and Institute for the Care of Mother and Child in Prague. All the data are scored by an experienced neurologist.







Previous page: Projects in Neurophysiology
Next page: Long-Term EEG