AT505339B1 - METHOD FOR EVALUATING AND / OR PREPARING A MULTIVARIATE SIGNAL - Google Patents

Description

Austrian Patent Office AT505 339B1 2012-10-15

Description: The invention relates to a method according to the preamble of claim 1.

In particular, the invention relates to a method for processing multivariate brain signals (electroencephalograms (EEG), electrocorticograms (ECoG) or derived from deep electrodes signals) derived simultaneously from multiple electrodes, amplified and sampled for digital processing and converted analog-to-digital ,

The aim of the invention is primarily the definition of a measure that allows a more accurate evaluation of a multivariate brain signal with regard to the occurrence of certain events, about which the signal contains information. Such events may e.g. be caused by thinking processes or by movements or movement attempts of the body, so that switching processes can be triggered in an exact evaluation of the associated electrical signals.

The objectives of the invention are achieved by a method of the type mentioned above with the features mentioned in the characterizing part of claim 1.

With the procedure according to the invention, the originating from the event or caused by the event signals can be processed for further processing or recovery in a manner that is the cause of the event justice. With the method according to the invention, evaluation measures for the signals are created, whereby the significance of the signals or the reliability of the assessment and classification of these signals is increased. These advantages are achieved by the special combination of the sampled signals and the fact that in the obtained evaluation measures for the multivariate signal corresponding design options are available by selecting coefficients or parameters which can be varied by using appropriate determination methods or algorithms. Depending on the desired validity of the evaluation measure one of the two listed in claim 1 alternative procedures can be selected.

It should be noted that the inventive method for the evaluation of multivariate signals of any origin can be used; in view of the selected formation of the evaluation measure, the method according to the invention is particularly suitable for the evaluation of electrical signals derived from a human being, e.g. Brain or cardiac or muscle signals, in particular for use in brain-computer interfaces (BCI). The obtained evaluation measures can be used in particular to determine a switching operation, e.g. for switching a device to trigger.

In the following the method according to the invention will be explained by way of example with reference to the drawing. The drawing shows a block diagram for the calculation of a [n].

The sampled multivariate signal x [n \ = [x; [n],..., XL [n]] is composed of L recorded signals x; [n], / = 1,. L together. First, these signals are classified into A ^ signal groups x (k) [n] = [xW ___ (χ [n]] T k = Ι,.,., Κ (1), in particular of different size L (/ c), with x (fe) [n] = xP (feJ) [n] according to a defined assignment P (k, l).

For each signal χ ^ \ η] in (1) of each signal group £, a prediction function fiKn.) Is formed, ie, a total of ££ = iL (/ c) functions, each representing past values in the group x (fe ) [n - MJ, ..., x (fe) [n - l] and additionally existing past prediction values for the group x (k) [n-M2], ..., x (k) [nl], with x (fe) [n] = [x [k) [n], ..., x ^) [n]] T, to a value 1/5 Austrian Patent Office AT 505 339 B1 2012-10-15 x ^ k \ n] = f ^ k, n \ x ^ [n - MJ, ..., x ^ [n - l]; x ^ [n - M2], ..., x ^ [n - Z]) depict. Such prediction functions are described in the literature e.g. in the context of linear prediction or autoregressive models (see, for example, S. M. Kay, Modern Spectral Estimation, Englewood Cliffs (NJ): Prentice Hall, 1988). Thus, x; (fc) [n] represents an estimate value of a brain signal value calculated from the previous received signal values and the previous M2 prediction values.

In order to judge the accuracy of this prediction, a quality measure p (fc'n) is defined for each prediction function, defined e.g. as the expected value of the quadratic distance p (M0 _ _ x; (fc) [n]) 2j or as the expected value of the distance amount ρ [Κ7ί) = E j, etc.

The prediction functions are selected from a defined function space F such that the value for the quality measure becomes optimal, that is fik'n) = arg min p \ k'n) (2)

f * F

It should be noted that the examples given for plk, n) measure the prediction error, i.e., small values correspond to high accuracy. Of course, quality functions or measures can also be defined, which conversely measure the match between the prediction value and the signal, so that large values are sought.

In general, the prediction function will be time-dependent (time index n), because at least the detection of instationarities in the signal, caused by the event to be detected, is made possible in the first place.

For practical implementations of (2), adaptation algorithms may be used, such as e.g. in "Adaptive Filter Theory" (Haykin, Adaptive Filter Theory, Englewood Cliffs (NJ): Prentice Hall, 4th ed., 2002). These determine (adapt) f (k, n) jn ^ 2) approximately and recursively from p ^ k'n ~ Q \ ..., p [k'n ~ 1 ^. The optimization is also carried out in adaptive algorithms with regard to a function from a defined function space F.

As an alternative to recursive adaptation methods, moving-window techniques can also be used to decompose the signal into sections ("Windows") within which ffk, n ^ are time-invariant and are chosen such that the corresponding quality measure is optimized on average.

Finally, either the evaluation measure KL (k) a [n \ = Yb (k) log Σαϊ *] ΡΪ * 'η) k = 1 1 = 1 or the evaluation measure KL (from the quality measures p [k, n] k) -Wi- (3) Z ^ iog ΣΛ, (Μ) k = l 1 = 1, which is the basis for the evaluation of the multivariate brain signal.

For this purpose, the parameters a [fc) and / or the parameters a \ k \ Mfc), c; (fc) and 2/5

Claims (4)

Austrian Patent Office AT505 339 B1 2012-10-15 are selected / determined, which significantly influence the properties of a [n]. These may e.g. by statistical learning techniques (see, eg, VN Vapnik, Statistical Learning Theory, New York: Wiley, 1998, CM Bishop, Neural Networks for Pattern Recognition, Oxford: Clarendon Press, 1995), or other suitable methods, such that a [n] a reliable detection of the requested events allows. In the drawing, the determination method for a [n] is schematically illustrated in a block diagram. At the top left is the input of the system, for the multivariate signal x [n], from this signal the K signal groups x (fc) [n] are formed. Each of these signal groups is then each processed in a separate block. A signal group x (fc) [n] is supplied together with previous prediction values x (fc) [n] to the predicators // fc, n + 1), which generate prediction values xw [n + 1]. In block p, the quality criterion ρ ^ 'η) is calculated from the prediction values x (fe) [n] of the previous time and the currently recorded signal group value x (fc) [n]. First, the predictor is adapted with this measure p (fc'n). Secondly, it is used together with the quality measures of all other signal groups for the calculation of the evaluation measures a [n]. A simple example will now show a possibility for the determination of the parameters a [k \ b (k \ c; (fe) and in a [n], which is derived from a multivariate brain signal x [n] = [ x, [η], ..., x6 [n]] T, that is, there are, for example, six signals which naturally form two groups xjn], ..., x3 due to the anatomical arrangement of their derivative positions. Accordingly, four signal groups can be formed, consisting of the two anatomical groups * (1) [n] = [xjn], ..., x3 [n ]] r and x (2) [n] = [x4 [n], ..., x6 [n]] T, and their combinations λγ ^ 3 ^ [π] = [ΛΓ (1) Γ [η], Λ: (2) 7 '[η]]' Γ and χ (4) [n] = [χ®τ [η], χ (1 ') 7' [η]] 7 '. [0020] For these signal groups Let us now form the evaluation measure a [n] which permits the most reliable detection of the required events, and the necessary parameters a [k \ b ^ k \ c; (fe) and d (fe) in (3) could be used in this Example through a statistical learning process It is advantageous, but not absolutely necessary, to specify initial values for the parameters in order to ensure the fastest possible convergence to an (generally local) optimum. In the given example good results can be achieved with the following choice: For a \ k) choose positive initial values for k = 1, ..., 4 and l = 1,2,3. For k = 3.4 and l = 4.5.6 choose zero and c, (fc) = a [fc). The parameters M4), and should be initialized positive and M2), M3), and d (3) negative. The statistical learning method optimizes the parameters in a [n] such that in many cases a reliable detection of the requested events is made possible. 1. A method for evaluating and / or processing a multivariate, preferably electrical, signal, in particular a brain electrical signal containing information about at least one event to be assessed, preferably an EEG and / or ECoG and / or depth electrode signal, in particular for use in brain-computer interfaces, characterized in that - the multivariate signal (x (t) = [xx (t), ..., xL (t)] T is sampled and the or a selected number of in the sampled multivariate signal (x [n] = [xjn], ..., xL [n]] T contained signals (x; [nj) to a predetermined number (K) of signal groups (x (fe) [n] = [x® [n], ..., x ^ k ^ [n]] T, k = 1, ..., K), - that for each of the signal groups a number of prediction values (x; (fc ) [n]) is determined by the already sampled signal values (xj (fc) [nm], m = 1,.,..., Μχ) assigned to the respective signal group and optionally a number of existing or fü the respective signal groups of predicted values (xf ^ ln - m], m = 1, ..., M2) already determined by at least one prediction function (// fc, n)) which predicates a signal value (xj-fc) [n] immediately following the already sampled signal value (xf ^ Di-Zj), that for each signal (x; (fc) [n]) of the individual Signaling group used prediction function (s) (// fe, n)) from a given function space (F), eg linear function space, such that the selected prediction function obtains a measure of goodness (P; (fe, n)) for the prediction value (x; (fc) [n]), e.g. Expected value of the quadratic distance to the actual signal value (x; (fc) [n]), optimized, that for each signal group the logarithm of a weighted linear combination (XazWP? 'N)) of the obtained quality measures is calculated and groups for the individual signal groups obtained values are combined to form a further linear combination p \ k, n)) kl, and that either the coefficients or parameters {a [k \ Mfc)] of the further linear combination by application of determination methods and / or algorithms, eg statistical learning, neural networks or the like, such that the reliability of the classification of the time-variant characteristics of the multivariate signal and / or the separability and / or extraction possibility of the information contained in the signal are optimized, and these with the formed coefficients or Parameters used linear combination as a measure of evaluation (a [n]) is used for the multivariate signal, or - for each signal group, the logarithm of a different or with newly selected coefficients or parameters weighted linear combination pf'n)) derter- Lived quality measures calculated and these values obtained for the individual signal groups are combined to form an additional linear combination p.sub.k, n)) which is different from the linear combination obtained, the quotient k1 of the obtained and the additional linear combination is formed, all coefficients or parameters (a [fc b ^ k \ c ^ k \ of the obtained and the additional linear combinations together by application of detection methods and / or algorithms, e.g. statistical learning, neural networks or the like, in such a way that the reliability of the classification of the time-variable characteristics of the multivariate signal and / or the separability and / or extraction possibility of the information contained in the signal are optimized, and with the formed coefficients or Parameters received quotient is used as evaluation measure (a [n]) for the multivariate signal. 2. Method according to claim 1, characterized in that initial values are specified for the determination methods of the coefficients or parameters. 3. The method according to claim 1 or 2, characterized in that the determined evaluation measure (a [n]) is used to make a statement as to whether the multivariate signal is considered to be sufficiently substantiated for a switching operation of a device. A computer program product having program code means stored on a computer readable medium for performing the method of any one of claims 1 to 3 when the program product is executed on a computer. 1 sheet of drawings 4/5