JP6559095B2 - Sleep model creation device, sleep stage estimation device, method and program - Google Patents

Description

  The present invention relates to a sleep model creation apparatus, a sleep stage estimation apparatus, a method, and a program for estimating a person's sleep stage.

  The sleep stage is an index representing the depth of sleep, and is defined in 2 to 6 stages according to international standards. The sleep progress diagram showing the sleep stages in time series is important information in grasping the sleep state in the medical and healthcare fields. The sleep stage is determined every time interval (about 30 seconds) based on the R & K method, which is an inspection determination rule, by a professional engineer using electroencephalogram data or electromyogram data obtained by a polygraph examination. However, since this polygraph test has a large physical load on the user, in recent years, research on a sleep stage estimation technique using heart rate data that can be acquired by a measurement device with a small physical load has been actively conducted.

  The conventional sleep stage estimation method using the heart rate data, for example, generates a plurality of feature quantities related to the sleep stage from the heart rate data, and estimates using machine learning (support vector machine, random forest, etc.). (See Non-Patent Document 1, for example).

  In addition, a technique has been proposed in which the sleep stage estimation accuracy is improved in consideration of the property that the sleep stage distribution and transition probability change depending on the elapsed time from bedtime (for example, non-patented). Reference 2).

Xiao, M et al., “Sleep stages classification based on heart rate variability and random forest” Biomedical Signal Processing and Control, p624-633, (2013). “Time-dependent sleep stage transition model using heart rate variability” DEIM Forum 2015 G6-5, http://db-event.jpn.org/deim2015/paper/24.pdf (confirmed on October 22, 2015).

  However, the autonomic nervous activity reflected in the heartbeat changes due to influences other than brain activity during sleep, such as blood pressure adjustment, body temperature adjustment, body movement, and breathing. For this reason, with the technique described in Non-Patent Document 1, it is difficult to estimate the sleep stage obtained from the brain activity in complete agreement with the heartbeat.

  Moreover, in the technique described in the said nonpatent literature 2, there existed a malfunction that the estimation precision at the time of the REM sleep state and awakening state in the intermediate part of sleep was low. As one of the reasons, the technique described in Non-Patent Document 2 considers the property that the sleep stage distribution and the transition probability change depending on the elapsed time from bedtime.

  However, as shown in FIG. 19A, the sleep pattern of the subject 1 and FIG. 19B, the sleep pattern of the subject 2, the sleep stage distribution may vary greatly depending on the person even in the same time zone C1. As it is, it becomes a factor which reduces the estimation precision at the time of the REM sleep state and awakening state in the intermediate part of sleep.

  The present invention has been made paying attention to the above circumstances. The purpose of the present invention is to provide a sleep model creation device capable of estimating the sleep stage more accurately even when the heartbeat is strongly influenced by other than brain activity. An object of the present invention is to provide a sleep stage estimation device, method and program.

  In order to achieve the above object, according to one aspect of the present invention, a heart rate variability feature amount calculating unit that calculates a heart rate variability feature amount for each first unit time from a set of sleep electrocardiograms, and the heart rate variability feature amount calculating unit calculates A heart rate variability storage unit for storing heart rate variability information obtained by aggregating the obtained heart rate variability feature values for each sleep stage, and a heart rate change for calculating a heart rate change value feature amount from a deviation of the heart rate variability feature amount calculated by the heart rate variability feature amount calculation unit Value feature quantity calculation unit, heart rate change value feature quantity storage unit for storing heart rate change value feature quantity information obtained by totaling heart rate change value feature quantities calculated by the heart rate change value feature quantity calculation unit for each sleep stage transition, and the sleep A model determination unit that determines a model based on the heartbeat change value feature amount for each time period obtained by dividing the time axis of the electrocardiogram by a second unit time longer than the first unit time, and the divided times Every obi, and A time / model-specific sleep stage appearance frequency analysis unit for analyzing the appearance frequency of the sleep stage for each first unit time for each model determined by the model determination unit, and the time / model-specific sleep stage appearance frequency An appearance frequency storage unit that stores appearance frequency information for each time and model analyzed by the analysis unit, and for each time period that analyzes the frequency of sleep stage transitions for each of the first unit times for each divided time zone A sleep stage transition frequency analysis unit; and a transition frequency storage unit that stores transition frequency information analyzed by the hourly sleep stage transition frequency analysis unit.

  According to the present invention, it is possible to provide a sleep model creation device, a sleep stage estimation device, a method, and a program capable of estimating the sleep stage more accurately even when the heartbeat is strongly influenced by other than brain activity. .

The figure which illustrates the change of the elapsed time from bedtime, a heart rate deviation, and a sleep stage. The figure which showed an example of the appearance frequency of a sleep stage according to the time slot | zone. The figure which shows an example of the appearance probability of a sleep stage transition pattern. The block diagram which shows the function structure of the sleep stage estimation apparatus which concerns on 1st Embodiment. The figure which illustrates the contents of the heart rate change value feature-value table concerning the embodiment. The figure which illustrates the contents of the appearance frequency table concerning the embodiment. The figure which illustrates the contents of the heart rate variability table concerning the embodiment. The figure which illustrates the contents of the transition frequency table concerning the embodiment. The figure which illustrates the contents of the acceleration table concerning the embodiment. The flowchart which shows the process sequence and process content in the sleep stage estimation apparatus in the 1st learning phase which concerns on the embodiment. The figure which shows an example of the observation method of the time series heart rate data which concerns on the embodiment, and the output method of a sleep stage estimated value. The figure which shows an example of the Lorentz plot which concerns on the same embodiment. The figure which displayed the list of the kind of heart rate variability feature-value based on the embodiment. The flowchart which shows the process sequence and process content in the sleep stage estimation apparatus in the 2nd learning phase which concerns on the embodiment. The figure which shows the content of the determination process of the hourly model which concerns on the same embodiment. The flowchart which shows the content of the determination process of the hourly model which concerns on the embodiment. The flowchart which shows the content of the determination process of the hourly model which concerns on the embodiment. The flowchart which shows the process sequence and process content in the estimation phase which concerns on the embodiment. The figure which shows the elapsed time from bedtime and the distribution of a sleep stage which change with test subjects.

Embodiments according to the present invention will be described below with reference to the drawings.
(principle)
The present invention focuses on the fact that the appearance frequency and transition probability of the sleep stage change according to the elapsed time after falling asleep, so that even if the heartbeat is strongly influenced by other than brain activity, a more likely sleep stage can be estimated. It is a thing. For example, although there is a low possibility of sudden awakening after falling asleep and falling into a deep sleep again, the state may be estimated with a probability of 0.52 when estimated only with the heartbeat. In the present invention, the estimation accuracy is improved by lowering the estimation probability for such a rare transition or when the estimation confidence level is low.

  In addition, in the present invention, as shown in FIG. 1, focusing on the fact that the heart rate changes greatly at the transition of the sleep stage, the estimation accuracy is further increased by considering the probability value of the heart rate change value feature value at the transition of the sleep stage. Is to improve. FIG. 1 exemplifies the elapsed time from bedtime, heart rate deviation, and change in sleep stage. As shown by the dotted circle in the figure, the heart rate changes greatly at the transition of the sleep stage. I understand.

  Furthermore, since the appearance frequency of the sleep stage according to the elapsed time after falling asleep has individual differences as shown in FIG. 19, the transition of the sleep stage is preliminarily made for each time zone using the heart rate change value feature amount. It is divided into a group that is likely to have more and a group that is likely to be less, and by generating and using the probability value of the appearance frequency of the sleep stage within each group, individual differences are absorbed and estimation accuracy is improved.

  FIG. 2 shows an example of the result of calculating the number of appearances of each sleep stage every 30 minutes using the sleep stage data for 45 persons. As is clear from this example, the number of appearances of REM (REM sleep) in 30 minutes from the start of sleep (FIG. 2 (a)) is 180 minutes and 360 minutes from sleep as shown in FIGS. 2 (b) and 2 (c), respectively. Less than the number of appearances of REM in 30 minutes after elapse.

  FIG. 3 also shows an example of the result of calculating the appearance probability of each sleep stage transition pattern every 30 minutes using the sleep stage data for 45 people. As shown in the figure, the transition probability from REM to NREM (non-REM sleep) in 30 minutes after the onset of sleep (FIG. 3 (a)) is 150 minutes from the sleep shown in FIGS. 3 (b) and 3 (c), respectively. It turns out that it becomes higher than the transition probability from REM to NREM in 30 minutes after and after 300 minutes.

  Therefore, in the present invention, the appearance frequency and transition probability of the sleep stage according to the elapsed time are incorporated into the sleep stage generation model by the learning phase, and the feature quantity of the user's heartbeat data is obtained in the estimation phase, and the feature quantity of the heartbeat data is obtained. The sleep stage is estimated based on the sleep stage generation model. This makes it possible to estimate the sleep stage more accurately even when the heartbeat is strongly influenced by other than brain activity.

  In addition, in the learning phase of the appearance frequency of the sleep stage according to the elapsed time, the generation model to be incorporated is changed depending on whether the heart rate change value feature amount is large or small for each elapsed time, and in the estimation phase, every elapsed time The generation model to be used is selected from the change value of the heart rate data of the user, and the sleep stage is estimated. By doing in this way, it becomes possible to absorb the individual difference of the appearance frequency of a sleep stage, and to improve a precision more.

  Furthermore, paying attention to the fact that the heartbeat changes greatly at the transition of the sleep stage, and incorporating the probability value of the heartbeat change value feature quantity at the transition of the sleep stage into the model, the accuracy can be further improved.

[First Embodiment]
<Configuration>
FIG. 4 is a block diagram showing a functional configuration of the sleep stage estimation apparatus according to the embodiment of the present invention.
The sleep stage estimation apparatus 1 according to the present embodiment includes, for example, a server computer or a personal computer, and includes a control unit 10, a storage unit 20, and an input / output interface unit 30.

  Among these, the input / output interface unit 30 captures the electrocardiogram data output from the electrocardiograph 2 and the acceleration data output from the triaxial accelerometer 5 attached to the chest, and also inputs an input unit such as a keyboard or a touch panel. 3 and the display unit 4 such as a liquid crystal device are used to input / output operation data and display data.

  The storage unit 20 uses a non-volatile memory capable of writing and reading, such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive), as a storage medium, and a storage unit necessary for implementing this embodiment As shown, a model database 21 and an estimated data storage unit 22 are provided.

  The model database 21 includes a heart rate variability table 211, a heart rate change value feature amount table 212, an appearance frequency table 213, a transition frequency table 214, and an acceleration table 215.

  As shown in FIG. 7, the heart rate variability table 211 is used to store the heart rate feature quantity level and the appearance frequency in association with the heart rate feature vector identification information (heart rate feature vector ID) for each sleep stage.

  As shown in FIG. 5, the heart rate change value feature amount table 212 is associated with the heart rate change value feature vector identification information (heart rate change value feature vector ID) representing the heart rate change for each sleep stage transition pattern. Used to store levels and their frequency of appearance.

  As shown in FIG. 6, the appearance frequency table 213 is used to store, for each model, sleep stages and appearance frequencies that appear in each time zone.

  As shown in FIG. 8, the transition frequency table 214 is used to store sleep stage transition patterns that occur in each time zone and their appearance frequencies.

As shown in FIG. 9, the acceleration table 215 is used to store the acceleration feature quantity level and the appearance frequency in association with the identification information (acceleration feature vector ID) of the acceleration feature vector for each sleep stage.
The estimated data storage unit 22 is used to store data representing the sleep stage estimation result.

  The control unit 10 includes a central processing unit (CPU) and a memory. As a control processing function necessary for carrying out the present embodiment, a heart rate variability feature amount calculation unit 11, a heart rate change value, and the like. Feature amount calculation unit 12, hourly model determination units 13 and 14, model creation unit 10A, acceleration feature amount calculation unit 15, posture determination unit 16, body vertical determination unit 17, sleep stage probability estimation unit 18 and an estimated data output control unit 19.

  The model creation unit 10A includes a sleep stage-specific fluctuation analysis unit 10A1, a sleep stage transition-specific heart rate change value feature quantity analysis unit 10A2, a time / model-specific sleep stage appearance frequency analysis unit 10A3, and a time-specific sleep stage transition frequency analysis. Part 10A4 and acceleration analysis part 10A5 classified by sleep stage.

  All of these control processing functions are realized by causing the CPU to execute a program stored in a program memory (not shown).

  The heart rate variability feature quantity calculation unit 11 and the heart rate change value feature quantity calculation unit 12 calculate the feature quantity of the heart rate variability based on the electrocardiogram data acquired from the electrocardiograph 2 via the input / output interface unit 30.

  Specifically, an RRI (R-R Interval) representing a peak interval is detected from an electrocardiogram waveform, and the detected RRI is divided every n seconds to obtain a frame length N [sec] and a slide width n [sec]. ], The heartbeat variability feature vector ht = {ht1, ht2,..., Htk} and the heartbeat change feature vector dht = {dht1, dht2,. Here, t indicates an elapsed time, and is counted every n seconds such as t = {n, 2 * n, 3 * n,.

  The heart rate variability feature vector ht calculated by the heart rate variability feature quantity calculator 11 is sent to the sleep stage variation analyzer 10A1 and sleep stage probability estimator 18 of the model generator 10A.

  The heart rate change feature vector dht calculated by the heart rate change value feature quantity calculation unit 12 is sent to the hourly model determination units 13 and 14 and the sleep stage transition-specific heart rate change value feature quantity analysis unit 10A2 of the model creation unit 10A. .

  For each time period, the hourly model determination units 13 and 14 perform model determination processing using the result of the heart rate change value feature quantity calculation unit included in the time period. The time zone indicates that the time series data is divided every q minutes (for example, 30 minutes) and each divided time zone m.

The determination result of the hourly model determination unit 13 is sent to the sleep stage probability estimation unit 18.
The determination result of the hourly model determination unit 14 is sent to the time / model-specific sleep stage appearance frequency analysis unit 10A3 in the model creation unit 10A.
In addition, the hourly model determination units 13 and 14 may perform model determination processing by clustering heart rate change value feature values for each time zone. For clustering, k-means (k is the number of models) in which the number of models is set may be used, the topic model is a latent variable (topic), and the heart rate change value feature quantity is an observed value (word). Model determination may be performed using.

  Based on the time information t, the sleep stage variation analysis unit 10A1 in the model creation unit 10A uses the heartbeat variation feature vector ht calculated by the heartbeat variation feature amount calculation unit 11 in the first learning phase described later. In addition, it is associated with the correct answer label of the sleep stage every n seconds input in the input unit 3, and the distribution of each heartbeat feature amount of each sleep stage is calculated. Then, the distribution of each heartbeat feature amount calculated for each sleep stage is stored in the heartbeat fluctuation table 211 in the model database 21.

  The heart rate change value feature quantity analysis unit 10A2 for each sleep stage transition uses the heart rate change value feature vector dht calculated by the heart rate change value feature quantity calculation unit 12 based on time information t in a first learning phase to be described later. In addition, it is associated with the correct answer label of the sleep stage every n seconds input in the input unit 3, and the distribution of each heart rate change value feature quantity of each sleep stage transition is calculated. Then, the distribution of each heart rate change value feature amount calculated for each sleep stage is stored in the heart rate change value feature amount table 212 in the model database 21.

  The sleep stage appearance frequency analysis unit 10A3 classified by time / model, in a first learning phase described later, sets time series data of sleep stage correct labels every n seconds input in the input unit 3 every q minutes (for example, 30 minutes). And the number of appearances of each sleep stage i in each divided time zone m is calculated. At this time, the number of appearances is calculated for each model determined by the hourly model determination unit. Then, the sleep stage appearance frequency analysis unit 10A3 for each time / model performs a process of storing the calculated number of appearances of each sleep stage i in the appearance frequency table 213 in the model database 21 for each model.

  In the first learning phase, which will be described later, the hourly sleep stage transition frequency analysis unit 10A4 performs n in each time zone m divided every q minutes in the same manner as the time / model sleep stage appearance frequency analysis unit 10A3. The number of appearances of each sleep stage transition per second is calculated. Then, the hourly sleep stage transition frequency analysis unit 10A4 performs a process of storing the calculated number of appearances of each sleep stage transition in the transition frequency table 214 in the model database 21.

  The sleep stage-specific acceleration analysis unit 10A5 calculates an acceleration feature amount based on acceleration data acquired from the triaxial accelerometer 5 via the input / output interface unit 30 in a second learning phase to be described later.

Specifically, the D value and the arousal determination value for each predetermined section are calculated using the algorithm of Cole et al. The D value and the arousal determination value are values obtained by the algorithm of Cole et al., And these are acceleration feature vectors at = {at1, at2}.
Roger J. Cole, Daniel F. Kripke, William Gruen, Daniel J. Mullaney, J. Christian Gillin. “Automatic Sleep / Wake Identification From Wrist Activity” Sleep, 15 (5): 461-469, (1992). The unit 16 performs posture determination based on acceleration data captured from the triaxial accelerometer 5 via the input / output interface 30 in an estimation phase described later. Specifically, binary determination is performed for the sleeping posture or the posture of raising the body. For example, when the accelerometer 5 is worn on the chest so that the y-axis of the triaxial accelerometer 5 is horizontal with respect to the subject's body and the lower part of the body is positive, the y-axis gravitational acceleration value is If it exceeds a certain level, it is determined as the posture that caused the body to wake up. Based on the determination result of the posture determination unit 16, the upper body vertical determination unit 17 determines whether or not the upper body of the body is vertical, and sends the determination result to the sleep stage probability estimation unit 18.

  The sleep stage probability estimation unit 18, in the estimation phase to be described later, the heart rate variability feature vector calculated by the heart rate variability feature amount calculation unit 11 based on the electrocardiogram data input from the electrocardiograph 2, and the heart rate change value. Based on the heart rate change value feature vector calculated by the feature amount calculation unit 12, the result determined by the hourly model determination unit 13, and the acceleration data input from the accelerometer 5, the acceleration feature amount calculation unit 15 The arousal state determined by the body vertical determination unit 17 from the calculated acceleration feature vector, information stored in the tables 211, 212, 213, 214, and 215 of the model database 21 and acceleration data from the posture determination unit 16 Are used to estimate the sleep stage St every n seconds. And the process which stores the data showing the estimation result in the said estimated data memory | storage part 22 is performed.

  The estimated data output control unit 19 reads out data representing the sleep stage estimation result in the corresponding time zone of the corresponding user from the estimated data storage unit 22 according to the output instruction input in the input unit 3. Then, display data for displaying the read estimated data is generated, and the generated display data is output to the display unit 4 via the input / output interface unit 30 and displayed.

<Operation>
Next, the operation of estimating the sleep stage by the apparatus configured as described above will be described.
(1) Learning phase
In the first and second learning phases, a measurement person such as a medical worker visually observes the sleep stage of the subject every n seconds while measuring the brain wave, electromyogram, electrocardiogram waveform, and acceleration of the subject. The result is input by the input unit 3 as the “sleep phase correct answer label”.

  Then, the sleep stage estimation device 1 calculates the feature quantity of heart rate variability, the feature quantity of heart rate change, and the feature quantity of acceleration from the electrocardiogram waveform and acceleration, and based on the inputted correct label of the sleep stage , Calculating the appearance frequency of the feature value of the heart rate variability in each sleep stage, the appearance frequency of the feature value of the acceleration in each sleep stage, and the appearance frequency of the feature value of the heart rate change in each sleep stage transition, While storing in the acceleration table 215 and the heart rate change value feature quantity table 212, the sleep stage appearance frequency and the sleep stage transition frequency for each time and model are calculated, and the calculated values are represented as the appearance frequency table 213 and the transition frequency, respectively. Store in table 214.

FIG. 10 is a flowchart showing a processing procedure and processing contents in the sleep stage estimation device 1 in the first learning phase.
When the measurement in the first learning phase is started, first, the electrocardiogram data of the subject output from the electrocardiograph 2 in step S11 is taken in via the input / output interface unit 30 under the control of the heart rate variability feature quantity calculation unit 11. . In step S12, the feature amount of heart rate variability is calculated as follows based on the acquired electrocardiogram data.

  That is, first, RRI (R-R Interval) representing the peak interval of the electrocardiogram waveform is detected from the electrocardiogram data, and the detected RRI is divided every n seconds. Then, the heartbeat variability feature vector ht = {ht1, ht2,..., Tk} is calculated as the frame length N [sec] and the slide width n [sec]. Here, t indicates an elapsed time, and is counted every n seconds such as t = {n, 2 * n, 3 * n,. A specific example is shown in FIG.

  Examples of the heart rate variability feature amount include, for example, an average value of heart rate variability RRI (R-R Interval), a standard deviation of RRI, a root mean square of a difference between preceding and following RRIs (RMSSD), and a temporal difference between preceding and following RRIs. Occurrence rate (pNN50), the long side length L of the shape drawn by the preceding and following RRI Lorentz plots, the short side lengths T and L drawn by the RRI Lorentz plot, and Sympathetic index (CSI) obtained from T, parasympathetic activity index (CVI) obtained from L and T, high frequency component (HF) of power spectrum waveform of RRI, or low frequency Is it at least one of the components (LF), or their features? Other feature amounts calculated from the above are used.

  An example of the RRI Lorentz plot is shown in FIG. 12, and the definition of each heart rate variability feature amount is shown in FIG. The detailed calculation method of the heart rate variability feature amount is described in Non-Patent Document 1, for example.

  The value of each feature value of the heart rate variability feature vector ht is preferably normalized to a range of, for example, 0 to 1 and then made discrete at an L level. For example, the first element of the feature vector when t = n (first n-second interval) is expressed as hn, 1 = {0, 1,..., L}.

  Further, the heart rate change value feature amount is calculated from the deviation of the heart rate variability feature amount. That is, the heart rate change value feature amount is obtained by, for example, the standard deviation of the heart rate variability feature amount, the moving average of the standard deviation, and the difference value before and after the heart rate variability.

  As a specific example, a standard deviation value obtained by inputting a value before discretization of the first element of each of the heart rate variability feature vectors h1 to h10 included in the range from t = 1 to t = 10 is expressed as dh1. it can. At this time, dh2 is a standard deviation value with the values before the discretization of the first element of the heart rate variability feature vectors h2 to h11 as inputs.

  Like the heart rate variability feature vector ht, the value of each feature amount of the heart rate change value feature vector dht may be normalized to a range of, for example, 0 to 1 and converted to a discrete value at the L ′ level.

In parallel with the first learning phase, the sleep stage estimation device 1 also executes processing of the second learning phase.
FIG. 14 is a flowchart showing a processing procedure and processing contents in the sleep stage estimation device 1 in the second learning phase.
When measurement in the second learning phase is started, under the control of the acceleration feature quantity calculation unit 15, first, the subject's acceleration data output from the accelerometer 5 in step S 21 is taken in via the input / output interface unit 30. In step S22, the acceleration feature amount calculation unit 15 calculates the acceleration feature amount as follows based on the acquired acceleration data.

  That is, first, a D value and a wakefulness determination value, etc., every n seconds are calculated from acceleration data using the algorithm described in Non-Patent Document 1, and an acceleration feature vector at = {at1, at2,... Ask for.

  The value of each feature value of the acceleration feature vector at may be normalized to, for example, a range of 0 to 1 and made discrete at the Z level. For example, the first element of the feature vector when t = n (first n-second interval) is expressed as an, 1 = {0, 1,..., Z}.

  Subsequently, the sleep stage acceleration analysis unit 10A5 of the model creation unit 10A analyzes the sleep stage acceleration as follows based on the acceleration feature vector at calculated by the acceleration feature quantity calculation unit 15. That is, the calculated acceleration feature vector at is associated with the correct answer label of the sleep stage every n seconds input by the input unit 3 based on the time information t. Then, the distribution of each acceleration feature at each sleep stage is calculated.

  On the other hand, in the first learning phase, the fluctuation analysis unit 10A1 for each sleep stage of the model creation unit 10A calculates the heartbeat fluctuation for each sleep stage based on the heartbeat fluctuation feature vector ht calculated by the heartbeat fluctuation feature quantity calculation unit 11. Analyze as follows. That is, the calculated heart rate variability feature vector ht is associated with the correct answer label of the sleep stage every n seconds input by the input unit 3 based on the time information t. Then, the distribution of each heartbeat feature amount in each sleep stage is calculated.

  Furthermore, in the heart rate change value feature quantity analysis unit 10A2 for each sleep stage transition, the heart rate change value feature quantity for each sleep stage transition is calculated based on the heart rate change value feature vector dht calculated by the heart rate change value feature quantity calculation unit 12. Analysis is performed in the same manner as the sleep stage-specific fluctuation analysis unit 10A1. Then, the distribution of each heart rate change value feature quantity of each sleep stage transition is calculated.

  The hourly model determination units 13 and 14 determine the model based on the magnitude of the heartbeat change value feature amount for each time zone.

  15 to 17 show the processing contents. In FIG. 15A, when time-series data of heart rate change value feature values every n seconds is input (step S1-1), the input time-series data of heart rate change value feature values is every q minutes (for example, 30 minutes). In each divided time zone m, the time when the heartbeat change value feature amount exceeds a certain value as shown in FIG. 15B is counted (step S1-2).

  It is determined whether or not the count number exceeds a certain number (step S1-3).

  Further, instead of the process of FIG. 15A, the hourly model determination units 13 and 14 may increase the number of models by setting some thresholds of the count number by a process as shown in FIG. In the figure, time-series data of heart rate change value feature values are input (step S2-1), and the heart rate change value feature values are not less than a certain value in each time zone obtained by dividing the input time-series data of heart rate change value feature values. Is counted (step S2-2).

  If the count number is equal to or less than the threshold value count1, it is determined as model A (step S2-3). If the count number is greater than the threshold value count1, it is further determined whether it is greater than the threshold value count2 and equal to or less than the threshold value count3 (count3> count2> count1). (Step S2-4). Here, it is assumed that the model B is determined to be larger than the threshold value count2 and equal to or smaller than the threshold value count3, and the model C is determined to be larger than the threshold value count3.

  In addition, the hourly model determination units 13 and 14 may perform the model determination process based on whether or not a plurality of predetermined heart rate change value feature values exceeding a threshold value exceed a majority. .

  Further, instead of the processing of FIGS. 15 and 16, the hourly model determination units 13 and 14 may perform model determination by clustering heart rate change value feature values for each time zone.

  FIG. 17 shows the processing process, and after inputting one or more of the heart rate change value feature quantities in each time zone obtained by dividing the time series data among the time series data of the heart rate change value feature quantities. (Step S3-1), clustering is executed (Step S3-2).

  As a clustering method, k-means in which the number of models is set (k is the number of models) may be used, and the topic is a latent variable (topic), and the heart rate change value feature quantity is an observed value (word). Model determination may be performed using a model.

  The hourly model determination units 13 and 14 output one of the models A, B,... As a result obtained by the determination.

  Next, in step S13, the appearance frequency of each sleep stage for each time zone (m minutes) is tabulated as follows by the time / model-specific sleep stage appearance frequency analysis unit 10A3. That is, first, the time series data of the sleep stage correct answer labels every n seconds input in the input unit 3 are divided every q minutes (for example, 30 minutes), and the number of appearances of each sleep stage i in each divided time zone m. Is calculated for each model of the hourly model determination result. At this time, m = 1 is a section from 0 minute to (q × 1), and m = 2 is a section from (q × 1) to (q × 2).

  Then, the calculated number of appearances of each sleep stage i is stored in the appearance frequency table 213 together with information indicating the sleep stage in association with the time zone and the model, for example, as shown in FIG.

  Subsequently, in step S14, the frequency of appearance of each sleep stage transition pattern for each time zone (m minutes) is tabulated as follows by the hourly sleep stage transition frequency analysis unit 10A4. That is, first, in each time slot m divided every q minutes, the number of appearances of each sleep stage transition every n seconds is calculated.

  For example, in a certain time zone m, when the sleep stage of the first 0-n second section is “wake (WAKE)” and the sleep stage of the n-2 n second section is “REM sleep (REM)”, the n-2n second section The transition of the sleep stage is counted as a transition pattern of “wake (WAKE) → REM sleep (REM)”.

  In addition, the transition pattern of the sleep stage of the first 0-n second section does not need to be counted. Then, the number of appearances of each sleep stage transition calculated as described above is stored in the transition frequency table 214 together with information representing the sleep stage transition pattern in association with the time zone, for example, as shown in FIG.

  Subsequently, in step S15, the fluctuation analysis unit 10A1 for each sleep stage totals the heartbeat fluctuation feature amount for each sleep stage, and the appearance frequency as the calculation result is represented by, for example, the heartbeat feature vector ID and the heartbeat feature as shown in FIG. Together with the quantity level, it is stored in the heart rate variability table 211 in association with each sleep stage, and the first learning phase is thus completed.

  For example, assuming that the total time when the correct answer label of the sleep stage is “wake (WAKE)” is t_wake = {n, 5 * n, 8 * n}, the count is ht_wake, 1 = 1 for all times. Is memorized. At this time, since the discrete value of the heart rate variability feature quantity is 0 ≦ l ≦ L, each value is counted.

  In the other second learning phase, finally, in step S23, the acceleration feature quantity for each sleep stage is tabulated by the sleep stage acceleration analysis unit 10A5, and the appearance frequency as the calculation result is shown in FIG. 9, for example. Thus, together with the acceleration feature vector ID and the acceleration feature amount level, each sleep stage is associated with each sleep stage and stored in the acceleration table 215. Thus, the second learning phase is completed.

  For example, if the total time when the correct answer label of the sleep stage is “Wake” is t_wake = {n, 5 * n, 8 * n}, the count at which at_wake, 1 = 1 for all times Remember. At this time, since the discrete value of the acceleration feature amount is 0 ≦ l ≦ M, each value is counted.

(2) Estimation phase
In the estimation phase, only the electrocardiogram data and acceleration data of the person to be estimated are measured, and the heart rate variability feature amount, heart rate change value feature amount, hourly model determination result, and acceleration calculated from the measured electrocardiogram data and acceleration data Based on the feature amount and the model data stored in the heart rate variability table 211, the heart rate change value feature amount table 212, the appearance frequency table 213, the transition frequency table 214, and the acceleration table 215, the sleep stage St of the estimation target person Is estimated every n seconds.

  FIG. 18 is a flowchart showing the processing procedure and processing contents. First, in step S31, from the electrocardiograph 2 and the accelerometer 5 via the input / output interface unit 30 by the heart rate variability feature quantity calculation unit 11, the heart rate change value feature quantity calculation unit 12, and the acceleration feature quantity calculation unit 15. The electrocardiogram data and acceleration data in one sleep of the estimation target person are captured.

  In step S32 and step S33, the heart rate variability feature quantity calculation unit 11, the heart rate change value feature quantity calculation unit 12, and the acceleration feature quantity calculation unit 15 perform the calculation every n seconds from the electrocardiogram data and acceleration data captured in parallel. A heart rate variability feature amount, a heart rate change value feature amount, and an acceleration feature amount are calculated. The calculation method is the same as that in the learning phase.

  On the other hand, in step S38 executed in parallel, the posture determination unit 16 determines the posture every n seconds from the acquired acceleration data. That is, a binary determination is made of the posture of sleeping or the posture of raising the body. Awakening is the case where the body is raised.

  In a succeeding step S39, it is determined whether or not the result of the binary determination is “WAKE” indicating awakening.

  On the other hand, following step S33, under the control of the sleep stage probability estimation unit 18, the sleep stage of the estimation target person is estimated as follows. That is, first, at step S34, sleep stage variables every n seconds are initialized at random. Also, the Gibbs sampling counter g is initialized to “1”.

  Next, in step S35, the sleep stage variable every n seconds is updated by Gibbs sampling, and the following processing is repeated until it is determined in step S36 that the Gibbs sampling counter g has reached the preset number of Gibbs sampling repetitions. .

(A) As initial values, time t = 0 and sleep stage S0 = “WAKE” are set.
(B) While the time t is smaller than the sleep time T, the following processing is executed.
(B-1) If the posture determination unit 17 does not determine awakening in step S39, in step S35, the sleep stage S_ {t, g} is changed to the sleep stage St-1, time t-1. The value of P (St = i | ht, dht, at, St-1 = j, Ct = m) using the band Ct, the heart rate variability feature amount ht at time t, the heart rate change value feature amount dht, and the acceleration feature amount at. Sampling according to
(B-2) If it is determined in step S39 that the posture determination unit 17 has awakened, in the subsequent step S40, the sleep stage S_ {t, g} is set to “WAKE”.
(B-3) Time t = t + 1.
(C) The Gibbs sampling counter g is incremented by “+1”.

  Subsequently, in step S37, for each time t, the appearance distribution of each sleep stage is totaled from the g sleep stage sampling history sets {S_ {t, g}} and output as a sleep stage probability distribution at time t. When outputting the sleep stage, the sleep stage that is the mode value in the sampling history set {S_ {t, g}} is stored in the estimation data storage unit 22 as an estimation result.

  P (St = i | ht, dht, at, St-1 = j, Ct = m) is obtained by the following equation.

  Here, θmi represents the probability that the sleep stage is i in the time zone m, and is calculated by the following equation.

  Here, Mmi represents the number of times the sleep stage i appears in the time zone m.

  Also, λmji represents the probability that the sleep stage transitions from j to i in the time zone m, and is calculated by the following equation.

  Here, Mmji represents the number of times that a state in which the sleep stage transitions from j to i appears in the time zone m.

  Φihtk represents the appearance probability of each heartbeat feature amount in the sleep stage i,

Calculated by
Furthermore, πiatw represents the sleep probability of each acceleration feature amount in the sleep stage i,

Calculated by
When the heart rate variability feature vector kth element htk newly obtained at time t at sleep stage i is lk, Milk refers to the model database, and the appearance frequency when htk satisfies the value lk Represents the number of times.

  When the acceleration feature vector wth element atw newly obtained at time t at the sleep stage i is zw, Mizm refers to the model database 21 and the appearance frequency when atw satisfies the value zw. Represents the number of times.

  Note that the data representing the estimation result of the sleep stage of the estimation target person stored in the estimation data storage unit 22 is read from the estimation data storage unit 22 under the control of the estimation data output control unit 19. Then, display data is generated based on the read estimated data, and this display data is supplied to the display unit 4 via the input / output interface unit 30 and displayed. An example of the display result is shown in the lower part of FIG.

<Effect of embodiment>
As described in detail above, in one embodiment, in the learning phase, the heart rate variability feature amount for each sleep stage, the heart rate change value feature amount for each sleep stage transition, and the acceleration feature amount are calculated from the electrocardiogram data and acceleration data of the subject. The heart rate variability table 211, the heart rate change value feature quantity table 212, and the acceleration table 215 are stored, and the model determination for each time according to the time course using the heart rate change value feature quantity is performed. The appearance frequency of the sleep stage according to the time and the transition frequency according to the passage of time are obtained, and these are stored in the appearance frequency table 213 and the transition frequency table 214, respectively.

  In the estimation phase, the heart rate variability feature amount, the heart rate change value feature amount, the acceleration feature amount calculated from the electrocardiogram data and acceleration data of the estimation target person, the model determination result by the hourly model determination units 13 and 14, and the posture determination unit 16, and the sleep stage of the estimation target person is estimated based on the sleep determination model data stored in the respective tables 211, 213 to 215 and the heart rate change value feature quantity table 212. Yes.

  Therefore, even when the heartbeat changes due to influences other than brain activity during sleep, such as blood pressure adjustment, body temperature adjustment, body movement, breathing, and the like, the sleep stage can be accurately estimated. Moreover, since the body movement is large at the time of awakening, it becomes possible to estimate the awakening more accurately by considering the acceleration. And by incorporating the arousal determination by the posture determination, it is possible to reliably detect the awakening state when the body is awakened from the bed.

  Furthermore, by providing the learning phase, it is possible to construct an optimal heart rate fluctuation table 211, acceleration table 215, appearance frequency table 213, and transition frequency table 214.

  Considering the case where the sleep stage estimation device 1 is realized by a portable information device such as a smartphone or a tablet personal computer and an application program implemented thereon, the user of the portable information device that is a subject is awakened. When the detection accuracy in the state is lowered, there is a possibility that the reliability of the user with respect to the application program is significantly lowered.

  Therefore, by reliably determining that the user is in the awake state from the body movement and posture based on the acceleration data, it is possible to reliably suppress a decrease in the reliability of the user with respect to the application program.

[Other Embodiments]
In addition, this invention is not limited to the said embodiment. For example, in the embodiment, the case where the storage unit 20 is provided in the sleep stage estimation apparatus 1 has been described as an example. However, the storage unit 20 is provided in a database server or the like provided in the cloud, and the sleep stage estimation apparatus 1 and its database are provided. Data may be written and read by communicating with the server.

  In the embodiment, the heart rate variability feature amount is calculated based on the electrocardiogram data obtained by the electrocardiograph. However, the pulse wave is measured and the heart rate variability feature amount is calculated based on the measurement data. You may do it.

  Further, the data representing the estimation result may be transmitted to another terminal or server via the communication line. Other control functions provided in the control unit, processing procedure and processing contents by the control function, display of the estimation results The format and the like can be variously modified and implemented without departing from the gist of the present invention.

  In short, the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, you may combine the component covering different embodiment suitably.

DESCRIPTION OF SYMBOLS 1 ... Sleep stage estimation apparatus, 2 ... Electrocardiograph, 3 ... Input part, 4 ... Display part, 5 ... Triaxial accelerometer, 10 ... Control unit, 10A ... Model creation part, 10A1 ... Sleep stage fluctuation | variation analysis part, 10A2 ... heart rate change value feature quantity analysis unit by sleep stage transition, 10A3 ... sleep stage appearance frequency analysis unit by time / model, 10A4 ... sleep stage transition frequency analysis unit by time, 10A5 ... acceleration analysis unit by sleep stage, 11 ... heart rate Fluctuation feature amount calculation unit, 12 ... heart rate change value feature amount calculation unit, 13 ... hourly model determination unit, 14 ... hourly model determination unit, 15 ... acceleration feature amount calculation unit, 16 ... posture determination unit, 17 ... upper body Vertical determination unit, 18 ... sleep stage probability estimation unit, 19 ... estimation data output control unit, 20 ... storage unit, 21 ... model database, 211 ... heart rate variability table, 212 ... heart rate change value feature quantity table, 213 ... output Frequency table, 214 ... transition frequency table, 215 ... acceleration table, 22 ... estimation data storage unit, 30 ... input-output interface unit.

Claims (13)

A heart rate variability feature amount calculating unit that calculates a heart rate variability feature amount for each first unit time from a set of sleep electrocardiograms;
A heart rate variability storage unit that stores heart rate variability information obtained by counting the heart rate variability feature amounts calculated by the heart rate variability feature amount calculation unit for each sleep stage;
A heart rate change value feature quantity calculating unit that calculates a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated by the heart rate fluctuation feature quantity calculating unit;
A heart rate change value feature amount storage unit for storing heart rate change value feature amount information obtained by totaling the heart rate change value feature amounts calculated by the heart rate change value feature amount calculation unit for each sleep stage transition;
Said time axis of sleep electrocardiogram for each time slot obtained by dividing a long second unit of time than the first unit time, the heartbeat change value, wherein determining model determining unit models based on the amount,
For each divided time zone and for each model as a result of the determination by the model determination unit, a sleep stage appearance frequency analysis unit for each time / model for analyzing the appearance frequency of the sleep stage for each first unit time, and
An appearance frequency storage unit for storing time / model-specific appearance frequency information analyzed by the time / model sleep stage appearance frequency analysis unit;
For each of the divided time zones, an hourly sleep stage transition frequency analysis unit that analyzes the frequency of sleep stage transitions for each first unit time, and
A transition frequency storage unit that stores the transition frequency information analyzed by the hourly sleep stage transition frequency analysis unit;
A sleep model creation device comprising: An acceleration analysis unit for each sleep stage that analyzes an acceleration feature value for each sleep stage from the set of sleep accelerations for each first unit time; and
An acceleration storage unit that stores an acceleration feature amount for each sleep stage analyzed by the sleep stage acceleration analysis unit;
The sleep model creation device according to claim 1, further comprising: A heart rate variability storage unit for storing heart rate variability information by summing heart rate variability features calculated for each first unit time from a set of sleep electrocardiograms for each sleep stage;
A heart rate change value feature amount storage unit for storing heart rate change value feature amount information obtained by summing heart rate change value feature amounts calculated from deviations of the heart rate variability feature amounts for each sleep stage transition;
The first unit time for each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time and for each model determined based on the heartbeat change value feature quantity Appearance frequency storage unit that stores appearance frequency information for each time and model that analyzes the appearance frequency of each sleep stage,
For each divided time zone, a transition frequency storage unit that stores transition frequency information that analyzes the frequency of sleep stage transitions for each first unit time;
From a sleep electrocardiogram acquired from a subject to be estimated, a heart rate variability feature amount calculation unit that calculates a feature amount of heart rate variability for each first unit time for each time period divided by the second unit time;
A heart rate change value feature quantity calculating unit that calculates a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated by the heart rate fluctuation feature quantity calculating unit;
A model determination unit that determines a model based on the heart rate change value feature amount calculated by the heart rate change value feature amount calculation unit;
The heart rate change value feature amount calculated by the heart rate change value feature amount calculation unit, the model determined by the model determination unit, the heart rate fluctuation storage unit, the heart rate change value feature amount storage unit, the appearance frequency storage unit, and the transition frequency storage An estimation unit that estimates a transition probability distribution of the sleep stage of the subject based on each information stored in the unit;
A sleep stage estimation device comprising: A heart rate variability storage unit for storing heart rate variability information by summing heart rate variability features calculated for each first unit time from a set of sleep electrocardiograms for each sleep stage;
A heart rate change value feature amount storage unit for storing heart rate change value feature amount information obtained by summing heart rate change value feature amounts calculated from deviations of the heart rate variability feature amounts for each sleep stage transition;
The first unit time for each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time and for each model determined based on the heartbeat change value feature quantity Appearance frequency storage unit that stores appearance frequency information for each time and model that analyzes the appearance frequency of each sleep stage,
For each of the divided time zones, a transition frequency storage unit that stores transition frequency information obtained by analyzing the frequency of sleep stage transitions for each first unit time;
An acceleration storage unit for storing an acceleration feature quantity for each sleep stage obtained by analyzing an acceleration feature quantity for each sleep stage from the set of accelerations during sleep;
From a sleep electrocardiogram acquired from a subject to be estimated, a heart rate variability feature amount calculation unit that calculates a feature amount of heart rate variability for each first unit time for each time period divided by the second unit time;
A heart rate change value feature quantity calculating unit that calculates a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated by the heart rate fluctuation feature quantity calculating unit;
A model determination unit that determines a model based on the heart rate change value feature amount calculated by the heart rate change value feature amount calculation unit;
A posture determination unit that determines the posture from the acceleration acquired from the subject;
The heart rate change value feature amount calculated by the heart rate change value feature amount calculation unit, the model determined by the model determination unit, the acceleration acquired from the subject, the posture determined by the posture determination unit, and the heart rate variability memory An estimation unit that estimates a transition probability distribution of the sleep stage of the subject based on the information stored in the unit, the heart rate change value feature amount storage unit, the appearance frequency storage unit, the transition frequency storage unit, and the acceleration storage unit,
A sleep stage estimation device comprising: A heart rate variability feature amount calculating unit that calculates a heart rate variability feature amount for each first unit time from a set of sleep electrocardiograms;
A heart rate variability storage unit that stores heart rate variability information obtained by counting the heart rate variability feature amounts calculated by the heart rate variability feature amount calculation unit for each sleep stage;
A heart rate change value feature quantity calculating unit that calculates a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated by the heart rate fluctuation feature quantity calculating unit;
A heart rate change value feature amount storage unit for storing heart rate change value feature amount information obtained by totaling the heart rate change value feature amounts calculated by the heart rate change value feature amount calculation unit for each sleep stage transition;
For each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time, and for each model determined based on clustering of the heartbeat change value feature quantity, An appearance frequency storage unit that stores appearance frequency information for each time and model that analyzes the appearance frequency of the sleep stage for each unit time,
For each of the divided time zones, an hourly sleep stage transition frequency analysis unit that analyzes the frequency of sleep stage transitions for each first unit time, and
A transition frequency storage unit that stores the transition frequency information analyzed by the hourly sleep stage transition frequency analysis unit;
A sleep model creation device comprising: An acceleration analysis unit for each sleep stage that analyzes an acceleration feature value for each sleep stage from the set of sleep accelerations for each first unit time; and
An acceleration storage unit that stores an acceleration feature amount for each sleep stage analyzed by the sleep stage acceleration analysis unit;
The sleep model creation device according to claim 5, further comprising: A heart rate variability storage unit for storing heart rate variability information by summing heart rate variability features calculated for each first unit time from a set of sleep electrocardiograms for each sleep stage;
A heart rate change value feature amount storage unit for storing heart rate change value feature amount information obtained by summing heart rate change value feature amounts calculated from deviations of the heart rate variability feature amounts for each sleep stage transition;
For each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time, and for each model determined based on clustering of the heartbeat change value feature quantity, An appearance frequency storage unit that stores appearance frequency information for each time and model that analyzes the appearance frequency of the sleep stage for each unit time,
For each divided time zone, a transition frequency storage unit that stores transition frequency information that analyzes the frequency of sleep stage transitions for each first unit time;
From a sleep electrocardiogram acquired from a subject to be estimated, a heart rate variability feature amount calculation unit that calculates a feature amount of heart rate variability for each first unit time for each time period divided by the second unit time;
A heart rate change value feature quantity calculating unit that calculates a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated by the heart rate fluctuation feature quantity calculating unit;
A model determination unit that determines a model based on clustering of heart rate change value feature values calculated by the heart rate change value feature value calculation unit;
The heart rate change value feature amount calculated by the heart rate change value feature amount calculation unit, the model determined by the model determination unit, the heart rate fluctuation storage unit, the heart rate change value feature amount storage unit, the appearance frequency storage unit, and the transition frequency storage An estimation unit that estimates a transition probability distribution of the sleep stage of the subject based on each information stored in the unit;
A sleep stage estimation device comprising: A heart rate variability storage unit for storing heart rate variability information obtained by summing heart rate variability features calculated for each first unit time from a set of sleep electrocardiograms for each sleep stage;
A heart rate change value feature amount storage unit for storing heart rate change value feature amount information obtained by summing heart rate change value feature amounts calculated from deviations of the heart rate variability feature amounts for each sleep stage transition;
For each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time, and for each model determined based on clustering of the heartbeat change value feature quantity, An appearance frequency storage unit that stores appearance frequency information for each time and model that analyzes the appearance frequency of the sleep stage for each unit time,
For each divided time zone, a transition frequency storage unit that stores transition frequency information that analyzes the frequency of sleep stage transitions for each first unit time;
An acceleration storage unit for storing an acceleration feature quantity for each sleep stage obtained by analyzing an acceleration feature quantity for each sleep stage from the set of accelerations during sleep;
From a sleep electrocardiogram acquired from a subject to be estimated, a heart rate variability feature amount calculation unit that calculates a feature amount of heart rate variability for each first unit time for each time period divided by the second unit time;
A heart rate change value feature quantity calculating unit that calculates a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated by the heart rate fluctuation feature quantity calculating unit;
A model determination unit that determines a model based on clustering of heart rate change value feature values calculated by the heart rate change value feature value calculation unit;
A posture determination unit that determines the posture from the acceleration acquired from the subject;
The heart rate change value feature amount calculated by the heart rate change value feature amount calculation unit, the model determined by the model determination unit, the acceleration acquired from the subject, the posture determined by the posture determination unit, and the heart rate variability memory An estimation unit that estimates a transition probability distribution of the sleep stage of the subject based on the information stored in the unit, the heart rate change value feature amount storage unit, the appearance frequency storage unit, the transition frequency storage unit, and the acceleration storage unit,
A sleep stage estimation device comprising: A heart rate variability feature amount calculating process for calculating a heart rate variability feature amount calculated for each first unit time from a set of sleep electrocardiograms;
A heart rate variability storage process for storing heart rate variability information obtained by counting the heart rate variability feature amounts calculated in the heart rate variability feature amount calculation process for each sleep stage;
A heart rate change value feature quantity calculating process for calculating a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated in the heart rate fluctuation feature quantity calculating process;
A heart rate change value feature amount storage process for storing heart rate change value feature amount information obtained by totaling heart rate change value feature amounts calculated in the heart rate change value feature amount calculation process for each sleep stage transition;
The time axis of the sleeping electrocardiogram for each time slot obtained by dividing a long second unit of time than the first unit of time, and the model determination process determines model on the basis of the heart rate variation value feature quantity,
For each of the divided time zones and for each model of the result determined in the model determination process, the sleep stage appearance frequency analysis process for each time and model for analyzing the appearance frequency of the sleep stage for each first unit time,
Appearance frequency storage process for storing appearance frequency information for each time and model analyzed in the time and model sleep stage appearance frequency analysis process,
For each of the divided time zones, an hourly sleep stage transition frequency analysis process for analyzing the frequency of sleep stage transitions for each first unit time;
Transition frequency storage process for storing transition frequency information analyzed in the sleep phase transition frequency analysis process by time,
A sleep model creation method characterized by comprising: A heart rate variability storage process for storing heart rate variability information by summing heart rate variability features calculated for each first unit time from a set of sleep electrocardiograms for each sleep stage;
A heart rate change value feature amount storage process for storing heart rate change value feature amount information obtained by summing heart rate change value feature amounts calculated from deviations of the heart rate variability feature amounts for each sleep stage transition;
The first unit time for each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time and for each model determined based on the heartbeat change value feature quantity Appearance frequency storage process for storing appearance frequency information by time and model that analyzed the appearance frequency of each sleep stage,
For each divided time zone, a transition frequency storage process for storing transition frequency information analyzing the frequency of transition of the sleep stage for each first unit time;
From a sleep electrocardiogram acquired from a subject to be estimated, a heart rate variability feature amount calculation process for calculating a feature amount of heart rate variability for each first unit time for each time period divided by the second unit time;
A heart rate change value feature quantity calculating process for calculating a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated in the heart rate fluctuation feature quantity calculating process;
A model determination step of determining a model based on the heart rate change value feature amount calculated in the heart rate change value feature amount calculation step;
The heart rate change value feature quantity calculated in the heart rate change value feature quantity calculation process, the model determined in the model determination process, the heart rate variability storage process, the heart rate change value feature quantity storage process, the appearance frequency storage process, and the transition frequency storage An estimation process for estimating a transition probability distribution of the sleep stage of the subject based on each information stored in the process;
A sleep stage estimation method characterized by comprising: A heart rate variability feature amount calculating process for calculating a heart rate variability feature amount for each first unit time from a set of sleep electrocardiograms;
A heart rate variability storage process for storing heart rate variability information obtained by counting the heart rate variability feature amounts calculated in the heart rate variability feature amount calculation process for each sleep stage;
A heart rate change value feature quantity calculating process for calculating a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated in the heart rate fluctuation feature quantity calculating process;
A heart rate change value feature amount storage process for storing heart rate change value feature amount information obtained by totaling heart rate change value feature amounts calculated in the heart rate change value feature amount calculation process for each sleep stage transition;
For each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time, and for each model determined based on clustering of the heartbeat change value feature quantity, Appearance frequency storage process that stores appearance frequency information by time and model that analyzes the appearance frequency of sleep stages per unit time,
For each of the divided time zones, an hourly sleep stage transition frequency analysis process for analyzing the frequency of sleep stage transitions for each first unit time;
Transition frequency storage process for storing transition frequency information analyzed in the sleep phase transition frequency analysis process by time,
A sleep model creation method characterized by comprising: A heart rate variability storage process for storing heart rate variability information by summing heart rate variability features calculated for each first unit time from a set of sleep electrocardiograms for each sleep stage;
A heart rate change value feature amount storage process for storing heart rate change value feature amount information obtained by summing heart rate change value feature amounts calculated from deviations of the heart rate variability feature amounts for each sleep stage transition;
For each time zone obtained by dividing the time axis of the sleep electrocardiogram by a second unit time longer than the first unit time, and for each model determined based on clustering of the heartbeat change value feature quantity, Appearance frequency storage process that stores appearance frequency information by time and model that analyzes the appearance frequency of sleep stages per unit time,
For each divided time zone, a transition frequency storage process for storing transition frequency information analyzing the frequency of transition of the sleep stage for each first unit time;
From a sleep electrocardiogram acquired from a subject to be estimated, a heart rate variability feature amount calculation process for calculating a feature amount of heart rate variability for each first unit time for each time period divided by the second unit time;
A heart rate change value feature quantity calculating process for calculating a heart rate change value feature quantity from a deviation of the heart rate fluctuation feature quantity calculated in the heart rate fluctuation feature quantity calculating process;
A model determination process for determining a model based on clustering of heart rate change value feature values calculated in the heart rate change value feature value calculation process;
The heart rate change value feature quantity calculated in the heart rate change value feature quantity calculation process, the model determined in the model determination process, the heart rate variability storage process, the heart rate change value feature quantity storage process, the appearance frequency storage process, and the transition frequency storage An estimation process for estimating a transition probability distribution of the sleep stage of the subject based on each information stored in the process;
A sleep stage estimation method characterized by comprising:   A program that causes a computer included in the apparatus to execute processing performed by each unit included in the apparatus according to claim 1.