音声ブラウザご使用の方向け: SKIP NAVI GOTO NAVI

DETECTION OF EVENT-RELATED POTENTIALS AS THE BASIS FOR A DIRECT BRAIN INTERFACE

Jane E. Huggins, M.S.,1,2 Simon P. Levine, Ph.D.,1,2 Spencer L. BeMent, Ph.D.,1,3 Ramesh K. Kushwaha, Ph.D.,4 Lori A. Schuh, M.D.,4 Mitchell M. Rohde, M.S.1,3 1Bioengineering Program, 2Rehabilitation Engineering Program, 3Electrical Engineering and Computer Science Department, 4Neurology, University of Michigan

ABSTRACT

Electrocorticograms (ECoG) were collected from five subjects performing simple voluntary activities. Templates of event-related potentials (ERPs) corresponding to the activities were created from half of the ECoG using triggered averaging. The templates were then cross-correlated with the other half of the ECoG. Detections were identified each time the cross-correlation statistic exceeded an experimentally determined threshold. Detections which occurred within one second of an actual movement were considered valid detections. Three of the subjects had at least one activity that could be detected with greater than 85% accuracy and for which less then 5% of the detections were considered to be false-positive. These results show that detection of voluntary cognitive activity using cross-correlation of ERP templates with ECoG data could form the basis of a direct brain interface.

BACKGROUND

The term direct brain interface refers to an interface that detects specific voluntary cognitive activities and triggers appropriate external responses. Such an interface could provide people with severe disabilities with an option for control of assistive technology that does not require physical movement. Over the longer term, a direct brain interface might also be useful for people with less severe limitations. The work described here focuses on movement-related cognitive activities because the occurrence of movements can be precisely determined. This allows accurate evaluation of methods developed to detect voluntary cognitive signal patterns. Although a potential user of a direct brain interface may be unable to perform physical movements of the type studied here, the methods developed to detect cognitive signal patterns related to movement are expected to be applicable to the detection of signal patterns related to attempted movement or other cognitive activities. Cross-correlation of a signal template with electroencephalogram (EEG) [2],[5] or electrocorticogram (ECoG) [5] has been used to detect human sensory evoked potentials. Cross-correlation methods have also been used to detect signal patterns related to movement in animals with up to 83% accuracy [6]. The use of cross-correlation for the detection of signal patterns related to movement has not yet been demonstrated with human subjects.

RESEARCH QUESTION

The research hypothesis is that a method based on cross-correlation of a signal pattern template with ECoG can detect ERPs associated with voluntary movement in humans with sufficient accuracy for the operation of a direct brain interface. The accuracy of detection is measured in terms of percentages of valid and false positive detections.

METHODS

Data collection: The subjects for this research are patients in an epilepsy surgery program who have cortical surface electrodes implanted for diagnostic purposes prior to ablation surgery. Data was collected from five subjects, each of whom performed approximately fifty repetitions of up to four distinct activities, resulting in ECoG data sets from 18 subject-activity pairs. Each data set contains ECoG from 20 - 31 cortical locations. The electrodes are 4 mm in diameter and separated by 1 cm center-to-center. Electrodes are arranged either in one dimensional strips of 4 to 6 electrodes or in two dimensional grids of up to 64 electrodes. The preferred electrode configuration for this research is electrode grids positioned over sensory-motor cortex. Electrode configuration is chosen solely for diagnostic purposes, however, and is most commonly over the temporal lobe, with a few electrodes over tongue sensory-motor areas. A summary of electrode placement and configuration for the five research subjects is presented in Table 1.

Subject Electrode Type Electrode Location
Grid Temporal Lobe
DT JR Strips Temporal Lobe
DV Grid Sensory Motor Cortex
JP Grid Temporal Lobe
DP Strips Temporal Lobe

Table 1: Electrode type and location by subject.

Triggered averaging of the first half of each data set is used to create a signal pattern template for every electrode location. The triggered averaging is based on a switch closure activated by the given movement. Details of the triggered averaging have been previously described [3]. Templates begin 2.5 seconds before the trigger and end 1.5 seconds after.

ERP identification: The templates produced by the triggered averaging are identified as valid ERPs through the comparison of a pre-activity interval and a peak signal interval (see Figure 1). The pre-activity interval is defined to be from 2.5 to 2.0 seconds prior to the trigger, well before the earliest reported time for signals related to the initiation of movement [4]. The data from the pre-activity interval is used to calculate the pre-activity mean and the 95% predictive interval for future observations, based on the Student T-test. The template is then shifted by this mean (to prevent biasing) and rectified. The position of the maximum value in the rectified template is determined and a peak signal interval a half second in duration is selected centered around the maximum value. The signal mean over this peak interval is calculated. The ratio of the peak signal mean to the 95% predictive interval amplitude is used to select ERP templates for cross-correlation analysis. Templates with ratios greater than two were used.

Cross-correlation: The signal pattern template and the ECoG from the second half of the data set are cross-correlated. A detection is identified when the cross-correlation statistic exceeds an experimentally determined detection threshold. An identified detection is accepted as valid if it is within one second before or after the time at which the movement actually occurred (as recorded by a mechanical switch). The detection threshold can be adjusted to provide a balance between the percentage of movements detected (hit percentage) and the number of incorrect detections (false-positive percentage). These performance statistics vary based on the level of the detection threshold (see Figure 2). Three different detection thresholds are selected for each data set in order to more fully explore the detection achievable with cross-correlation. The three thresholds are determined by a function which minimizes the false-positive percentage (F) and maximizes the hit percentage (H) based on weightings of H versus F. H/F weighting ratios of 1/1, 10/1, and 1/10 were used.

RESULTS

Valid ERP templates were identified for 15 of the 18 subject-activity data sets (Table 2). For each subject, at least two activities produced valid ERPs. Several of the subjects had at least one ERP corresponding to each of the activities performed. As would be expected due to the typical electrode placement, four of the five subjects had ERPs related to tongue protrusion. Subject DV had the best electrode configuration for research purposes, a grid placed over sensory motor cortex. At least one valid ERP template was identified for each activity performed by this subject with the largest number identified for tongue protrusion. All tongue related ERPs were recorded from nearly contiguous locations on the electrode grid (see Figure 3). Stimulation studies performed on this subject by the epilepsy surgery program showed that the grid was placed primarily over tongue sensory-motor areas.

Figure 1: ERP identification for averaged template.

Figure 2:. Hit percentage (H) and false-positive percentage (F) at the three detection thresholds. Asterisk (*) indicates time of first 10 movements. Activity Subject DT JR DV JP DP Tongue Protrusion 0/24 10/20 7/31 3/31 11/20 Facial Movement 1/24 Verbalization 1/24 3/31 Sip 0/24 3/20 Naming 1/31 Thumb to Palm 1/31 4/20 Wrist Extension 6/31 5/20 Arm Rotation 2/31 Puff 3/31 Finger Extension 2/20 Table 2: Valid ERPs identified from the first half of each subject/activity data set. Entries are the number of valid ERPs templates over the total number of templates. For the detection thresholds selected based on to an H/F weighting ratio of 1/1, three of the five subjects had at least one activity that could be detected with greater than 85% hit percentage and less than 5% false-positives percentage (see Table 3). The remaining two subjects also had high hit percentages, but these were coupled with false-positive percentages of 40% or greater.

Figure 3: Locations of valid ERPs for subject DV. Activity Subject DT JR DV JP DP Tongue Protrusion 92/2 100/4 89/16 72/40 Facial Movement 100/62 Verbalization 100/70 72/15 Sip 100/48 Naming 82/21 Thumb to Palm 12/18 96/64 Wrist Extension 77/22 88/64 Arm Rotation 36/5 Puff 89/2 Finger Extension 84/54 Table 3: Detection accuracy with detection thresholds selected with an H/F weighting ratio of 1/1. Entries are in the form H/F. When multiple ERPs were identified for the subject-activity data set, the electrode location with the maximum (H-F) difference is presented. The best detection accuracies for the 1/1 H/F weighting were 92/2 (JR), 100/4 (DV), and 89/2 (JP) (see Table 3). When the detection threshold was based on a H/F weighting ratio of 1/10, these accuracies changed to 77/0 (JR), 62/0 (DV), and 43/0 (JP). When the hit percentage was emphasized with a H/F weighting ratio of 10/1, the accuracies changed to 100/41 (JR), 100/4 (DV), and 96/38 (JP).

DISCUSSION

The detection accuracy obtained for three of the five subjects shows that it may be possible to detect ECoG signal patterns related to voluntary cognitive activities with sufficient accuracy to operate a direct brain interface. These results are especially impressive because the location of the electrodes could not be specified. Improved accuracy should be possible when the electrode locations can be chosen for the purpose of detecting ECoG signals related to a specific activity. A portion of the false-positive percentage may be a result of the measurement techniques. The identification of detections as either valid hits or false-positives is based on their proximity to the recorded triggers. However, the trigger channel only records activities which result in the activation of the mechanical switch. Thus, activities which are actually performed (or just planned) but do not activate the switch could add to the false-positive percentage. The dramatic drop in the hit percentages for the three best detection accuracies when a H/F weighting ratio of 1/10 was used is consistent with the hypothesis that some of the false-positives were actual movements which did not result in switch activation. Future experiments will address this uncertainty by the use of an improved trigger such as electromyography. Detection of voluntary cognitive activity by means of cross-correlation relies on the availability of an ERP template specific to a particular subject, electrode location and activity. There are several possible methods by which such templates may be obtained for people with severe disabilities. First, templates could be established during the course of a progressive degenerative illness. Second, advanced signal processing methods may be able to identify templates based on the content of the ECoG signal alone [1]. Third, templates representative of typical ERPs could be used. Cross-correlation between signal templates and ECoG could therefore form the basis of a direct brain interface for people with severe disabilities.

REFERENCES

[1] Birch, G.E., Lawrence, P.D., Hare, R.D.; Single-Trial Processing of Event-Related Potentials Using Outlier Information, IEEE Trans. Biomed. Eng., 1993, 40(1):59-73.

[2] Cilliers, P.J., Van Der Kouwe, A.J.W.: A VEP-Based Computer Interface for C2-Quadriplegics, Proc IEEE Engineering in Medicine and Biology, 1993, 15(pt 3):1263-1264.

[3] Huggins, J.E., Levine, S.P., Kushwaha, R., BeMent, S., Schuh, L.A., Ross, D.A.: Identification of Cortical Signal Patterns Related to Human Tongue Protrusion, Proc. of the RESNA'95 Annual. Conf., 1995, pp. 670-672.

[4] Kitamura, J., Shibasaki, H., Takagi, A., Nabeshima, H., Yamaguchi, A., Enhanced Negative Slope of Cortical Potentials Before Sequential as Compared with Simultaneous Movement of Two Fingers, Electroenceph. Clin. Neurophysiol., 1993, 86:176-182.

[5] Sutter, E.E.: The Brain Response Interface: Communication Through Visually-Induced Electrical Brain Responses. J. Microcomputer Appli., 15:31-45, 1992.

[6] Woody, C.D., Nahvi, M.J.; Application of Optimum Linear Filter Theory to the Detection of Cortical Signals Preceding Facial Movements in Cat, Exp. Brain Res., 1973, 16:455-465. Address Jane E. Huggins 1C335 University of Michigan Hospital Ann Arbor, MI 48109-0032

BASIS FOR DIRECT BRAIN INTERFACE

CROSS-CORRELATION FOR BRAIN INTERFACE