WO2005022398A1

WO2005022398A1 - Method for selecting parameters from an eeg signal for monitoring narcosis

Info

Publication number: WO2005022398A1
Application number: PCT/EP2004/009593
Authority: WO
Inventors: Gudrun Stockmanns; Michael Ningler; Eberhard Kochs; Gerhard Schneider; Hans-Dieter Kochs
Original assignee: B. Braun Melsungen Ag
Priority date: 2003-08-28
Filing date: 2004-08-27
Publication date: 2005-03-10
Also published as: DE10339713A1

Abstract

The invention relates to a method for selecting parameters from a biosignal, especially EEG and EP signals, for monitoring narcosis, said method comprising the following steps: time segments (nos. 1 to 8) are extracted from the detected biosignal; numerical values respectively representing pre-determined parameters (Median, SEF95, ..) in the time segments (nos. 1 to 8) are determined; said parameter values are classified into pre-determined categories (low /middle/high); the categorised parameter values of the time segments (nos. 1 to 8) are compared in order to establish which parameters are dispensable for the differentiation of one time segment from the other; and the parameters that are dispensable for the differentiation of the time segments (nos. 1 to 8) are eliminated.

Description

Method for selecting parameters from an EEG signal for anesthesia monitoring

The present invention relates to a method for selecting parameters for anesthetic monitoring from a biosignal, in particular from EEG signals and evoked signals for anesthetic monitoring.

EEG and evoked potentials are increasingly being used as a biosignal to detect intraoperative alertness during general anesthesia. For automated anesthesia monitoring, it is necessary to extract suitable parameters from the EEG signal or from the evoked potentials (EP) that differentiate different anesthetic conditions as well as possible, e.g. Allow "awake" and "passed out". However, based on the evaluation of isolated individual parameters, this distinction is only partially successful.

The object of the present invention is to enable an improved method for anesthetic monitoring. In particular, it is the object of the present invention to provide a method with which several parameters which can be used efficiently for the detection of anesthesia states can be found.

This object is achieved by a method with the features mentioned in claim 1. In this method, preselected individual parameters, which can be relevant in a biosignal, in particular EEG and EP signals (signals from evoked potentials), generally for a distinction between different anesthetic states, for example "awake" and "unconscious" reduced a smaller number of individual parameters. For this purpose, periods of time in which the one anesthesia state (for example “awake”) is present and those in which another anesthesia state (for example “unconscious”) is present are extracted from the recorded biosignal. Then the individual parameter values of the time segments are determined for the preselected individual parameters and categorized into predetermined categories (eg low / medium / high). The categorized individual parameter values of the time segments determine whether there are dependencies between combinations of the total of the categorized individual parameter values, ie whether there are any Parameters for differentiating the time periods are unnecessary. In order to reduce the number of individual parameters, the unnecessary individual parameters are eliminated.

By comparing combinations of the total of the categorized individual parameter values, the time periods and the subsequent combination of the indispensable individual parameters, this method can also be used to use parameters that would be classified as unusable parameters in the isolated evaluation of individual parameters. Furthermore, the method is a transparent method in which, for example, a treating doctor is able to understand on the basis of which individual parameters a distinction, such as "awake" or "unconscious", was determined. This also increases the safety of anesthesia monitoring, since the automatically determined distinction can also be verified if necessary.

According to another aspect, the above-mentioned object is achieved by a method with the features mentioned in claim 2. In contrast to the method according to claim 1, in this method the parameter values of the time segments are determined for the selected parameters and categorized into predetermined categories (e.g. low / medium / high). The categorized parameter values of the time segments are used to determine whether there are dependencies between the parameters, i.e. whether individual parameters are unnecessary for differentiating the time periods. In order to reduce the number of parameters, the unnecessary parameters are eliminated.

The invention is explained below with reference to the drawing. Show it:

1 shows an exemplary decision table in which calculated and categorized parameters and the associated patient status for eight EEG time segments are shown,

2 shows the example table of FIG. 1 in reduced form,

3 shows a table with results of a practical exemplary embodiment of the method and averaged classification rates,

4 shows a table in which three objects are assigned numerical values for three condition attributes,

5 shows a diagram for the assignment of the values of the condition attributes from FIG. 4, 6 shows a table in which the assignment of the numerical values of the condition attributes of the scheme from FIG. 5 is shown, FIG. 7 shows a table in which several objects are generated from an object from FIG. 4 and which are shown in the table from FIG. 6 contains combinations of this object, FIG. 8 a table in which a weighting is given for the object from FIGS. 4 and 7, FIG. 9 a correspondingly derived table for another object from FIG. 4, FIG. 10 also Correspondingly derived table for a further object from FIGS. 4 and 11 is a table in which objects of the objects shown in FIGS. 8 to 10 are summarized.

In one embodiment, which is explained with reference to FIGS. 1 and 2, EEG signals or signals from evoked potentials (EP) are recorded before and during a general anesthesia, from which temporal EEG sections are extracted. Parameters for assessing the state of anesthesia are determined, which are selected so that they generally have characteristics relevant to the assessment. These parameters, which represent complex signal contents, are converted into numerical values by calculation, which reflect the special properties of the signal. In order to reduce the number of parameters to be considered, those parameters are eliminated which are not necessary or cannot be used to distinguish between the "awake" and "unconscious" anesthetic states.

A method is described below with which parameters can be combined, superfluous parameters can be eliminated and rules for classification can be derived. This method examines the extent to which objects (here different EEG sections) can be assigned to a class (here a decision between the states "awake" and "unconscious") using their known properties. For objects with the same properties but different decisions, the decision cannot be clearly derived from the properties. The decisions can therefore only be roughly described by the known properties. The properties are represented by condition attributes, the decisions by (at least) one decision attribute, which is given here by the states "awake" and "unconscious".

With this method, the number of condition attributes can be reduced. For this purpose, the condition attributes can be reduced in two reduction steps to the condition attributes that are necessary for the differentiation of the objects without Consideration of the decision (first reduction step) and the assignment of objects to decisions (second reduction step).

Classification rules for the assignment of the objects are formed with the remaining condition attributes.

1 shows an exemplary decision table that calculates and subsequently categorizes parameters from an EEG signal and the empirically determined associated patient status for, for example, eight EEG signal sections, which form objects Nos. 1 to 8 in FIG. 1. For example, selected as parameters

-Media (i.e. the frequency value that divides the power spectrum of the EEG section into two parts of the same power),

SEF95 (i.e. the frequency value that divides the power spectrum so that the sum of the band powers over all frequencies f <SEF95 is 95% of the total power), -Crestf. (Crest factor, i.e. the maximum amplitude / rms value) and -RMS (rms value, i.e. the root mean square over all amplitudes of the EEG section).

These parameters expressed in numerical values form condition attributes which are categorized into the values “low”, “high” and “middle”. The states “awake” and “unconscious” (“conscious” in FIG. 1) represent the decision attribute ,

Each of the condition attributes divides the set of objects Nos. 1 to 8, ie the eight EEG sections, into groups, so-called equivalence classes, which cannot be distinguished by considering this condition attribute alone. In the table of FIG. 1, for example, the condition attribute median divides the objects into three equivalence classes {1, 5, 6}, {2, 3, 4, 7}, {8} corresponding to the three median values "low", "high" and "medium", the equivalence class {1, 5, 6} being assigned the median "low", the equivalence class {2, 3, 4, 7} being the median "high", and the equivalence class {8} being the median "medium" is. Equivalence classes can be determined in the same way for the other condition attributes, for the entire set of condition attributes and for subsets of the condition attributes. If all condition attributes median, SEF95, crest factor and RMS in the table of FIG. 1 are viewed simultaneously, which can all assume the values “low”, “high” and “medium”, one finds the equivalence classes {1, 6}, {2 , 4}, {3}, {5}, {7} and {8} for the total set of condition attributes. In general, an equivalence class is thus formed by the set of objects or EEG sections, which cannot be distinguished with regard to the set of condition attributes considered in each case.

With the aid of a comparison on the basis of equivalence classes, the number of condition attributes is reduced in the exemplary embodiment in preferably two reduction steps.

First reduction step

In the first reduction step, regardless of the decision (awake or conscious), it is determined which of the condition attributes are not necessary for the differentiation of the objects. For example, it would not be necessary a condition attribute that takes on the same values for all objects, or a condition attribute that distinguishes individual objects, but this distinction is already clearly given by other condition attributes.

In general terms, the first reduction step can be carried out, for example, by forming the equivalence classes from each subset of the condition attributes and comparing them with the equivalence classes that are found while considering all condition attributes at the same time. A subset of the condition attributes, which forms the same equivalence classes as the entire set of condition attributes, is then a subset (hereinafter also reduced) to which the set of confirmation attributes is reduced in the first reduction step.

The set of condition attributes remaining after the first reduction step is called the reduced. In general, reductions are the subsets of the condition attributes that divide the objects into the same equivalence classes as the set of all condition attributes. In the example from FIG. 1, for example, the condition attributes Mediän, SEF95 and RMS form a reduct and divide the objects into the equivalence classes {1, 6}, {2, 4}, {3}, {5}, {7} and { 8}, which are also found when considering all condition attributes median, SEF95, crest factor and RMS.

In the example shown in the table in FIG. 1, it follows from the first reduction step that the condition attribute crest factor can be omitted because the Condition attribute Crest factor distinguishes object No. 5 from all other objects, but this object is already clearly identified by Mediän and SEF95. The condition attribute crest factor is therefore not necessary to differentiate the objects and can therefore be omitted.

Algorithms can be used for the first step, by means of which subsets of the condition attributes are formed and by which it is checked whether the same equivalence classes are formed.

Second reduction step

In the second reduction step, taking into account the decision, i.e. Whether the condition is classified as "awake" or "unconscious" determines which condition attributes of the reduct (or all condition attributes ^ are not required for an assignment of the objects to a decision attribute. For example, if a condition attribute only distinguishes those objects from each other that have the same value for the decision attribute, this condition attribute is not required for the assignment to a decision attribute.

The second reduction step is preferably carried out by counting the objects to all subsets of the condition attributes that cannot be clearly assigned to a value of the decision attribute (i.e. a decision) with the aid of the condition attributes. This assignment is not possible if objects of an equivalence class are assigned to different decisions and thus contradict each other. A subset of the condition attributes in which the number of objects that cannot be assigned to a decision does not increase (hereinafter referred to as a relative reduction) is sufficient to assign the objects to a decision.

In the example shown in the table in FIG. 1, the second reduction step shows that the condition attribute RMS can be omitted. In this exemplary embodiment, objects No. 1 and No. 6, which belong to the equivalence class {1, 6} and are assigned to different states awake and unconscious, contradict each other. The reduced median, SEF95, and RMS (and also the set of all condition attributes median, SEF95, crest factor and RMS) thus means that two objects are not clearly assigned to a decision awake / unconscious can be. The subset Mediän and SEF95 does not mean that other objects cannot be clearly assigned to a decision. Leaving out the condition attribute RMS therefore does not result in a higher number of objects that cannot be assigned to a decision than the reduct.

In other words, RMS is necessary to distinguish object no. 3 from object no. 2 and 4. However, since all three objects belong to the "awake" state, this distinction is not necessary for an assignment to the decision attribute. The set of condition attributes remaining after the second reduction step, that is to say the relative reduction, is thus median and in the example from FIG. 1 SEF95 (see also Fig. 2).

FIG. 2 shows the decision table from FIG. 1 in reduced form, which is reduced to the relative reduced, namely the condition attributes Mediän and SEF95, and in which objects Nos. 2, 3 and 4, and thus the associated EEG sections, are summarized are.

Resel-based classification

Rules for assigning the objects to an anesthetic state and a decision system can be derived directly from a decision table reduced to the reduced or to the relative reduced. The rules are derived according to the general scheme

"IF [parameter 1] = [categorized value of parameter 1 in the object] AND [parameter 2] = [categorized value of parameter 2 in the object] AND [parameter 3] = [categorized value of parameter 3 in the object] AND

[Parameter n] = [categorized value of parameter n in the object] THEN [decision attribute] = [decision in the object] "

the parameters used (for example parameter 1: median and parameter 2: SEF95) must be inserted in the parentheses before the equal sign and the categorized values of these parameters used in the parentheses after the equal sign.

For example, the following is obtained from the decision table in FIG. 2, reduced to the relative reduct, for object no. 8: "IF Mediän = medium AND SEF95 = high THEN condition = moved."

This rule can be simplified because the one shown in FIGS. 1 and 2

Example, the condition attribute Mediän only takes the value "medium" in object no. 8. This is the rule

"IF median = medium THEN condition = proven" also valid

In the same way, rules can be derived for objects no. 2 to 5 and 7. No valid rules can be derived from objects no. 1 and 6, since these cannot be distinguished in terms of their condition attributes, but are assigned to different decisions. In general, several valid rules can be derived from an object.

With these rules, which can be carried out by a microprocessor, an assignment of the extracted EEG sections using the parameter values to the anesthetic states is then “awake” and

"Unconscious" possible.

The method described above for selecting parameters for anesthesia monitoring from a biosignal is thus based on the principle of first extracting time segments nos. 1 to 8 from the biosignal recorded, in order to then determine numerical values, the predetermined individual parameters such as median, SEF95, crest factor , RMS, in each of these periods and assign categories (low / medium / high) to these individual parameter values. In order to determine which individual parameters are unnecessary for distinguishing one time segment from the other time segment, a comparison of combinations of the total of the categorized individual parameter values of the time segments No. 1 to 8 is carried out. Individual parameters which are determined to be indispensable for differentiating the time periods are combined.

Figure 3 shows in a table results of an investigation in which the method described above was applied to data from a study of 40 patients who underwent surgery under general anesthesia. In this practical embodiment, EEG signals are continuously derived both when awake and under anesthesia. From the EEG, segments of 8 s length are selected in equal parts from the states "awake" and "unconscious" and 52 parameters are calculated for each segment. The parameters are divided into four groups: 24 basic parameters (eg RMS RMS value), 20 spectral parameters (eg power in frequency bands), 4 statistical parameters ("Statistics"("Stat." In Fig. 3) eg skewness, excess ), 5 complexity measures ("complexity"("compl." In Fig. 3) e.g. Shannon entropy). To categorize the parameter values, the value range of each parameter is divided into intervals and each parameter value is assigned to an interval. Three different categorizations (4, 6, and 8 intervals) are examined. The distribution of the parameter values among the states is taken into account when determining the interval limits.

In order to shorten the computing times for a simultaneous examination of all 52 parameters, three calculations are carried out, separated according to parameter groups, the groups “statistics” and “complexity” being combined, as can be seen from the table in FIG. 3. In a further calculation for the combination of the groups, the condition attributes of the relative reduct of the three groups are used.

For the application of the method, the available objects, i.e. 2/3 of the segments divided into a training set and 1/3 into a test set. Using the rules found, a state is determined for the objects in the test set and the classification rate is calculated. Three calculations are carried out with different training and test quantities.

The table in Fig. 3 shows i.a. the number of condition attributes in the relative reducts, the total number of rules generated in the practical embodiment and the classification rates achieved for the test quantities, which are each averaged over the three calculations. The number of incoming condition attributes is given in brackets for the "Combination" parameter group.

The number of rules calculated is relatively large in all cases. A rule often describes only a single object of the training set (“support” of the rule = 1). The last two columns of the table in FIG. 3 show the number of rules and the classification rate after the rules have been omitted with support 1.

The investigation shows that the method presented significantly reduces the number of parameters and that the rules generated achieve satisfactory classification rates. Restricting the rules to those with support greater than 1 has a minimal impact on the classification rates. As the number of intervals for categorization increases, the number of parameters and rules required decreases, while the classification rate hardly changes. The large number of rules is mainly due to the fact that all rules must be valid with regard to the amount of training, so that small disturbances, such as noise, prevent the generation of a few simple rules. An improvement can be achieved by using a model in which a slight misclassification of the objects is permitted.

Categorical condition attributes are required for the application of the method, i.e. those that can only take a small number of different attribute values. Real-value condition attributes, such as voltage values, must be appropriately categorized.

Fuzzy categorization of the condition attributes

A variant of the categorization (fuzzy categorization) is explained below using an example with reference to FIGS. 4 to 9, in which a deliberately fuzzy categorization of the condition attributes is carried out before the first reduction step.

4 shows a table in which three objects Nos. 1 to 3 (i.e. here EEG signal sections or EP signal sections) are assigned numerical values for three condition attributes B1, B2 and B3. A decision attribute D, which here takes on the values 0 and 1, is set to 1 for objects No.1 and No.2 and to 0 for object No.3.

5 shows, for example on the basis of object 1, a scheme for assigning the values of the condition attributes B1, B2 and B3 to the elements of the basic sets 1 to 6, which are determined, for example, by a customary, sharp categorization (as described above), and Membership functions μ for the elements of the basic set. Each real-valued parameter value is assigned to two elements of the basic set (two categories) using the membership functions μ. These elements are determined as follows: A membership function is determined for each element of the basic set (category) of each parameter. This assigns a numerical value between 0 and 1 to the real parameter value, which represents a measure of how much (or to what extent) a (real) parameter value is to be assigned to the element of the basic set under consideration. The membership functions assume the function value (with the standardization selected here) μ = 1 in the middle between the interval limits. As the real values increase, μ drops linearly down to the middle between the limits of the next larger element μ = 0 is reached. For further increasing values (or for larger values) μ = 0 remains. The same applies to decreasing values (triangular functions). Only the membership functions of adjacent elements overlap, as shown in FIG. 5, and the sum of their function values always results in them, ie 1 for each parameter value.

The assignment of the condition attributes B1, B2 and B3, represented by number lines in FIG. 5, takes place via the membership functions. As shown in Fig. 5, each condition attribute value (at most) can be assigned to two elements of the basic set, i.e. μ ≠ 0 is only for two elements of the basic set.

For object 1 you get e.g. the assignment of the shown in a table in Fig. 6

Numerical values of the condition attributes B1, B2 and B3 ("attributes") for the elements of the

Basic set with μ ≠ 0. The smaller element from basic set 1 to 6 is called element x, the larger element from basic set 1 to 6 with element y. Are further

Degrees of membership μ _x and μ _{y are given} for the two elements.

The value 1.7 for Bl with element of membership μ _x = 0.3 becomes element 1 of the basic set

(Elements x), and with the degree of membership μ _y = 0.7 element 2 of the basic set

(Elements y) assigned (see Fig. 4 and dashed line I in Fig. 5).

Corresponding to the value 26 for B2, the degree of membership μ _x = 0.4 with respect to the element

2 of the basic set (elements x), and the degree of membership μ _y = 0.6 with respect to the element

3 assigned to the basic set (element y) (see FIG. 4 and dashed line II in FIG. 5). In the same way, the value 320 for B3 is assigned the degree of membership μ _x = 0.8 with respect to element 3 of the basic set (element x) and the degree of membership μ _y = 0.2 with respect to element 4 (element y) of basic set 4 (see FIG. 4 and dashed line HI in Fig. 5).

Tables can be created for the other objects no. 2 and 3 in the same way, in which each condition attribute value (one or) is assigned two elements of the basic sets.

FIG. 7 shows a table in which an object, here, for example, No. 1; several objects, here, for example, Nos. 1.1 to 1.8, are generated, which contain each of the combinations of the categories (elements of the basic set) of the object No. 1 contained in the table from FIG. 6. For this the table in FIG. 7 has columns for the assignment of the elements of the basic sets to the condition attributes Bl, B2 and B3 and columns μsu μm and μß ₃ for values of μ, which are respectively assigned to the condition attributes Bl, B2 and B3 in the basic sets ,

Objects 1.1 to 1.4 relate to the combinations from FIG. 6, in which B1 is assigned the value 1 of the basic set (field at the top left in FIG. 6 and left column, upper half in FIG. 7). In this case, μ _x = 0.3, as can be seen from the table in FIG. 6. μsi is therefore set to 0.3 for objects no. 1.1 to 1.4. Objects 1.5 to 1.8 relate to the combinations from FIG. 6, in which B1 is assigned to element 2 of the basic set (FIG. 6). In these, μ _y = 0.7, as can be seen from the table in FIG. 6. μm is therefore set to μsi = 0.7 for objects no. 1.5 to 1.8.

As can also be seen from the table in FIG. 6 (second row), B2 can be assigned to elements 2 or 3 of the basic set, the values μ _x = 0.4 and μ _y = 0.6 being assigned in FIG. 6. Accordingly, μs _{2 is set} to 0.4 for the assignment to element 2 of the basic set and to 0.6 for the assignment to element 3 of the basic set (fourth column in FIG. 7). The table in FIG. 6 also shows (third row) that B3 can be assigned to element 3 of the basic set or element 4 of the basic set, the values μ _x = 0.8 or μ _y = 0.2 in FIG. 6 assigned. Accordingly, μe3 is set to 0.8 for element 3 and to 0.2 for element 4 (sixth column in FIG. 7).

A categorization is thereby obtained, which delivers eight objects Nos. 1.1 to 1.8 from the object No. 1 of FIG.

For these objects Nos. 1.1 to 1.8, the sum of the values of the membership degrees μsi, μβ ₂ and μs3 is formed (Sum μ in the right column of FIG. 7).

Corresponding tables can be created in the same way for objects No. 2 and 3 from FIG. 4, whereby a categorization is obtained in each case, which generates eight objects 2.1 to 2.8 or 3.1 to 3.8 from objects 2 and 3, respectively.

FIGS. 8 to 10 each show tables in which the objects obtained in this way are shown with their parameter values categorized via the respective assignment to elements of the basic set, and a weighting of the objects is indicated. 8 shows the table for object no. 1, or the objects no. 1.1 to 1.8 generated therefrom. For the weighting carried out here, the sum values Sum μ from FIG. 7 are divided by the total sum of the sum values for all objects (here, for example, 12, see last row of the table in FIG. 7). Furthermore, in the table in FIG. 8, for each of these objects, the assignment to elements of the basic set of parameter values for the parameters B1, B2 and B3, as well as that assigned to object No. 1 in FIG. 4, is entered in the table of FIG. 7 Value of decision attribute D specified.

FIG. 9 shows a correspondingly derived table for object no. 2 from FIG. 4, the objects derived from object no. 2 being designated 2.1 to 2.8 in FIG. 9.

A corresponding table for object no. 3 from FIG. 4 is created in FIG. 10.

Fig. 11 shows a table in which the newly created tables for objects Nos. 1 to 3 (Fig. 8 to 10) are combined.

With regard to objects No. 1 and No. 2, which have the same values of the decision attribute D = 1, all objects of the table from FIG. 9 that are different in at least one attribute value B1 to B3 are inserted into the table from FIG. 8. If indistinguishable objects occur, only the weight of the table in FIG. 8 is increased.

In the example described here, objects 1.5 and 2.4 in FIGS. 8 and 9 cannot be distinguished because they both have the values 2, 2, 3, 1. Therefore object 2.4 is not newly inserted into the table of FIG. 11, but the weight of 5 '(originally 1.5) is increased by the weight of 2.4. All other objects in the table of FIG. 9 are added to the table of FIG. 8.

With regard to object no. 3, which has a different value of the decision attribute D = 0, the rows of the table from FIG. 10 are inserted in the last eight rows in FIG. Since object 3 has a different value of the decision attribute (D = 0) than objects no. 1 and 2 from FIG. 4 (D = 1), object no. 3 supplies eight further new objects. Specifically, object 3.2 has the same values for attributes B1, B2 and B3 as object 1.3. But since the value of the decision attribute D for object no. 3 differs from the values of the Decision attribute D of objects No.1 and No.2, objects 3.2 and 1.3 are not summarized.

In the final table, the sum of the weights gives the number of original objects.

The table obtained in this way can be processed further with a program which processes an input table which contains an attribute with the number of objects described by a line and takes this number into account in the calculation. (In the table described above, the weight also describes the number of objects, but in the form of a non-integer value).

Instead of the triangular functions described in connection with FIG. 5 and shown there, other membership functions μ can be used. For example, functions in the form of a ttape can be used in which a central section is constant (for example = 1) and only overlap with falling areas on the side. Similarly, unlike in the example shown in FIG. 5, a non-equidistant division into elements of the basic set can take place. In addition, asymmetrical membership functions can also be used. Such variants can be used to adapt the categorization to the parameters or parameter values to be categorized in each case.

The basic set and the membership functions are generally defined for each parameter. Therefore, unlike the example described above, the number of elements contained in the basic set can be different for different parameters. For example, the basic set can only contain four elements for B1 and eight elements for B3.

In summary, with this categorization method, a categorization of the parameter values will be obtained, which takes into account the overlap of the basic quantities, how close a parameter value is to the interval limit of a basic quantity, because the values of the membership functions μ _x and μ _y obtained when assigning a parameter value into one Weighting factor.

This weighting factor can then be taken into account when applying rules that have been derived from the respective object (or, in the exemplary embodiment, time period). modifications

In the exemplary embodiment described above, the second reduction step is used, in which condition attributes are determined which cannot be used to distinguish between the states "awake" and "unconscious" in order to remove these condition attributes from the decision table. This second reduction step can also be omitted, with the result that u. U. redundant rules can be derived.

Likewise, the rule-based classification used in the above-described embodiments may be omitted and another type of classification may be used instead.

The first reduction step can also be omitted.

The method described is also suitable for treating more than two states. The method can therefore also be used to distinguish between several stages of anesthesia and is not limited to the distinction between the states "awake" and "unconscious".

In the exemplary embodiment shown in FIGS. 1 and 2, no valid rules can be generated from objects 1 and 6, since these cannot be distinguished with regard to their condition attributes, but are assigned to different decisions. No rules are generated for these.

In the exemplary embodiment, EEG signals are used for anesthetic monitoring. The method is not limited to this, but can also be carried out accordingly with other biosignals, such as blood pressure values.

Claims

claims

1.Procedure for selecting parameters for anesthesia monitoring from a biosignal, in particular from EEG and EP signals, comprising the steps: extracting time segments (No. 1 to 8) from the recorded biosignal, determining numerical values, the pre-determined individual parameters ( Accept Mediän, SEF95, ..) in the time periods (No. 1 to 8), classify these individual parameter values into predetermined categories (low / medium / high), compare combinations of the total of the categorized individual parameter values of the time periods (No. 1 to 8 ) to determine which individual parameters are unnecessary for differentiating a time period from the other time periods, and a combination of the individual parameters that are not necessary for differentiating the time periods (No. 1 to 8).

2. A method for selecting parameters for anesthesia monitoring from a biosignal, in particular from EEG and EP signals, comprising the steps: extracting time segments (No. 1 to 8) from the acquired biosignal, determining numerical values, the predetermined parameters ( Mediän, SEF95, ..) in each of the time periods (No. 1 to 8), classify these parameter values into predetermined categories (low / medium / high), compare the categorized parameter values of the time periods (No. 1 to 8) to determine which parameters are unnecessary for differentiating a time period from the other time periods and eliminating the parameters which are unnecessary for differentiating the time periods (No. 1 to 8).

3. The method of claim 1 or 2, wherein from each subset of the condition attributes sets of time periods (equivalence classes) are formed, which are indistinguishable with respect to this subset, in order to be compared with the set of time periods (equivalence classes), while considering all Condition attributes are found, and if the sets of time periods (equivalence classes) match, it is found that a parameter not included in this subset is unnecessary for distinguishing a signal time period from the other time periods.

4. The method of claim 1, 2 or 3, wherein for each extracted time period (No. 1 to 8) a state of anesthesia (awake / unconscious) is determined during the respectively associated time period, the categorized parameter values of the time periods (No. 1-8) which are assigned to the same anesthetic state are compared in order to determine which parameters are required for the assignment of the anesthetic state, and parameters which are not required for the assignment of the anesthetic state are eliminated (second reduction step).

5. The method of claim 4, wherein for all subsets of the condition attributes are counted the time periods that can not be assigned a value of the decision attribute (anesthetic state) because they are assigned to an equivalence class and different values of the decision attribute (anesthetic state), and if the Number of time periods that cannot be assigned to a value of the decision attribute (anesthesia state), it is determined that a parameter not included in this subset is not necessary for the assignment of the anesthetic state.

6. The method according to any one of the preceding claims, wherein a decision system (based on rules) is derived from the categorized parameter values of one or more time periods, which enables an assignment of parameter values of a respective corresponding time period to an anesthetic state (awake / unconscious).

7. The method according to any one of the preceding claims, wherein a fuzzy categorization is used for the classification of parameter values.

8. The method according to any one of the preceding claims, wherein the time periods are assigned weights.

9. The method according to claim 8, wherein when the rules are formed, a weighting is assigned to the rules formed, which weighting is determined from the weightings of the time segments.