The present invention relates to a sleep determination apparatus, a sleep determination method, and storage medium.
In the related art, a sleep polysomnography (PSG) test based on an electroencephalogram, a respiratory test, an eye movement test, an electrocardiogram, and a blood pressure test, and other test items is known to determine a sleep stage.
However, the PSG test involves measurement in a facility, such as a hospital, such that a PSG device having a large number of electrodes is attached to the body of a user, and imposes constraints such as heavy physical and mental loads. Another issue is that such a PSG test needs to be conducted by a skilled expert to determine a sleep stage from the test items obtained using the PSG device, which incurs a large cost.
Accordingly, methods have been devised in recent years for easily estimating a sleep stage, without using a PSG device, from a heart rate signal, electrocardiogram information, and so on obtained using a wristwatch-type measurement device or the like (see NPLs 1, 2, and 3).
However, the sleep stage estimation methods of the related art can determine only four out of five sleep stages including the REM sleep stage and first to fourth non-REM sleep stages. In addition, the maximum precision in the determination of the four sleep stages is about 0.69 (69%). Accordingly, there is an issue of insufficient accuracy.
A sleep determination apparatus according to an aspect of the present invention includes: a heart rate data obtainer for obtaining, based on data regarding a heart rate of a body of a user, a very low frequency component (VLF), a ratio of a low frequency component (LF) to a high frequency component (HF), a mean heartbeat interval (RRI), and a standard deviation of each heartbeat interval (RRI) of the heart rate as heart rate variability parameters indicating a heart rate state; and a sleep state determiner for determining which of the first non-REM sleep stage or the second non-REM sleep stage the heart rate variability parameters of the user indicate, in accordance with a first determination condition that is set based on a correlation between the heart rate variability parameters and the first non-REM sleep stage and a second determination condition that is set based on a correlation between the heart rate variability parameters and the second non-REM sleep stage.
A sleep determination apparatus according to an aspect of the present invention comprising: a heart rate data obtainer for obtaining, based on data regarding a heart rate of a body of a user, a very low frequency component (VLF), a ratio of a low frequency component (LF) to a high frequency component (HF), a mean heartbeat interval (RRI), and a standard deviation of each heartbeat interval (RRI) of the heart rate as heart rate variability parameters indicating a heart rate state; and a sleep state determiner for determining which of a plurality of sleep stages including a first non-REM sleep stage and a second non-REM sleep stage the heart rate variability parameters of the user indicate, wherein the sleep state determiner determines a sleep stage from the heart rate variability parameters, the determination being based on learning result data that is obtained by learning, from training data for which the sleep stages are known, using the heart rate variability parameters as inputs.
A sleep determination apparatus according to an aspect of the present invention comprising: a heart rate data obtainer for obtaining, based on data regarding a heart rate of a body of a user, a very low frequency component (VLF), a ratio of a low frequency component (LF) to a high frequency component (HF), a mean heartbeat interval (RRI), and a standard deviation of each heartbeat interval (RRI) of the heart rate as heart rate variability parameters indicating a heart rate state; and a sleep state determiner for determining which of the sleep stages the user is in, the determination being based on a function indicating a probability of being in each of the plurality of sleep stages, the function having the heart rate variability parameters as variables, the plurality of sleep stages including the first non-REM sleep stage and the second non-REM sleep stage.
According to the present invention, it can be possible to perform an accurate and detailed determination of a sleep stage.
In the following, a sleep determination apparatus, a sleep determination method, and a sleep determination program according to this embodiment of the present invention, and an embodiment of a recording medium will be described in detail with reference to the drawings. It should be noted that this embodiment is not intended to limit the present invention.
First, a configuration of a sleep determination apparatus 100 according to this embodiment will be described, and, then, processing and the like in a sleep determination method according to this embodiment will be described in detail.
In
The various databases and tables (such as a heart rate variable file 106a and a determination technique file 106b) stored in the storage unit 106 are storage means, such as a fixed disk device, and configured to store various programs, tables, files, databases, web pages, and so on that are used for various processing operations.
Among them, the heart rate variable file 106a is a heart rate variability parameter storage means for storing heart rate variability parameters. As an example, the heart rate variable file 106a may store the very low frequency component (VLF), low frequency component (LF), and high frequency component (HF) of heart rate variability, the mean heartbeat interval (RRI), the standard deviation of each heartbeat interval (RRI), the ratio of the low frequency component to the high frequency component (LF/HF ratio), and so on as heart rate variability parameters.
The determination technique file 106b is a determination technique storage means for storing a determination technique for determining a sleep stage using heart rate variability parameters (e.g., VLF, LF, HF, mean RRI, standard deviation of each RRI, and LF/HF ratio). For example, the determination technique file 106b may store a mathematical formula, an algorithm, or the like that takes heart rate variability parameters as input values and that outputs evaluation values of sleep stages. As a more specific example, the determination technique file 106b may store algorithm data or the like regarding learning results that are obtained by learning, from training data for which sleep stages are known, using the heart rate variability parameters as inputs.
Alternatively, the storage unit 106 may store heart rate variability data detected from the user. For example, the storage unit 106 may store heart rate variability data input from the input unit 112 or heart rate variability data transmitted from an external device 200. Further, the storage unit 106 may store training data for which sleep stages are known. More specifically, the storage unit 106 may store a dataset that associates a sleep stage determined by an expert using a PSG test with heart rate variability parameters set at that time. The storage unit 106 may also store a result of determining a sleep stage by a sleep stage determination unit 102b described below.
In
In
Among them, the variable extraction unit 102a is a heart rate data obtaining means for obtaining a heart rate variability parameter indicating a heart rate state from the heart rate variability data stored in the storage unit 106. For example, the variable extraction unit 102a may extract the very low frequency component (VLF), low frequency component (LF), and high frequency component (HF) of heart rate variability, the mean heartbeat interval (RRI), the standard deviation of each heartbeat interval (RRI), the ratio of the low frequency component to the high frequency component (LF/HF ratio), and so on as heart rate variability parameters.
As an example, the variable extraction unit 102a may filter the heart rate signal data stored in the storage unit 106 to remove noise, and detect a peak interval signal from a filtered heart rate signal. Further, the variable extraction unit 102a may calculate a power spectral density from the peak interval signal and calculate the HF value, the LF value, the HF/LF ratio, the mean or standard deviation of each RRI, and so on from the calculated power spectral density. More specifically, the peak interval signal is a signal having a variable that is an interval between waveforms (R-waves) around peak intensities of a heart rate signal. In heart rate variability analysis, the peak interval signal is known as an R-R interval (RRI) signal indicating an interval between adjacent R-wave peaks. The variable extraction unit 102a may calculate the mean or standard deviation of each RRI on the basis of the peak interval (RRI). As an example, the variable extraction unit 102a performs frequency analysis, such as Fourier transform, on the detected peak interval signal to calculate a power spectral density. The power spectral density shows a different distribution depending on the state of the autonomic nervous system. In this embodiment, the VLF value is a local maximum value in a range less than about 0.04 Hz, the LF value is a local maximum value in a range of about 0.04 Hz to 0.15 Hz, and the HF value is a local maximum value in a range of about 0.15 Hz to 0.40 Hz. The variable extraction unit 102a can calculate the LF/HF ratio from the determined LF and HF values. Alternatively, the variable extraction unit 102a may extract the heart rate variability parameters from the heart rate variability data by using a known parameter extraction technique.
The sleep stage determination unit 102b is a sleep stage determining means for determining a sleep stage using the heart rate variability parameters extracted by the variable extraction unit 102a. As an example, the sleep stage determination unit 102b may determine light sleep among the sleep stages. For example, the sleep stage determination unit 102b may determine a light sleep stage including first and second non-REM sleep stages (N1 & N2), or determine the first non-REM sleep stage (N1) or the second non-REM sleep stage (N2) by distinguishing in detail between the first non-REM sleep stage (N1) and the second non-REM sleep stage (N2). Further, the sleep stage determination unit 102b may determine a sleep stage by distinguishing, as sleep stages, five stages, including the REM sleep stage (REM), the first non-REM sleep stage (N1), the second non-REM sleep stage (N2), the third non-REM sleep stage (N3), and the fourth non-REM sleep stage (N4). Alternatively, the sleep stage determination unit 102b may determine a sleep stage by distinguishing, as sleep stages, four stages, including the REM sleep stage (REM), the light sleep stage including the first and second non-REM sleep stages (N1 & N2), the third non-REM sleep stage (N3), and the fourth non-REM sleep stage (N4).
The sleep stage determination unit 102b may determine a sleep stage by calculating evaluation values or the like of the sleep stages from the heart rate variability parameters using the mathematical formula or algorithm stored in the determination technique file 106b. For example, the sleep stage determination unit 102b may obtain, from the storage unit 106, learning result data that is obtained by learning, from training data for which sleep stages are known, using the heart rate variability parameters as inputs, and determine a sleep stage on the basis of the obtained learning result data.
Alternatively, the sleep stage determination unit 102b may calculate a parameter indicating which of the sleep stages (N1 to N5) including the first non-REM sleep stage (N1) and the second non-REM sleep stage (N2) the heart rate variability parameters to be used to determine a sleep stage indicate, and determine a sleep stage identified by the parameter with high accuracy. For example, the sleep stage determination unit 102b may calculate the probability σrem(τ) of being in the REM sleep stage (REM) at time τ, the probability σn1(τ) of being in the first non-REM sleep stage (N1) at the time τ, the probability σn2(τ) of being in the second non-REM sleep stage (N2) at the time τ, the probability σn3(τ) of being in the third non-REM sleep stage (N3) at the time τ, and the probability σn4(τ) of being in the fourth non-REM sleep stage (N4) at the time τ using Eqs. (1) to (5) below, and determine that a sleep stage indicated by the heart rate variability parameters is the sleep stage having the highest probability.
In Eqs. (1) to (5), a1 to a5, b1 to b5, c1 to c5, d1 to d5, and e1 to e5 are constants and may be statistically determined from measurement data of a plurality of subjects or determined from the result of determining measurement data of an individual subject by an expert using PSG. Further, RR(τ) denotes the mean heartbeat interval at the time τ, RRstd(τ) denotes the standard deviation of the heartbeat interval at the time τ, VLF(τ) denotes the very low frequency component of heart rate variability at the time τ, LF(τ) denotes the low frequency component of heart rate variability at the time τ, and HF(τ) denotes the high frequency component of heart rate variability at the time τ. In the calculation of the probability σrem(τ) of being in the REM sleep stage (REM) at the time τ, the probability σn1(τ) of being in the first non-REM sleep stage (N1) at the time τ, the probability σn2(τ) of being in the second non-REM sleep stage (N2) at the time τ, the probability σn3(τ) of being in the third non-REM sleep stage (N3) at the time τ, and the probability σn4(τ) of being in the fourth non-REM sleep stage (N4) at the time τ, the respective probabilities may be calculated in such a manner as to reflect data obtained a predetermined time ago.
In this embodiment, furthermore, the first non-REM sleep stage (N1) and the second non-REM sleep stage (N2) may be distinguished and determined, the determination being based on a first determination condition that is set based on a correlation between the heart rate variability parameters and the first non-REM sleep stage (N1) and a second determination condition that is set based on a correlation between the heart rate variability parameters and the second non-REM sleep stage (N2). For example, the following determination technique may be used.
Specifically, threshold values th1 to th4, each for distinguishing between the first non-REM sleep stage (N1) and the second non-REM sleep stage (N2), are set for the very low frequency component (VLF) of heart rate variability, the ratio of the low frequency component (LF) to the high frequency component (HF) (LF/HF ratio), the mean heartbeat interval (RRI), and the standard deviation of each heartbeat interval (RRI), respectively. The threshold values th1 to th4 may be statistically determined from measurement data of a plurality of subjects or determined from the result of determining measurement data of an individual subject by an expert using PSG.
Then, for the heart rate variability parameters to be used to determine a sleep stage, the condition for the respective threshold values th1 to th4 (whether the heart rate variability parameters are larger than the threshold values or smaller than the threshold values) is determined.
For example, if the respective values of the very low frequency component (VLF) of heart rate variability, the ratio of the low frequency component (LF) to the high frequency component (HF) (LF/HF ratio), and the standard deviation of each heartbeat interval (RRI) are larger than the set threshold values th1, th2, and th4, respectively, the first non-REM sleep stage (N1) may be determined. If the value of the mean heartbeat interval (RRI) is smaller than the set threshold value th3, the first non-REM sleep stage (N1) may be determined.
From the determination results described above, it can be determined which of the first non-REM sleep stage (N1) and the second non-REM sleep stage (N2) the heart rate variability parameters to be used to determine a sleep stage indicate. For example, which of the first non-REM sleep stage (N1) and the second non-REM sleep stage (N2) the four elements, including the very low frequency component (VLF) of heart rate variability, the ratio of the low frequency component (LF) to the high frequency component (HF) (LF/HF ratio), the mean heartbeat interval (RRI), and the standard deviation of each heartbeat interval (RRI), indicate can be determined by majority vote to determine a sleep stage.
The learning unit 102c is a machine learning means that takes the heart rate variability parameters as inputs and that learns from training data for which sleep stages are known, which is stored in the storage unit 106. The learning unit 102c may use a known learning technique, such as a neural network (NN), a support vector machine (SVM), a random forest (RF), a recurrent neural network (RNN), or a hidden Markov model (HMM), as an algorithm for the learning. A specific example of the learning technique will be described below.
In
The sleep determination apparatus 100 may be communicably connected to the external device 200, which provides various databases such as for training data, heart rate signals, and heart rate variability parameters, an external program such as a sleep determination program according to the present invention, and so on, via the network 300. Alternatively, the sleep determination apparatus 100 may be communicably connected to the network 300 via a communication device, such as a router, and a wired or wireless communication line, such as a dedicated line.
In
When determining a sleep stage, the sleep determination apparatus 100 can use the determination technique described above solely or use a plurality of determination techniques in combination.
The description of the configuration of the sleep determination apparatus 100 according to this embodiment is complete.
Next, an example of processing performed by the sleep determination apparatus 100 according to this embodiment having the configuration described above will be described in detail hereinafter with reference to
As illustrated in
As illustrated in
Here, a feature matrix is represented by the following expression.
where HR denotes the heart rate, LF denotes the total power value of the LF component, n denotes the number of observations, and HF denotes the total power value of the HF component.
The following table is a chart illustrating the rate of identification (%) when four classes out of five stages including wake (WK), REM sleep (REM), the first non-REM sleep stage (N1), the second non-REM sleep stage (N2), and the third non-REM sleep stage (N3) are classified. It is confirmed that the random forest achieves relatively accurate classification.
Alternatively, the learning unit 102c may derive learning results using a known learning technique such as a support vector machine (SVM) or a hidden Markov model (HMM), and set a determination technique.
Referring back to
Then, the variable extraction unit 102a in the sleep determination apparatus 100 extracts heart rate variability parameters to be used for the determination technique set in step SB-1 from the heart rate variability data stored in the storage unit 106 (step SB-2). For example, the variable extraction unit 102a may extract the VLF, the LF, the HF, the mean RRI, the standard deviation of each RRI, the LF/HF ratio, and so on by using a known heart rate variability parameter extraction technique.
Then, the sleep stage determination unit 102b in the sleep determination apparatus 100 determines a sleep stage using, as inputs, the heart rate variability parameters extracted by the variable extraction unit 102a, in accordance with the determination technique set in step SB-1 (step SB-4). As an example, the sleep stage determination unit 102b may determine light sleep among the sleep stages using one-vs-rest classification. For example, the sleep stage determination unit 102b may discriminate between the light sleep stage including the first and second non-REM sleep stages (N1 & N2) and the other sleep stages. The following table is a diagram illustrating precisions (%) obtained as a result of one-vs-rest classification for N2 and the others.
Alternatively, the sleep stage determination unit 102b may perform multiclass classification. The sleep stage determination unit 102b may perform five-class classification using the REM sleep stage (REM), the first non-REM sleep stage (N1), the second non-REM sleep stage (N2), the third non-REM sleep stage (N3), and the fourth non-REM sleep stage (N4). Alternatively, the sleep stage determination unit 102b may perform four-class classification using, as sleep stages, the REM sleep stage (REM), the light sleep stage including the first and second non-REM sleep stages (N1 & N2), the third non-REM sleep stage (N3), and the fourth non-REM sleep stage (N4). The processing described above is an example of a sleep determination process performed by the sleep determination apparatus 100 according to this embodiment.
The following table is a chart that compares precisions obtained as a result of four-class classification and five-class classification according to this embodiment (“this study” in the table) with precisions obtained using the techniques of the related art (NPLs [1], [2], and [3]).
As in the above chart, it is confirmed that, compared to four-class classification in the techniques of the related art, four-class classification according to this embodiment achieves accurate classification with a precision of 72%. It is further confirmed that, in this embodiment, five-class classification, which has failed in the techniques of the related art, also achieves accurate classification with a precision of 66%. In this embodiment, accordingly, if the heart rate and heart rate variability can be accurately extracted, five-stage determination for sleep stages can be performed using only this information. In the techniques of the related art, only four-stage determination can be performed. In this embodiment, five-stage determination can be performed with a high precision of 0.72, compared to four-stage determination with a maximum precision of 0.69 in the technique of the related art using the heart rate and respiratory flow. In addition, since whether breathing is present can be determined, sleep apnea syndrome can be detected. The heart rate can be measured every day, without using a PSG device, which is useful for prediction of the onset of cardiac disease such as angina pectoris.
In other words, the methods of the related art provide four-class classification in which classification is performed such that N1 and N2 are grouped into a single class. If light sleep is identified through classification with separation of N1 (light sleep) and N2, the rate of sleep disorders is determined, leading to the discovery of a clinically significant category. In addition, N3 tends to decrease with age (reference: Hideto HIRASAWA and Yoshikata ATSUMI, “Polysomnographic Changes in Normal Human Aging”, Japanese Journal of Geriatrics, Vol. 34, No. 6 (June 1997)), and N1 tends to increase with age. A new relationship between age and N1 can be found. It is said that risk for hypertension or diabetes mellitus with an increase in N1 increases, and such diseases may be prevented (see Md Basunia, et al., “Relationship of symptoms with sleep-stage abnormalities in obstructive sleep apnea-hypopnea syndrome”, J Community Hosp Intern Med Perspect. 2016; 6(4):32170, and Denis Martinez et al., “Dimensions of sleepiness and their correlations with sleep-disordered breathing in mild sleep apnea”, J. bras. pneumol. vol. 35 no. 6, June 2009). The method for classifying N1 and N2 previously requires attachment of a PSG device and testing in the hospital. In contrast, a sleep stage can be determined only using the heart rate, and the need for such capability is expected to grow.
In the related art, furthermore, there has been a technique for sleep stage determination using body movement or the like in which REM/non-REM sleep is determined only using information indicating whether there is a movement. In this embodiment, in contrast, the heart rate is accurately measured, and a correlation between determination values for sleep-stage determination and heart rates, which is obtained from PSG, is used to determine a sleep stage through learning. As a result, it is possible to identify five stages, including REM, non-REM, non-REM 1 to 3, and wake, with a precision of 72% being achieved currently because of an extremely small number of subjects having N1 data. Adaptive learning of N1 data expects further improvements in accuracy. That is, each time data of an individual is collected, the data is adaptively learned online, resulting in an expectation that the recognition rate is increasingly improved.
In this embodiment, furthermore, to classify these stages, heart rate feature values are determined and selected as appropriate. Training data with correct answers obtained through sleep-stage determination in a PSG test using only heart rates that are actually measured during sleep with attachment of a PSG device is used. Various kinds of noise are included, examples of which include noise caused by the extrinsic factor such as body movement or detachment of sensors during sleep, and noise caused on the device side when heart rate variability is obtained. To obtain the true heart rates from heart rate data containing these kinds of noise, it has taken 15 years or more for the present inventors to develop a technique for separating a signal obtained from a biological signal from noise. As a result of intensive studies, the present inventors have created this embodiment.
According to this embodiment, the use of a genetic algorithm among these techniques to perform five-class classification leads to the discovery of the importance of each feature value. It has been discovered that to perform five-class classification, the mean RRI, the standard deviation of each RRI, the VLF (0.003-0.04 Hz), the LF (0.04-0.15 Hz), the HF (0.15-0.4 Hz), and the LF/HF ratio are important. This embodiment further makes it possible to simultaneously classify these feature values for all of NN (Neural Network), SVM (Support Vector Machine), RF (Random Forest), RNN (Recurrent Neural Network), and HMM (Hidden Markov Model) classifiers. These learning algorithms can be used differently for different purposes. For example, RF is optimum to increase the rate of classification only for N2, and makes it possible to achieve high recognition, as illustrated in the chart described above, [Table 3]. In contrast, RNN is the most suitable for uniform recognition. According to this embodiment, as illustrated in the chart described above, [Table 4], the result of 66% or more can be obtained.
As described above, once the heart rate is obtained, this information is tagged with sleep information obtained from the PSG device, thereby enabling tagging such that the heart rate feature in sleep stage A is A1, the heart rate feature in sleep stage B is B1, and so on. Accordingly, it is possible to determine a sleep stage only using a heart rate feature. In addition, these results are learned using deep learning, and can thus be applied to various people.
Next, an example according to the embodiment described above was conducted. The example may be similar to part of the embodiment described above.
The human heartbeat interval varies based on the action of sympathetic and parasympathetic nerves. The changes in heart rate variability are typically compared using the LF/HF ratio. With a human's transition from the wake state to the sleep state, the LF/HF ratio decreases. The recovery of the body during sleep decreases the basal metabolic rate and activates the parasympathetic nerves. As a result, the heart rate and its variability are considered to decrease, with LF/HF decreasing. The LF/HF ratio varies from one sleep stage to another. The LF/HF ratio in REM sleep is significantly larger than that in any other sleep stage. It is considered that during REM sleep, the basal metabolic rate increases, the sympathetic nerves are activated, and the LF/HF ratio increases. Additionally, it is said that there is a correlation between heart rate variability and EEG (electroencephalogram) in sleep.
Some researchers (Dumont et al.) showed that there was a cross-correlation between heart rate variability and EEG (electroencephalogram) during respiration by using SL (Synchronization Likehood) (see M. Dumont, F. Jurysta, J. Lanquart, P. Migeotte, P. Borne, and P. Linkowski, “Interdependency between heart rate variability and sleep EEG: linear/non-linear?”, Clinical Neurophysiology, Vol. 115, No. 9, pp. 2031-2040, (2004)).
Other researchers (Yeh et al.) found that there was a fractal property between heart rate variability and EEG during sleep by using DFA (Detrended Fluctuation Analysis) (see J. R. Yeh, C. K. Peng, M. T. Lo, C. H. Yeh, S. C. Chen, C. Y. Wang, P. L. Lee, and J. H. Kang, “Investigating the interaction between heart rate variability and sleep EEG using nonlinear algorithms”, Journal of Neuroscience Methods, Vol. 219, No. 2, pp. 233-239, (2013)). Since a sleep stage is determined mainly from EEG (electroencephalogram), any evidence that there is a correlation between heart rate variability and EEG index suggests the successful determination of a sleep stage using heart rate variability, which was examined in this example.
First, an analysis procedure will be described for determining a sleep stage using ECG (electrooculogram). In the first stage of the analysis procedure, a bandpass filter from 5 to 15 Hz is applied to ECG to remove noise. Thereafter, the ECG is segmented into 30-second zones, and then, to enhance the R-wave peaks, the signal is squared to identify the R-wave. The R-wave is identified by shifting a window having a width of 0.5 seconds in increments of 0.25 seconds. After the R-wave is identified, the RRI is computed to generate RRI time-series data. Since the RRI is not a uniform interval, resampling is performed at a sampling frequency of 4 Hz by using spline interpolation. Then, a power spectral density (PSD) is computed from the RRI time-series data by using an AR model (autoregressive model) (see H. Akaike, “Fitting autoregressive models for prediction”, Annals of the Institute of Statistical Mathematics, Vol. 21, No. 1, pp. 243-247, (1969)).
The power spectral density (PSD) is used to obtain the VLF, the LF, the HF, and the LF/HF ratio. The AR model provides frequency conversion, even for short time-series data, with higher frequency resolution than Fourier transform. In addition, the AR model is also suitable to obtain the deviation in peak frequency. In this example, the PSD is used for these reasons. The AR model is constructed using the Yule-Walker method, and the order of the AR model is selected on the basis of the following Akaike information criterion.
AIC=−2 log f(x,θ)+2p
(where x denotes the observation variable, θ denotes the deviation in prediction error, f(x, θ) denotes the likelihood of the variable, and p denotes the number of variables.)
The above equation is minimized to determine an AR model.
The PSD (power spectral density) is computed by the following equation.
PSD(f)=σ2/|1−Σj=1Maje−2πiff|2 [Math. 5]
(where a denotes the AR (autoregressive) coefficient, and M denotes the order of the AR model: in this example, M was experimentally set to 2.)
After the RRI time-series data is subjected to frequency conversion, 0.003 to 0.04 Hz (VLF), 0.004 to 0.15 Hz (LF), 0.15 to 0.4 Hz (HF), and the LF/HF ratio are computed from a set of power spectral densities. Through the analysis procedure described above, the RRI mean, the standard deviation of each RRI, the VLF, the LF, the HF, and the LF/HF ratio are computed as feature variables.
To obtain a determination criterion in this example, PSG was performed on 50 subjects and a sleep experiment was conducted. Unfortunately, the equipment was separated from five subjects in the middle, and data for the five subjects was missing and excluded. In this example, a sleep stage was estimated using only ECG (electrocardiogram). The estimated sleep stage was matched against correct answer data to compute a recognition rate. Current data that was correctly estimated was used as the determination criterion described above.
It is generally said that, in normal sleep, REM, N2, and N3 occupy 25%, 50%, and 20% of the total sleep time, respectively. In this example, as shown in the following table, REM occupied 21%, N2 occupied 60%, and N3 occupied 10%. In a PSG test, the occurrence ratio of N3, which is deep sleep, appears to be kept low.
The R-wave and the RRI time-series data were identified from the ECG data (see
Subsequently, as illustrated in
Subsequently, a sleep stage was determined using the computed feature values. In this example, RNN, HMM, neural network (NN), support vector machine (SVM), and random forest (RF) were used as classifiers for sleep-stage determination. Whereas RNN and HMM are methods that take into account time-series information of observed values, NN, SVM, and RF handle each observed value as an independent value. The HMM requires vector quantization of feature values, and clustering and vector quantization are carried out using the SOM (see T. Kohonen, “Self-organizing formation of topologically correct feature maps”, Biological Cybernetics, Vol. 43, pp. 59-69, (1982)). Thereafter, learning was performed by the Baum-Welch algorithm, and Viterbi algorithm was used to obtain outputs (see A. J. Viterbi, “Error bounds for convolutional codes and an asymptotically optimum decoding algorithm”, IEEE Transactions on Information Theory, Vol. 13, No. 2, pp. 260-269, (1967)). One of the subjects was used as test data to perform leave-one-out cross-validation (LOOCV), and the remaining subjects were used as training data to evaluate the classifiers. Finally, the recognition rate of each sleep stage was determined to evaluate the classifiers. The recognition rate is expressed in the ratio of the number of correctly classified classes to the total number of classes.
The recognition results of the sleep stages using heart rate variability are illustrated in
Tables 6 and 7 below show results of recognition rates separately for learning and testing. RNN and HMM were applied to the test data.
Furthermore, NN, SVM, and RF were applied for comparison with the method according to this embodiment. As a result, outcomes shown in Table 7 below were obtained.
The outcomes indicate that the method according to this embodiment (“Proposed” in the table) achieved relatively higher recognition accuracy than using RNN as a classifier. Compared to various methods, a higher recognition rate was obtained than that in a method that does not take into account a time information method. Since sleep stages do not markedly change, the use of time-series information may improve the recognition rate. In NN and SVM, learning was performed to minimize classification errors in the entire data, with the result that other classes were erroneously classified into N2 for which the number of pieces of data is large. In this example, the number of pieces of N2 data was large, whereas the number of pieces of N1 data and the number of pieces of N3 data were small. The increase in the number of pieces of data leaded to an increase in recognition rate.
As a result of comparison between the method according to this embodiment and the methods of the related art, the outcomes shown in Table 7 were obtained (see NPLs 1 to 3 and T. Willemen, D. V. Deun, V. Verhaert, M. Vandekerckhove, V. Exadaktylos, J. Verbraecken, S. V. Huffel, B. Haex, and J. V. Sloten, “An evaluation of cardiorespiratory and movement features with respect to sleep-stage classification”, IEEE Journal of Biomedical and Health Informatics, Vol. 18, No. 2, pp. 661-669, (2014)). RIP (respiratory inductance plethysmography) measures the respiratory flow rate, and ACT (actigraphy) measures the movement of the body, which is measured by accelerometers on the arms or ankles.
There is no reported criterion for five stages of sleep using ECG. Many determinations have been based on four stages of sleep, in which N1 and N2 are combined into a single stage. N1, which is light sleep, is important as a sleep stage. Some studies have reported that Apnea-hypopnea index (AHI) and N2 have a significant correlation and have a correlation with the maximum dimension of excessive daytime sleepiness EDS. Furthermore, the results of multivariate regression analysis have reported that AHI-N1 has a significant correlation with PSG factors (sleep efficiency, sleep urgency) and that AHI-N2 has a significant correlation with the chief complaint such as daytime sleepiness (drowsiness) or insomnia (see Denis Martinez, Magali Santos Lumertz, and Maria do Carmo Sfreddo Lenz, “Dimensions of sleepiness and their correlations with sleep-disordered breathing in mild sleep apnea”, J Bras Pneumol. 2009; 35(6):507-514). It is therefore considered that recognition of five sleep stages, in which the N1 and N2 sleep stages are separate, is important.
In this example, five sleep stages were recognized using only ECG. The methods of the related art are methods for four-stage recognition in which N1 and N2 are combined into a single stage. In contrast, the method according to this embodiment can identify five stages in which N1 and N2 are separate. According to this embodiment, as a result of identifying four stages in a similar way to that in the methods of the related art that use RIP and ACT data in addition to ECG, outcomes with higher recognition rate than those in the methods of the related art can be obtained. According to this embodiment, furthermore, it is possible to identify five stages in a similar way to that in the methods of the related art for identifying four stages. One of the state-of-the-art research findings indicates that the REM sleep stage (REM), the first non-REM sleep stage (N1), the second non-REM sleep stage (N2), the third non-REM sleep stage (N3), and the wake state can be identified with accuracies of 0.86, 0.93, 0.69, 0.90, and 0.93, respectively, as the recognition rates for the sleep stages. There has been no report for an identification method for five stages of sleep using ECG. Many methods of the related art merely provide identification of four stages in which N1 and N2 are regarded as a single stage. For sleep disorders or EDS, N1 and N2 need to be separated, and therefore the five-stage identification method according to this embodiment has high value in use. In this example, the recognition rate for N1 was low because the number of pieces of N1 data was not so large, whereas the recognition rate for N2 was as high as 90% or more. It is considered that the use of data including a large number of pieces of N1 data further improves the recognition rate for N1.
The description of this embodiment is complete.
While an embodiment of the present invention has been described, the present invention may be carried out in various different embodiments, other than the embodiment described above, without departing from the scope of the technical idea defined in the claims.
For example, the sleep determination apparatus 100 has been described as performing processing in a standalone mode. However, the sleep determination apparatus 100 may be configured to perform processing in response to a request from a client terminal (such as the external device 200) and to return the result of the processing to the client terminal.
Furthermore, among the processes described in the embodiment, all or some of the processes described as being automatically performed may be performed manually, or all or some of the processes described as being manually performed may be performed automatically using a known method.
In addition, the processing procedures, the control procedures, the specific names, the information including registered data and parameters, such as search conditions, for each processing operation, the example screens, and the database configurations, which are provided in the literatures described above or illustrated in the drawings, may be changed, as desired, unless otherwise stated.
The illustrated components of the sleep determination apparatus 100 are functionally conceptual, and need not be physically configured as illustrated in the drawings.
For example, the processes and functions of the devices in the sleep determination apparatus 100, in particular, the processes and functions performed by the control unit 102, may be implemented in part or in whole using a CPU (Central Processing Unit) and a program interpreted and executed by the CPU, or may be implemented in part or in whole as wired-logic-based hardware. The program is recorded on a non-transitory computer-readable recording medium described below, which includes instructions programmed to cause a computer to perform a method according to the present invention. The program is mechanically loaded into the sleep determination apparatus 100, as necessary. That is, the storage unit 106, such as a ROM or an HDD (Hard Disk Drive), has recorded thereon a computer program that cooperates with an OS (Operating System) to provide instructions to the CPU to perform various processing operations. The computer program is executed by being loaded into the RAM. The computer program cooperates with the CPU to form a control unit.
The computer program may be stored in an application program server connected to the sleep determination apparatus 100 via any desired network 300, and may be downloaded in full or in part, as necessary.
A program according to the present invention may be stored in a computer-readable recording medium, or may be configured as a program product. The “recording medium”, as used here, includes any “portable physical medium” such as a memory card, a USB memory, an SD card, a flexible disk, a magneto-optical disk, a ROM, an EPROM, an EEPROM, a CD-ROM, an MO, a DVD, and a Blu-ray (registered trademark) Disc.
The “program” refers to a data processing method written in any desired language or writing method, and may be in the format of source code or binary code or in any other format. The “program” is not necessarily limited to a program configured as a single unit, and includes a program configured in a distributed manner as a plurality of modules or libraries, or includes a program that cooperates with a separate program, such as an OS (Operating System), to achieve its function. A specific configuration of each device provided in the embodiment for reading a recording medium, a reading procedure, or an installation procedure or the like after reading may be implemented using a known configuration or procedure. The program may be configured as a program product recorded on a non-transitory computer-readable recording medium, and may constitute the present invention.
Various databases and the like (such as the heart rate variable file 106a and the determination technique file 106b) stored in the storage unit 106 are storage means such as a memory device such as a RAM or a ROM, a fixed disk device such as a hard disk, a flexible disk, and an optical disk, and store various programs, tables, databases, web page files, and so on to be used to perform various processing operations or provide a website.
The sleep determination apparatus 100 or the external device 200 may be configured as a known information processing apparatus such as a personal computer or a workstation, or may be configured such that any peripheral device is connected to the information processing apparatus. Alternatively, the sleep determination apparatus 100 or the external device 200 may be implemented by loading, to the information processing apparatus, software (including a program, data, and so on) for implementing the method according to the present invention.
Moreover, the devices may be distributed or integrated in any form other than that illustrated in the drawings. All or some of the devices may be configured to be distributed or integrated functionally or physically in any unit in accordance with various loads or the like or in accordance with the functional load. That is, any combination of embodiments among the embodiments described above may be implemented, or an embodiment may be selectively implemented.
As described above in detail, according to the present invention, it is possible to provide a sleep determination apparatus, a sleep determination method, and a sleep determination program that provide an accurate and detailed determination of a sleep stage.
Number | Date | Country | Kind |
---|---|---|---|
2017-110495 | Jun 2017 | JP | national |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/JP2018/021420 | 6/4/2018 | WO | 00 |