JPH08206085A - Autonomic activity-classifying device - Google Patents

Abstract

PURPOSE: To make high speed classification of the autonomic activity conditions by using a neural network model of statistic structure incorporation type. CONSTITUTION: An autonomic index data frequency distribution preparing part 16 prepares the data concerning the frequency distribution of the autonomic indices using the autonomic index data, and a frequency distribution form parameter presuming part 17 presumes the parameter which characterizes the shape of the frequency distribution, and an initial value data preparing part 18 prepares the initial value data pertaining to the posterior identification probability of the autonomic activity. A neural network model part 19 of statistic structure incorporation type represents the probability distribution of the population of the autonomic index data using a finite number of components having a specific probability distribution, presumes by study the parameters of those components, and emits the candidate of the condition of autonomic activity corresponding to the autonomic index data. An autonomic activity classifying part 20 classifies the autonomic activities of the subject into several stages.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an autonomic nerve activity classification device, and more particularly to an autonomic nerve activity classification device for automatically and rapidly classifying autonomic nerve activity using autonomic nerve index value data.

[0002]

2. Description of the Related Art Conventionally, as an autonomic nerve activity analyzing apparatus of this kind, for example, the 8th Human Interface
Symposium Proceedings, Oct. 21-23, 1992,
There is known a device as described in a paper published on pages 471 to 474 by Mr. Suzuki and Mr. Takizawa entitled "Evaluation of blood pressure fluctuation characteristics by mental work load by spectrum analysis". This has a basic configuration as shown in FIG. 7, and the finapress as the sensor for measuring autonomic nerve index 51 is for blood pressure measurement, the electrode as the sensor for autonomic nerve index measurement 52 is for electrocardiography, and the sensor for autonomic nerve index measurement. The thermistor as 53 is used for respiration measurement. The sensors 52 and 53 for measuring the autonomic nerve index are an amplifier (HiCut 30 Hz, time constant 0.1 msec) for the living body and an amplifier (HiCut 5H) as the autonomic nerve index measuring unit 41, respectively.
z, time constant 5.2 msec), and is recorded in the data recorder as the autonomic nerve index signal recording / reproducing unit 42. The autonomic nerve index analysis unit 43 includes a signal processor and a workstation. The signal processor performs A / D conversion using the autonomic nerve index signal 421 recorded in the autonomic nerve index signal recording / reproducing unit 42. The workstation performs data compression processing or spectrum analysis on the blood pressure data after A / D conversion, for example.

All the results obtained by the autonomic nerve index analysis unit 43 are graphed as a function of time or frequency. Therefore, the experimenter only reads the time when the autonomic nerve activity changes and the shift of the main frequency band due to the change of the autonomic nerve activity, by observing the graph.

As an attempt to automate the classification of autonomic nerve activity, a method of applying a probability distribution model to autonomic nerve index value data can be considered. This method finds the parameters of the probability distribution model that minimizes the log-likelihood calculated from the assumed probability distribution model.If we try to apply the conventional optimization method to this minimization problem, the logarithmic A partial derivative (first and second order) with respect to the likelihood parameter has to be calculated. For example, a probability distribution model is assumed to be a mixed normal distribution model having components of the number of active states to be classified. In this case, the relationship between the number of activity states K to be classified, the number P of unknown parameters to be estimated, and the number of partial differential calculation TC is P = 3 (K-1) +2, TC = P + P (P + 1) / 2, The number of partial differential calculations increases as the number of states increases. Furthermore, when the number of states is large, it is unlikely that the conventional optimization method will converge easily, and the calculation time required for parameter estimation of the probability distribution model becomes enormous.

[0005]

In the above-mentioned conventional autonomic nerve activity analysis device, only the graph relating to the autonomic nerve activity is output, and the experimenter only reads the change in the autonomic nerve activity by the inspection. There was a problem that it did not exist.

Further, when a probability distribution model is applied to the autonomic nerve index value data, if the number of states of the autonomic nerve activity to be classified increases, the calculation time required for parameter estimation of the probability distribution model becomes enormous. was there.

An object of the present invention is to make a calculation required for parameter estimation of a probability distribution model based on autonomic nerve index value data even if the number of states of activities to be classified increases by using a statistical structure-embedded neural network model. It is an object of the present invention to provide an autonomic nerve activity classification device that requires less time and enables high-speed classification.

[0008]

The autonomic nerve activity classification device of the present invention is an autonomic nerve activity classification device using autonomic nerve index value data, wherein the autonomic nerve index value data is used to determine the frequency of the autonomic nerve index value. Autonomic nerve index value data using the autonomic nerve index value data frequency distribution creating means for generating data on the distribution and the autonomic nerve index value data frequency distribution data generated by the autonomic nerve index value data frequency distribution creating means A frequency distribution shape parameter estimating means for estimating parameters characterizing the shape of the frequency distribution of
Using the parameters that characterize the shape of the frequency distribution of the autonomic nerve index value data estimated by this frequency distribution shape parameter estimation means, initial value data generation means for generating initial value data relating to the posterior identification probability of autonomic nerve activity,
The learning mode is provided, and the input layer, the intermediate layer, and the output layer are composed of three types, and the input layer includes data on the frequency distribution of the autonomic nerve index value generated by the autonomic nerve index value data frequency distribution creating means and the initial value. The initial value data concerning the posterior discrimination probability of the autonomic nerve activity generated by the data generating means is given, and the probability of the population of the autonomic nerve index value data is obtained by using a finite number of components having a certain probability distribution in the output layer. A statistical structure-embedded neural network model means for expressing a distribution, estimating the parameters of each component learning-wise, and outputting data representing candidates for the state of autonomic nerve activity corresponding to the autonomic nerve index value data, and statistics. Parameters of the probability distribution model output from the neural network model means with built-in statistical structure and Using data representative of a candidate of the state of the autonomic nervous activity corresponding to the nerve index data, and an autonomic nerve activity classification means for classifying the several stages the state of the autonomic nervous activity of the subject. Further, based on the classification result by the autonomic nerve activity classification means, it is judged whether the classification result is a desirable result for the experimenter, and if necessary, the section boundary value of the amplitude value data of the autonomic nerve index is changed by the autonomic nerve. An autonomic nerve index measurement / analysis procedure control means for instructing the index value data frequency distribution creating means to re-perform classification of autonomic nerve activity is provided.

Further, the autonomic nerve activity classification apparatus of the present invention is
In an autonomic-nerve activity classification device using autonomic-nerve index value data, using the autonomic-nerve index value data, an autonomic-nerve index value data frequency distribution creating unit that generates data relating to the frequency distribution of the autonomic-nerve index value; An initial value data generating means for generating initial value data relating to the posterior identification probability of autonomic nerve activity using random numbers in the range of 1 and a learning mode are provided, and the input layer, the intermediate layer and the output layer are provided. , Data on the frequency distribution of the autonomic nerve index values generated by the autonomic nerve index value data frequency distribution creating means and initial value data on the posterior discrimination probability of the autonomic nerve activity generated by the initial value data generating means in the input layer , And the probability distribution of the population of the autonomic nerve index value data is output to the output layer using a finite number of components having a certain probability distribution. A statistical structure-embedded neural network model means, which outputs the data representing the candidates of the state of the autonomic nerve activity corresponding to the autonomic nerve index value data, while estimating the parameters of each component learning-wise, and the statistical structure embedding. Type neural network model means for outputting the parameters of the probability distribution model and the autonomic nerve activity state candidates corresponding to the autonomic nerve index value data to determine the autonomic nerve activity state of the subject at several stages. Based on the autonomic nerve activity classification means to be classified into, and the classification result by this autonomic nerve activity classification means, it is judged whether the classification result is a desirable result for the experimenter, and if necessary, the section of the amplitude value data of the autonomic nerve index. Instructing the autonomic nerve index value data frequency distribution creating means to change the boundary value And a autonomic nervous index measurement and analysis procedure control means for re-performing the classification of the dynamic.

[0010]

In the autonomic nerve activity classification device of the present invention, the autonomic nerve index value data frequency distribution creating means uses the autonomic nerve index value data to generate data relating to the frequency distribution of the autonomic nerve index values, and the frequency distribution shape parameter The estimation means estimates the parameters characterizing the shape of the frequency distribution of the autonomic nerve index value data by using the data on the frequency distribution of the autonomic nerve index value data generated by the autonomic nerve index value data frequency distribution creating means, and the initial value data Utilizing the parameter that characterizes the shape of the frequency distribution of the autonomic nerve index value data generated by the frequency distribution shape parameter estimation means by the generation means,
Initial value data relating to the posterior discrimination probability of autonomic nerve activity is generated, the statistical structure-embedded neural network model means is provided with a learning mode, and is composed of three types of an input layer, an intermediate layer, and an output layer. Given the data about the frequency distribution of the autonomic nerve index values generated by the index value data frequency distribution creating means and the initial value data about the posterior identification probability of the autonomic nerve activity generated by the initial value data generating means, and in the output layer A finite number of components with a specific probability distribution are used to express the probability distribution of the population of the autonomic nerve index value data, the parameters of each component are learned, and the autonomic nerves corresponding to the autonomic nerve index value data are estimated. Data representing the candidates of activity states is output, and the autonomic nerve activity classifying means outputs a statistical structure-embedded neural network. The state of the autonomic nerve activity of the subject is classified into several stages by using the data representing the candidates of the state of the autonomic nerve activity corresponding to the parameters of the probability distribution model and the autonomic nerve index value data output from the network model means. To do.

Further, in the autonomic-nerve activity classification device of the present invention, the autonomic-nerve index value data frequency distribution creating means uses the autonomic-nerve index value data to generate data relating to the frequency distribution of the autonomic-nerve index values, and to generate an initial value. The data generation means uses random numbers in the range of 0 to 1 to generate initial value data relating to the posterior discrimination probability of autonomic nerve activity, and the statistical structure-embedded neural network model means has a learning mode and an input layer. , An intermediate layer and an output layer, and in the input layer, data relating to the frequency distribution of the autonomic nerve index value generated by the autonomic nerve index value data frequency distribution creating means and the autonomic nerve generated by the initial value data generating means. Given the initial value data for the posterior identification probability of an activity, and autonomously using a finite number of components with a certain probability distribution in the output layer It expresses the probability distribution of the population of the index data, estimates the parameters of each component in a learning manner, and outputs data representing the candidates of the state of autonomic nerve activity corresponding to the autonomic nerve index value data. The activity classification means outputs the candidates of the autonomic nerve activity state corresponding to the parameters and the autonomic nerve index value data of the probability distribution model output from the statistical structure-embedded neural network model means, and the autonomic nerve of the subject is used. The state of activity is classified into several stages, the autonomic nerve index measurement / analysis procedure control means judges whether the classification result is a desirable result for the experimenter based on the classification result by the autonomic nerve activity classification means, and it is necessary. If so, it instructs the autonomic nerve index value data frequency distribution creating means to change the section boundary value of the amplitude value data of the autonomic nerve index. To re-implement the classification of law neural activity.

[0012]

The present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an autonomic nerve activity classification apparatus according to an embodiment of the present invention. The autonomic nerve activity classification device of the present embodiment includes an autonomic nerve index measurement sensor 10, an autonomic nerve index measurement unit 11, an autonomic nerve index signal recording control unit 12, and an autonomic nerve index signal recording / reproducing unit 13.
An autonomic nerve index value data conversion unit 14, a conversion data recording / reproduction unit 15, an autonomic nerve index value data frequency distribution creation unit 16, a frequency distribution shape parameter estimation unit 17, an initial value data generation unit 18, Neural network model unit 19 with built-in statistical structure and autonomic nerve activity classification unit 2
0 and an autonomic nerve index measurement / analysis procedure control unit 21.

The autonomic nerve index measuring unit 11 amplifies the autonomic nerve index input via the autonomic nerve index measuring sensor 10 and performs preprocessing such as filtering for noise removal, if necessary, to perform autonomic nerve index measurement. The signal 111 related to the index is output.

The autonomic nerve index signal recording control unit 12 outputs the signal 2 relating to the autonomic nerve index measurement start time and end time output from the autonomic nerve index measurement / analysis procedure control unit 21.
11 and the signal 111 related to the autonomic nerve index output from the autonomic nerve index measurement unit 11 until the time when the recording of the autonomic nerve index starts and the time when the autonomic nerve index recording ends. A signal 121 related to the autonomic nerve index is output to 13.

The autonomic nerve index signal recording / reproducing unit 13 inputs the signal 121 relating to the autonomic nerve index output from the autonomic nerve index signal recording control unit 12, and records, holds and reproduces it.

The autonomic nerve index value data conversion unit 14 uses the signal 131 regarding the autonomic nerve index recorded in the autonomic nerve index signal recording / reproducing unit 13 to convert into a signal 141 regarding the amplitude value of the autonomic nerve index.

The converted data recording / reproducing unit 15 inputs the signal 141 relating to the amplitude value of the autonomic nerve index output from the autonomic nerve index value data converting unit 14, and records, holds and reproduces it.

Autonomic nerve index value data frequency distribution creating unit 16
Uses the signal 151 regarding the amplitude value of the autonomic nerve index recorded in the conversion data recording / reproducing unit 15 to generate data 161 regarding the frequency distribution of the autonomic nerve index value data.

The frequency distribution shape parameter estimation unit 17 uses the signal 161 relating to the frequency distribution of the autonomic nerve index value data output from the autonomic nerve index value data frequency distribution generation unit 16 to determine the parameter that characterizes the shape of the frequency distribution. presume.

The initial value data generation unit 18 uses the parameters (signals) 171 that are output from the frequency distribution shape parameter estimation unit 17 and characterize the shape of the frequency distribution, and the initial value data 181 relating to the posterior identification probability of autonomic nerve activity. To calculate.

The statistical structure-embedded neural network model unit 19 includes, in the input layer, data 161 relating to the frequency distribution of the autonomic nerve index value output from the autonomic nerve index value data frequency distribution generation unit 16 and the initial value data generation unit 18. A signal 181 relating to the initial value data of the posterior discriminant probability of the autonomic nerve activity output is given, and a finite number of components having a certain probability distribution are used in the output layer to generate the probability distribution of the population of the autonomic nerve index value data. And the parameter 191 of each component is learned and estimated, and a signal 192 relating to data representing a candidate for the state of autonomic nerve activity corresponding to the autonomic nerve index value data is output.

The autonomic nerve activity classification unit 20 outputs a signal 191 relating to the parameters of the probability distribution model output from the statistical structure-embedded neural network model unit 19.
And the signal 192 regarding the candidate state of the autonomic nerve activity is utilized to classify the state of the autonomic nerve activity of the subject into several stages.

Autonomic nerve index measurement / analysis procedure control unit 21
Determines whether the classification result is a desirable result for the experimenter, based on the signal 201 regarding the classification result output from the autonomic nerve activity classification unit 20. If necessary, the signal 212 relating to the change of the section boundary value of the amplitude value data of the autonomic nerve index is sent to the autonomic nerve index value data frequency distribution creating unit 16
, And classify the autonomic nerve activity again.

The physiological index reflecting the activity of the autonomic nervous system is
Pupil area, blood flow, phonocardiogram, electrocardiogram, magnetocardiogram, heart rate,
Heartbeat interval, heartbeat fluctuation, finger pulse wave, blood pressure, salivary output,
Salivary secretion rate, respiratory flow rate, exhaled breath component, skin resistance value, gastric juice pH, body temperature, respiratory rate, sweat rate, sweat rate, EEG, magnetoencephalogram, blood glucose level, blood hormone level, etc. Below,
An actual usage pattern including a measurement part of an autonomic nerve index value and a display part of a classification result will be described by taking as an example a case where a fingertip pulse wave is used as input data.

For example, a pulse pickup (for fingertips) 45261 manufactured by NEC Sanei Co., Ltd. is used as the autonomic nerve index measuring sensor 10, and a bioelectric phenomenon amplifier such as a bioelectric amplification unit 1253A is used as the autonomic nerve index measuring unit 11. To use. The sensor 10 for measuring the autonomic nerve index is attached to the tip of a finger (for example, the index finger). With these, it is possible to obtain the fingertip pulse wave as a voltage change. Therefore, this voltage change is applied to A / A of Canopus Electronic Co., Ltd. ADX-98E mounted on a personal computer such as PC-9821Af manufactured by NEC Corporation. It can be converted into a digital signal 111 via a D converter board.

When the preparation for the fingertip pulse wave measurement is completed, the fingertip pulse wave measurement is performed. The measurement is autonomic nerve index measurement
It proceeds according to the procedure set in the analysis procedure control unit 21. The set time (for example,
The signal of the fingertip pulse wave for 30 minutes) is recorded by the autonomic nerve index signal recording / reproducing unit 1 via the autonomic nerve index signal recording control unit 12.
3, for example, recorded on a magnetic disk device, a magnetic tape or the like.

When the measurement is completed, the autonomic-nervous-index-value data converting unit 14, for example, the workstation or the personal computer, causes the autonomic-nerve index signal recording / reproducing unit 1 to operate.
The fingertip pulse wave data recorded in 3 is converted into amplitude value data. Amplitude value x _{i at} each time t (i = 1, 2, ...,
T; t = Δf · i; where Δf is the sampling frequency and T is the number of samples) is, for example, the maximum value and the minimum value of the fingertip pulse wave data at a certain time (for example, 2 seconds) until that time. It can be calculated as the difference. The signal 141 relating to the amplitude value of fingertip pulse wave data is recorded in the converted data recording / reproducing unit 15, for example, a magnetic disk device, a magnetic tape or the like.

Next, the autonomic nerve index value data frequency distribution creating unit 16, for example, a workstation or a personal computer, based on the fingertip pulse wave amplitude value data 151 recorded in the conversion data recording / reproducing unit 15, The frequency distribution of fingertip pulse wave amplitude value data is created by the procedure of.

A section boundary value B _n (n = 0, 1, ..., N) for dividing the autonomic nerve index value into N stages is set in advance. Input the autonomic nerve index value data x _i in order, and if B _n ≤x _i ≤B _{n + 1} (n = 0, 1, ..., N-1), increase the value of the distribution frequency h (n) by 1. . By performing the same procedure for all x _i (i = 1, 2, ..., T), the distribution frequency h (n) (n = 1, 2, ..., N)
Can be requested. Further, using the distribution frequency h (n), the relative frequency is calculated by Equation 1.

[0031]

[Equation 1]

In FIG. 3, the distribution frequency h is obtained by applying the above-described method to the fingertip pulse wave data actually measured under a certain experimental condition.
It is a graph of (n). In this experiment, brain potential data was simultaneously measured in order to accurately obtain a visual evoked potential according to the state of autonomic nerve activity. In order to obtain high-quality visual evoked potentials according to each autonomic nervous activity state,
I would like to judge whether each section of the brain potential data corresponds to "tensed" or "not tense". To that end, we used fingertip pulse wave data to monitor the autonomic nervous activity of the subjects. Therefore, in this experimental design, the number of autonomic nerve activities to be classified is two: "tensed" or "not strained". Therefore, in the following, in consideration of the result of FIG. 3, the frequency of the fingertip pulse wave amplitude value obtained by the autonomic nerve index value data frequency distribution creating unit 16 in the mixed normal distribution model in which two normal distributions overlap each other. Approximate the distribution.

The frequency distribution shape parameter estimation unit 17, for example, a workstation or a personal computer, uses the signal 161 (section boundary value) concerning the frequency distribution of the fingertip pulse wave amplitude values obtained by the autonomic nerve index value data frequency distribution creation unit 16. B _n (n = 0, 1, ..., N) and distribution frequency h
(N) (n = 1, 2, ..., N) or relative frequency

[Outside 1] Using (n) (n = 1, 2, ..., N), parameters that characterize the shape of the frequency distribution are estimated according to the following procedure.

4 and 5 are views exemplifying the mixed normal distribution model defined by the equation (2).

[0035]

[Equation 2]

In FIG. 4, α = 0.3, μ ₁ = 0.5, μ ₂
= 1.2, σ ₁ = 0.2, σ ₂ = 0.2, and in FIG. 5, α =
0.3, μ ₁ = 0.5, μ ₂ = 1.0, σ ₁ = 0.2,
σ ₂ = 0.2. FIG. 4 exemplifies a case where two normal distributions are apart (hereinafter referred to as type A), and FIG. 5 illustrates a case where two normal distributions are close to each other (hereinafter referred to as type B). FIG. 3 shows an example of type B. Here, the graphs of the mixed normal distributions of FIGS. 4 and 5 are regarded as continuous functions. At this time, 2 in type A (Fig. 4)
The two maximum points P ₁ and P ₂ approximate the average μ ₁ and μ ₂ , respectively, and in the type B (FIG. 5), one inflection point P ₁ and one maximum point P ₂ each have the average μ ₁ and μ ₂ , respectively. You can think of it as approximating μ ₂ .

In general, any differentiable continuous function f
The maximum point of (x) at x = c is defined by Equation 3, and the inflection point is Equation 4 and d ² f (x) / d between the left side and the right side of x = c.
It is defined by the sign of x ² being different.

[0038]

(Equation 3)

[0039]

[Equation 4]

Since the actual frequency distribution is discrete data, the first derivative and the second derivative of the distribution frequency h (n) (n = 1, 2, ..., N) are shown by equations 5 and 6 ( Forward) Calculated by the difference formula. However, b (n) is defined by Equation 7.

[0041]

(Equation 5)

[0042]

(Equation 6)

[0043]

(Equation 7)

Further, b (N) -b (N-1) = b (N-
If 1) −b (N−2) = ... = b (2) −b (1) = Δη, Equations 5 and 6 are simplified to Equations 8 and 9.

[0045]

(Equation 8)

[0046]

[Equation 9]

Including the above procedure, the frequency distribution shape parameter estimation unit 17 determines the relative frequency.

[Outside 2] The following calculation is performed using the data of (n).

(1) The distribution frequency h (n) is converted into a relative frequency by the expression 10.

[0049]

[Equation 10]

(2) The maximum point or the inflection point is searched by the following procedure while increasing n by 1. However,
The number 11 "maximum points at n = n _C", "inflection point at n = n _C" is the number 12, can be detected.

[0051]

[Equation 11]

[0052]

(Equation 12)

(3) If the shape of the frequency distribution is type A
If so, the first maximum point

[Outside 3] ₁ and the second maximum

[Outside 4] _{Assume 2} . If the shape of the frequency distribution is type B, the first detected inflection point or maximum point

[Outside 5] ₁ and the second detected maximum point or inflection point

[Outside 6] _{Assume 2} .

(4) Solving Equation 13,

[Outside 7] Ask for. However, the number is 14.

[0055]

(Equation 13)

[0056]

[Equation 14]

(5) Initial values of σ ₁ and σ ₂ ,

[Outside 8] ₁ and

[Outside 9] ₂ is given by Equation 15. However, K is the number of states of autonomic nerve activity to be classified.

[0058]

(Equation 15)

Next, the initial value data generator 18, for example,
The parameters estimated by the frequency distribution shape parameter estimation unit 17 by the workstation or personal computer

[Outside 10] ,

[Outside 11] ₁ ,

[Outside 12] ₂ ,

[Outside 13] ₁ ,

[Outside 14] By using ₂ , the initial value W _nkm ⁽⁰⁾ (n = 1, 2, ..., N; k = 1, 2,
, K; m = 1, 2, ..., M) is calculated by the equation 16.

[0060]

[Equation 16]

However,

[Outside 15] _km is a weighting coefficient, which is given so as to satisfy the expression 17.

[0062]

[Equation 17]

N is the number of data (the number of stages of the autonomic nerve index value data), K is the number of states of autonomic nerve activity to be classified,
M represents the number of components forming the probability distribution.

In the present embodiment, K = 2 and M = 1, and Eq.

[0065]

(Equation 18)

Next, the frequency structure of the fingertip pulse wave amplitude value data obtained by the autonomic nerve index value data frequency distribution creating unit 16 by the statistical structure-embedded neural network model unit 19, for example, a workstation or a personal computer. Signal 161 (section boundary value B _n (n
= 0, 1, ..., N), relative frequency

[Outside 16] (N) (n = 1, 2, ..., N) and the initial value data 181 (W _nkm ⁽⁰⁾ (n = 1, 1 ⁾ regarding the posterior discrimination probability of the autonomic nerve activity obtained by the initial value data generation unit 18. 2, ...,
N; k = 1,2, ..., K; m = 1,2, ..., M)), and using a finite number of components having a certain probability distribution, The probability distribution is expressed and the parameters of each component are estimated by learning. The operation principle and structure of the statistical structure-embedded neural network model unit 19 are as follows.

A statistical structure-embedded neural network model having a mixed normal distribution model in which a normal distribution is used for each of the above components is, for example, “Neu”.
ral Networks, Vol. 4, 1991,
89-102 ". I. Perlovsky,
M. M. McManus says "Maximum Like
lihood Neural Networks fo
r Sensor Fusion and Adapt
Details of the paper published under the title "Ive Classification".

Now, the d-dimensional feature vector

[Outside 17] ∈

[Outside 18] Consider the problem of identifying (or classifying) ^d into K classes. As is well known, in order to minimize the loss associated with identification, the Bayes decision method may be used. That is, if
If the number is 19, class k is adopted. However, k, l =
1, 2, ..., K, l ≠ k, p (k |

[Outside 19] ) Is the posterior identification probability.

[0069]

[Formula 19]

Assuming that each class k is composed of M components, equation 20 is obtained.

[0071]

(Equation 20)

Using Bayes' theorem, p (k, m |

[Outside 20] ) Is the number 21.

[0073]

[Equation 21]

Here,

[Outside 21] Probability density function p (

[Outside 22] ) Is given by Equation 22. However, p (

[Outside 23] | K, m) is the probability density function of each component.

[0075]

[Equation 22]

Further, the prior probability p (k, m) is written as α _km, and θ _km is a parameter of the probability density function of the component km, and p (

[Outside 24] │k, m) Ψ (

[Outside 25] , Θ _km ), Equation 21 becomes Equation 23.

[0077]

(Equation 23)

Therefore, if the parameters α _km and θ _km can be obtained for each component,
Using Equations 20 and 23, the feature vector

[Outside 26] Can be identified.

The calculation of the posterior discrimination probabilities of Expressions 20 and 23 can be expanded to the three-layer hierarchical neural network model shown in FIG.

The input layer is composed of K × M units, and based on the probability density function of each component, a feature vector which is an input to this neural network.

[Outside 27] Output the conditional probability under class k, component m of. Where the probability density function of each component is the mean

[Outside 28] _km ∈

[Outside 29] ^d , feature vector

[Outside 30] Assuming that each element of is a d-dimensional normal distribution with independent and equal variance σ _km ² , the input-output relationship of the input layer is
It can be expressed as. Here, k = 1, 2, ..., K; m
= 1, 2, ..., M, o _km ⁽ⁱ⁾ is the unit km of the input layer
Represents the output of.

[0081]

[Equation 24]

The intermediate layer receives via the weighting factor α _km of the output of the input layer, and according to Eq.

[Outside 31] The posterior probability of the class k and the component m under is calculated by Equation 25. However, k = 1, 2, ..., K; m =
1, 2, ..., M.

[0083]

(Equation 25)

Further, the unit of the output layer is a linear unit, and the feature vector

[Outside 32] The posterior classification probability of class k under

[0085]

(Equation 26)

Therefore, the parameter α _{km of the} neural network,

[Outside 33] _{If km} and σ _km can be learned, the feature vector

[Outside 34] Can be classified into K classes.

The neural network uses N observation samples.

[Outside 35] Learning is performed according to the following procedure based on _n (n = 1, 2, ..., N).

First, for a given observation sample,
The parameter α _km using the maximum likelihood method,

[Outside 36] Consider finding _km and σ _km . That is, the parameter α _km that maximizes the log likelihood of the entire sample shown in Expression 27 under the constraint condition shown in Expression 28,

[Outside 37] _km, seek _{km σ.}

[0089]

[Equation 27]

[0090]

[Equation 28]

The maximum likelihood estimator of each parameter is given by equations 29 to 32, as is well known.
However, the number is 33.

[0092]

[Equation 29]

[0093]

[Equation 30]

[0094]

[Equation 31]

[0095]

[Equation 32]

[0096]

[Expression 33]

In this way, _if the initial value W _nkm ^{(0) of} W _nkm is given, the procedures of _equations 29 to 32 may be applied repeatedly. This iterative algorithm can be regarded as a kind of unsupervised learning. For the convergence determination of learning, for example, Bhattachar of the posterior identification probability shown in Expression 34 is used.
The yya distance may be used. However, 1 shows the number of times of learning.

[0098]

(Equation 34)

If learning is finished when this distance Bh becomes equal to or less than ε, it is necessary for the operation of the neural network.

[Outside 38] _km ,

[Outside 39] _km ,

[Outside 40] _km has been acquired in the neural network.

In the present embodiment, the fingertip pulse wave is used to classify the autonomic nerve activity into two states, "tensed" and "not tense". The observation sample is one-dimensional and the number of classes is two. Also, each class is represented by one normal distribution. Therefore, in the above neural network, this corresponds to d = 1, K = 2, M = 1.

Thus, the statistical structure-embedded neural network model unit 19 follows the procedure below.
Parameter of mixed normal distribution model α _km ,

[Outside 41] Acquire _km and σ _km (k = 1, 2; m = 1).

(1) Using the initial value W _nkm ⁽⁰⁾ obtained by the initial value data generation unit 18, the _expressions 29 to 32 are calculated, and the parameters

[Outside 42] _km ,

[Outside 43] _km ,

[Outside 44] _Ask for _km . However, in Expression 16, b (n) =

[Outside 45] _n (n = 1, 2, ..., N).

(2) The above estimated parameters

[Outside 46] _km ,

[Outside 47] _km ,

[Outside 48] Calculate the equations (24) and (25) using _km . At this time, the value of _Equation 25 gives the value of W _nkm ⁽¹⁾ .

(3) The value of W _nkm ⁽⁰⁾ and W _nkm ⁽¹⁾ is used to calculate equation 34. At this time, if the value of the distance Bh is ε (for example, 10 ⁻⁴ ) or less, the learning ends. Otherwise, W _nkm ⁽⁰⁾ = W _nkm ⁽¹⁾ is set and the above procedure is repeated until the value of the distance Bh becomes ε or less.

(4) Using the parameters obtained at the end of learning, the parameters on the right side of Equation 2 are α = α ₁₁ (α ₂₁ = 1−α ₁₁ ), μ ₁ =

[Outside 49] ₁₁ , μ ₂ =

[Outside 50] ₂₁ , σ ₁ = σ ₁₁ , σ ₂ = σ ₂₁ .

Next, the autonomic nerve activity classification unit 20, for example,
The workstation or personal computer uses the parameters α, μ ₁ , μ ₂ , σ ₁ , σ ₂ of the mixed normal distribution model obtained by the statistical structure-embedded neural network model unit 19 and follows the procedure described below. Classify the autonomic nervous activity of. It is difficult to determine which state belongs to the vicinity where the two normal distributions intersect. Therefore, it is desirable to specify an area around the average of each distribution for each state. Therefore, if x is the value of the autonomic nerve index value data, for example, as a classification rule, (μ ₁ −σ ₁ ≦) x ≦ μ ₁ + σ ₁ is in a “tensed” state, μ ₂ −σ ₂ ≦ It is possible to define x (≦ μ ₂ + σ ₂ ) as a “non-tensioned” state.
Examples of this classification rule are shown in FIG. 4 (type A) and FIG. 5 (type B).

However, it is possible that the two means are very close to each other, and μ ₂ −σ ₂ <μ ₁ + σ ₁ . In such a _{case, x ≦ μ 1 + σ 1} /2, may be changed classification law and _{μ 2 -σ 2/2 ≦ x} . In this way, it is possible to classify autonomic nervous activity into two states. The autonomic nerve activity classification unit 20 outputs a signal 201 related to the above classification result.

Autonomic nerve index measurement / analysis procedure control unit 21
Receives the signal 201 relating to the classification result output from the autonomic nerve activity classification unit 20 and determines whether the classification result is satisfactory to the experimenter. If the classification result is not satisfactory, the signal 2 regarding the change of the interval boundary value
12 is output.

Autonomic nerve index measurement / analysis procedure control unit 21
Based on the signal 212 regarding the change of the section boundary value output from the autonomic nerve index value data frequency distribution creating unit 16
Is the new interval boundary value

[Outside 51] _n (n = 1, 2, ...,

[Outside 52] ) And create a new frequency distribution. Hereinafter, the frequency distribution shape parameter estimation unit 17 and the initial value data generation unit 1
8. The autonomic nerve activity classification unit 20 sets a new classification rule via the statistical structure-embedded neural network model unit 19.

FIG. 2 is a block diagram showing the configuration of the autonomic nerve activity classification device according to the second embodiment of the present invention. The autonomic-nervous-activity classification device of the present embodiment is similar to the autonomic-nervous activity classification device of the first embodiment in that the frequency distribution shape parameter estimation unit 17
It has a configuration excluding. Therefore, the parts corresponding to those of the autonomic-nerve activity classification device of the first embodiment are designated by the same reference numerals, and detailed description thereof will be omitted.

In the apparatus for classifying autonomic nervous activity of the second embodiment, instead of removing the frequency distribution shape parameter estimating unit 17, an initial value data generating unit 18, for example, a workstation or a personal computer, is used for post-excitation of autonomic nervous activity. A random number that satisfies 0 <W _nkm ⁽⁰⁾ <1 is used as the initial value data of the identification probability. Random number generation can be easily achieved by incorporating a commercially available random number generation software package into a workstation or personal computer.

Hereinafter, the processes in the statistical structure-embedded neural network model unit 19, the autonomic nerve activity classification unit 20, and the autonomic nerve index measurement / analysis procedure control unit 21 are performed by the autonomic nerve activity classification apparatus of the first embodiment. Same as the case.

In the apparatus for classifying autonomic nerve activity of the second embodiment having such a configuration, the initial value data generation unit 18 does not use a parameter that characterizes the frequency distribution of the autonomic nerve index value data, but uses a random value. Since the random number is used as initial value data, the statistical structure-embedded neural network model unit 19 and the autonomic nerve activity classification unit 2
After the processing of 0, the autonomic nerve index measurement / analysis procedure control unit 2
In No. 1, the classification result is not always the desired result for the experimenter. Therefore, it is necessary to instruct the autonomic nerve index value data frequency distribution creating means to change the section boundary value of the amplitude value data of the autonomic nerve index to repeatedly perform the classification of the autonomic nerve activity to obtain the desired classification result. Become.

[0114]

As described above, according to the present invention, autonomic nerve index value data frequency distribution creating means, frequency distribution shape parameter estimating means, initial value data generating means, statistical structure-embedded neural network model means and autonomic nerve activity. A classifying unit is provided to generate data relating to the frequency distribution of the autonomic nerve index values by using the autonomic nerve index value data, and estimate the parameters characterizing the shape of the frequency distribution of the autonomic nerve index value data by using the generated data. Then, using the estimated parameters, initial value data for the posterior classification probability of autonomic nerve activity is generated, and the probability distribution of the population of autonomic nerve index value data is calculated using a finite number of components with a certain probability distribution. To estimate the parameters of each component in a learning manner and to correspond to the autonomic nerve index value data. Output the data representing the candidates of the state of autonomic nerve activity, and use the data representing the candidates of the state of the autonomic nerve activity corresponding to the parameters of the probability distribution model and the data of the index of the autonomic nerves to determine the state of the autonomic nerve activity of the subject. By classifying into several stages, unlike the conventional optimization method, it is not necessary to calculate the partial differentials of the first and second orders of the evaluation function in advance. The calculation time required for the parameter estimation of the probability distribution model based on this is small, and the state classification of autonomic nerve activity can be performed faster than before. In particular, even if the number of states of autonomic nerve activity to be classified increases, there is an advantage that the parameters of the probability distribution model can be estimated faster than the conventional optimization method.

Further, an autonomic nerve index measurement / analysis procedure control means is provided, and based on the classification result by the autonomic nerve activity classification means, it is judged whether or not the classification result is a desirable result for the experimenter, and if necessary, the autonomic nerve index. By instructing the autonomic nerve index value data frequency distribution creating means to change the interval boundary value of the amplitude value data to re-execute the classification of the autonomic nerve activity, once the autonomic nerve index value data is input, There is an effect that it can be classified into a predetermined stage based on the experimental plan and the subject's introspection report, and that the data analysis of the autonomic nerve index value can be efficiently performed.

Further, an autonomic nerve index value data frequency distribution creating means, an initial value data generating means, a statistical structure-embedded neural network model means, an autonomic nerve activity classifying means, and an autonomic nerve index measuring / analyzing procedure control means are provided. Data about the frequency distribution of the autonomic nerve index value is generated by using the autonomic nerve index value data, and initial value data about the posterior identification probability of the autonomic nerve activity is generated by using the random number in the range of 0 to 1, A finite number of components with a specific probability distribution are used to express the probability distribution of the population of the autonomic nerve index value data, the parameters of each component are learned, and the autonomic nerves corresponding to the autonomic nerve index value data are estimated. Outputs data representing activity state candidates and supports probability distribution model parameters and autonomic nerve index value data The state of the autonomic nerve activity of the subject is classified into several stages by using the data representing the candidate of the state of the autonomic nerve activity, and based on the classification result, it is judged whether the classification result is a desirable result for the experimenter, If necessary, by instructing to change the section boundary value of the amplitude value data of the autonomic nerve index and re-classifying the autonomic nerve activity, a parameter characterizing the shape of the frequency distribution of the autonomic nerve index value data is set. Without the estimation, the calculation time required for parameter estimation of the probability distribution model based on the autonomic nerve index value data is short, and the state classification of autonomic nerve activity can be performed faster than before. In particular, even if the number of states of autonomic nerve activity to be classified increases, there is an advantage that the parameters of the probability distribution model can be estimated faster than the conventional optimization method.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an autonomic nerve activity classification device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of an autonomic nerve activity classification device according to a second embodiment of the present invention.

FIG. 3 is a diagram exemplifying a frequency distribution created from amplitude value data of actually measured fingertip pulse waves.

FIG. 4 is a diagram exemplifying a mixed normal distribution model based on two normal distributions with different means and a classification rule thereof.

FIG. 5 is a diagram illustrating a mixed normal distribution model based on two normal distributions whose averages are close to each other and a classification rule thereof.

FIG. 6 is a diagram showing an example of a structure of a statistical structure-embedded neural network model unit in FIG. 1;

FIG. 7 is a block diagram showing an example of a conventional autonomic nerve activity classification device.

[Explanation of symbols]

10 Sensor for measuring autonomic nerve index 11 Autonomic nerve index measuring unit 12 Autonomic nerve index signal recording control unit 13 Autonomic nerve index signal recording / reproducing unit 14 Autonomic nerve index value data converting unit 15 Conversion data recording / reproducing unit 16 Autonomic nerve index value Data frequency distribution creation unit 17 Frequency distribution shape parameter estimation unit 18 Initial value data generation unit 19 Statistical structure embedded neural network model unit 20 Autonomic nerve activity classification unit 21 Autonomic nerve index measurement / analysis procedure control unit

Claims (3)

[Claims] 1. An autonomic nerve activity classifying apparatus using autonomic nerve index value data, wherein the autonomic nerve index value data is used to generate data relating to the frequency distribution of autonomic nerve index values. And a frequency distribution shape for estimating a parameter that characterizes the shape of the frequency distribution of the autonomic nerve index value data by using the data relating to the frequency distribution of the autonomic nerve index value data generated by the means An initial value for generating initial value data relating to the posterior identification probability of autonomic nerve activity using the parameter estimating means and the parameter characterizing the shape of the frequency distribution of the autonomic nerve index value data estimated by the frequency distribution shape parameter estimating means. It is equipped with a data generation means and a learning mode, and consists of three types: an input layer, an intermediate layer and an output layer. Data on the frequency distribution of the autonomic nerve index values generated by the autonomic nerve index value data frequency distribution creating means and initial value data on the posterior identification probability of autonomic nerve activity generated by the initial value data generating means, In the output layer, a finite number of components with a certain probability distribution are used to represent the population probability distribution of the autonomic nerve index value data, and the parameters of each component are learned and estimated A statistical structure-embedded neural network model means for outputting data representing a candidate for a state of autonomic nerve activity corresponding to the index value data, and a parameter of a probability distribution model output by the statistical structure-embedded neural network model means and Represents candidates for the state of autonomic nerve activity corresponding to the autonomic nerve index value data Using the chromatography data, autonomic nerve activity classification device and having an autonomic nervous activity classification means for classifying the several stages the state of the autonomic nervous activity of the subject. 2. Based on the classification result by the autonomic nerve activity classification means, it is judged whether the classification result is a desirable result for the experimenter, and if necessary, the section boundary value of the amplitude value data of the autonomic nerve index is changed. The autonomic nerve activity classification device according to claim 1, further comprising autonomic nerve index measurement / analysis procedure control means for instructing the autonomic nerve index value data frequency distribution creating means to re-perform autonomic nerve activity classification. 3. An autonomic nerve activity classifying apparatus using autonomic nerve index value data, wherein the autonomic nerve index value data is used to generate data relating to the frequency distribution of autonomic nerve index values. Means, initial value data generating means for generating initial value data relating to the posterior discrimination probability of autonomic nerve activity by using random numbers in the range of 0 to 1, and a learning mode, and an input layer, an intermediate layer and an output layer Data on the frequency distribution of the autonomic nerve index value generated by the autonomic nerve index value data frequency distribution creating means and the posterior identification of the autonomic nerve activity generated by the initial value data generating means in the input layer. Given the initial value data related to the probability, a population of autonomic nerve index value data using a finite number of components with a certain probability distribution in the output layer. Statistical structure-embedded neural network model means for expressing the probability distribution of the group, learning-estimating the parameters of each component, and outputting data representing candidates for the state of autonomic nerve activity corresponding to the autonomic nerve index value data And the data representing the candidates of the state of the autonomic nerve activity corresponding to the parameters of the probability distribution model and the autonomic nerve index value data output by the statistical structure-embedded neural network model means, using the autonomic nerve activity of the subject. Based on the autonomic nerve activity classifying means for classifying the states of the above into several stages and the result of classification by this autonomic nerve activity classifying means, it is judged whether or not the result is desirable for the experimenter, and if necessary, the autonomic nerve activity The section boundary value of the amplitude value data of the index is changed by the above-mentioned autonomic nerve index value data frequency distribution creation procedure. Autonomic nerve activity classification device and having an autonomic nervous index measurement and analysis procedure control unit instructs re implementing the classification of the autonomic nervous activity.