Classification of EEG Signals from Four SubjectsDuring Five Mental TasksCharles W. Anderson and Zlatko Sijerˇci´cDepartment of Computer ScienceColorado State UniversityFort Collins, CO 80523fanderson,[email protected]: 970-491-7491, FAX: 970-491-2466AbstractNeural networks are trained to classify half-second segments of six-channel, EEG data into one of fiveclasses corresponding to five cognitive tasks performed by four subjects. Two and three-layer feedfor-ward neural networks are trained using 10-fold cross-validation and early stopping to control over-fitting.EEG signals were represented as autoregressive (AR) models. The average percentage of test segmentscorrectly classified ranged from 71% for one subject to 38% for another subject. Cluster analysis of theresulting neural networks’ hidden-unit weight vectors identifies which EEG channels are most relevant tothis discrimination problem.1 IntroductionVisual inspection of multiple time series of EEG signals in their unprocessed form is still the predom-inant way of discriminating and classifying EEG patterns in the medical community and requires highlytrained medical professionals. Since the early days of automatic EEG processing, representations based ona Fourier transform have been most commonly applied. This approach is based on earlier observations thatthe EEG spectrum contains some characteristic waveforms that fall primarily within four frequency bands—delta (1–3 Hz), theta (4–7 Hz), alpha (8–13 Hz), and beta (14–20 Hz). Such methods have proved beneficialfor various EEG characterizations, but the Fourier Transform and its discrete version, the FFT, suffer fromlarge noise sensitivity. Numerous other techniques from the theory of signal analysis have been used to ob-tain representations and extract the features of interest for classification purposes. For an overview of thesetechniques, see [5].Neural networks and statistical pattern recognition methods have been applied to EEG analysis; some ofthis work is briefly reviewed in this section. The experiments reported in this article use neural networks tolearnEEG classification functions, but go beyondthis typical use by applyingcluster analysis to the resultingweight vectors. This analysis reveals some of the relationshipsbetween EEG featuresand theirclassificationthat exist in the learned classification function. It also provides information that can support and generatehypothesesabout brainactivity underlyingthe studied cognitive behaviors. The representationsused to codethe informationin EEG signals greatlyinfluenceswhatcan be learnedfromthe analysis oftheneuralnetworkclassifiers.Some recent work that deals with the problem of EEG representations finds time domain methods basedon parametric models very useful for EEG feature extraction. Tseng, et al., [21] evaluated different para-metric models on a fairly large database of EEG segments. Using inverse filtering, white noise tests, andone-second EEG segments, they found that autoregressive (AR) models of orders between 2 and 32 yieldedthe best EEG estimation. For a method which avoids the use of signal segmentation and provides an on-lineAR parameterestimation that fits nonstationarysignals, like EEG, see [6]. In a problem of classifying EEGsof normal subjects from those with psychiatric disorders, Tsoi, et al., [22] used AR representations in a pre-processing stage and artificial neural networks in the classification stage. Inouye, et al. [9], used the entropyof the power spectra and a mutual information measure to determine directional EEG patterns duringmentalarithmetic and a resting state. Rotation and change in size of mental images and its corresponding patternsof cerebral activations are considered in [18].Finding a suitable representation of EEG signals is the key to learning a reliable discrimination and tounderstanding the extracted relationships [1, 2]. In this article, the coefficients of sixth-order AR models areused to represent the EEG signals. Standard, feed-forward neural networks are trained as classifiers usingerrorbackpropagationwith early stopping and ten-fold crossover. The representationand training procedureare defined in Section 2. Results are presented in Section 3. Section 4 contains a description and results ofthe cluster analysis performed on trained networks. Section 5 summarizes the conclusions and limitationsof the classification experiments.2 Method2.1 EEG Data Acquisition and RepresentationAll data used in this article was obtained previously by Keirn and Aunon [13, 12] using the followingprocedure. The subjects were seated in an Industrial Acoustics Company sound controlled booth with dimlighting and noiseless fans for ventilation. An Electro-Cap elastic electrode cap was used to record from po-sitions C3, C4, P3, P4, O1, and O2, defined by the 10-20 system of electrode placement [10]. The electrodeswere connected through a bank of Grass 7P511 amplifiers and bandpass filtered from 0.1–100 Hz. Data wasand recorded at a sampling rate of 250 Hz with a Lab Master 12 bit A/D converter mounted in an IBM-ATcomputer. Eye blinks were detected by means of a separate channel of data recorded from two electrodesplaced above and below the subject’s left eye.For this paper, the data from four subjects performing five mental tasks was analyzed. These tasks werechosen by Keirn and Aunon to invoke hemispheric brainwave asymmetry [16]. The five tasks are: the base-line task, for which the subjects were asked to relax as much as possible; the letter task, for which the sub-jects were instructed to mentally compose a letter to a friend or relative without vocalizing; the math task,for which the subjects were given nontrivial multiplication problems, such as 49 times 78, and were askedto solve them without vocalizing or making any other physical movements; the visual counting task, forwhich the subjects were asked to imagine a blackboard and to visualize numbers being written on the boardsequentially; and the geometric figure rotation, for which the subjects were asked to visualize a particularthree-dimensional block figure being rotated about an axis. Data was recorded for 10 seconds during eachtask and each task was repeated five times per session. Most subjects attended two such sessions recordedon separate weeks, resulting in a total of 10 trials for each task. With a 250 Hz sampling rate, each 10 secondtrial produces 2,500 samples per channel. These are divided into