Biological Diagnosis of Brain Disorders
The Future of the Brain Sciences
Proceedings of the 5-th International Conference held at New York Academy of Medicine, October 2-3,1972
Edited by Samuel Bogoch
Spectrum Publications Inc., Flushing, N.Y., 1973
Pages 275-280

The Computer Analysis of Multiunit Activity of Neuronal Populations of the Human Brain

P.V. Bundzen, Yu.L. Gogolitsin and A.S. Kaplunovsky


The method of implanted electrodes makes it possible to investigate directly the electrical activity of neuronal populations of deep brain structures and to detail their role in the realization of mental activity (Bechtereva, 1971). The development of precise methods of study of the neurodynamics of these populations in connection with the type of activity being realized has thus become of great importance. One of the methods of such study is the investigation of the dynamics of multi-unit activity of neuronal populations (Bechtereva et al., 1971; Bundzen et al., 1971). The complicated noise-like form of the signal under consideration and its wide frequency spectrum (from 500 Hz to 3000 Hz) make any manual processing of the data practically impossible and the use of digital and analog computers for that purpose is desirable.

In this paper a complex method of analysis of multi-unit activity is described, which can be used for the investigation of background activity, functional reorganization of activity under different conditions, and the search of patterns formed by different stimuli. This method was used in the investigation by Bechtereva and Bundzen (see the results published in this book).

The process consists of two stages: first the discrimination analysis stage, and second the digital computer stage, several programs being used during the latter stage.

Discrimination analysis is carried out with the help of "MH-14" analog computer and consists of separating the recorded signal into several amplitude levels. The signal at the output of each level appears only in case the amplitude of spike is limited by the upper and lower thresholds of the level (Fig. 1). In such cases the level separates from the whole recording the signals of the comparatively small group of neurons, the size of which is determined by the distance between the thresholds. The number of groups separated is equal to the number of levels and can be changed.

Fig.1 Method of amplitude discrimination. The separation of five neuronal groups from the recording of multi-unit activity is shown. The abscissa axis - time.

The measurement of current frequency of each level is carried out simultaneously with the discrimination analysis, i.e., the number of impulses per fixed sequential time intervals is counted, time intervals may be chosen within the interval of 20-60 msec.

The second stage of processing consists of factor analysis, dynamic selective correlation analysis, and classification analysis. For that purpose a digital computer "Minsk-32" is being used.

From the large number of factor analysis methods available principal component analysis was chosen (Lawley and Maxwell, 1963), and was used to investigate links and relationships existing between different discrimination levels. In other words, it makes it possible to investigate the interactions of chosen neuronal groups. During the factor analysis procedure not only the correlation coefficients between the discrimination levels are calculated, but also groups of factors, defining the interconnections of different levels.

For example, factor analysis was applied for investigation of relations between neuronal groups when they formed a specific trace (Fig. 2). In such cases the current frequency was analyzed for a period of 500-800 msec, sampling time intervals being 20 msec.

Fig.2 Results of factor analysis. The abscissa axis - the numbers of discriminant levels; the ordinate axis - the coefficient of interconnection. Curves show the main and second-order interconnections between different levels.

Principal component analysis could be applied not only to the correlation matrix, but also to the information matrix calculated after Shannon theory. It should be noted that this form of analysis deals only with the informational properties of interaction between neuronal groups, but not with the way of coding and processing of external signals by these groups. The described methods of mathematical analysis also permit the reception of detailed characteristics of the functional structure of neuronal populations under the action of different kinds of stimuli, and permit a statistical description of the trace to be given.

Investigation of the dynamics of traces is carried out with the help of the dynamic selective correlation method, the aim of which is to calculate the correla- tion coefficients between the specially chosen fragment of activity, usually recorded at the moment of a stimulus presentation, and the fragments of the same length taken from the whole recording of the process under consideration, step by step with a comparatively small delay (Fig. 3). Instead of the special fragment of activity, it is also possible to use some other fragments, such as the time dependence of different parameters of stimulus. These parameters could be, for example, the current frequency or some other characteristics of verbal stimulus. This method proved the existence of a high degree of correlation between the frequency of multi-unit activity at the moment of presentation of the verbal signal to a patient and spectral characteristics of this signal (Bechtereva et al., 1971).

Fig.3 The Method of dynamic selective correlation. (A) Fragment of activity under consideration, (B) a shorter fragment, which was registered during the presentation of the signal, (C) sequence of correlation coefficients. The abscissa axis - time. The clearly seen maxima on C prove the existence of the high degree of correlation at several moments.

After some fragments from the recording of impulse activity, where specific traces were shown to exist, have been separated by the dynamic selective correlation method, these particular fragments should be analyzed with the help of classification analysis.

The methods of classification analysis determine the existence of classes having a high degree of resemblance of dynamics of frequency within a large number of fragments of multi-unit activity recordings. It helps to find periodical changes of current frequency and to investigate the reproducibility and specificity of patterns of multi-unit activity during the presentation of different signals.

The method of automatic classification of multiparametric experimental data (Klimenko et al., 1972) does not require any a priori information about the structure and number of classes due to the fact that all these parameters are completely defined by the data when this method is used. Central points of classes which are considered to be characteristic representatives of the structures found are calculated automatically and described with the aid of multi-dimensional vectors. It is also possible to use the principal components method for the classification, and in that case one can separate groups using several methods. Each one clarifying the resemblance of the data from a different point of view.

The third classification method, which may also be used during the second stage of data processing, is the potential function classification. It is based on the investigation of the resemblance of classified objects, with the aid of their geomet- ric description. For example, as was used for the investigation of multi-unit activity during the presentation of short-term memory tests, unknown English words were presented to a patient before and after he had learned the meaning of them. and the fragments of multi-unit activity during presentation, retention, and verbal response stages were chosen for analysis (Gogolitsin, 1972). When the meaning of a word was unknown to the patient, the frequency patterns appeared to he quite stable during all three stages of the test. After learning, when an engram was formed in the long-term memory, the same pattern was found to reproduce during the presentation stage only. hut during the retention and verbal response it appeared to he quite different (Fig. 4).

Fig.4 The stability of patterns of multi-unit activity in the case of presentation of unknown English words, which were learned afterwards: (I) before learning, (II) and (III) alter learning. Curve III shows the stability of changed pattern. The abscissa axis displays the stages of test: (A) presentation of word, (B) retention, (C) verbal response. The ordinate axis displays the degree ol pattern stability.

The methods described above make it possible to investigate not only the dynamics of traces in the central nervous system, but to understand the reasons defining the appearance of specific patterns ol multi-unit activity of neuronal populations of human brain as well.


1. Bechtereva, N.P. Neurophysiological aspects of mental activity of man. Meditsina, 1971.
2. Bechtereva, N.P., Bundzen, P.V., Matveev, Yu. K., and Kaplunovsky, A.S. "Functional organization of activity of cerebral neuronal assemblies in man during short-term memory." Sechenov Physiol. Journ. USSR 57, (12), 1971.
3. Bundzen, P.V., Vassilevsky, N.N., Kaplunovsky, A.S., and Shabaev, V.V. "Factor analysis in studies of functional organization of the brain electrical activity." Sechenov Physiol. Journ. USSR 57, (7), 1971.
4. Gogolitsin, Yu. L. Report at the symposium "Neurophysiological Mechanisms of Mental Activity" (Leningrad. July 2-5. 1972, to be published).
5. Klimenko, V.M., Kaplunovsky, A.S., and Neroslavsky, N.A. "Automatic Classification of Multiparametric Experimental Data." Sechenov Physiol. Journ. USSR 58, (4), 1972.
6. Lawley, D.N., and Maxwell, A.E. Factor Analysis as a Statistical Method. Butterworths, London, 1963.

WebMaster RK