JP4876988B2 - Brain computer interface device - Google Patents

Description

The present invention relates to a brain computer interface device for transmitting a user's intention to the outside world by observing a user's brain activity (such as an electroencephalogram), and particularly, even an electroencephalogram measured under a single trial. The present invention also relates to a brain computer interface device that can distinguish the trial difference from the measured electroencephalogram.

  The Brain-Computer Interface (hereinafter referred to as BCI) refers to the user's intention to generate electrical signals that can be detected directly from the human scalp, (cerebral) cortex surface, or brain, reflecting brain activity. Is converted to an output that conveys to the outside world (Non-Patent Document 1). In BCI described in Non-Patent Document 1, a raw EEG (raw electroencephogram) measured at specific electrode positions (C3 and C4) for specifying an electroencephalogram is subjected to frequency analysis, and a specific frequency band ( A two-dimensional cursor movement on the display screen is to be realized by a weighted linear sum of amplitude values of μ wave: 8 to 12 Hz or β wave: 18 to 26 Hz. However, the weighting factor must be updated for each trial and requires considerable training to be able to achieve proper cursor movement.

  In the BCI described in Non-Patent Document 2, raw EEG is measured with multiple channels when performing an exercise image task. However, in the method described in Non-Patent Document 2, before applying independent component analysis (hereinafter abbreviated as ICA) and dipole analysis to raw EEG, Laplacian spatial filtering and time-frequency analysis are performed. And lacks real-time performance. Furthermore, when the dipole is estimated in the somatosensory / motor cortex on the same side as the hand to be imaged from the dipole analysis of only the first independent component, the correct answer is never good.

In Japanese Translation of PCT International Publication No. 2006-524157 (WO 2004/083972) (Patent Document 1), in order to support driving of a vehicle or the like, a current signal in the brain of the driver of the vehicle is detected to start a brake, for example. It is disclosed. Japanese Laid-Open Patent Publication No. 2006-280421 (Patent Document 2) discloses a brain function information monitoring apparatus that removes unnecessary components contained in a detection signal so that measurement using a brain function signal can be performed with high accuracy. . Japanese Patent Laying-Open No. 2003-84793 (Patent Document 3) discloses an independent component analysis method and apparatus suitable for application to an electroencephalogram signal.
Special table 2006-524157 (WO 2004/083972) JP 2006-280421 A JP2003-84793A Urpo J.R., McFarland D.J, "Control of a Dimensional Movement Signal by a Non-Invasive Brain-Computer Interface in Humans (Control of a two-dimensional movement signal by a noninvasive brain-computer interface in humans), Proceedings of the National Academy of Sciences of the United States of America, 2004. Year, Vol. 101, p. 17849-17854 Quinn L, Din L, Ha B, “Motor imagery classification by means of source analysis for brain-computer interface applications (Motor imagery classification by means of source analysis for brain-computer interface applications), Journal of Neural Engineering, 2004, Vol. 1, p. 135-141 Cardoso J. -F, Thoreau Miac A, "Blind beam-forming for Gaussian signals", EI Proceedings-F, 1993, Vol. 140, p. 362-370 Hiberainen A, Oja E, “A fast fixed-point algorithm for independent component analysis”, Neural Computation, 1997 Vol. 9, no. 7, p. 1483-1492 Amari S, Chen T, Chichokki A, “Nonholonomic orthogonal learning algorithm for blind source separation”, Neural Computation, 2000, Vol. . 12, p. 1463-1484 Choi S, Chichokki A, Amari S, “Flexible independent component analysis”, Journal of BUISI Signal Processing, 2000, Vol. 26, no. 1/2, p. 25-38 Friedman N, Murphy K, Russell S, "Learning the structure of dynamic probabilistic networks", Proceedings of Fourteenth Conference On Answer In Artificial Intelligence (UAI 98), 1998, p. 139-147 Schwartz G. “Estimating the dimension of a model”, Anals of Statistics, 1978, p. 461-464 Friedman N, “The Bayesian structural EM algorithm”, Fourteenth Conference on Thirty in Artificial Intelligence (UAI-98), 1998 Takahiro Yamanoi, Sakai Traveler, “About one method of quantifying 3-way data”, Hokkaido University research report, No. 132, 1986, p. 155-160

The object of the present invention is to eliminate the above-mentioned problems of the prior art in the brain computer interface, and to deal with each different trial using brain waves measured under a single trial. It is an object of the present invention to provide a brain computer interface device capable of detecting a spatiotemporal pattern of brain activity.

The brain computer interface apparatus of the present invention includes an independent component analysis execution unit that applies independent component analysis to digital brain wave data of a plurality of channels measured by a user, and each independent component output from the independent component analysis execution unit. Is projected onto each channel, and the independent component projection execution unit that calculates the amplitude of the EEG for each independent component at the electrode position at the time of EEG measurement, and the dipole analysis on the multichannel EEG data projected by the independent component projection execution unit Based on the results obtained in the applied dipole analysis unit, independent component analysis execution unit, independent component projection execution unit, and dipole analysis unit, it consists of brain regions where trial, time, and dipole are estimated for each independent component Independent component analysis / dipole analysis result storage unit for generating 3D data and brain activity pattern for learning 3D data It has a learning section, a.

The brain computer interface device of the present invention further includes an electroencephalogram measurement unit that measures a user's electroencephalogram with a plurality of channels using a plurality of electrodes and converts the measured electroencephalogram signal into analog / digital data to obtain digital electroencephalogram data. It may be provided. In addition, brain activity learning unit is equipped with a dynamic probability network construction unit to learn the brain activity patterns by using a dynamic probability network Luke, or, the brain activity patterns by using a 3-way category data analysis that features a 3-way category data analysis unit to be learned.

Hereinafter, each component constituting the brain computer interface device of the present invention will be described.

  (1) Electroencephalogram measurement unit: For example, a plurality of electrodes for electroencephalogram measurement are bonded to the surface of the human scalp, or a cap (so-called electrode cap) with an electroencephalogram measurement electrode attached is attached to the human head. An analog signal measured through each electrode is amplified by a biological signal amplifier, converted from analog to digital (A / D), and stored as digital signal data (digital brain wave data x).

  (2) Independent component analysis execution unit: Applies independent component analysis to digital electroencephalogram data x. Specifically, a matrix W that satisfies u = Wx is calculated. However, u is an independent component. If n is the number of electrodes, p is the number of sampling points, and k is the number of independent components, x is an n × p-order matrix and u is a k × p-order matrix.

(3) Independent component projection execution unit: The amplitude value of the electroencephalogram at the original electrode position to which the independent component u _j contributes is calculated by x _j = W ⁺ u _j . Here, W ⁺ is a generalized inverse matrix of W, and u _j is only the j-th row of the matrix u, and the rest are all zeros.

  (4) Dipole analysis execution unit: Applies dipole analysis to projected multichannel EEG data. The analysis results are the location, orientation, and magnitude of brain activity at each sampling point.

  (5) Independent component analysis / dipole analysis result storage unit: The results obtained from the independent component analysis execution unit, the independent component projection execution unit, and the dipole analysis execution unit can be obtained by ignoring the estimated dipole direction and size. It can be set as three-dimensional data composed of trials, time (sampling points), and intracerebral sites where dipoles are estimated for independent components. The independent component analysis / dipole analysis result storage unit generates such three-dimensional data.

  (6) Brain activity pattern learning unit: The brain activity pattern obtained by the independent component analysis / dipole analysis result accumulation unit is learned using a dynamic probability network or using 3-way category data analysis.

In the BCI (Brain Computer Interface) device of the present invention, an ICA and a dipole analysis are immediately applied to a raw EEG measured in multiple channels at the time of performing a single trial. By grasping it as a function of time and trial and applying a dynamic probability network or multivariate analysis to this three-dimensional data, the spatiotemporal pattern of brain activity corresponding to each trial is learned. Thereby, even in an electroencephalogram measured under a single trial, it is possible to accurately and efficiently discriminate the trial difference from the measured electroencephalogram.

According to the present invention, it is possible to accurately and efficiently detect and identify a spatiotemporal pattern of intracerebral activity corresponding to a single trial even in an electroencephalogram measured under a single trial in which noise contamination is sufficiently considered. it can. The present invention is useful not only for practical application of BCI devices , but also for research progress on new BCI devices .

  Next, a preferred embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a BCI (Brain Computer Interface) device used in the first embodiment of the present invention. This BCI device uses a plurality of electrodes to measure a user's brain wave in a plurality of channels, and applies an independent component analysis to the brain signal measuring unit 11 that uses the measured brain wave signal as digital signal data. Independent component analysis executing unit 12 and independent component output from independent component analysis executing unit 12 are projected onto each channel, and an independent component for calculating an amplitude value of an electroencephalogram for each independent component at an electrode position during electroencephalogram measurement Obtained by the projection execution unit 13, the dipole analysis unit 14 that applies dipole analysis to the multichannel EEG data projected by the independent component projection execution unit, the independent component analysis execution unit, the independent component projection execution unit, and the dipole analysis unit Based on the results, independent component analysis that generates three-dimensional data consisting of brain regions where trial, time, and dipole are estimated for each independent component Dipole analysis result storage unit 15, and a brain activity pattern learning unit 16 for learning the three-dimensional data.

  In the following, the three types of line drawing stimuli shown in FIG. 2 (referred to as “rare target”, “rare non-target” and “frequent non-target”, respectively) are randomly assigned 20% (rare target), 20 % (Rare non-target), 60% (frequent non-target) time on the display screen. When the line drawing stimulus is rare target, the subject shakes his hand with the right hand of the line drawing. When the line drawing stimulus is rare non-target, he shakes his left hand with the left hand of the line drawing. Instruct subjects to do nothing at -target. During each trial run, raw EEG is measured from above the subject's scalp using a multi-channel (eg, 32 channels) electrode cap attached to the subject's head. A method of learning the spatiotemporal pattern of brain activity corresponding to the right-hand image and the left-hand image using the single-trial EEG data obtained in this way, and identifying the difference between trials (right-hand / left-hand image) will be described. A specific configuration for electroencephalogram measurement is shown in FIG.

  In the example shown in FIG. 3, a subject is placed in an electromagnetic shield room and an eye movement diagram (EOG) and an electromyogram (EMG) are acquired in addition to an electroencephalogram (EEG). The electroencephalogram is acquired by attaching an electrode cap or a plurality of electrodes to the subject's head, amplified by a biological signal amplifier, and A / D (analog / digital) converted. The eye movement diagram and electromyogram are amplified by the polygraph and A / D converted. In addition, an electrode position sensor for measuring the position of the electrode is provided. In order to give a line drawing stimulus to a subject, an AV taxoscope (instant exposure device) is provided. In general, an EEG measurement system consisting of a personal computer (PC) or workstation (WS) is provided. The trigger signal from the AV taxiscope, the signal from the electrode position sensor, the digitized electroencephalogram, and the eyeball The exercise diagram and the electromyogram signal are input to this EEG measurement system. Therefore, the EEG measurement system functions as the independent component analysis execution unit 12, the independent component projection execution unit 13, the dipole analysis unit 14, the independent component analysis / dipole analysis result storage unit 15, and the brain activity pattern learning unit 16 in the present embodiment. Will be fulfilled. FIG. 3 also shows the arrangement of electrodes on the scalp when performing 32-channel electroencephalogram measurement.

  Next, each of the electroencephalogram measurement unit 11, independent component analysis execution unit 12, independent component projection execution unit 13, dipole analysis unit 14, independent component analysis / dipole analysis result storage unit 15 and brain activity pattern learning unit 16 will be described in detail. explain.

EEG measurement unit 11:
As the electrode cap, for example, Electrocap manufactured by ECI can be used. The analog signal measured from each electrode is amplified by a biological signal amplifier (for example, Biotop 6R12-4 manufactured by Japan GE Marquette) and A / D converted to obtain digital signal data x (see FIG. 3). Here, it is assumed that the number of electrodes is 32, and in each trial, an electroencephalogram is measured from 100 ms before onset of line drawing stimulation to 600 ms after onset, and the sampling frequency in the electroencephalogram measurement is 1 kHz. At this time, the digital signal data x is a 32 × 700 order matrix for each trial.

Independent component analysis execution unit 12:
Next, the independent component analysis execution unit 12 applies ICA (independent component analysis) to the digital signal data x output from the electroencephalogram measurement unit 11. ICA is an algorithm that searches for mutual independence based on higher-order order statistics. Specifically, u = Wx
As a result, the matrix W is determined so that u becomes an independent component. If the number of independent components is k, W is a k × n-order matrix. As algorithms for obtaining W, JADE (robust joint approximate diagonalization of eigen matrices) (Non-patent document 3), Fast ICA (non-patent document 4), SANG (self-adaptive natural gradient algorithm with nonholonomic constraints) (Non-patent document 5) Various types such as NG-FICA (natural gradient-flexible ICA) (Non-Patent Document 6) are known. Academic-free software (for example, ICALAB for Signal Processing, http://www.bsp.brain.riken.jp/ICALAB/ ICALABSignalProc /) that can be executed on the computer software MATLAB Labbox can be used. .

From the result of the above independent component analysis, the original matrix x is x = W ⁺ u
Can be reconfigured. However, since W ⁺ is generally k ≠ n (k is p at maximum), it is a generalized inverse matrix (n × k order) of W, and u is a k × p order matrix.

Independent component projection execution unit 13:
Independent component projection execution unit 13, to the determined by independent component analysis execution unit 12 matrix W, then perform a project operation, obtaining the multi-channel EEG data x _j. The amplitude value of the electroencephalogram at the original electrode position to which the independent component u _j (j = 1,..., K) contributes is
x _j = W ⁺ u _j
Can be obtained. However, u _j is a matrix in which all but the j-th row of the matrix u are set to 0.

Dipole analysis execution unit 14:
The dipole analysis execution unit 14 applies dipole analysis to the multichannel electroencephalogram data x _j projected by the independent component projection execution unit 13 for each independent component (j). For example, NEC dipole analysis software “SynaCenter” can be used for the dipole analysis.

Independent component analysis / dipole analysis result storage unit 15:
The independent component analysis / dipole analysis result accumulation unit 15 accumulates the results of ICA and dipole analysis, and obtains a spatiotemporal pattern of brain activity based on the results.

  Assuming unconstrained dipoles as the dipole model, dipoles can be estimated for each of all sampling points. However, as illustrated in FIG. 4, the time change of each independent component immediately after application of ICA tends to have one or two extreme values at a certain time, and therefore, only one at those times. Alternatively, dipole analysis is applied assuming two unconstrained dipoles. At this time, the estimated time was assigned to 0 to 100 ms, 101 ms to 200 ms, 201 ms to 300 ms, 301 ms to 400 ms, 401 ms to 500 ms, 501 ms to 600 ms, that is, any of the six sections. Further, since the number of electrodes is 32, the number of independent components is 32 at most. However, according to experiments by the present inventors, as a result of applying ICA to actual data, it has been found that the number of meaningful (that is, not noise) independent components is at most 20. After all, assuming that dipole analysis is performed for each trial in all the above time widths, 20 (the number of independent components) × 1-2 (the number of dipoles) × 6 (the number of time widths) = 120 ˜240 estimation results (regions in the brain) will be obtained. Thus, a spatiotemporal pattern of brain activity is obtained for each trial. These results are modeled as shown in FIG.

Brain activity pattern learning unit 16:
The brain activity pattern learning unit 16 corresponds to two types of trials (right hand image and left hand image) using the spatiotemporal pattern of brain activity for each trial obtained in the independent component analysis / dipole analysis result accumulation unit 15. A dynamic probabilistic network (hereinafter abbreviated as DPN) is constructed. In the first embodiment, a dynamic probability network constructing unit 21 for constructing a DPN is provided in the brain activity pattern learning unit 16. DPN is an extended expression of a probability network (probabilistic network, hereinafter abbreviated as PN) that describes a probability distribution on a fixed set of random variables in order to model a temporal transition process (non-patented). Reference 7).

  Here, before explaining PN and DPN in detail, some preparations for notation are made below.

Consider a distribution on a set X of discrete random variables. However, each variable X _i takes a value from one finite set and writes it as Val (X _i ). The size of Val (X _i ) is denoted as ‖X _i ‖. The variable names are represented by capital letters such as X, Y, and Z, and the values taken by these variables are represented by small letters such as x, y, and z. Given a network, X = {X ₁ ,..., X _n } represents a network variable having a topological order, and P represents a joint probability distribution on the variables in X.

PN is an annotated DAG (directed acyclic graph) that encodes a joint probability distribution on X. Formally, the PN for X is expressed as B = (G, Θ). However, G is DAG and the vertex corresponds to random variables X ₁ ,..., X _n . These random variables encode the following set of conditional independence assumptions. This assumption is that when a parent Pa (X _i ) is given in G, each variable X _i is independent of its non-descendants. Θ represents a set of parameters that quantify the network. In the simplest case, it takes each possible value k _{i of} X _i and each possible value of Pa (X _i ) for a set of j _i parameters

including. Given G and Θ, one PN, B, uniquely defines a joint probability distribution on X, which is

Given in.

The DPN is an expression obtained by extending the PN in order to model a time process. For simplicity, assume that changes occur between discrete time points indexed by non-negative integers. In this case, it is assumed that X = {X ₁ ,..., X _n } is a set of attributes whose process changes. X _i [t] is the random variable and represents the value of the attribute X _i at time t, and X [t] is a set of random variables X _i [t].

In order to express what is believed about the trajectory that the process can take, a probability distribution on the random variable X [0] ∪X [1] ∪X [2] ∪. Of course, one such distribution is likely to be very complex. Therefore, the process has a Markov property with respect to X, that is,
P (X [t + 1] | X [0],..., X [t]) = P (X [t + 1] | X [t])
Assuming Further, it is assumed that the process is stationary, that is, the transition probability P (X [t + 1] | X [t]) does not depend on t.

Under the assumption of Markovism and stationarity, a DPN that represents the joint distribution over all possible trajectories of a process is composed of two parts:
B ₀ : specifies a distribution on the prior network, initial state X [0];
B _→ : Transition network, distribution on variable X [0] ∪X [1] for specifying transition probability P (X [t + 1] | X [t]) for all t.

  FIG. 6 shows a simple example. In a transit network (but not a prior network), the variables in X [0] have no parent. The transition probability implied by such a network (NW) is

Given in.

The DPN defined by (B ₀ , B _→ ) corresponds to a semi-infinite network on variables X [0],..., X [∞]. In practice, only finite intervals 0,. For this purpose, the DPN structure can be conceptually “unrolled” to one PN on X [0],..., X [T]. In slice 0, the parent of X _i [0] is the parent identified in prior network B ₀ ; in slice t + 1, the parent of X _i [t + 1] is the slice t and slice t + 1 Corresponding to the parent of X _i [1] in B 1 _→ . Similarly, copy the conditional distribution for these variables. FIG. 6B shows a result of unrolling the network (NW) of FIG. 6A with respect to three time slices. Given one DPN model, the simultaneous distribution on X [0], ..., X [T] is

Given in. However,

Is obtained in an obvious way from the transition model.

By comparing FIG. 5 and FIG. 6, in the present invention, X ₁ , X ₂ ,... Correspond to different brain regions such as the visual cortex and the visual precortex, and t = 0, 1,. For example, it can be seen that it corresponds to ˜100 ms, ˜200 ms,.

  In the following, the data to be learned will be described separately for (a) complete data and (b) incomplete data.

(A) Procedure for learning DPN from complete data The problem of learning PN is formulated as follows:
“When a learning set D consisting of instances of X is given, find the network B = (G, Θ) that best matches D”
The concept of best match is defined using a score function. The most frequently used score functions are BIC (Bayesian Information Criteria) and BDe. Both of these score functions are the likelihood of data according to the network L (B: D) = log Pr (D | B)
Combined with some penalty related to the complexity of the network. Complexity penalties are essential when learning the structure of a PN. This is because the maximum likelihood network is usually a fully connected network.

BIC and BDe scores are derived from the posterior probabilities of the network structure. Assume that the random variable G is in the range of possible network structures that may in fact be obtained in the real world. At this time, using Bayes rule, the posterior distribution on G is
Pr (G | D) ∝Pr (D | G) Pr (G)
Meet. However, given a network structure, the likelihood of the data can be calculated by making it conditional on the relevant network parameters:
Pr (D | G) = ∫Pr (D | G, Θ) Pr (Θ | G) dΘ (2)
It can be clearly difficult to identify the prior parameters and evaluate this integral. One approach to avoiding the complete calculation of the integral of equation (2) is to examine the asymptotic behavior of this term. Given the majority of data points, the posterior probability is not sensitive to prior probability selection (assuming that the prior probability does not give a probability of 0 for any event). Non-Patent Document 8 derives the following asymptotic estimator for well-behaved prior probabilities:

However,

Is a parameter setting for G that maximizes the likelihood of the data. N is the number of learning instances, #G is the dimension of G (in the case of complete data, it is the number of parameters), and O (1) is a constant term independent of N and G. The BIC score uses equation (3) to rank the candidate network structure. This obviates the need for prior parameters, and structural prior probabilities result in counting parameters.

(i) BIC score calculation method for DPN Assume that a learning set consisting of N _seq complete observation sequences is given. It has such an l-th column length N _l and specifies a value for the variables x ^l [0],..., X ^l [N _l ]. One such data set contains N _seq instances of the initial slice and the transition

Give instances. B ₀ can be learned from the former instance and B _{1 →} can be learned from the latter instance.

  Here are some notations.

  Similarly, for t = 1,.

  Also, for each family, you need some notation for the statistics enough:

However, I (•; x ^l ) is an indicator function, which takes a value of 1 if an event occurs in the sequence x ^l and takes a value of 0 otherwise.

  By using equation (1) and rearranging each term, just like PN, we find that the likelihood function is decomposed according to the structure of DPN as follows:

Thus, the log likelihood is

Given in.

  Using the standard maximum likelihood estimator for the multinomial distribution,

Get the following expression for:

In the case of a transition network, it is obtained similarly. In the case of conditional probability tables, the number of parameters in the network is # G = # G ₀ + # G _→
Given by. However,

The same applies to the case of the transition network.

Finally, by substituting into equation (3), the BIC score is
BIC (G: D) = BIC ₀ + BIC _→ (6a)
Given in. However,

(ii) Method for calculating BDe score for DPN Since it is described in detail in Non-Patent Document 8, it is omitted here.

(B) Procedure for learning DPN from incomplete data The main difficulty with learning from partial observation is that it no longer has the nature of the score decomposition of equation (4). This means that the selection of optimal parameters in one part of the network depends on the parameter selection in other parts of the network. The most commonly used method for alleviating this problem is the EM (Expectation-Maximization) method. The E step of EM uses the currently estimated parameters to complete the data by calculating the expected value of the count. The M step then re-estimates the maximum likelihood parameter value as if the expected value of that count was a true observed count.

After all, the procedure for obtaining a DPN that best matches complete data or incomplete data D when using BIC as a score function can be summarized as follows:
<< 1 >> Select (B ₀ ⁰ , B _→ ⁰ ) as randomly as possible;
<< 2 >> Hereinafter, n = 0, 1,... Is repeated until (B ₀ ^{n + 1} , B _→ ^{n + 1} ) = (B ₀ ⁿ , B _→ ⁿ ) is satisfied. Using the EM ^{_{method, (B 0 n, B →}} n) modifying the parameters ^{_{of, (B 0 n, B →}} n) based on the count number of the expected value calculated under the possible DPN Structure All Explore. Specifically, from the analogy of the results of Non-Patent Document 9, in the case of incomplete data, the magnitude relationship of the expected value of the DPN score guarantees the magnitude relationship of the true DPN score. Substitutes these expected values. That is, when D is complete data, the equations (6a)-(6c) are used as they are for the calculation of BIC, but when D is incomplete data, equations (5 '), (5 "), (6b) and (6c)

Replace with the expected values given below:

<< 3 >> In << 2 >>, the DPN having the best BIC score value is set as (B ₀ ^{n + 1} , B _→ ^{n + 1} ). If (B ₀ ^{n + 1} , B _→ ^{n + 1} ) = (B ₀ ⁿ , B _→ ⁿ ), return (B ₀ ^{n + 1} , B _→ ^{n + 1} ) and end, otherwise Set n to n + 1 and return to << 2 >>.

Next, a second embodiment of the present invention will be described. The BCI device of the second embodiment is the same as that of the first embodiment, but the brain activity pattern learning unit 16 performs 3-way categorical data analysis instead of constructing the DPN. Thus, it is different from that of the first embodiment. Therefore, the BCI apparatus according to the second embodiment is provided with a 3-way categorical data analysis unit 22 instead of the dynamic probability network construction unit 21 in the brain activity pattern learning unit 16 in the configuration shown in FIG. It has been.

  Hayashi's quantification theory is one method of categorical data analysis. The purpose of this method is to analyze so-called 2-way data consisting of an individual and an item category with one attribute. This is a data analysis method related to the viewpoint of an individual and one attribute. However, generally obtained categorical data does not necessarily have one type of attribute. For this reason, Non-Patent Document 10 proposes a three-way categorical data analysis method that is an extension of quantification type II.

The number of samples is measured from two sets of viewpoints (Q _j , R _k ) (j = 1,..., L k = 1,..., M) for n individuals P _i (i = 1,..., N). Assume that a simultaneous pattern δ _i (j, k) is given. Where δ _i (j, k) is

It is defined as Here, it is assumed that n individuals are divided into g groups. Consider the following issues for this 3-way data:
<Problem> Based on the given simultaneous pattern δ _i (j, k) and group information, individuals P _h having a simultaneous pattern d _h (j, k) whose group is unknown are classified into groups.

In this embodiment, P _i is each trial, Q _j is a different brain region, R _k is a time interval, a group is a right / left hand image, and d _h (j, k) is measured as test data. (single-trial) This is a result of applying ICA and dipole analysis to EEG, and corresponds to data to be classified into right-handed images, left-handed images, or other groups.

  In preparation for solving this problem, we define the following quantities:

D _i (α) ≡δ _i (j, k)

α≡m (j−1) + k
L≡l × m
Note that the ranges that α, β, and γ can take are α = 1,..., L; β = 1,..., M (= 1 + m); Furthermore, the dummy variable for each cell of the simultaneous pattern is b _jk , and the dummy variable for the row and column of the simultaneous pattern is c _j and d _k , respectively. Further, if the dummy variables and a _gamma for each group, these by, and defines the value by group y _i regarding individual P _i, the value z _i by simultaneous pattern as follows.

  In addition, the following quantities are introduced as dummy vectors.

a≡ [a ₁ , a ₂ ,..., a _g ]
b≡ [b ₁₁ , b ₁₂ , ..., b _1m , b ₂₁ , ..., b _2m , ... b _l1 , b _l2 , ..., b _lm ]
The purpose of quantification, as y _i, correlation z _i is the maximum, be determined dummy vector b, it allows the quantity of the z _i by (6.2.2) below. Since the correlation in this case is a canonical correlation,

To take. Where Σ ₁₁ , Σ ₂₂ and Σ ₂₁ are variance / covariance matrices defined as follows:

However, δ _i (γ)

There is a rank drop in the above variance / covariance matrix. Therefore, it is necessary to remove arbitrary elements one by one. Therefore, “g” with respect to δ _i (γ) and a, and L with respect to D _i (α) and b with the L-th element removed are represented by adding “˜” on the character representing the vector or matrix. I will do it. Then, equation (6.2.3) becomes

It becomes. This formula has the following conditions:

Think about maximizing under. Using Lagrange's undetermined multipliers λ and μ

You can. This

From

Get. In consideration of conditional expressions (6.2.5) and (6.2.6), λ = μ. After all, this is

Can be reduced to a general eigenvalue problem. From this formula

Together with (6.2.7) and (6.2.8)

Get. This and (6.2.2) formula, quantification for simultaneous pattern P _h is

It is obtained as However, the pattern vector is

It is.

It is a block diagram which shows the structure of the BCI (brain computer interface) apparatus used in the 1st and 2nd embodiment of this invention. It is a figure which shows three types of line drawing stimuli utilized in the electroencephalogram measurement experiment. It is a figure which shows an example of the arrangement | positioning for an electroencephalogram measurement. It is a figure which shows the example of the result of having projected the example of application of independent component analysis, and an independent component. It is a figure which shows the example of the result of having further performed dipole analysis to the independent component analysis. It is a figure which shows the example of a dynamic probability network model.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 EEG measurement part 12 Independent component analysis execution part 13 Independent component projection execution part 14 Dipole analysis execution part 15 Independent component analysis and a dipole analysis result storage part 16 Brain activity pattern learning part 21 Dynamic probability network construction part 22 3 way categorical Data analysis department

Claims (3)

Brain computer interface device using brain waves,
An independent component analysis execution unit that applies independent component analysis to digital electroencephalogram data of multiple channels measured by a user;
Each independent component output from the independent component analysis execution unit is projected onto each channel, and an independent component projection execution unit that calculates an amplitude value of an electroencephalogram for each independent component at the electrode position at the time of electroencephalogram measurement,
A dipole analysis unit that applies dipole analysis to the multi-channel EEG data projected by the independent component projection execution unit;
Based on the results obtained by the independent component analysis execution unit, the independent component projection execution unit, and the dipole analysis unit, three-dimensional data including trial, time, and a portion of the brain where the dipole is estimated for each independent component Independent component analysis / dipole analysis result storage unit that generates
A brain activity pattern learning unit for learning the three-dimensional data;
I have a,
The brain activity learning unit uses the dynamic probabilities network comprises a dynamic random network construction unit that learns the brain activity patterns, brain computer interface device. Brain computer interface device using brain waves,
An independent component analysis execution unit that applies independent component analysis to digital electroencephalogram data of multiple channels measured by a user;
Each independent component output from the independent component analysis execution unit is projected onto each channel, and an independent component projection execution unit that calculates an amplitude value of an electroencephalogram for each independent component at the electrode position at the time of electroencephalogram measurement,
A dipole analysis unit that applies dipole analysis to the multi-channel EEG data projected by the independent component projection execution unit;
Based on the results obtained by the independent component analysis execution unit, the independent component projection execution unit, and the dipole analysis unit, three-dimensional data including trial, time, and a portion of the brain where the dipole is estimated for each independent component Independent component analysis / dipole analysis result storage unit that generates
A brain activity pattern learning unit for learning the three-dimensional data;
Have
The brain activity learning unit includes a 3-way categorical data analysis unit utilizing the 3-way category data analysis learns the brain activity patterns, blanking rain computer interface device. Measuring brain waves of the user in a plurality of channels using a plurality of electrodes, the measured electroencephalogram signal by the analog / digital converter further comprises a brain wave measuring unit to the digital EEG data, according to claim 1 or 2 Brain computer interface device .