Significance probability mapping: An aid in the topographic analysis of brain electrical activity

Electroencephalography and Clinical Neurophysiology, 1981, 51 : 455--462

455

© Elsevier/North-Holland Scientific Publishers, Ltd.

SIGNIFICANCE PROBABILITY MAPPING: AN AID IN THE TOPOGRAPHIC ANALYSIS OF BRAIN ELECTRICAL ACTIVITY FRANK H. DUFFY *, PETER H. BARTELS ** and JAMES L. BURCHFIEL *

• Seizure Unit and Developmental Neurophysiology Laboratory, Department of Neurology, Children's Hospital, Medical Center and Harvard Medical School, Boston, Mass. 02115, and ** Department of Microbiology and Optical Sciences Center, University of Arizona, Tucson, Ariz. 85721 (U.S.A.)

(Accepted for publication: December 30, 1980)

Electroencephalography (EEG) has proven to be an important diagnostic tool in the identification of certain neurolOgic diseases that produce focal and/or paroxysmal abnormalities (e.g. abscess, epilepsy). With a few exceptions, however, EEG has proven less useful for identifying neurologic diseases that cause more subtle alterations of background activity (e.g. mental retardation, presenile dementia). EEG's cousin, the long latency cortical evoked potential (EP), has failed even to achieve the level of clinical utilization afforded EEG. We hypothesize that these apparent diagnostic limitations of EEG and EP do not reflect a lack of sensitivity to underlying brain abnormality. On the contrary, we suggest that such measures of brain electrical activity contain not too little but too much information to be easily grasped by unaided visual inspection of polygraphic records. To assist clinical appraisal of such data we have recently developed a system for the topographic mapping and computerized display of scalp recorded signals referred to as brain electrical activity mapping or BEAM (Duffy et al. 1979). This methodology takes advantage of pioneering works in topographic display techniques (Walter and Shipton 1951; Bechtereva et al. 1963; Vaughn and Ritter 1970; Gotman et al. 1975; Bickford 1976; Petsche 1976) and of latest developments in solid state electronics. As routinely used in our clinical laboratory, BEAM images are con-

structed from both EEG and EP data. For EEG data, we display the spatial distribution of activity in the classic delta, theta, alpha and beta frequency bands. For EP data, we visualize the dynamic change of electrical activity with time by sequential display of images; the display technique produces an animation effect highlighting the spread of EP activity over the scalp. These methods condense and summarize the spatio-temporal information obtained from multielectrode recordings and assist the clinician by providing him with immediate visual perception of the data. The field of 'image processing' (Grasselli 1969) has produced numerous techniques for enhancing the visibility of information contained in displays such as the BEAM. Most of these techniques have involved optical to optical transformations such as the enhancement of gradients or sharpening of edges. We have been exploring the utility of another type of image processing, statistical to optical transformations, in which statistical information derived from an image or set of images is itself transformed into an image. This technique, known as 'significance probability mapping' or SPM, was first introduced by Bartels and Subach (1976) as a method to extract information from light microscopic and Sonar images. We have adapted this technique to the analysis of BEAM data and report its potential value in the localization of anatomical brain lesions and functional brain asymmetries.

456

Methods

(A ) Data acquisition and formation of BEAM images Data were recorded from 20 EEG electrodes in the International 10-20 format including standard locations Oz, Pz, Cz, Fz, O2, P4, C4, F4, W6, T4, Fs, O1, P3, C3, Fa, Ts, T3, F~, Fpl, and Fe:. For the results described in this paper, all recordings were monopolar, referenced to linked ear electrodes. In other instances, bipolar montages can and have been used. Signals were recorded on analog magnetic tape and later processed on a PDP-12 or PDP-11/60 computer using programs of SIGSYS biomedical software package. Segments o f raw data were inspected on the computer video screen prior to analysis. ThOse containing eye movement, muscle artifact, or other interference were eliminated. EEG data were Fourier transformed via a conventional fast algorithm (FFT) to yield a spect r u m over the range 0--32 c/sec. Average visual evoked potential (VEPs) were formed to over 500 strobe flashes (Grass model PS2 at intensity 4 placed 30 cm from subject's open eyes; analysis time 512 msec). A single spectrum or VEP was derived for each o f the 20 electrodes. Next the set of 20 spectra or EPs was processed into topographic maps of electrical field distribution using the BEAM technology under SIGSYS control as previously described (Duffy et al. 1979). First, single values were obtained or calculated from each of the 20 electrodes. Such values were either the EP amplitude at some point in time o r the a m o u n t of energy in a spectral analyzed frequency band. Next, these 20 values were mapped onto a 64 × 64 numerical matrix. Each interpolated value was based upon the 3 nearest real values. For the study described in this report topographic maps of spectral alpha (8--12 c/sec) activity were formed for visual display on a television monitor and the underlying 64 × 64 numerical matrices were retained in computer memory for subsequen.t analysis. Likewise, for EP analysis, images and matrices of EP data were

F.H. DUFFY ET AL.

retained. However, instead of a single BEAM spectral image, an EP data set consisted of a series of 128 images, each representing the average voltage distribution o f VEP activity during a 4 msec epoch.

(B) Formation maps (SPMs)

of significance probability

In the analysis of BEAM images we formed two types of SPM. t statistic SPM. The picture elements (pixels) of this SPM represent individually calculated t statistics between the mean pixel values of two different sets of BEAM matrices (Fig. 1). Typically, one would be comparing a control set, in which each individual matrix entering into the mean would be derived from a control subject, with an experimental set, in which matrices would be derived from exper-

SET 1

/ m a p 1/ +

/map 2/ -t-

SET 2

Sum

/map 1/

Individual "1/map 2/ Maps Point- by - point +

+

/map

set

mean

+

n/

1~,

Calculate Mean / m a p and Variance Make Maps

j/variance j

/

n/

set 2 /__.

/mean

/7

/variance/ Calculate t and

Statistic

/

t-map statist~'c/ / SET 1 vs SET 2 Fig. 1. Formation of significance probability map by Student's t test (see text for explanation).

SIGNIFICANCE PROBABILITY MAPPING

imental subjects. The t statistic SPM is formed as follows. First the beam matrices for each set are summed on a point-by-point basis, i.e. the values of homologous points in the individual matrices of a set are added together. Next, the mean and variance are calculated for each point and t w o new matrices are formed for each set representing these respective population statistics. Finally, Student's t statistic (Downie and Heath 1974) is calculated for each matrix point (based upon the mean and variance matrices of the t w o sets) and a matrix is formed of these t values. The spatial framework of the pixels within the head outline is retained. Thus, the t statistic SPM reveals regions in which the t w o populations statistically differ from one another. z statistic SPM. The z transform expresses the number of standard deviations an individual observation lies from the mean of a set of similar observations (Downie and Heath 1974). Thus, it measures the magnitude, in a statistical sense, b y which an individual observation differs from the mean of a reference set. The z transform is calculated for each point in a BEAM matrix derived from an individual subject in comparison to the mean and

REFERENCE SET

UNKNOWN SUBJECT

/mean (~)/~ /variance(s2)/

/ individual/ / m a p (x) / Calculate

Point - by - point and

z - statistic~

map

/

UNKNOWNvs REFERENCE Fig. 2. F o r m a t i o n o f significance p r o b a b i l i t y m a p b y z t r a n s f o r m (see t e x t for e x p l a n a t i o n ) .

457

variance BEAM matrices derived from a reference population (Fig. 2). The result of this point-by-point z transformation is a new matrix of z values retaining the spatial framework of the original BEAM images. Thus the z statistic SPM defines regions in which brain electrical activity from an individual subject differs statistically from that of a reference population. The goal of SPM formation is to delineate regional topographic differences in brain activity using the z or t transformation as measurement. The SPM transformation, itself, does not necessarily assess the overall level of group or individual difference. Overall significance of group difference is best assessed b y multivariate statistical techniques (Duffy et al. 1980b).

Results

We will illustrate the utility of SPM with two examples. In the first, the z statistic SPM was used to identify supratentorial brain tumors b y transforming BEAM images of VEP data. In the second, the t statistic SPM was used to define differences in the distribution of EEG alpha activity when subjects were listening to speech and music, respectively. This was done for two populations: b o y s with specific reading disability (dyslexia) and normal boys. From the data already generated, t statistic SPMs were then formed to highlight differences b e t w e e n the two populations. (A ) Brain tumor identification VEPs were derived from 12 patients with confirmed supratentorial brain tumors and from 18 normal control subjects. For each subject these VEP data were processed into a set of BEAM images (128 images each representing average VEP distribution over 4 msec). For the control population a set of mean and variance matrices was calculated and then based u p o n this reference, a set of z statistic SPMs was derived for each t u m o r subject. A clinical neurophysiologist with experience in

458

F.H. D U F F Y ET AL.

459

SIGNIFICANCE PROBABILITY MAPPING TABLE I Assignment of subjects to diagnostic categories based upon different electrophysiological data formats. Data format used as basis for assignment

Raw VEP wave forms BEAM VEP images z statistic SPM

Assigned to tumor category

Assigned to normal category Number correctly assigned

Number incorrectly assigned

Percent correctly assigned

Number correctly assigned

Number incorrectly assigned

Percent correctly assigned

14 17 18

4 1 0

78 94 100

5 8 11

7 4 1

42 67 92

assessing VEP and BEAM, data examined, on a blind basis, the 20 individual raw VEPs, the set of 128 VEP BEAM images, and the set o f 128 SPM images. These 3 data types were separately evaluated. Assignment of subjects into diagnostic categories based upon these examinations is given in Table I. Eleven of the 12 p a t i e n t s with supratentorial brain tumors demonstrated focal deviations from normal as illustrated in Fig. 3A and 3B. In every case, the z value exceeded 2.5 standard deviations for at least 75 msec. In most instances, the m a x i m u m z value reached or exceeded 5. The one t u m o r not identified b y SPM was a subtle left frontoparietal lesion in a patient with a normal EEG. The CT scan was initially read as normal b u t

this opinion was subsequently revised. No histological confirmation is available for this patient. For 8 o f the 11 patients, the region of high z value coincided with the locus of the tumor, its surrounding edema, or b o t h . In 2 cases (deep left parietal and right occipital), equal bilateral regions were found on SPM. In one case {left occipital-temporal) the SPM localization was slightly more contralateral than ipsilateral to the anatomical lesion.

(B) Alpha distribution during speech and music EEG data were recorded from 13 normal and 11 dyslexic b o y s ranging in age from 9 to 11 years. Dyslexia was defined according to the criteria of Rutter and Yule (1971) as

Fig. 3. Clinical use of SPM. A: CT scan of a 59-year-old male with a confirmed glioblastoma multiforme of t h e right parietal-posterior temporal region. There is considerable edema surrounding the tumor. B: z transform SPM of same patient. The SPM is shown in vertex view, nose toward top of page, left to viewer's left. The temporal regions appear exaggerated due to the method of projection, which is designed to give equal representation to all scalp regions. The SPM represents the sustained deviation from normal of the visual evoked potential (VEP) over the latency interval 290--366 msec. Each color of the display represents a particular value of z from 1 (dull red) to 7 (mint green). Note how the SPM highlights the VEP abnormality and the close agreement between the region of maximum z and the location of the tumor on CT scan. Note also that the physiologically derived SPM identifies a more extensive area of abnormality surrounding the tumor; this probably represents functional deficits due to edema. C--D: t statistic SPMs showing differences in alpha activity (8--12 Hz) within a group of normal subjects depending upon whether they are listening to speech (C) or music (D). In each case the t statistic is derived in comparison to alpha activity recorded during the state of restful alertness, eyes closed with no auditory stimulation. In these and subsequent plots, different colors represent different values of t as follows: light blue-green, t > 1 . 9 6 , P (two-tailed) <0.05; red, t > 2.33, P < 0.02; pink, t > 2.58, P < 0.01. Note that listening to speech significantly alters alpha activity over the left temporal region (C) whereas listening to music affects right temporal alpha (D). E--F: t statistic SPMs showing differences in alpha activity in a group of dyslexic boys listening to speech (E) or music (F). This is the same paradigm as shown in C and D for normal controls. Note that although the dyslexic subjects show well localized alpha differences during speech and music, that the locations of the differences differ from those shown by the control subjects.

460 being more than 1.5 years below expected reading level. In addition, all dyslexics fulfilled the criteria for 'dyslexia-pure' established by Hughes and Denckla (1978): they were of normal I.Q., were not hyperactive, were neurologically normal, and were free of other evidence of learning disability. For every subject, a 2 min epoch of EEG was analyzed for each of 3 states: (1) resting but alert, with no external stimulation; (2) listening to a tape recording of speech (e.g., cumming's fairy tale, 'The Elephant and the Butterfly'); and (3) listening to a tape recording of music (The Dance of the Sugax-plum Fairy from Tchaikovsky's Nutcracker Suite). BEAM images were formed of the distribution of spectral alpha (8--12 Hz) energy during each state. Mean and variance matrices for each state w e r e calculated for the control population and this was repeated for the dyslexics. Finally, t statistic SPMs were derived. Two specific comparisons were made: (1) between alpha distribution resting and alpha distribution listening to speech, and (2) between alpha distribution resting and alpha distribution listening to music. This was done separately for each population. Thus the t statistic SPM was used to define the changes induced by speech and music, respectively, upon the resting pattern of alpha activity, and to examine these changes separately for normal readers and dyslexics. Inspection of the t statistic SPM revealed clear differences between the changes in alpha distribution induced by listening to music and those induced by listening to speech. In addition, these state-induced alpha distributions were different for dyslexics and normal readers. For normal subjects, the greatest speech-induced change in alpha activity occurred over the left temporal region (Fig. 3C). In the same subjects listening to music produced significant chanSes in alpha over the right temporal area (Fig. 3D). By contrast, t statistic SPM comparison within the dyslexic population revealed a markedly different pattern of alpha changes during speech and music presentation. Speech-induced changes were

F.H. DUFFY ET AL. maximal in the left posterior-frontal region and did not significantly affect the temporal lobes (Fig. 3E). While listening to music, dyslexic subjects did show changes in alpha activity in the right temporal region as did normal readers, but in addition, they showed significant changes in the midline occipitalparietal area (Fig. 3F). In every instance, for both normals and dyslexics, the regional changes delineated by SPM represented a decrease of spectral energy in the alpha band. We wish to emphasize that these patterns demonstrated by SPM were not evident from examination of the BEAM images of alpha distribution. Random fluctuations in alpha distribution from subject-to-subject completely obscured any perception of a consistent change in the individual images. Even in the average images of the normal population where it was clear that speech tended to suppress alpha in the left hemisphere and music in the right, precise localization of these changes was possible only by SPM.

Discussion Significance probability mapping (SPM) consists of the replacement of an image by the results of a statistical manipulation of the numerical data that underly the original image (Bartels and Subach 1976). Results of such data abstraction axe presented within the coordinate system or framework of the original data. The goal is to make visible data characteristics that might otherwise remain obscured. Results of the tumor study reaffirm the ability of topographic mapping of brain electrical activity to assist the clinician in the diagnosis of supratentorial brain tumors. When SPM is performed, the lesion becomes more visible as evidenced by the correct diagnostic assignment of 29 of 30 subjects. The reason for this is of some interest. Normal VEP topographic studies contain transient asymmetries. Tumors also induce asymmetries. Identification of patients with a tumor

SIGNIFICANCE PROBABILITY MAPPING from among a population of normals requires the clinician to differentiate tumor-induced asymmetries from normal asymmetries. By statistically delineating the magnitude and location of deviation from normal, SPM assists the clinician in just this differentiation, thereby improving diagnostic success. The finding o f deviations from normal contralateral to the t u m o r in 3 subjects is curious. Possible explanations include the existence of a second lesion, edema, or some sort of deafferentation effect. In any event, these observations illustrate that neurophysiological data analysis often provides additional information to the anatomical definition of the CT scan. Results o f the speech-music study demonstrate the ability of SPM to define regions of functional brain activation within both a normal and pathologic population. Due to variability in the normal distribution of resting alpha activity across all subjects, it was difficult to see in the raw data distinct patterns of change when speech or music were introduced. SPM, however, clearly delineated such changes (see Fig. 3C--F). This, in part, is attributable to the sensitivity of the t statistic to variance as well as difference in mean values. Decreases in alpha activity are often taken to indicate cortical activation. If so, then for our normal boys, speech induced activation of the left temporal lobe and music the right temporal lobe. This is in accord with classic neuropsychological t h e o r y (Hecaen and Albert 1978). The dyslexics, however, showed a different pattern. Whereas speech activated a well defined region, it was in the 'wrong' place, i . e , the left posterior frontal lobe. Similarly, although music activated portions of the right temporal lobe, additional 'abnormal' activation of the occipital lobe was noted. Whether these differences from the normal group signal a different brain organization, a compensatory process, or a different microanatomical brain structure is not known. Results o f this study confirm previous reports of electrophysiological differences in dyslexia

461 and provide additional evidence for regional specificity of dyslexic pathophysiology. Moreover, t h e y demonstrate the utility of SPM in the delineation of regional brain function. SPM in general need not be limited to the t statistic or z transform. One can employ virtuaUy any statistical hypothesis such as variance ratios and non-parametric tests. Furthermore, the serial application of 2 or more statistical transforms may prove useful in that several hypotheses can be tested at once and displays produced of only those regions fulfilling all of them. Analysis need not stop at visualization of SPM. From them, one could develop single numerical measures of abnormality such as the integrated value of the entire SPM matrix or a similar integration limited to the area above a certain significance level. Regions delineated by SPM can serve to indicate regions from which to take measurements on individual subjects. We studied SPM formed by group t statistic comparisons between normals and dyslexic boys (Duffy et al. 1980a,b). Using the regions the SPM outlined, we integrated the raw EEG and EP data from sets of subjects, applied formalized diagnostic rulemaking procedures, and tested the resultant rules on u n k n o w n subjects. With these measures we were 80--90% successful in classifying a group o f subjects not used to generate the decision rules. In conclusion, we have applied SPM to the detection of physical lesions and functional asymmetries of brain. Its relative success in these applications suggests that it may prove to be clinically useful in the analysis of topographic maps of brain electrical activity and thereby enhance the diagnostic value of EEG and EP data.

Summary We illustrate the application of significance probability mapping (SPM) to the analysis of topographic maps of spectral analyzed EEG

462 and visual evoked potential ( V E P ) a c t i v i t y from patients with brain tumors, b o y s with dyslexia, and control subjects. When the VEP topographic plots of t u m o r patients were displayed as number of standard deviations from a reference mean, more subjects were correctly identified than b y inspection of the underlying raw data. When topographic plots o f EEG alpha activity obtained while listening to speech or music were compared b y t statistic to plots o f resting alpha activity, regions of cortex presumably activated b y speech or muSic were delineated. Different regions were defined in dyslexic b o y s and controls. We propose that SPM will prove valuable in the regional localization of normal and abnormal functions in other clinical situations.

Rdsum~

Cartographie de probabilitd de significativitd: aide d l'analyse topographique de l'activitd dlectrique cdrdbrale Les auteurs illustrent l'application de la cartographie de probabilitd de significativitd (CPS) ~ l'analyse des cartes topographiques des activit6s EEG et des potentiels 6voqu6s visuels (PEV) obtenues par analyse spectrale chez des malades atteints de tumeur c6r6brale, des gargons dyslexiques et des sujets de contrble. Les aires topographiques des PEV, dessindes chez les malades atteints de tumeur par nombre de ddviations standards par rapport une r6fdrence moyenne, permettent d'identifier correctement plus de sujets que l'inspection des donndes brutes sous-jacentes. En comparant par test statistique t les aires topographiques d'activitd EEG alpha obtenues au cours de l'6coute de langage ou de musique aux aires d'activitd alpha au repos, on peut ddlimiter les r6gions du cortex que l'on suppose activ6es par le langage ou la musique. Des aires diff6rentes sont ddfinies chez les garqons dyslexiques et les contrSles. Les auteurs sugg6rent que le CPS pourrait constituer une m6thode valable de localisation r6gionale des fonctions normales ou anormales dans d'autres situations cliniques.

F.H. DUFFY ET AL. References Bartels, P.H. and Subach, J.A. Automated interpretation of complex scenes. In: E. Preston and M. Onoe (Eds.), Digital Processing of Biomedical Imagery. Academic Press, N e w York, 1976: 101-114. Bechtereva, N.P., Vvedenskaia, I.V., Dubikaitis, Y.V., Stepanova, T.S., Ovnatov, B.S. and Usov, V.V. Localization of focal brain lesions by electroencephalography. Electroenceph. clin. Neurophysiol.,1963, 15: 177--196. Bickford, R.G. N e w trends in clinical electroencephalography. Practitioner, 1976, 217: 100--107. Downie, N.M. and Heath, R.W. Basic Statistical Methods. Harper and Row, N e w York, 1974. Duffy, F.H., Burchfiel, J.L. and Lombroso, C.T. Brain electrical activity mapping (BEAM): a new method for extending the clinical utility of E E G and evoked potential data. Ann. Neurol., 1979, 5: 309--321. Duffy, F.H., Denckla, M.B., Bartels, P.H. and Sandini, G. Dyslexia: regional differences in brain electrical activity by topographic mapping. Ann. Neurol., 1980a, 7: 412--420. Duffy, F.H., Denckla, M.B., Bartels, P.H., Sandini, G. and Kiessling, L.S. Dyslexia: automated diagnosis by computerized classification of brain electrical activity. Ann. Neurol., 1980b, 7: 421--428. Gotman, J., Gloor, P. and Ray, W.F. A quantitative comparison of traditional reading of the EEG and interpretation of computer-extracted features in patients with supratentorial brain lesions. Electroenceph, clin.Neurophysiol., 1975, 38: 623--639. Grasselli, A. (Ed.) Automated Interpretation and Classification of Images. Academic Press, N e w York, 1969. Hecaen, H. and Albert, M. H u m a n Neuropsychology. Wiley-Interscience, N e w York, 1978. Hughes, J.R. and Denckla, M.B. Outline of a pilot study of electroencephalographic correlates of dyslexia. In: A.L. Benton and D. Pearl (Eds.), Dyslexia -- an Appraisal of Current Knowledge. Oxford University Press, London, 1978: 205--250. Petsche, H. Topography of the EEG: survey and prospects. Clin. Neurol. Neurosurg., 1976, 79: 15--28. Rutter, M. and Yule, W. The concept of specific reading retardation. J. child. Psychol. Psychiat., 1971, 12: 91--113. Shipton, H.W. A new frequency-selective toposcope for electroencephalography. Med. Electron. Biol. Engng, 1963, 1: 483--495. Vaughn, Jr., H.G. and Ritter, W. The sources of auditory evoked responses recorded from the human scalp. Electroenceph. clin. Neurophysiol., 1970, 28: 360--378. Walter, W.G. and Shipton, H.W. A new toposcopic display system. Electroenceph. clin. Neurophysiol., 1951, 3: 281--292.