JP2014073237A - Sleep monitoring system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sleep monitoring system which increases estimation accuracy of a sleep state, using peak intervals of respiration.SOLUTION: A sleep monitoring system extracts respiratory information from biological information of an object person detected by biological information detection means, and estimates a sleep state from the respiratory information of the object person. The sleep monitoring system calculates peak intervals from peaks in respiration, interpolates time series data of the peak intervals to convert discrete data of the peak intervals into continuous data CD, resamples the continuous data CD in an arbitrary time period to acquire the peak intervals I, I, I, ... in the period, calculates feature quantities of the respiration on the basis of the peak intervals I, I, I, ..., and estimates a sleep state of the object person on the basis of the feature quantities of the respiration.

Description

  The present invention relates to a sleep monitoring system that estimates a sleep state from a subject's breathing.

  Various systems for estimating a sleep state from a subject's breathing during sleep have been proposed. For example, Patent Document 1 discloses an apparatus that extracts a respiratory signal from an output signal of a biological information sensor, calculates a peak and a peak interval from the respiratory signal, and estimates a sleep state based on the peak and the peak interval. . An exhalation band is applied as this biological information sensor, and an elastic breathing band with a strain gauge is wrapped around the chest and abdomen of the human body, and the voltage change according to the expansion and contraction of the strain gauge due to the breathing movement of the human body is measured. ing. In calculating the peak and peak interval, a positive threshold value and a negative threshold value are provided, and the maximum value from when the respiratory signal (voltage change data) exceeds the positive threshold value until reaching the negative threshold value is positive. It is calculated as a peak, and the time between the calculated peaks is calculated as a peak interval.

Japanese Patent No. 3733133 WO09 / 150744 pamphlet International Publication No. 09/150765 Pamphlet JP 2011-115188 A

  As described above, in order to detect the peak of respiration, a threshold is provided as a dead zone countermeasure for the output of a strain gauge or the like, and filtering is performed. When such a filtering process is performed, the measured voltage may not exceed the threshold if the expansion / contraction becomes smaller than expected due to the movement of the subject (turning over), the attachment position of the strain gauge, or the like. In such a case, since the peak cannot be extracted and the peak interval cannot be obtained, the estimation accuracy of the sleep state decreases. Therefore, when the sleep state is estimated using the breathing peak or peak interval, there is room for improvement in terms of estimation accuracy.

  Then, this invention makes it a subject to provide the sleep monitoring system which improved the estimation precision of the sleep state using the peak space | interval of respiration.

  The sleep monitoring system according to the present invention is a sleep monitoring system that extracts respiratory information from the biological information of the subject detected by the biological information detection means, and estimates the sleep state of the subject from the respiratory information. Peak interval calculation means for calculating the peak interval of respiratory peaks from information, and time-series data of peak intervals calculated by the peak interval calculation means are interpolated to convert discrete data of peak intervals into continuous data, and continuous data can be arbitrarily selected A peak interval acquisition means for re-sampling at a time interval and acquiring a peak interval at an arbitrary time, and a sleep state estimation for estimating a sleep state of the subject based on the peak interval at an arbitrary time acquired by the peak interval acquisition means Means.

  In this sleep monitoring system, the biological information of the subject is detected by the biological information detection means, and the respiratory information of the subject is extracted from the biological information. In the sleep monitoring system, the peak interval calculation means extracts the respiration peak from the respiration information of the subject and calculates the peak interval between successive peaks. Furthermore, in the sleep monitoring system, the time interval data (discrete data) of the peak interval is interpolated and converted into continuous data by the peak interval acquisition means, and the continuous data is resampled at an arbitrary time interval for an arbitrary time. Get the peak interval at. In this way, by converting the discrete data of the peak interval obtained by actual detection of the subject into continuous data, the peak interval at an arbitrary time is obtained from the continuous data even at an arbitrary time without a peak. be able to. Even if the peak or peak interval cannot be obtained by actual detection of the subject, the peak interval at an arbitrary time can be obtained from the continuous data. In the sleep monitoring system, the sleep state (sleep stage, sleep depth, etc.) of the subject is estimated based on the peak interval at an arbitrary time obtained by resampling from the continuous data by the sleep state estimation means. In this sleep monitoring system, discrete data of peak intervals are interpolated and converted into continuous data, and the continuous data is resampled to obtain a peak interval at an arbitrary time. The state estimation accuracy can be improved. In particular, since the peak interval at any time without a peak can be obtained, the time resolution of estimation can be improved. In the case where the peak interval is acquired by actually detecting from the biological information of the subject by the biological information detection means, it is necessary to have time to extract the peak at least twice. Since the peak interval in time can be obtained, the time resolution of estimation can be increased. Further, even when the respiration peak of the subject cannot be extracted from the subject's biometric information by the biometric information detection means (and thus the peak interval cannot be obtained), the peak interval can be obtained.

  The sleep monitoring system of the present invention includes a peak extraction unit that extracts a breathing peak at a timing when the inspiration of the subject changes from inspiration to expiration, and the peak interval calculation unit uses the peak extracted by the peak extraction unit. It is preferable to calculate the peak interval.

  In this sleep monitoring system, the peak is extracted by the peak extraction means at the timing when the inspiration obtained from the breath information of the subject changes to the expiration. For example, by extracting the place where the differential value of the breathing information (breathing waveform) becomes 0 or the inflection point, the timing at which the inspiration changes to the expiration can be extracted. In the sleep monitoring system, the peak interval between the extracted continuous peaks is calculated by the peak interval calculation means. In this sleep monitoring system, the peak of respiration can be detected with high accuracy by extracting the timing at which breathing information changes from inspiration to expiration as a peak.

  The sleep monitoring system of the present invention includes a feature amount calculating unit that calculates a respiratory feature amount using a peak interval at an arbitrary time acquired by a peak interval acquiring unit, and the sleep state estimating unit is a feature amount calculating unit. It is preferable to estimate the sleep state of the subject based on the calculated feature amount of respiration.

  In this sleep monitoring system, the feature quantity calculation means calculates the feature quantity of respiration from the peak interval at an arbitrary time acquired by resampling from the continuous data. This feature quantity is a feature quantity capable of estimating a sleep state, and includes, for example, a respiratory rate (average value, standard deviation, variation coefficient, etc.) and a breathing time (average value, standard deviation, variation coefficient, etc.). In the sleep monitoring system, the sleep state of the subject is estimated by the sleep state estimation means based on the feature amount of the breath. In this sleep monitoring system, the sleep state can be estimated with high accuracy using the respiration feature amount by deriving the respiration feature amount from the respiration peak interval.

  The sleep monitoring system according to the present invention includes a body motion detection unit that detects the body motion of the subject, and the sleep state estimation unit takes into account the detection result of the body motion in the body motion detection unit, and the sleep state of the subject Is preferably estimated.

  In this sleep monitoring system, the body motion of the subject is detected by the body motion detection means. When the subject has a body motion, the subject is awake or sleeps poorly, and in particular, the subject is clearly awake during a predetermined body motion (such as conversation). Therefore, in the sleep monitoring system, the sleep state of the subject is estimated by the sleep state estimation means in consideration of the detection result of the subject's body movement. In this sleep monitoring system, the sleep state can be estimated with high accuracy by estimating the sleep state in consideration of the body movement of the subject.

  The sleep monitoring system of the present invention may include a sleep quality estimation unit that estimates the sleep quality of the subject based on the time series data of the sleep state estimated by the sleep state estimation unit. This sleep monitoring system can estimate the sleep quality of the subject with high accuracy using the sleep state estimated with high accuracy.

  According to the present invention, the discrete data of the peak interval is interpolated and converted into continuous data, and the continuous data is resampled, whereby the peak interval at an arbitrary time can be obtained and the sleep state estimation accuracy is improved. Can be made.

It is a block diagram of the sleep monitoring system which concerns on this Embodiment. It is explanatory drawing of the phosphorus sampling method of a peak space | interval, (a) is an output signal (respiration signal) of a respiration sensor, (b) is a predetermined area | region of a respiration signal, (c) is axis-converted from a respiration signal. (D) is obtained by resampling the continuous data. It is explanatory drawing of a support vector machine. It is explanatory drawing of the extraction method of the peak of respiration. It is a classification table of body movement by a body movement filter. It is explanatory drawing of the selection method of a feature-value, (a) is a feature-value candidate (example in the case of eight), (b) is four feature-values extracted by the principal component analysis, (c ) Is the five feature quantities newly added, and (d) is the six feature quantities newly added. It is explanatory drawing of parameter optimization. It is an example of the result of feature quantity selection and parameter optimization, (a) is a graph of the correct answer rate with respect to the number of feature quantities, and (b) is a table showing the feature quantity number, correct answer rate, and optimum parameters. It is an example of the time change of a sleep stage. It is an example of the appearance probability of a δ wave in a standard sleep rhythm model. It is a flowchart which shows the flow of the sleep stage estimation process in the sleep stage estimation part 4 of FIG. It is a flowchart which shows the flow of the learning process in the sleep stage estimation part 4 of FIG.

  Hereinafter, embodiments of a sleep monitoring system according to the present invention will be described with reference to the drawings. In addition, the same code | symbol is attached | subjected about the element which is the same or it corresponds in each figure, and the overlapping description is abbreviate | omitted.

  The sleep monitoring system according to the present embodiment is a system that detects respiratory information of a user (subject) and estimates a sleep stage (sleep state) from the respiratory information of the user. The user is not particularly limited, and is, for example, a general person who sleeps in daily life or a driver of a vehicle such as a vehicle for the purpose of preventing a dozing. The sleep monitoring system according to the present embodiment estimates a sleep stage while learning a personal fitting for a user using a support vector machine (SVM [Support Vector Machine]) as a learning base model. The quality of sleep is also estimated using the estimated data. The sleep monitoring system according to the present embodiment incorporates a standard model learned in advance using data of a large number of subjects by SVM, and learns using the user's data after the start of user use. Personal fitting for the user from the model. In addition, although breathing is voluntary (conscious) at the time of awakening, it becomes involuntary (unconscious) when sleeping, so it changes drastically and has a good feature. Also, considering the robustness of measurement, since the S / N is good, the maintainability is also good.

  With reference to FIGS. 1-10, the sleep monitoring system 1 which concerns on this Embodiment is demonstrated. FIG. 1 is a configuration diagram of a sleep monitoring system according to the present embodiment. FIG. 2 is an explanatory diagram of a phosphorus sampling method for peak intervals. FIG. 3 is an explanatory diagram of the support vector machine. FIG. 4 is an explanatory diagram of a method for extracting a respiration peak. FIG. 5 is a classification table of body movements by the body movement filter. FIG. 6 is an explanatory diagram of a feature amount selection method. FIG. 7 is an explanatory diagram of parameter optimization. FIG. 8 is an example of the result of feature quantity selection and parameter optimization. FIG. 9 is an example of a time change in the sleep stage. FIG. 10 is an example of the appearance probability of a δ wave in a standard sleep rhythm model.

  The sleep monitoring system 1 calculates a respiration feature amount from the user's respiration signal (respiration waveform), and estimates a sleep stage (sleep depth) based on the respiration feature amount. As the feature amount, a feature amount obtained from at least the peak interval of respiration is used. In particular, the sleep monitoring system 1 interpolates the discrete data of the peak interval obtained by actual detection from the user and converts it into continuous data so that the peak interval can be obtained at all times. Sampling.

  The sleep monitoring system 1 includes a respiratory sensor 2, a biological information processing unit 3, a sleep stage estimation unit 4, a sleep quality estimation unit 5, and an output unit 6. The sleep monitoring system 1 having such a configuration may have any configuration as a system form. For example, the sleep monitoring system 1 may be a dedicated system for estimating a sleep stage, or input an output signal from a respiratory sensor and install an application program. By doing so, a system using a general-purpose computer (such as a personal computer, a tablet terminal, or a smartphone) may be used, or a system mounted on a vehicle or the like may be used.

  In the present embodiment, the respiration sensor 2 corresponds to the biological information detection means described in the claims, and the processing of the respiration sensor 2, the biological information processing unit 3, and the sleep stage estimation unit 4 is included in the claims. It corresponds to the body movement detection means described, and each process of the sleep stage estimation unit 4 is applied to the peak extraction means, the peak interval calculation means, the peak interval acquisition means, the feature amount calculation means, and the sleep state estimation means described in the claims. The processing of the sleep quality estimation unit 5 corresponds to the sleep quality estimation means described in the claims.

  The respiration sensor 2 can be a well-known respiration sensor, and is a sensor composed of a piezoelectric element, for example. The respiration sensor 2 is a sensor that is attached to the user's chest or / and abdominal clothing and is not in contact with the user's background. The respiration sensor 2 outputs a respiration waveform, which is an operation of the user's chest or / and abdomen, as a voltage change. FIG. 2A shows an example of the output signal of the respiration sensor 2.

  The biological information processing unit 3 performs preprocessing for extracting a respiratory signal from the output signal of the respiratory sensor 2. As preprocessing, for example, noise is removed by filtering, or an analog voltage signal is converted into a digital voltage signal. The preprocessed respiration signal (respiration waveform) is buffered for a predetermined time. In addition, when the output signal of the respiration sensor 2 can be used as it is as a respiration signal, the biometric information processing unit 3 may be omitted.

As shown in FIG. 2 (b), from the respiration signal (respiration waveform) BW, respiration peaks P ₁ and P ₂ (in this example, positive peaks), continuous peaks P ₁ and P _2, and Extracting the peak interval (respiration time for one cycle of breathing), the amplitude that is the difference (voltage difference) between the continuous positive peak and negative peak, the average value, etc. Can do. A feature value of respiration can be obtained from each of these values.

  The sleep stage estimation unit 4 is configured by a computer including a CPU [Central Processing Unit], a ROM [Read Only Memory], a RAM [Random Access Memory], and the like, and is realized by executing a sleep stage estimation program on the computer. The In the sleep stage estimation unit 4, a support vector machine estimator (hereinafter referred to as an SVM estimator) is constructed. The sleep stage estimation unit 4 extracts a peak from the buffered respiration signal for a predetermined time, interpolates discrete data at intervals of the peak to convert it into continuous data, resamples the continuous data, and performs arbitrary sampling Get peak interval in time. When the peak interval can be acquired by re-sampling from the continuous data (after a predetermined time has elapsed after activation), the sleep stage estimation unit 4 calculates the feature amount of respiration from the acquired peak interval and the respiratory signal itself, The user's body movement is detected from the respiration signal. Then, the sleep stage estimation unit 4 estimates the user's sleep stage (three stages of deep sleep, light sleep, and arousal) using the feature amount and body movement of the breath. The sleep stage estimation unit 4 uses the feature quantity of breathing for a predetermined period and the user's actual sleep stage (correct answer value) to select and use the feature quantity of personal fitting used for sleep stage estimation. Learns the optimal values of parameters.

  The sleep monitoring system 1 includes means for obtaining the user's actual sleep stage (correct answer value). For example, there is a means for the user himself / herself to input the state when the user woke up (such as when an alarm clock alarm sounded), and “a clean wake-up (that is, when waking up from a light sleep)”, “a clean-up” This is a button or the like for inputting “not awake (that is, when waking up from deep sleep)” or “was awake (ie, when awakening)”, and the correct value of the sleep stage is obtained from this input value. Moreover, there is a simple electroencephalogram sensor, and the correct value of the sleep stage is obtained from the user's electroencephalogram detected by the electroencephalogram sensor.

  A support vector machine (SVM) will be briefly described with reference to FIG. SVM is one of the learning machine identification methods proposed in the framework of statistical learning theory, and is a two-class pattern classifier. SVM is characterized by the ability to classify unknown data that has not been used for learning, and has high general-purpose capabilities. The SVM uses a mathematical technique called a kernel trick. The kernel includes “linear”, “polynomial”, “Gaussian”, and the like, and the sleep monitoring system 1 uses a Gaussian kernel. FIG. 3 shows a case where learning data of two feature quantities (two-dimensional) is used for easy understanding, and training points T1, T2, which are located in the vicinity of a class boundary called a support vector in the learning data. With reference to T3, the identification boundary B is set so that the distance L becomes the largest (margin maximum). The optimum values of parameters (C (soft margin cost), γ (Gauss kernel)) for defining the identification boundary are determined. The sleep stage estimation unit 4 determines identification boundary parameters that divide the two classes of deep sleep and others (shallow, awakening) and identification boundary parameters that separate the two classes of shallow sleep and wakefulness.

  With reference to FIG. 4, the extraction process of the peak from a respiration signal is demonstrated. The sleep stage estimation unit 4 detects the timing (the apex of the respiration waveform) when the respiration waveform BW changes from inspiration (voltage increasing portion) to inspiration (voltage decreasing portion), as in the right respiration waveform of FIG. The timing is extracted as a respiration peak. As a method of detecting the apex of the respiration waveform, a search is made for a place where the differentiation of the respiration signal becomes 0 or an inflection point. However, when the respiration waveform is disturbed, there are cases where two vertices are formed as in the respiration waveform on the left side of FIG. If such vertices are extracted as peaks, accurate peak intervals cannot be obtained.

Accordingly, the sleep stage estimation unit 4, each for detecting the vertex from the respiratory waveform, the time difference between the time t _i-1 of the vertices detected last from the time t _i of the current peaks detected (t i _{-t i-1} ) Is calculated, and it is determined whether or not the time difference (t _i −t _i−1 ) is greater than the threshold value TH. The threshold value TH is set in advance in consideration of an average respiration peak interval. When the time difference (t _i −t _i−1 ) is larger than the threshold value TH, the sleep stage estimation unit 4 extracts the vertex detected this time as a respiration peak. Further, in the sleep stage estimation unit 4, when the time difference (t _i -t _i-1 ) is equal to or smaller than the threshold value TH, an intermediate between the currently detected vertex time t _i and the previously detected vertex time t _i-1. The point (t _i -t _i-1 ) / 2 is extracted as a peak. In the case of the example in FIG. 4, (t ₂ −t ₁ ) is less than or equal to the threshold value TH for the two vertices of time t ₁ and time t ₂ , so the time of (t ₂ −t ₁ ) / 2 is the current time. (FIG. 4 also shows a peak at time t _n when it is assumed that the respiration waveform has not been disturbed). Since the time difference between the current peak time (t ₂ −t ₁ ) / 2 and the next detected vertex time t _{n + 1} is larger than the threshold value TH, the vertex detected at time t _{n + 1} is directly extracted as a peak. .

Processing for calculating the peak interval extracted from the respiratory signal will be described. In sleep stage estimation unit 4, for each combination of the extracted successive peaks, it calculates the i-th time difference between the time P _{i + 1} of the time P _i and the next peak of the (P i _{+ 1 -P i),} the time difference Let (P _{i + 1} −P _i ) be the peak interval. This peak interval is the breathing time for one cycle. Since the peak interval is a value obtained from two peaks, the time series data of the peak interval is discrete data.

The conversion process from discrete data of peak intervals to continuous data will be described. The sleep stage estimation unit 4 uses the time series data of the peak time and the peak interval to calculate the x coordinate in a two-dimensional orthogonal coordinate system (for example, the coordinate system of the xy plane including the x axis and the y axis). i-th and time _{P i} of the peak, generates each coordinate point of the y-coordinate and the time _{P i + 1} of the next (i + 1) -th peak _{(P i, P i + 1).} For example, as shown in FIG. 2C, the coordinate point (P ₁ , P ₂ ), the second peak time P ₁ is the x coordinate, and the second peak time P ₂ is the y coordinate. A coordinate point (P ₂ , P ₃ ) with the peak time P ₂ as the x coordinate and the third peak time P ₃ as the y coordinate. Thus, as shown in FIGS. 2B to 2C, the peak interval indicated by the value in the x-axis (horizontal axis) direction is indicated by the value in the y-axis (vertical axis) direction. The axis is converted. Further, interpolated between coordinate points on the two-dimensional orthogonal coordinate system (time-series discrete data indicating the peak interval) with splines, straight lines, etc., to obtain continuous data CD as shown in FIG. For example, discrete data are smoothly connected by cubic natural spline interpolation.

  As for the continuous data, the continuous data may be generated only once using a respiration signal buffered for a predetermined time after activation, or new data that is detected even after a predetermined time has elapsed after activation. The continuous data may be updated sequentially using the respiration signals. The predetermined time for obtaining this continuous data is a time at which a sufficient number of respiration peaks can be detected and high-accuracy continuous data (smooth interpolation is possible) can be obtained from discrete data at peak intervals. Is set.

A resampling process from continuous data will be described. As shown in FIG. 2D, the sleep stage estimation unit 4 resamples the continuous data CD at a constant interval, thereby obtaining peak intervals I ₁ , I ₂ , I ₃ ,... At an arbitrary time. calculate. The fixed time interval for resampling may be set to an arbitrary time. In this way, by converting the time-series data of actual peak intervals into axes and generating continuous data from discrete data, it is possible to obtain peak intervals of arbitrary time by resampling the continuous data. it can.

  The body movement detection process will be described with reference to FIG. The sleep stage estimation unit 4 detects (estimates) body movement by pattern analysis or the like with respect to a respiratory signal (respiration waveform). By using body movement, the accuracy of sleep stage estimation is improved. For example, it is prevented to estimate that sleep is deep even though there is a body motion indicating an awake state, or to estimate that it is awake even though there is a body motion indicating a rollover. The body motion detection method includes the period (corresponding to the peak interval), amplitude, autocorrelation (identity between the waveform obtained by shifting the respiratory waveform by an arbitrary time and the original respiratory waveform), and reproducibility (respiratory waveform). Fluctuations of the minimum value for each period of time) and the like are detected, and when each of these values exceeds a predetermined threshold value set in advance, it is detected as a user's body movement. For example, using a body motion filter as shown in FIG. 5, body motion is detected by combining the period, amplitude, autocorrelation and reproducibility of the respiratory waveform. When a person sits back, changes appear in the amplitude, autocorrelation and reproducibility of the respiratory waveform. When a person extends his hand upwards, changes appear in the amplitude and reproducibility of the respiratory waveform. When people talk, changes appear in the autocorrelation and reproducibility of the respiratory waveform. When a person takes a deep breath, changes appear in the period and amplitude of the respiration waveform. In addition to these, body movements such as rearrangement of leg or lack of extension can be detected by using a combination of the period, amplitude, autocorrelation and reproducibility of the respiratory waveform. Details of this body motion detection are disclosed in International Patent Application No. PCT / JP2010 / 0669847.

  As a method for detecting body movement, in addition to using breathing, another method may be used. For example, an acceleration sensor is provided on the environment side where the user sleeps (bet, driver's seat, etc.), and body movement is detected (estimated) by pattern analysis for the acceleration waveform detected by the acceleration sensor, or A user is imaged with a camera, and the captured image is analyzed to detect body movement.

  The feature amount calculation process will be described. The sleep stage estimation unit 4 calculates a feature amount of respiration using a peak interval obtained by resampling continuous data. Examples of the feature quantity obtained from the peak interval include the respiratory rate, the average value for the breathing time, the rate of change, the standard deviation, and the coefficient of variation. As this calculation method, for example, in the case of the average value of respiration rate, the average of the reciprocal of the peak interval (respiration time) of unit time (for example, 1 minute) is calculated. In the case of the average value of respiration time, the average of the peak interval of unit time is calculated. In the case of the change rate of the breathing time, the change rate (difference) of the average breathing time at an arbitrary time interval (for example, several tens of seconds to 60 seconds) is calculated. In the case of the standard deviation of the respiratory rate, the standard deviation of the reciprocal of the peak interval is calculated. In the case of standard deviation of breathing time, the standard deviation of peak interval is calculated. In the case of the coefficient of variation of the respiratory rate, the average value per unit time of the reciprocal of the peak interval is divided by the standard deviation in that interval. In the case of the coefficient of variation of the breathing time, the average value per unit time of the peak interval is divided by the standard deviation in that section. The sleep stage estimation unit 4 calculates a feature quantity of respiration without using a peak interval from a respiration signal (respiration waveform). Examples of the feature amount include an inspiratory amount, an expiratory amount, an average value for the expiratory amount / inspired amount, a standard deviation, and a change coefficient. The feature quantity of respiration mentioned here is an example, and other feature quantities may be used.

  The sleep stage estimation process will be described. The sleep stage estimation unit 4 (SVM estimator) identifies the sleep stage from the respiratory feature amount. In particular, for the discrimination of shallow / wakefulness, the detection is also performed in consideration of the detection result of body movement. Specifically, first, the sleep stage estimation unit 4 uses the optimum value of the parameter for identifying deep / others (the identification boundary is defined by the optimum value of this parameter (C, γ)), / Identifies whether the value for each feature quantity selected to identify others (the feature quantity value calculated in the feature quantity calculation process above) is in the deep sleep class or other class To do. When it is identified as a deep sleep class, the sleep stage estimation unit 4 estimates that this sleep stage is a deep sleep state. When discriminating from other classes, the sleep stage estimation unit 4 estimates the current sleep stage as an arousal state in the case of body movement indicating that it is occurring from the detection result of the body movement. In the case of other body motions or when no body motion is detected, the sleep stage estimation unit 4 is selected to identify shallow / wakefulness using the optimum value of the parameter for identifying shallow / wakefulness. It is identified whether the value for each feature value falls into the shallow sleep class or the awakening class. If the sleep class is identified as a shallow sleep class, the sleep stage estimation unit 4 estimates the current sleep stage as a shallow state. When it is identified as the awakening class, the sleep stage estimation unit 4 estimates the current sleep stage as the awakening state.

  Note that the optimum value of the parameter is initially the optimum value of the parameter of the standard model, but when updated by the following learning process, the optimum value of the parameter for personal fitting for the user is used. The number of feature quantities selected for use in identification and the type of each feature quantity are initially the number of feature quantities and their respective feature quantities used in the standard model. The number of individual fitting features and their respective features are used.

  The learning process will be described. The sleep stage estimation unit 4 performs feature selection and parameter optimization using an SVM estimator for personal fitting to the user. The selection of the feature amount and the optimization of the parameters are performed respectively for identifying deep sleep and others and for identifying shallow sleep and awakening.

  Here, as the learning data and the evaluation data, all feature amounts obtained from the respiration waveform detected from the user (particularly, all feature amounts for a predetermined time (eg, 60 minutes, 90 minutes) before waking up) are obtained. The correct value of the sleep stage input by the user with the value and input means is used. Learning is performed when a predetermined amount (such as one week or one month) of such data can be collected. Further, the feature quantity candidates to be selected are selected from the feature quantities shown above. An SVM estimator is created by learning using the learning data / evaluation data and the feature quantity of the feature quantity candidate.

  A feature amount selection method will be described with reference to FIG. In this selection method, several feature quantities are first extracted from the feature quantity candidates using cross validation, and the remaining feature quantity candidates are analyzed and added one by one. In advance, feature quantity candidates are prepared from the respiratory feature quantities. In the example of FIG. 6, the number of feature amount candidates is 8, and feature amount candidates include feature amount A to feature amount H.

  First, four feature quantities are first extracted (selected) by principal component analysis. Start learning from 4D features. You may start from a dimension other than the 4th dimension. In the example of FIG. 6, as shown in FIG. 6B, first, a feature quantity A, a feature quantity B, a feature quantity C, and a feature quantity D are extracted. Then, feature quantities that can be used for estimation are selected one by one from the remaining feature quantity candidates.

  Each time a new feature value is selected (when the first four feature values are extracted or one feature value is added one by one), the SVM estimator uses the remaining feature value candidates one by one. To extract. The SVM estimator uses the learning data and the evaluation data to estimate the sleep stage using each of the selected feature quantities and each extracted feature quantity, and the estimated sleep stage and sleep The correct answer value of the stage is compared to calculate the correct answer rate. At this time, in the parameter optimization processing described below, the above processing is performed while sequentially changing the parameters (C, γ) of the identification boundary, and the parameter (optimum value) with the highest correct answer rate is selected. The SVM estimator performs the above processing while sequentially replacing the remaining feature amount candidates. Then, the SVM estimator selects (adds) a feature quantity candidate having the highest correct answer rate among the remaining feature quantity candidates as a new feature quantity. In this way, feature quantities are sequentially added to increase the dimension of the feature quantity. In the case of the example in FIG. 6, the feature amount F is newly added in FIG. 6C to make the feature amount five-dimensional, and in FIG. 6D, the feature amount E is newly added to make the feature amount six-dimensional. Become.

  Each time a feature amount is added, the SVM estimator determines whether the correct answer rate is not saturated or decreased. If the correct answer rate is saturated or decreased when the feature amount is added, the SVM estimator stops adding the feature amount and determines the selected feature amount.

A parameter optimization method will be described with reference to FIG. Each time feature values are extracted one by one from the remaining feature value candidates in the above processing, the SVM estimator sequentially changes the identification boundary parameters (C, γ), and learning data based on the changed parameters. And using the evaluation data to estimate the sleep stage using each extracted feature value and each extracted feature value, and compare the estimated sleep stage and the correct value of the sleep stage, Calculate the correct answer rate. In the SVM estimator, the C value and γ value of the parameter when the correct answer rate is the highest are defined as the discriminating ability of the selected feature quantity and the extracted one feature quantity, and the parameter Select the optimum value. Incidentally, when changing the parameters, ² -5 range of ^C, 2 -4, · · ^·, and ^{2 15,} the range of the gamma ^{2 -15,} 2 -14, · · ·, and ^{2 3.}

  The example shown in FIG. 7 is a case of prior learning, and there are samples (approximate population) collected from many subjects. For example, 10 sample sets (100) are prepared from the specimens, 9 of which are learned, and the calculation of the correct answer rate is repeated 10 times while changing the parameters of the remaining 1 set. Then, the parameter having the highest average value of correct answers is set as the optimum value. In the case of personal fitting for the user in the sleep stage estimation unit 4, learning is performed by the SVM estimator using a predetermined amount of data collected from the user as one week or one month as learning data or evaluation data. To do.

  In addition, when generating a standard model by prior learning, the respiratory waveform is detected by a respiratory sensor during sleep from a large number of subjects (male / female, each age group), and respiratory feature values are obtained from the respiratory waveform of each subject. In addition, an electroencephalogram waveform is detected by a high-precision electroencephalogram sensor during sleep, and the correct answer value of the sleep stage (three stages of deep sleep, shallow sleep, and awakening) is acquired from the electroencephalogram waveform of each subject. Then, a sample for a large number of subjects as shown in FIG. 7 is generated from these data, learning data and evaluation data are prepared from the sample, and learning processing similar to the above is performed using the learning data and evaluation data. I do. Then, the feature quantity (number of feature quantities and each feature quantity) selected to identify deep / other and shallow / wakefulness as the standard model and the selected feature quantities are used. Obtain the optimal values of the parameters (C, γ) for defining the discrimination boundary for discriminating deep / other and shallow / wakefulness.

  FIG. 8 shows the result of selection of feature amounts and parameter optimization in advance learning. In the example of FIG. 8, there are 16 feature quantity candidates, and feature quantities are sequentially added from the 4 feature quantities extracted by the principal component analysis. As can be seen from FIGS. 8A and 8B, the correct answer rate is increased or maintained until the number of feature quantities is 13, but when the 14th feature quantity is added, the correct answer rate decreases. Accordingly, 13 feature values are finally selected, and the optimum value of the parameter in the case of the 13-dimensional feature value is C value 1 and γ value 0.5.

  The sleep quality estimation unit 5 is configured by a computer including a CPU, a ROM, a RAM, and the like, and is realized by executing a sleep quality estimation program on the computer. The sleep quality estimation unit 5 uses the time series data of the sleep stages for a predetermined interval (for example, one day, one week, one month) estimated by the sleep stage estimation unit 4 to determine the sleep quality of the user. Estimate quality.

  The sleep quality estimation process will be described with reference to FIGS. 9 and 10. In this estimation, the quality of sleep is estimated from the residence time when sleep is deep in the sleep stages estimated during sleep. That is, based on the dwell time of deep sleep (deep sleep), the content rate of δ waves is obtained probabilistically from the sleep rhythm model, and the quality of sleep is estimated. It has been found that the δ wave, which greatly affects the quality of sleep, has a high correlation with the deep sleep.

Assume that the time and depth of sleep observed mainly in overnight sleep have a shape as shown in FIG. The deep sleep residence time appearing in each sleep cycle C1, C2,. In FIG. 10, it is assumed that the appearance probability of the δ wave fluctuates in the time of staying in the deep sleep stage because the Ultradian rhythm, which is the human ecologicalism in each cycle, is one cycle in about 90 minutes. That is, the sleep quality Q can be described by the model shown in the following equation (1). P in Formula (1) is an arbitrary constant (gain).

  The sleep quality estimation unit 5 uses the time series data of the sleep stage estimated by the sleep stage estimation unit 4 to extract the stay time of the deep sleep stage in each sleep cycle, and the δ wave appearance rate t1, t2,.・ ・ Calculate each. And the sleep quality estimation part 5 calculates the sleep quality Q by Formula (1) using the (delta) wave appearance rate t1, t2, .... The details of the sleep quality estimation are disclosed in International Patent Application No. PCT / JP2012 / 054661.

  The output unit 6 outputs a sleep stage estimation result and a sleep quality estimation result. As the output unit 6, for example, a display for displaying the estimation result and a printer for printing out the estimation result are used.

  Operation | movement of the sleep monitoring system 1 of the said structure is demonstrated. In particular, the sleep stage estimation process in the sleep stage estimation unit 4 will be described along the flowchart of FIG. 11, and the learning process will be described along the flowchart of FIG. 12. FIG. 11 is a flowchart showing the flow of the sleep stage estimation process in the sleep stage estimation unit 4. FIG. 12 is a flowchart showing the flow of the learning process in the sleep stage estimation unit 4.

  The sleep monitoring system 1 is activated by the user. The respiration sensor 2 detects a user's respiration and outputs a change in respiration as a voltage change (output signal). The biological information processing unit 3 pre-processes the output signal from the respiration sensor 2 and extracts a respiration signal that can be processed by post-processing.

  The sleep stage estimation unit 4 extracts a respiration peak (particularly, peak time) from a respiration signal (respiration waveform) for a predetermined time. And the sleep stage estimation part 4 calculates a peak space | interval using the time of each peak for every extracted continuous peaks. Furthermore, the sleep stage estimation unit 4 performs axis conversion on the peak interval using the time series data of the time and peak interval of each peak, and interpolates the discrete data of the converted peak interval to obtain continuous data.

  The sleep stage estimation unit 4 resamples the continuous data at a predetermined time interval, and acquires a peak interval at an arbitrary time. The sleep stage estimation unit 4 calculates a plurality of respiratory feature quantities using the acquired time-series data of peak intervals (S10). Further, the sleep stage estimation unit 4 calculates a plurality of other respiratory feature quantities (feature quantities not using peak intervals) from the respiratory signal (respiration waveform) (S10). The sleep stage estimation unit 4 detects the user's body movement by the body movement filter using the feature quantity of respiration (S11). Note that the calculated feature amount data is accumulated because it is used for learning. Since learning also uses feature amount data other than the feature amount used for estimation, all feature amounts of the feature amount candidates are calculated.

  The sleep stage estimation unit 4 uses the feature value selected for identification to identify whether the sleep is deep or not based on the identification boundary based on the optimum value of the parameter for identifying deep / others ( S12). When it is identified that the sleep is deep in S12, the sleep stage estimation unit 4 estimates the current sleep stage of the user as a deep state (S14). When it is identified that the sleep is not deep in S12, the sleep stage estimation unit 4 identifies whether or not the user is awake based on the detection result of the body movement (S13). When the awakening is identified from the body movement, the sleep stage estimation unit 4 estimates the current sleep stage of the user as the awakening state (S16). When it is not possible to discriminate from wakefulness from body movement, the sleep stage estimation unit 4 uses the value of each feature quantity selected for discrimination to sleep based on the discrimination boundary based on the optimum value of the parameter for discriminating shallow / wakefulness. Is shallow or awake (S13). When it is identified that the sleep is shallow in S13, the sleep stage estimation unit 4 estimates the current sleep stage of the user as a shallow state (S15). When it is identified as awakening in S13, the sleep stage estimation unit 4 estimates the current sleep stage of the user as the awakening state (S16).

  The sleep stage estimation unit 4 repeats the acquisition of the peak interval, the calculation of the feature amount, the detection of the body movement, and the estimation of the sleep stage by the re-sampling at regular intervals during the user's sleep. The output unit 6 outputs sleep stage data estimated during the sleep. When the user wakes up, the state when the user wakes up (the correct value of the sleep stage) is input. Since the correct answer value of this sleep stage is used for learning, it is stored in association with the data of each feature amount described above. The estimated sleep stage data is accumulated because it is used for sleep quality estimation.

  When the above sleep stage data is accumulated for a predetermined period, the sleep quality estimation unit 5 extracts the stay time of the deep sleep stage in each sleep cycle using the sleep stage data, and the δ wave appears. Each rate is calculated. Then, the sleep quality estimation unit 5 calculates the sleep quality Q according to the equation (1) using the δ wave appearance rate. The output unit 6 outputs the estimated sleep quality.

  Further, when the above-described feature amount and correct value data of the sleep stage are accumulated for a predetermined period, the sleep stage estimation unit 4 extracts learning data and evaluation data from the accumulated data (S20). The sleep stage estimation unit 4 extracts four feature amounts from the feature amount candidates by principal component analysis (S21).

  Each time four feature values are extracted or one feature value is added in S21, the sleep stage estimation unit 4 uses the currently selected feature value and the SVM based on any feature value component in the remaining feature value candidates. An estimator is created (S22). Each time one feature amount is extracted from the remaining feature amount candidates, the sleep stage estimation unit 4 sequentially changes the parameters, and extracts the plurality of selected feature amounts using learning data and evaluation data. Estimate the sleep stage using each value of each feature quantity, compare the estimated sleep stage and the correct answer value of the sleep stage, calculate the correct answer rate for each parameter, the parameter when the correct answer rate is the highest (Optimum value) is selected (S23). And the sleep stage estimation part 4 compares and evaluates the correct answer rate of each feature-value of the remaining feature-value candidates, and selects the feature-value with the highest correct answer rate (S24). And the sleep stage estimation part 4 adds the feature-value with the highest correct answer rate newly (S25).

  The sleep stage estimation unit 4 compares the correct answer rate of the newly added feature amount with the previous correct answer rate, and evaluates whether or not a new feature amount may be added (S26). When it is evaluated that it may be added in S26, the sleep stage estimation unit 4 returns to the process of S22. When it is evaluated that it is not necessary to add in S26, the sleep stage estimation unit 4 uses the feature amount added in the previous process as the feature amount of the selection target, and the parameter (( (Optimum value) as a learning result. The sleep stage estimation unit 4 performs estimation using the feature quantity and parameter (optimum value) of the selection target obtained by this learning from the next sleep stage estimation. This learning is performed for sleepy / other identification and for sleep / awake identification.

  According to this sleep monitoring system 1, by converting discrete data of actually detected peak intervals into continuous data and re-sampling the continuous data, a peak interval at an arbitrary time can be obtained, The estimation accuracy of the sleep stage can be improved by using the peak interval that is always obtained. That is, since a situation in which the peak interval cannot be acquired can be prevented, the sleep stage can be estimated using the peak interval without leakage, and the estimation accuracy is improved. In particular, it does not require time to extract the actual breathing peak at least twice in order to obtain the peak interval, and the peak interval of any time without a peak can be obtained, and the respiration characteristics can be obtained from the peak interval. The quantity can be derived and the time resolution of estimation can be improved. By the way, when the sleep stage is estimated by taking into account other information such as heart rate, even if the interval for obtaining other information is shorter than the peak interval for breathing, the peak interval is adjusted to the short interval for obtaining other information. And the sleep stage can be estimated. Even when the user's respiration peak cannot be extracted by actual detection by the respiration sensor 2, the peak interval can be obtained.

  Moreover, according to the sleep monitoring system 1, the peak of respiration can be detected with high accuracy by extracting the timing at which the breathing waveform changes from inspiration to expiration as a peak. Furthermore, according to the sleep monitoring system 1, even when the respiratory waveform is disturbed (for example, when the apex of the respiration waveform appears at a very short interval), the peak of respiration is evaluated by evaluating the time interval between the apexes. Can be detected with high accuracy.

  Moreover, according to the sleep monitoring system 1, the sleep stage can be estimated with high accuracy using the respiration feature amount by deriving the respiration feature amount from the respiration peak interval. Moreover, 1 can estimate a sleep stage with high precision by 1 in this sleep monitoring system by estimating a sleep stage in consideration of a user's body movement. Moreover, according to the sleep monitoring system 1, the quality of a user's sleep can be estimated with high precision using the sleep stage estimated with high precision.

  Further, according to the sleep monitoring system 1, by applying the SVM estimator, the sleep stage can be identified with high accuracy, and the feature amount selection for identification and the optimization of the parameter are learned. Can do. In particular, learning can be performed using unknown data (data acquired from the user) that has not been used for prior learning, and it is possible to select a feature amount for personal fitting and optimize parameters for the user.

  As mentioned above, although embodiment which concerns on this invention was described, this invention is implemented in various forms, without being limited to the said embodiment.

  For example, in the present embodiment, a respiration sensor is applied as the biological information detection means, but a vital sensor other than the respiration sensor may be applied and a respiration signal may be extracted from the output signal of the vital sensor.

  In this embodiment, the body motion is detected using a body motion filter, and the sleep stage is estimated by taking the body motion into account. However, the sleep stage may be estimated without taking the body motion into account. . In the present embodiment, the sleep quality is also estimated. However, the sleep quality may not be estimated. Further, in the present embodiment, the learning is also performed. However, the learning may not be performed.

  In this embodiment, the sleep stage estimation and learning are performed using the learning-based model SVM. However, the sleep stage estimation and learning may be performed using other methods such as a conditional probability model. Good. Moreover, although an example of the sleep quality estimation method has been described in the present embodiment, the sleep quality may be estimated using other methods.

  In this embodiment, the sleep stage is estimated from the respiratory information of the subject. However, the sleep stage may be estimated in consideration of other biological information (for example, heartbeat) in addition to the respiratory information. .

  Further, in this embodiment, the sleep stage is configured to estimate three stages of deep, shallow, and awakening, but the sleep stage (sleep depth) may be four or more stages, or two stages of awakening / sleeping. . For example, use the international standard of Rechtschaffen & Kales sleep stage classification.

  In the present embodiment, an example of the respiration feature value obtained from the peak interval and the respiration feature value not using the peak interval has been shown, but various other feature values can be applied to the respiration feature value. In the present embodiment, the feature amount is calculated from the peak interval and the sleep stage is estimated based on the feature amount. However, the sleep step is directly estimated from the peak interval and the like without obtaining the feature amount. Also good.

  DESCRIPTION OF SYMBOLS 1 ... Sleep monitoring system, 2 ... Respiration sensor, 3 ... Biological information processing part, 4 ... Sleep stage estimation part, 5 ... Sleep quality estimation part, 6 ... Output part.

Claims (5)

A sleep monitoring system that extracts respiratory information from the biological information of the subject detected by the biological information detection means and estimates the sleep state of the subject from the respiratory information,
A peak interval calculating means for calculating a peak interval of a respiration peak from the respiration information of the subject;
The peak interval time-series data calculated by the peak interval calculation means is interpolated to convert the peak interval discrete data into continuous data, and the continuous data is resampled at an arbitrary time interval to obtain a peak interval at an arbitrary time. A peak interval acquisition means for acquiring;
Sleep state estimation means for estimating the sleep state of the subject based on the peak interval at an arbitrary time acquired by the peak interval acquisition means;
A sleep monitoring system comprising: Peak extraction means for extracting a breathing peak at a timing when the inspiration of the subject changes from inspiration to expiration,
The sleep monitoring system according to claim 1, wherein the peak interval calculation unit calculates a peak interval using the peak extracted by the peak extraction unit. A feature amount calculating means for calculating a respiration feature amount using a peak interval at an arbitrary time acquired by the peak interval acquiring means;
The sleep monitoring system according to claim 1, wherein the sleep state estimation unit estimates a sleep state of the subject based on a feature amount of respiration calculated by the feature amount calculation unit. A body motion detecting means for detecting the body motion of the subject person,
The said sleep state estimation means estimates a subject's sleep state in consideration of the detection result of the body movement in the said body movement detection means, The Claim 1 characterized by the above-mentioned. Sleep monitoring system.   5. The apparatus according to claim 1, further comprising a sleep quality estimation unit configured to estimate a sleep quality of the subject based on the time series data of the sleep state estimated by the sleep state estimation unit. The sleep monitoring system described.

Images