Classifier-directed signal processing in brain research

Adaptive BCI Based on Variational Bayesian Kalman Filtering: An Empirical Evaluation

This paper proposes the use of variational Kalman filtering as an inference technique for adaptive classification in a brain computer interface (BCI). The proposed algorithm translates electroencephalogram segments adaptively into probabilities of cognitive states. It, thus, allows for nonstationarities in the joint process over cognitive state and generated EEG which may occur during a consecutive number of trials. Nonstationarities may have technical reasons (e.g., changes in impedance between scalp and electrodes) or be caused by learning effects in subjects. We compare the performance of the proposed method against an equivalent static classifier by estimating the generalization accuracy and the bit rate of the BCI. Using data from two studies with healthy subjects, we conclude that adaptive classification significantly improves BCI performance. Averaging over all subjects that participated in the respective study, we obtain, depending on the cognitive task pairing, an increase both in generalization accuracy and bit rate of up to 8%. We may, thus, conclude that adaptive inference can play a significant contribution in the quest of increasing bit rates and robustness of current BCI technology. This is especially true since the proposed algorithm can be applied in real time.

Comparison between human and artificial neural network detection of Laplacian-derived electroencephalographic activity related to unilateral voluntary movements

A back-propagation artificial neural network (ANN) was tested to verify its capacity to select different classes of single trials (STs) based on the spatial information content of electroencephalographic activity related to voluntary unilateral finger movements. The rationale was that ipsilateral and contralateral primary sensorimotor cortex can be involved in a nonstationary way in the control of unilateral voluntary movements. The movement-related potentials were surface Laplacian-transformed (SL) to reduce head volume conductor effects and to model the response of the primary sensorimotor cortex. The ANN sampled the SL from four or two central channels overlying the primary motor area of both sides in the period of 80 ms preceding the electromyographic response onset in the active muscle. The performance of the ANN was evaluated statistically by calculating the percentage value of agreement between the STs classified by the ANN and those of two investigators (used as a reference). The results showed that both investigator and ANN were capable of selecting STs with the SL maximum in the central area contralateral to the movement (contralateral STs, about 25%), STs with considerable SL values also in the ipsilateral central area (bilateral STs, about 50%), and STs with neither the contralateral nor bilateral pattern ("spatially incoherent" single trials; about 25%). The maximum agreement (64-84%) between the ANN and the investigator was obtained when the ANN used four spatial inputs (P < 0.0000001). Importantly, the common means of all single trials showed a weak or absent ipsilateral response. These results may suggest that a back-propagation ANN could select EEG single trials showing stationary and nonstationary responses of the primary sensorimotor cortex, based on the same spatial criteria as the experimenter.

Monitoring working memory load during computer-based tasks with EEG pattern recognition methods

We assessed working memory load during computer use with neural network pattern recognition applied to EEG spectral features. Eight participants performed high-, moderate-, and low-load working memory tasks. Frontal theta EEG activity increased and alpha activity decreased with increasing load. These changes probably reflect task difficulty-related increases in mental effort and the proportion of cortical resources allocated to task performance. In network analyses, test data segments from high and low load levels were discriminated with better than 95% accuracy. More than 80% of test data segments associated with a moderate load could be discriminated from high- or low-load data segments. Statistically significant classification was also achieved when applying networks trained with data from one day to data from another day, when applying networks trained with data from one task to data from another task, and when applying networks trained with data from a group of participants to data from new participants. These results support the feasibility of using EEG-based methods for monitoring cognitive load during human-computer interaction.

A fuzzy clustering approach to EP estimation

The problem of extracting a useful signal (a response) buried in relatively high amplitude noise has been investigated, under the conditions of low signal-to-noise ratio. In particular, we present a method for detecting the "true" response of the brain resulting from repeated auditory stimulation, based on selective averaging of single-trial evoked potentials. Selective averaging is accomplished in two steps. First, an unsupervised fuzzy-clustering algorithm is employed to identify groups of trials with similar characteristics, using a performance index as an optimization criterion. Then, typical responses are obtained by ensemble averaging of all trials in the same group. Similarity among the resulting estimates is quantified through a synchronization measure, which accounts for the percentage of time that the estimates are in phase. The performance of the classifier is evaluated with synthetic signals of known characteristics, and its usefulness is demonstrated with real electrophysiological data obtained from normal volunteers.

Neural network approach in multichannel auditory event-related potential analysis

Even though there are presently no clearly defined criteria for the assessment of P300 event-related potential (ERP) abnormality, it is strongly indicated through statistical analysis that such criteria exist for classifying control subjects and patients with diseases resulting in neuropsychological impairment such as multiple sclerosis (MS). We have demonstrated the feasibility of artificial neural network (ANN) methods in classifying ERP waveforms measured at a single channel (Cz) from control subjects and MS patients. In this paper, we report the results of multichannel ERP analysis and a modified network analysis methodology to enhance automation of the classification rule extraction process. The proposed methodology significantly reduces the work of statistical analysis. It also helps to standardize the criteria of P300 ERP assessment and facilitate the computer-aided analysis on neuropsychological functions.

Neural network analysis of the P300 event-related potential in multiple sclerosis

Neural network analysis is sensitive to subtle changes in patterns of data. We hypothesized that a disease process which can cause impairment of cortical function such as multiple sclerosis (MS) would affect the P300 cognitive evoked potential (P300) in a manner detectable by a feedforward backpropagation neural network. Such a network was trained using a learning data set consisting of 101 P300 wave forms (from 26 MS patients and 26 normal controls). The network was then used to classify a randomly selected test data set of 20 studies (2 studies each of 5 MS patients and 5 controls) to which it had not been previously exposed, with an average accuracy (MS = abnormal, control = normal) of 81% for a single midline electrode, increasing to 90% using 3 midline electrodes in a jury system. Neural network analysis can be of help in distinguishing normal (control) P300 from abnormal (MS) P300.

Effects of prolonged mental work on functional brain topography

Topographic patterns of event-related covariance between electrodes were measured from subjects performing a difficult memory and fine-motor control task for 10-14 h. Striking changes occurred in the patterns after subjects performed the task for an average of 7-9 h, but before performance deteriorated. Pattern strength was reduced in a fraction-of-a-second-long response preparation interval over midline precentral areas and over the entire left hemisphere. By contrast, pattern strength in a succeeding response inhibition interval was reduced over all areas. The pattern changed least in an intervening interval associated with visual-stimulus processing. This suggests that, in addition to the well-known global reduction in neuroelectric signal strength, functional neural networks are selectively affected by sustained mental work in specific fraction-of-a-second task intervals.

Inverse filter computation of the neural impulse giving rise to the auditory evoked potential

An impulse response hypothesis for evoked potentials is tested. The auditory evoked potential (AEP) is shown to be the consequence of an impulse (the arrival of sensory signals in cortex) giving rise to an impulse response (the resonation of electrocortical activity in the form of group linear waves). To demonstrate this, pre- and post-stimulus EEG activity was recorded from subjects engaged in performance of an auditory odd-ball experiment. For each stimulus, the impulse required to account for the single auditory evoked potential (AEP) as a linear impulse response, was computed by use of the inverse of a filter obtained by autoregression analysis of the pre-stimulus EEG epoch. Single estimations of the impulse were then averaged. The average impulse exhibits a time course and topology consistent with the arrival of neural volleys in the cortex. The physical validity of the hypothesis is supported by a high lag correlation of the following values of the AEP to the average impulse. A further test calculation supports the linear additivity assumptions of the hypothesis.

Event-related covariances during a bimanual visuomotor task. II. Preparation and feedback

Event-related covariance (ERC) patterns were computed from pre-stimulus and feedback intervals of a bimanual, visuomotor judgment task performed by 7 right-handed men. Late contingent negative variation (CNV) ERC patterns that preceded subsequently accurate right- or left-hand responses differed from patterns that preceded subsequently inaccurate responses. Recordings from electrodes placed at left frontal, midline antero-central, and appropriately contralateral central and parietal sites were prominent in ERC patterns of subsequently accurate performances. This suggests that a distributed cortical 'preparatory network,' composed of distinct cognitive, integrative motor, somesthetic, and motor components, is essential for accurate visuomotor performance. ERC patterns related to feedback about accurate and inaccurate responses were similar to each other in the interval immediately after feedback onset, but began to differ in an interval spanning an early P300 peak. The difference became even greater in an interval spanning a late P300 peak. For both early and late P300 peaks, ERC patterns following feedback about inaccurate performance involved more frontal sites than did those following feedback about accurate performance. Together with the stimulus- and response-locked results presented in part I, results of this study on the preparatory and feedback periods suggest that ERCs show salient features of the rapidly shifting, functional cortical networks that are responsible for simple cognitive tasks. ERCs thus provide a new perspective on information processing in the human brain in relation to behavior--a perspective that supplements conventional EEG and ERP procedures.

Event-related covariances during a bimanual visuomotor task. I. Methods and analysis of stimulus- and response-locked data

A new method that measures between-channel, event-related covariances (ERCs) from scalp-recorded brain signals has been developed. The method was applied to recordings of 26 EEG channels from 7 right-handed men performing a bimanual visuomotor judgment task that required fine motor control. Covariance and time-delay measures were derived from pairs of filtered, laplacian-derived, averaged wave forms, which were enhanced by rejection of outlying trials, in intervals spanning event-related potential components. Stimulus- and response-locked ERC patterns were consistent with functional neuroanatomical models of visual stimulus processing and response execution. In early post-stimulus intervals, ERC patterns differed according to the physical properties of the stimulus; in later intervals, the patterns differed according to the subjective interpretation of the stimulus. The response-locked ERC patterns suggested 4 major cortical generators for the voluntary fine motor control required by the task: motor, somesthetic, premotor and/or supplementary motor, and prefrontal. This new method may thus be an advancement toward characterizing, both spatially and temporally, functional cortical networks in the human brain responsible for perception and action.

Dynamic functional topography of cognitive tasks

Improved neuroelectric recording and analysis tools are yielding increasingly specific information about the spatial and temporal features of neurocognitive processes. Such tools include recordings with up to 125 channels, digital signal processing techniques, and correlation of neuroelectric measures with anatomical information from magnetic resonance images. These tools, and their application to the study of cognitive functions, are presented in this paper.