WO2012018157A1

WO2012018157A1 - System for automatically classifying sleep stages on the basis of biosignals

Info

Publication number: WO2012018157A1
Application number: PCT/KR2010/006703
Authority: WO
Inventors: 유선국; 이충기
Original assignee: 연세대학교 산학협력단
Priority date: 2010-08-01
Filing date: 2010-09-30
Publication date: 2012-02-09
Also published as: KR101235441B1; KR20120012364A

Abstract

The present invention relates to a system for automatically classifying sleep stages on the basis of biosignals for predicting sleep stages using mutually different biosignals. For this purpose, the present invention relates to a system for automatically classifying sleep stages on the basis of biosignals, comprising: a biosignal detection unit using EEG, EOG, EMG, PPG, SKT, and GSR sensors to acquire various biosignals through mutually different channels; a signal-processing and characteristic extraction unit using a method suitable for extracting the characteristics of each biosignal to process the various biosignals acquired through each channel by the biosignal detection unit, and extract a certain characteristic from each of the biosignals; a subset-generating unit for generating biosignal groups based on the certain characteristics of each biosignal extracted by the signal-processing and characteristic extraction unit; and a performance evaluation unit for classifying the biosignals into sleep stage categories (i.e. REM, LS, and DS) using the biosignal groups formed by the subset-generating unit, and analyzing a performance evaluation of the groups for each stage, wherein for the analysis of sleep stages using mutually different biosignals, sleep stages can be accurately and conveniently predicted through the combination of biosignals in order to enable accurate measurement and convenience.

Description

Bio signal based automatic sleep stage classification system

The present invention relates to a biosignal-based automatic sleep stage classification system, and more particularly, to a system for biosignal combinations based on convenience of biosignal measurement using a quadratic classifier. Through the configuration and results of the present invention, the sleep stage can be more accurately analyzed and conveniently measured.

In other words, the present invention configures a subset of optimal biosignal combinations in consideration of the accuracy and convenience of measurement during sleep stage analysis using biosignals under different conditions, and constructs and repeats the data set independent of the experimental date. The present invention relates to an automatic sleep stage classification system based on bio-signals, which allows for accurate and convenient analysis of sleep stages through performance evaluation.

In addition, the present invention may be selectively supplemented by an individual optimization process using a gene-fuzzy algorithm (when high accuracy is required), thereby further improving its accuracy.

Sleep stages are generally divided into rapid eye movement (REM), shallow sleep, and deep sleep states, and

sleep stages

1 and 2 are classified into light sleep (LS) and sleep. Stages 3 and 4 are classified as Deep Sleep (DS).

The sleep classification method defined above includes a biosignal reading method (visual inspection, manual inspection) based on R & K techniques (EEG, EOG, EMG).

Recently, there is a limitation in automatic sleep pattern analysis due to individual differences in sleep biosignal responses in computer-based sleep stage classification. The standard criteria-based visual identification process used has been reported to be difficult even for professionals. have.

In addition, the probabilistic identification method for a particular frequency band is also ambiguous in its bounds. For example, according to the standard method of dividing the sleep stage, one period of the sleep stage (stage 3) should have a delta wave of more than 20% and less than 50%. However, if delta waves are less than 20%, they are considered a sleep stage (stage 2).

Therefore, it is necessary to develop a system capable of performing automatic signal classification based on biosignals in which quantitative performance evaluation is performed in order to distinguish these unclear boundaries.

With the growing interest in ubiquitous healthcare, automatic sleep pattern analysis techniques require the need for a monitoring system that can quantitatively evaluate how well people sleep in modern times.

The technical problem to be solved by the present invention is to configure a subset of the optimal bio-signal combination in consideration of the accuracy and convenience of the measurement in the sleep stage analysis using the bio-signals of different conditions to accurately and conveniently Make it predictable.

Embodiments of the present invention for achieving the above object, the biological signal for obtaining a variety of biological signals through different channels using at least one or more of a plurality of sensors, such as EEG, EOG, EMG, PPG, SKT, GSR sensor Signal processing for extracting feature points representing the characteristics of each biosignal from each biosignal by performing signal processing on the biosignals detected through the respective channels by the detector and the biosignal detector. Sleep step (REM, LS, DS) through a subset generator for generating a biosignal combination based on the feature points of each biosignal of the feature extractor, the signal processor and the feature extractor, and the biosignal combinations combined by the subset generator Performance evaluation unit for classifying and analyzing the performance evaluation of combinations according to each step, and individual optimization process through gene-fuzzy algorithm A bio-signal-based automatic sleep stage classification system comprising.

According to the biological signal-based automatic sleep stage classification system according to the present invention, a subset is composed of various biosignal combinations in consideration of the accuracy and convenience of measurement in the sleep stage analysis using the biosignals of different conditions, By dividing the sleep stages and optimizing them with a gene-fuzzy algorithm, the sleep stages can be easily predicted by accurately and predicting the sleep stages and reducing the number of measurement signals.

1 is a block diagram schematically showing the overall configuration of a biological signal-based automatic sleep stage classification system according to an embodiment of the present invention.

2 is a conceptual diagram illustrating the classification of the sleep stage classification process of the second order linear separator according to an embodiment of the present invention in more detail.

3 is a graph illustrating the results of evaluating the experimenter-dependent data set using the second order linear separator.

4 is a graph illustrating the results of evaluating the experimenter-independent data set using the second order linear separator.

FIG. 5A is a block diagram illustrating a gene-fuzzy optimizer configured to analyze a sleep stage by receiving a subset of separate biometric data from a subset generator as an embodiment of the present invention.

5B is a block diagram illustrating an individual optimization process of sleep stage classification using a gene-fuzzy algorithm.

FIG. 6 is a graph illustrating the relative power spectrum of the test subject and the normalization result of the sleep stage when the gene-fuzzy algorithm is used.

FIG. 7 is a graph showing differences between individualized normalized relative power spectral values of five frequency bands (alpha, theta, gamma, sigma, and beta) in successive sleep stages when using a gene-fuzzy algorithm. FIG. .

100: biosignal detection unit 110, 120: EEG sensor

130: EOG sensor 140: EMG sensor

150: PPG sensor 160: GSR sensor

170: SKT sensor 200: signal processing and feature extraction unit

300: subset generator 400: performance evaluation unit

500: gene-fuzzy optimization unit 600: comparison unit

Hereinafter, the configuration and operation of the biological signal based automatic sleep stage classification system according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

The terms or words used in this specification and claims are not to be construed as limiting in their usual or dictionary meanings, and the inventors may properly define the concept of terms in order to best explain their invention in the best way possible. It should be interpreted as meaning and concept corresponding to the technical idea of the present invention. Therefore, since the embodiments described in the specification and the configuration shown in the drawings are only one of the most preferred embodiments of the present invention, it is understood that there may be various equivalents and modifications that can replace them at the time of the present application. shall.

1 is a block diagram schematically showing the overall configuration of a biological signal-based automatic sleep stage classification system according to an embodiment of the present invention, Figure 2 is a detailed description of the sleep stage classification process performed in an embodiment of the present invention Illustrated classification conceptual diagram.

As shown in FIG. 1, the biosignal-based automatic sleep stage classification system according to an exemplary embodiment of the present invention includes a biosignal detection unit 100, a signal processing and feature extraction unit 200, a subset generator 300, and a performance evaluation unit. 400, and a performance optimizer 500.

The biosignal detection unit 100 is a block for measuring various biosignals from each part of the body using a 7-channel amplifier (Biopac MP150TM). The 7-channel amplifier includes two channels of electroencephalogram (EEG)

sensors

110 and 120, Electro-oculogram (EOG) sensor 130, electromyogram (EMG) sensor 140, photoplethysmography (PPG) sensor 150, galvanic skin response (GSR) sensor 160, and skin temperature (SKT) sensor 170 With a plurality of bio-signals detected from each sensor to obtain each through a different channel. Here it would be desirable to digitize the sampling rate of the 7-channel amplifier to 1000 Hz. Two-

channel EEG sensors

110 and 120 are attached to the body using conductive gels, and Grass electrodes are attached to the central side C3 and C4 of the scalp, respectively. The EOG sensor 130 for measuring the movement of the eyeball is attached to the left eye (+) and the right eye (-), respectively, and the ground electrode is attached to the forehead. An EMG sensor 140 for measuring EMG is attached under the chin. PPG sensor 150 is attached to a finger using a reflective IR sensor. The GSR sensor 160 is attached to the middle finger and the ring finger using two electrodes. SKT sensor 170 for skin temperature measurement is attached to the body using a calibrated sensor.

The signal processing and feature extractor 200 performs signal processing on various biosignals obtained through the respective channels by the biosignal detection unit 100 in a manner suitable for feature extraction of each biosignal, thereby extracting each biosignal from each biosignal. Extract a feature point representing the feature of. Here, the signal processing and feature extractor 200 extracts the feature points of the biosignal through the spectrum analysis of the detection waveforms of the EEG 1 sensor 110, the EEG 2 sensor 120, and the EOG sensor 130. The detection waveform of the PPG sensor 150 extracts feature points of the biosignal through the HRV analysis method, and the detection waveform of the GSR sensor 160 extracts the feature points of the biosignal through phasic analysis and detects the SKT sensor 170. The waveform extracts feature points of the biosignal using the mean value and the standard deviation value.

Signal processing for each biosignal detection waveform made by the signal processing and feature extraction unit 200 is as follows.

The EEG signal uses a fast Fourier transform (FFT) without noise reduction by covering the hanning windows in each epoch of consecutive single EEG signals read at C3-A2 according to the 10-20 International Electrode Position Guidelines for Body Position. To perform. The length of the Hanning window for performing this fast Fourier transform is, for example, 32,768 (2 ^ 15), and the bandwidths of the power spectrum are delta (1-4 Hz), theta (4-7.5 Hz), and alpha. (7.5-12 Hz), sigma (12-14 Hz) and beta (16-25 Hz). At this time, the relative power spectrum of each epoch is calculated by (power of each bandwidth) / (total power), and the power is calculated by summing the powers for each of the five epochs.

Each epoch represents the median of six cycles in each sequence (6 cycles * 30 seconds = 3 minutes). After such median adjustment, the relative power value normalized between 0 and 1 in the recording can be configured to be set as the input of a fuzzy classifier.

The EOG signal is used by removing the direct current (DC) level and removing the overall spectral power value to measure the movement of the pupil in the REM state.

Heart rate variability (HRV) is represented by a value that reflects the autonomic nervous system (ANS), and PPG and heart rate are used to determine the amount of change during sleep. You may also consider whether you have a transient heart rate or bradycardia. Changes in heart rate are also associated with awakening from sleep.

For HRV analysis using PPG, time-domain, frequency-domain, and non-linear methods are applied. Time domain analysis shows the mean HR (the mean of RR interval), the SDNN (the standard deviation of RR interval), the RMSSD (the root mean square of successive RR interval), and the NN50 (the number of successive intervals differing more than 50ms). Extract.

In short-term HRV analysis in frequency domain analysis, the commonly used features are LF / HF, HF (0.15Hz-0.4Hz), and LF (0.04-0.15Hz). In order to remove low frequency components (<0.04 Hz), low frequency components are removed with a least mean square based on a high frequency band filter. The successive three-minute lengths within each sleep phase are analyzed by the fourier method. HF and LF calculate the relative power to extract features (power per bandwidth) / (total power).

Poincar plots (SD1, SD2), which are commonly used in nonlinear HRV analysis, show the correlation between successive RR intervals. SD1 is associated with interpreting short-term variability related to breathing and SD2 is associated with long-term variability.

The change in GSR can be obtained by measuring the voltage between two electrodes when a low voltage current is supplied, which is represented by tonic and phasic waveforms, and the phasic component is particularly important for understanding temperature control during sleep. It is a signal. Therefore, in the present invention, after downsampling the sampling rate from 1000 Hz to 20 samples / second, the Bartlett window is used to distinguish each of them and to convolution continuously. The composition of phasic can be measured by two consecutive zero-crossings. Zero crossing rate refers to the detection of going from negative (+) to positive (+) or from positive (+) to zero with one epoch. The number of phasic component detections at the signal length of 3 minutes is extracted from the signal of the GSR sensor.

The SKT signal is calculated from the mean and standard deviation of the intervals. Skin maintains a balance between heat production and heat dissipation through thermoregulation mechanisms, which are affected by sleep stages, which closely correlate skin temperature with transient sleep stages. In other words, the average of the SKT signal is relatively lower than the shallow or deep sleep state in the REM state, and the variation is relatively large due to the temporary paralysis of the ANS (autonomic nervous system) within the interval (3 minutes of signal length).

The subset generator 300 configures a plurality of subsets required for sleep stage classification using a combination of biosignals under different conditions based on feature points of respective biosignals extracted by the signal processor and the feature extractor 200. At this time, the biosignal combination is an EEG-based biosignal combination and a PPG-based biosignal combination. The EEG-based biosignal combination is an EEG, EEG + EOG, EEG + EOG + GSR, EEG + GSR subset, and a PPG-based combination. The biosignal combinations are PPG, PPG + GSR + SKT, PPG + GSR, PPG + SKT, GSR + SKT, GSR, and SKT subsets. Since a plurality of signals exist based on a feature point within each signal, for example, in the case of an EEG signal, a normalized value of relative power spectrum values of a plurality of frequency bands alpha, theta, gamma, sigma, and beta bands is used as a subset combination. Can be.

The biosignals selected for the subset combination are preferably distinguished by focusing on the sleep prediction accuracy and the convenience of the biosignal measurement.

The data underlying the subset is preferably objective data that is as independent of external factors as possible. For example, the subsets are day-independent data sets that are experimental data sets and training data sets that are experimentally dated, and the day-independent data sets represent one day of three days within an experimenter. Three subjects, the subject-dependent data set composed of the remaining two days as a training data set, and one subject's randomly selected one day as the experimental data set, It consists of a subject-independent data set consisting of a training data set.

The performance evaluator 400 classifies the sleep stages REM, LS, and DS through the quadratic linear classfier, based on the combination of the bio signals based on the subset generator 300. We analyze the performance evaluation of subset combination according to. The performance evaluator 400 evaluates both the experimenter-dependent data set and the experimenter-independent data set capable of predicting the sleep stage in the same manner.

The second order linear separator is a tool developed by the present invention, which is a quadratic linear classifier based on Bayes's rule, which selects the posterior probability P (S _y | Classify _y _{= 1,2,3} (S ₁ : deep sleep, S ₂ : shallow sleep, S ₃ : REM). Equation for selecting the posterior probability maximum value is shown in Equation 1 below.

[Revision 17.12.2010 under Rule 91]
Equation 1

The secondary linear separator will be described in more detail as follows.

Referring to FIG. 2, assuming that each of the features x ₁ , x ₂ , x ₃ , ..., x _n is a normal density function, each density function is It can be defined as in Equation 2.

[Revision 17.12.2010 under Rule 91]
Equation 2

This is represented by Equation 3 below when the MAP (Maximum A Posterior) discriminant function is expressed according to the Bayes rule.

[Revision 17.12.2010 under Rule 91]
Equation 3

If the constant term is removed, (2π) ^{n / 2} and 1 / p (x) do not need to be redefined, as shown in Equation 4 below.

[Revision 17.12.2010 under Rule 91]
Equation 4

Taking the natural logarithm to Equation 4 may cause the exponential term to disappear and may be represented by Equation 5 below.

[Revision 17.12.2010 under Rule 91]
Equation 5

[Revision 17.12.2010 under Rule 91]
Here, the quadratic equation

It can be expressed as follows, and mapping each of them, it can be seen that the same as Equation 6 below.

[Revision 17.12.2010 under Rule 91]
Equation 6

Therefore, in the case of using such a linear linear separator it is possible to simply classify the sleep stages from the respective characteristics through the process as shown in FIG.

[Revision 17.12.2010 under Rule 91]
Where x1, x2, x3, ..., xn is characteristic and it has a multivariate normal density structure (Equation 2), the basic structure of the quadratic equation is

In comparison with the discriminant function,

Therefore, the result shown in Equation 5 can be represented.

Table 1

Vital signs	Feature Name	Sleep stage			ANOVAp value	post hoc multiple comparison not significant
		NREM		REM [mean ± SD]
		Deep Sleep State [mean ± SD]	Shallow sleep state [mean ± SD]	REM [mean ± SD]
EEG	Delta	0.8335 ± 0.08541	0.5165 ± 0.131	0.4294 ± 0.0962	<0.05
	Theta	0.3538 ± 0.1498	0.5160 ± 0.1592	0.712 ± 0.1363	<0.05
	Alpha	0.2776 ± 0.1024	0.5055 ± 0.1277	0.6261 ± 0.1473	<0.05
	Sigma	0.4093 ± 0.1560	0.6768 ± 0.1739	0.3721 ± 0.0956	<0.05
	Beta	0.3349 ± 0.0509	0.4975 ± 0.1256	0.7994 ± 0.0898	<0.05
EOG	power	(-) 0.0161 ± 0.1283	(-) 0.0223 ± 0.1411	0.0553 ± 0.1850	<0.05	Shallow water, deep water
PPG	meanHR	0.5779 ± 0.1352	0.5278 ± 0.1782	0.3537 ± 0.1754	<0.05
	SDNN	0.4202 ± 0.1399	0.5674 ± 0.1937	0.6212 ± 0.1701	<0.05
	RMSSD	0.5227 ± 0.1725	0.5489 ± 0.1901	0.4746 ± 0.2135	<0.05	Shallow water, deep water
	NN50	0.5506 ± 0.1735	0.5425 ± 0.1817	0.4191 ± 0.2026	<0.05	Shallow water, deep water
	SD1	0.5867 ± 0.174	0.503 ± 0.2115	0.4166 ± 0.2018	<0.05
	SD2	0.4133 ± 0.1740	0.4970 ± 0.2115	0.5834 ± 0.2018	<0.05
	LF / HF	0.3965 ± 0.1062	0.531 ± 0.1769	0.6227 ± 0.1599	<0.05
	power	0.4497 ± 0.1247	0.5739 ± 0.1746	0.6152 ± 0.1741	<0.05
	HF (nu)	0.7029 ± 0.1596	0.5171 ± 0.2074	0.3979 ± 0.1516	<0.05
	LF (nu)	0.2971 ± 0.1596	0.4829 ± 0.2074	0.6021 ± 0.1516	<0.05
GSR	phasic	0.6861 ± 0.1708	0.5446 ± 0.1392	0.474 ± 0.0845	<0.05
SKT	mean	0.5836 ± 0.1695	0.5115 ± 0.1778	0.3757 ± 0.2158	<0.05
SKT	SD	0.4417 ± 0.1187	0.5318 ± 0.1636	0.6076 ± 0.1814	<0.05

Table 1 above is a result of performing a one-way analysis of variance (ANOVA) analysis of the feature values extracted from each biological signal acquired from the experimenter using the sleep stage classification system according to the three sleep stages. 3 is a graph illustrating the results of evaluating the experimenter-dependent data set using the second-order linear separator. In the results related to the experimenter-dependent data set, all data generate three independent training data sets. The rest of the data was used to verify. The EEG-based sleep stage prediction result shows more than 85% accuracy, the PPG / GSR / SKT combination shows more than 80% accuracy, and the remaining biosignal combinations show more than 80% accuracy.

4 is a graph illustrating the results of evaluating the experimenter-independent data set using the second-order linear separator. EEG-based results show more than 87% of the results, and relatively high accuracy compared to the experimenter-dependent data set. Although low SD is shown, sleep stage evaluation using PPG, GSR, and SKT shows relatively low accuracy compared to the experimenter-dependent data set.

***** Optional, supplementary accuracy improvement by fuzzy inference *****

On the other hand, the sleep stage classification system according to the present invention can further improve the accuracy by selectively supplementing the analysis by fuzzy theory. FIG. 5A illustrates a gene-fuzzy optimizer 500 that analyzes a sleep stage by receiving a subset of separate bio data from a subset generator in addition to the performance evaluator 400 using a second linear separator as an embodiment of the present invention. 5B is a block diagram illustrating an individual optimization process of sleep stage classification using a gene-fuzzy algorithm.

The fuzzy-genetic optimization unit 500 applies a common type of fuzzy logic control called Mamdani to the fuzzy inference algorithm. The system consists of Xi's multiple input sets, output sets Y, and j to which a fuzzy algorithm is applied.

FIG. 6 illustrates a case where the subset signal input to the gene-fuzzy optimizer 500 is the simplest, and the power spectrum of each band of the delta, theta, alpha, sigma, and beta signals of the EEG signal is illustrated. Is a graph illustrating a state of normalizing a relative power value with respect to FIG. 6, and FIG. 6 (e) is a graph illustrating sleep stages classified by the RK law. The signal in each band of the EEG signal calculates the relative power in each band divided into delta, theta, alpha, sigma, and beta signals as the power / total power of each band, and the total power is the sum of each band in each band. Is obtained by calculating the power spectrum of each band.

Thus, in one embodiment, the system may consist of five input sets and five output sets of Xi, and j, to which a fuzzy algorithm is applied. Y represents a typical output and is classified into waking state (WA), shallow sleep (LS), deep sleep (DS), and REM. Xi, the input set, is the high interval, two fuzzy data sets.

And the lower interval

Divided by input vector

Is a membership function

Wow

Is converted to. The input set of the fuzzy inference algorithm is represented by Equation 7.

[Revision 17.12.2010 under Rule 91]
Equation 7

[Revision 17.12.2010 under Rule 91]
From the stomach

Is

Means an element in the input set.

[Revision 17.12.2010 under Rule 91]

Wow

Divided by the membership function of

In three

Generalizing this, the input range is between -1 and 1 and the output range is selected between 0 and 1, resulting in an appropriate membership map.

Related to Equation 8 to obtain the grade of the membership function is shown in Equation 8.

[Revision 17.12.2010 under Rule 91]
Equation 8

[Revision 17.12.2010 under Rule 91]

Wow

Mean relatively low and high intervals, respectively,

Represents the boundary line of the section that changes between the low section and the high section.

,

The matrix of is set as a membership function in each interval.

[Revision 17.12.2010 under Rule 91]
According to Mamdani fuzzy logic,

And low interval

Is combined with the inference system formal rules according to the above rules.

The rule is as shown in Equation (9).

[Revision 17.12.2010 under Rule 91]
Equation 9

[Revision 17.12.2010 under Rule 91]

Input stage of the front part

in

or

Is associated with one of the

Denotes one of the division of Y (WA, REM, shallow sleep, deep sleep), which is a step of determining a later sleep stage. Front

Is the occurrence of the number of signals (alpha, beta, theta, gamma, sigma)

Each rule becomes a rule that can determine the output of the fuzzy, and the process of selecting the rule is through training and reasoning. The generalization of each law and fuzzy law set assigned to the output set is determined by the fuzzy inference system.

After applying the appropriate fuzzy inference algorithms, the preceding part of each inference algorithm is computed by fuzzy combining with the redefined fuzzy set of result parts. When the preceding part of each inference algorithm is calculated, the disjunction of the fuzzy is calculated as the result aggregation. This execution is calculated in the same manner as (10).

[Revision 17.12.2010 under Rule 91]
Equation 10

[Revision 17.12.2010 under Rule 91]

Is

Means the result of the effect of the output appearing according to the law, combined with the output variable.

[Revision 17.12.2010 under Rule 91]

Where u (y) is the output variable represented by each law and the result of the total set combined with a single fuzzy set.

The crisp output of the fuzzy inference algorithm is evaluated by the defuzzification of u (y). In the present invention, the method of defuzzification is performed on the median of the maximum value representing the average value of the maximum values of all local membership functions. In one embodiment, the life stage is awake when the crisp output value is displayed as 1, rapid eye movement (REM) status when displayed as 2, shallow sleep when displayed as 3, and displayed as 4 At the same time, it can be determined by deep sleep.

As shown in FIG. 5A, the gene-fuzzy optimizer 500 that receives a plurality of subsets of the biosignal is defuzzified through a fuzzy inference algorithm and outputs a sleep stage, and the gene-fuzzy optimizer 600 compares the result. The sleep stage analyzed at 500 and the sleep stage analyzed at the performance evaluation unit 400 are compared. The comparison unit 600 helps the user determine the accuracy by determining and displaying whether the sleep stage analyzed by the gene-fuzzy optimizer 500 and the sleep stage analyzed by the performance evaluation unit 400 are the same. .

On the other hand, in constructing a fuzzy inference algorithm, finding a grade of a membership function and representing a set of fuzzy inference algorithms that are stably performed under given conditions are problematic. To solve this problem, we need to construct a genetic structure to determine the matrix parameters of the membership function and the set of fuzzy inference algorithms.

The structure of genes applied to fuzzy-genetic algorithms consists of binary sequences that define the rank of membership function C and the set of S fuzzy inference algorithms. C is expressed as in Equation (11).

[Revision 17.12.2010 under Rule 91]
Equation 11

[Revision 17.12.2010 under Rule 91]
here

Wow

In the relatively binary code

Input set of

,

Wow

Section (low section, high section, changing boundary section).

The length of is calculated as in Equation 12.

[Revision 17.12.2010 under Rule 91]
Equation 12

[Revision 17.12.2010 under Rule 91]
here

= 1, 2 and 3 which means

Means the lower, higher, and changing boundary.

The fuzzy inference algorithm set of the sequence S is expressed as in Equation 13.

[Revision 17.12.2010 under Rule 91]
Equation 13

[Revision 17.12.2010 under Rule 91]

Means the possible number from the preceding part,

and

Represents output variables.

silver

The rule defines the number of binary cases represented by '0' and '1'.

One of the most important in considering the successful generalization of the fuzzy-genetic optimizer 500 is the suitability function. The configuration of the system should match the target sleep level value. In this study, the fitness function for the mean error rate for four sleep stages is considered. The fit function equation for the average error rate is shown in Equation 14.

[Revision 17.12.2010 under Rule 91]
Equation 14

TP refers to the number of sleep steps accurately measured, and N represents the length of the sleep steps. K = 1, 2, 3, or 4 indicate each of the sleep stages WA, REM, SS (shallow sleep), and DS (deep sleep).

The genetic structure of the fuzzy-genetic optimizer 500 is an algorithm for generating population sizes of 20 and 50,

In the present invention, a typical selection method and a roulette wheel selection method are used. In the Roulette wheel selection, the gene has a probability on the selected ones directly proportional to the suitability. Two genes are randomly selected based on this probability and then produce offspring.

The genetic algorithm's crossover rate is 0.8 for scatter functions. This generated a random binary vector with '1' obtained from the first parent gene and '0' obtained from the second parent gene. The chromosomes obtained here combine the form with a new generation of genes. Gaussian modifications selected from the Gaussian distribution were applied to modify the parent chromosome. The degree of mutation relative to the distributed standard deviation decreased linearly with each new generation, finally reaching zero, the last step.

[Revision 17.12.2010 under Rule 91]
In chromosome representation,

Is encoded in 5 bits,

Is represented by 75 bits, and C is a set of five inputs.

With input

Based on the assumption of two fuzzy inference algorithm sets per input set of

The law of probability, and each of these laws

and

Is assigned to one of the output variables represented by. Then the,

By

The set of fuzzy inference algorithms from the stochastic laws is corrected. As a result, the chromosome is represented by 171 bits, which is by optimized membership function (75 bits) and fuzzy inference algorithm selection (96 bits).

FIG. 7 is a graph illustrating individual differences in normalizing relative power spectral values of five frequency bands (alpha, theta, gamma, sigma, and beta) in successive sleep stages. The delta and alpha bands in, theta bands in the deep sleep phase, the alpha and beta bands in the awake state, and the sigma bands in the shallow sleep phase.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above-described embodiments, which can be variously modified and modified by those skilled in the art to which the present invention pertains. Modifications are possible. Accordingly, the spirit of the present invention should be understood only by the claims set forth below, and all equivalent or equivalent modifications thereof will belong to the scope of the present invention.

Claims

A biosignal detection unit which acquires various biosignals through different channels using a biosignal measuring sensor;

Signal processing and features for extracting feature points representing the characteristics of each biosignal from each biosignal by performing signal processing on the biosignals detected through each channel by a method suitable for feature extraction of each biosignal. Extraction unit;

A subset generator which generates one or more biosignal combinations based on feature points of each biosignal of the signal processing and feature extractor;

And a performance evaluator for determining a sleep stage (REM, LS, DS) based on the biosignal combinations combined by the subset generator through a second linear separator.
The method of claim 1,

The biosignal measuring sensor includes at least one or more of an EEG, EOG, EMG, PPG, SKT, and GSR sensors.
The method of claim 2,

In the signal processing and feature extraction unit,

The detection waveform of the EEG and EOG sensors is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal through spectral analysis.
The method of claim 2,

In the signal processing and feature extraction unit,

The detection waveform of the PPG sensor is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal through the HRV analysis method.
The method of claim 2,

In the signal processing and feature extraction unit,

The detection waveform of the GSR sensor is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal through phasic analysis.
The method of claim 2,

In the signal processing and feature extraction unit,

The detection waveform of the SKT sensor is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal using the average value and the standard deviation value.
The method of claim 1,

And the subset generator is configured to generate a subset of the biosignals based on the prediction accuracy of the sleep stage and the convenience of measuring the biosignals.
The method of claim 1,

The subset generator,

Biological signal-based automatic sleep stage classification system, characterized in that for generating a subset by the combination of EEG-based bio-signal consisting of at least one of EEG, EEG + EOG, EEG + EO + GSR, EEG + GSR combination.
The method of claim 8,

The EEG combination is a biological signal-based automatic sleep stage classification system, characterized in that the combination of a plurality of frequency band signals of the EEG signal.
The method of claim 9,

And a plurality of frequency band signals of the EEG signal are a combination of at least two or more of a delta wave, theta wave, an alpha wave, a sigma wave, and a beta wave.
The method of claim 1,

The subset generator,

A biosignal-based automatic sleep stage classification system, characterized in that it generates a subset consisting of a PPG-based biosignal combination consisting of at least one of PPG, PPG + GSR + SKT, PPG + GSR, PPG + SKT, and GSR + SKT combinations. .
The method of claim 1,

The subset generator,

A subset of EEG, EEG + EOG, EEG + EOG + GSR, EEG + GSR, PPG, PPG + GSR + SKT, PPG + GSR, PPG + SKT, GSR + SKT, GSR, SKT Bio signal based automatic sleep stage classification system.
The method of claim 1,

The second linear separator of the performance evaluation unit,

Quadratic linear classifier based on Bayes's rule, which selects the maximum posterior probability P (S y │x) value to sleep stage S y = 1,2,3 (S 1 : deep sleep, S 2 : shallow sleep, S 3 : REM) biosignal based automatic sleep stage classification system characterized in that the classification.
[Revision 17.12.2010 under Rule 91]
The method of claim 13, wherein the second-order linear separator of the performance evaluation unit, using Equation 1 as the formula for selecting the maximum posterior probability, Equation 1 is [Equation 1]
Biological signal-based automatic sleep stage classification system, characterized in that defined as.
The method of claim 1,

The performance evaluation unit,

A biosignal-based automatic sleep stage classification system, characterized by evaluating performance on an experimenter-dependent / independent dataset capable of predicting sleep stages using a subset of biosignal combinations having different conditions.
The method according to any one of claims 1 to 15,

A gene-fuzzy optimizer for determining a sleep stage through fuzzy inference from the biosignal combination; And

And a comparison unit comparing the discrimination value of the performance evaluation unit and the discrimination value of the gene-fuzzy optimization unit.
A biosignal detection unit which acquires various biosignals through different channels using a biosignal measuring sensor;

Signal processing and features for extracting feature points representing the characteristics of each biosignal from each biosignal by performing signal processing on the biosignals detected through each channel by a method suitable for feature extraction of each biosignal. Extraction unit;

A subset generator which generates one or more biosignal combinations based on feature points of each biosignal of the signal processing and feature extractor;

A performance evaluation unit for determining a sleep stage (REM, LS, DS) of the biosignal combinations combined by the subset generator through a second linear separator;

A gene-fuzzy optimizer for determining a sleep stage through fuzzy inference from the biosignal combination;

And a comparison unit for comparing the determination value of the performance evaluation unit with the determination value of the gene-fuzzy optimization unit.
The method of claim 17,

The biosignal measuring sensor includes at least one or more of an EEG, EOG, EMG, PPG, SKT, and GSR sensors.
The method of claim 18,

In the signal processing and feature extraction unit,

The detection waveform of the EEG and EOG sensors is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal through spectral analysis.
The method of claim 18,

In the signal processing and feature extraction unit,

The detection waveform of the PPG sensor is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal through the HRV analysis method.
The method of claim 18,

In the signal processing and feature extraction unit,

The detection waveform of the GSR sensor is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal through phasic analysis.
The method of claim 18,

In the signal processing and feature extraction unit,

The detection waveform of the SKT sensor is a biological signal-based automatic sleep stage classification system, characterized in that to extract the feature points of the biological signal using the average value and the standard deviation value.
The method of claim 17,

And the subset generator generates a subset of biosignals based on the prediction accuracy of the sleep stage and the convenience of measuring the biosignals.
The method of claim 17,

The subset generator,

Biological signal-based automatic sleep stage classification system, characterized in that for generating a subset by the combination of EEG-based bio-signal consisting of at least one of EEG, EEG + EOG, EEG + EO + GSR, EEG + GSR combination.
The method of claim 24,

The EEG combination is a biological signal-based automatic sleep stage classification system, characterized in that the combination of a plurality of frequency band signals of the EEG signal.
The method of claim 25,

And a plurality of frequency band signals of the EEG signal are a combination of at least two or more of a delta wave, theta wave, an alpha wave, a sigma wave, and a beta wave.
The method of claim 17,

The subset generator,

A biosignal-based automatic sleep stage classification system, characterized in that it generates a subset consisting of a PPG-based biosignal combination consisting of at least one of PPG, PPG + GSR + SKT, PPG + GSR, PPG + SKT, and GSR + SKT combinations. .
The method of claim 17,

The subset generator,

A subset of EEG, EEG + EOG, EEG + EOG + GSR, EEG + GSR, PPG, PPG + GSR + SKT, PPG + GSR, PPG + SKT, GSR + SKT, GSR, SKT Bio signal based automatic sleep stage classification system.
The method of claim 17,

The second linear separator of the performance evaluation unit,

Quadratic linear classifier based on Bayes's rule, which selects the maximum posterior probability P (S y │x) value to sleep stage S y = 1,2,3 (S 1 : deep sleep, S 2 : shallow sleep, S 3 : REM) biosignal based automatic sleep stage classification system characterized in that the classification.
[Revision 17.12.2010 under Rule 91]
30. The method of claim 29, wherein the second order linear separator of the performance evaluator uses Equation 1 as an equation for selecting the posterior probability maximum value, and Equation 1 is [Equation 1]
Biological signal-based automatic sleep stage classification system, characterized in that defined as.
The method of claim 17,

The performance evaluation unit,

A biosignal-based automatic sleep stage classification system, characterized by evaluating performance on an experimenter-dependent / independent dataset capable of predicting sleep stages using a subset of biosignal combinations having different conditions.