WO2021230223A1

WO2021230223A1 - Sleep-waking determination system, sleep-waking determination device, and program

Info

Publication number: WO2021230223A1
Application number: PCT/JP2021/017807
Authority: WO
Inventors: 泰己上田; 晃士大出; 蕭逸史; 健太郎三井; 真知子香取
Original assignee: 国立大学法人東京大学
Priority date: 2020-05-11
Filing date: 2021-05-11
Publication date: 2021-11-18
Also published as: JPWO2021230223A1

Abstract

This sleep-waking determination system comprises: a respiration data acquisition device that is worn by a subject and that acquires respiration time series data regarding the subject; and a sleep-waking determination device that performs sleep-waking determination using the respiration time series data. The sleep-waking determination device comprises a feature amount acquisition unit for extracting at least one feature amount for use in sleep-waking determination from the respiration time series data and a sleep-waking determination unit that performs sleep-waking determination based on the extracted feature amount using a machine learning model.

Description

Sleep-wake determination system, sleep-wake determination device, and program

Cross-reference of related applications

This application is based on Japanese Patent Application No. 2020-083064 filed on May 11, 2020, and the contents of the description are incorporated herein by reference.

The present invention relates to a sleep / wake determination system, a sleep / wake determination device, and a program.

Dyssomnia has been shown to be closely related to mental illnesses such as depression and lifestyle-related diseases such as obesity, and the importance of healthy sleep is recognized. It has become clear that the amount of sleep required to lead a healthy life varies greatly among individuals, and for this reason, there is an increasing need to measure the sleep state of an individual easily and accurately. For example, to determine if you have sleep apnea, you need to know the length of the apnea episode in your sleep state, and it is important to make a rigorous determination of whether you are in a sleep state or a wakeful state. Is.
Highly accurate determination of sleep-wakefulness can be performed based on electroencephalogram EMG. However, the measurement of EEG EMG was not suitable for measurement at home because it requires assistance by a skilled person such as wearing multiple electrodes. On the other hand, a method for determining the sleep-wakefulness state without using an electroencephalogram EMG has been proposed. For example, Patent Document 1 describes estimating a sleep state based on a feature amount extracted from measurement data of respiratory movement and body movement measured by using a wearable sensor. Further, Patent Document 2 describes that a sleep state is determined by using a body motion signal in a respiratory motion detected by a body motion detection sensor.

Japanese Unexamined Patent Publication No. 2017-169884 Japanese Unexamined Patent Publication No. 2009-172197

However, with the conventional method, it was not possible to make a sufficiently accurate and robust sleep / wake determination using data based on the respiratory movements of the subject.

The present invention has been made in view of the circumstances described above, and one of the objects of the present invention is to perform a sleep / wake determination with high accuracy and robustness using respiratory data.

The sleep / awakening determination system according to the embodiment of the present invention is attached to a subject and has a breathing data acquisition device that acquires the breathing time series data of the subject, and a sleep / awakening that makes a sleep / awakening determination using the breathing time series data. The sleep / wake determination device includes a determination device, and the sleep / wake determination device is extracted from the breathing time series data by using a feature amount acquisition unit that extracts one or more feature amounts used for sleep / wake determination and a machine learning model. It is provided with a sleep-wake-up determination unit that performs sleep-wake-up determination based on the feature amount.

The sleep / awakening determination device according to the embodiment of the present invention is a sleep / awakening determination device that determines sleep / awakening, and sleeps from the breathing time series data of the subject acquired by the breathing data acquisition device attached to the subject. It is provided with a feature amount acquisition unit that extracts one or more feature amounts used for awakening determination, and a sleep / awakening determination unit that performs sleep / awakening determination based on the extracted feature amounts using a machine learning model. ..

The program according to the embodiment of the present invention has one or more features in which a computer for determining sleep / wakefulness is used for sleep / wakefulness determination from the respiratory time series data of the subject acquired by a respiratory data acquisition device attached to the subject. It functions as a feature amount acquisition unit for extracting a quantity and a sleep / wakefulness determination unit for performing a sleep / wakefulness determination based on the extracted feature amount using a machine learning model.

According to the present invention, it is possible to perform sleep / wake determination with high accuracy and robustness using respiratory data.

The figure which shows the structure of the sleep awakening determination system 1 which concerns on embodiment of this invention. The figure which illustrates the breathing data acquisition apparatus 20 which concerns on embodiment of this invention. The flowchart which shows the whole flow of the sleep awakening determination process which concerns on embodiment of this invention. The figure which exemplifies the correct answer rate at the time of making the sleep awakening determination using the 13 feature quantities which concerns on embodiment of this invention. The figure which exemplifies the correct answer rate at the time of making the sleep awakening determination using any 4 feature amounts of 13 feature amounts which concerns on embodiment of this invention. The flowchart of the procedure for acquiring the feature amount of the frequency domain corresponding to the respiratory rate of the subject which concerns on embodiment of this invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The same elements are designated by the same reference numerals, and duplicate description will be omitted.

Embodiment FIG. 1 is a diagram showing a configuration of a sleep / wake determination system 1 according to the present embodiment. As shown in FIG. 1, the sleep / wake determination system 1 includes a sleep / wake determination device 10 and a respiratory data acquisition device 20. The sleep / wake determination device 10 and the respiratory data acquisition device 20 are connected via the communication network N. The communication network N may be, for example, any of the Internet, LAN, leased line, telephone line, mobile communication network, Bluetooth (registered trademark), WiFi (Wireless Fidelity), other communication lines, and combinations thereof. , Wired or wireless. Note that data and the like may be exchanged between the sleep / wake determination device 10 and the respiratory data acquisition device 20 using a physical medium without going through the communication network N. For example, the measurement data acquired by the breathing data acquisition device 20 may be stored in a recording medium such as a USB memory or a flash card, and the stored data may be read by the sleep / wake determination device 10.

The sleep / wake determination device 10 makes a sleep / wake determination using the subject's breathing data (breathing time series data) measured by the breathing data acquisition device 20. The sleep / wake determination device 10 is a general-purpose computer, and may be composed of one computer or a plurality of computers distributed on the communication network N. The sleep / wake determination device 10 includes a control device 11 (feature amount acquisition unit, sleep / wake determination unit) and a storage device 12. The control device 11 includes, as hardware, a CPU, a memory such as a ROM or RAM, an input interface, an output interface, a communication interface, and a bus connecting them. The control device 11 realizes various functions by the CPU executing a program stored in a ROM or the like. The storage device 12 is a hard disk drive or the like, and stores various programs as well as respiratory data measured by the respiratory data acquisition device 20. The respiratory data such as respiratory data may be stored in a storage device provided in the respiratory data acquisition device 20.

The breathing data acquisition device 20 is a device for acquiring breathing data of a subject, and is a device for measuring breathing and respiratory effort such as a sleep evaluation device used in a sleep disorder test. Specifically, as shown in FIG. 2, the time-series changes in the movement of the respiratory muscles are measured by the belt B1 attached to the abdomen of the subject and the belt B2 (inductance type respiratory plethysmography band, etc.) attached to the chest. .. In this embodiment, as the breathing data, data on the movement of the chest and abdominal muscles (breathing effort) performed for breathing is used. It should be noted that, for reasons such as sleep apnea, even if breathing efforts are made, breathing may not actually be performed smoothly. Further, as the respiration data, the airflow and temperature data measured via the airflow sensor S1 attached to the nose of the subject and the airflow sensor S2 attached to the mouth may be used. The measured data may be recorded in the data recording device R.

FIG. 3 is a flowchart showing the overall flow of the sleep / wake determination process by the sleep / wake determination system 1 according to the present embodiment. As shown in FIG. 3, first, the respiratory data of the subject is measured by the respiratory data acquisition device 20 (step S101). The measurement is performed throughout a certain period of time (for example, all night). Next, the sleep / wake determination device 10 preprocesses the measured respiratory data (step S102). Specifically, the respiratory data measured over a fixed time is divided into epochs in a fixed time unit (for example, 30 seconds).

Next, the sleep / wake determination device 10 acquires the feature amount used for the sleep / wake determination from the respiratory data (step S103). In this embodiment, 13 types of feature quantities are used. These features can be divided into three groups: a feature amount in the time domain, a feature amount in the frequency domain, and a feature amount in the frequency domain according to the respiratory rate of the subject. Details of each feature will be described later. The feature amount used for determining sleep / wakefulness is not limited to that given in this embodiment, and may be any feature amount that can be extracted from the respiratory data.

Next, the sleep / wake determination device 10 makes a sleep / wake determination using a machine learning model based on the extracted features (step S104). As the machine learning algorithm, logistic regression, random forest, XGBoost, multi-layer perceptron and the like can be used. In creating the machine learning model, the feature amount extracted from the respiratory data collected by using the sleep / wake determination device 10 can be used as input data for training. Further, as the correct answer data (determination result of sleep or awakening) for the input data, for example, the result determined by the electroencephalogram electromyogram can be used. The machine learning model is an algorithm for sleep / wake determination generated in advance using machine learning, and when actually performing sleep / wake determination, the trained model is mounted on the sleep / wake determination device 10. There is.

Next, the feature amount used for sleep / wake determination will be described in detail. The feature amount used in the present embodiment can be divided into three groups: a feature amount in the time domain, a feature amount in the frequency domain, and a feature amount in the frequency domain according to the respiratory rate of the subject. Each will be described below.

(Features in the time domain)
Time domain features are associated with information such as the subject's respiratory amplitude. In the present embodiment, as the feature amount in the time domain, the average, the variance, the maximum value of the absolute value, and the number of zero crossings in one epoch of the respiratory data are used. In this embodiment, all of these values may be used as feature quantities, or some of them may be used.

(Frequency domain features)
Frequency domain features are associated with information such as the subject's respiratory rate. In the present embodiment, the feature quantities in the frequency domain include spectral entropy, spectral center of gravity, spectral flux, spectral roll-off, power in an ultra-low frequency band (for example, 0 to 0.06 Hz), and low frequency band (for example, 0 to 0.06 Hz) in one epoch of respiratory data. Power in the middle frequency band (for example, 0.15 to 0.5Hz), power in the middle frequency band (for example, 0.15 to 0.5Hz), and since many frequency components derived from respiration are observed in the middle frequency band, the peak frequency and peak in the middle frequency band. Use power at frequency.

Of these, spectral entropy, spectral centroid, spectral flux, and spectral roll-off are each defined by the following equations. Here, Z _F ^(k) is a series obtained by normalizing the absolute value of _{the series X F} ^(k) obtained by fast Fourier transforming the time series X _T ^{(k) of the kth epoch into the range [0,1].} ..

For the threshold used to calculate the spectral roll-off, leave-one-out cross-validation was performed using values of 0.50, 0.55, ..., and 0.95, respectively, and the accuracy of sleep-wake determination was the highest. Although 0.95 is adopted, the threshold value can be set to any value.

The spectral entropy, spectral center of gravity, spectral flux, and spectral roll-off are also mentioned in the document "Gu et al.," Intelligent sleep stage mining service with smartphones. "10.1145/2632048.2632084". In addition, the power in the ultra-low frequency band (for example, 0 to 0.06 Hz), the power in the low frequency band (for example, 0.06 to 0.15 Hz), the power in the middle frequency band (for example, 0.15 to 0.5 Hz), and the peak frequency in the middle frequency band. And the power at the peak frequency is also mentioned in the document "Tataraidze et al.," Bioradiolocation-based sleep stage classification. "10.1109 / EMBC.2016.7591321.".

In the present embodiment, the sleep / wakefulness determination was performed using 13 feature quantities including the above-mentioned feature quantities in the four time domains and the feature quantities in the nine frequency regions. FIG. 4 is a diagram illustrating the correct answer rate when the sleep / wake determination is performed using the 13 feature quantities.

In the example of FIG. 4, the judgment is made by the model that has been machine-learned by XGBoost. The horizontal axis of the graphs (1) to (3) in FIG. 4 indicates each feature amount. The graph of (1) shows the weight of each feature amount in XGBoost. The graph of (2) shows the accuracy rate when the judgment is made with 12 feature quantities other than the feature quantity on the horizontal axis. The graph of (3) shows the accuracy rate when the judgment is made only by the feature amount on the horizontal axis. In the graphs (2) and (3), the left end of the horizontal axis shows the correct answer rate when the judgment is made using all 13 feature quantities. As is clear from (2) of FIG. 4, when any one of the feature amounts was removed from the 13 feature amounts, a correct answer rate of 90% was obtained. Further, even when the judgment is performed using only one feature amount, as is clear from (2) of FIG. 4, a correct answer rate of 65% or more is obtained.

Further, FIG. 5 is a diagram showing the correct answer rate when a judgment is made using any four of the 13 feature quantities. In the example of FIG. 5, 18 patterns in which the correct answer rate is 90% or more out of all the combinations (715 patterns) in which 4 are selected from the 13 feature quantities are shown. The features marked with "○" in the matrix shown at the bottom of the graph are the features included in the four. From the example of FIG. 5, it can be seen that the combinations that obtain a high accuracy rate tend to include power (low frequency) and roll-off.

(Characteristics in the frequency domain according to the respiratory rate of the subject)
The feature amount in the frequency domain corresponding to the respiratory rate of the subject is a power spectrum around the peak frequency (corresponding to the respiratory rate of the subject) in the epoch determined that the subject is sleeping. In the present embodiment, it is determined whether or not the subject is sleeping by using a machine-learned classifier. In this embodiment, as an example, logistic regression that does not require parameter search and can be calculated at high speed is used as a classifier.

Using the flowchart of FIG. 6, a procedure for acquiring a feature amount in the frequency domain according to the respiratory rate of the subject from the respiratory data of the subject will be described.

First, the respiration data of one subject (subject A) is set as verification data, and the respiration data of a plurality of other subjects is set as training data (step S201). Next, for the training data, learning of the classifier is performed using the feature amount in the frequency domain corresponding to the respiratory rate of each subject (step S202).

The feature amount in the frequency domain according to the respiratory rate of the subject will be described. The respiratory data is divided into epochs (step S102 in FIG. 3), and data for a certain period of time before and after is added to the data of each epoch. This is to increase the frequency resolution of the fast Fourier transform. For example, to the data of 1 epoch 30 seconds, the data of 15 seconds before and after is added, and the data of 60 seconds X _T = (X ₁ , X ₂ , ..., X _N ). Here, N is N = 32 × 60 = 1920, for example, assuming that the sampling frequency of the respiratory data is 32 Hz.

Then, since the X _T excluding the DC component, X _1, X _2, ..., value average Xa of the X _N was subtracted from all the components of _{_{_{X T Xa T = (X 1}}} -Xa, X 2 -Xa , ..., X _N- Xa). Furthermore, Hanning window of the same length as _{Xa T W = (W1, W2} , ..., WN) prepared, the product of each element of Xa _T and W

Ask for.

Is a time series multiplied by a window function, and a series obtained by performing a fast Fourier transform on this.

Is a frequency series with a frequency interval of 32/1920 = 1 / 60Hz and a maximum frequency of 16Hz.

Of the above X _F _{, take 60 points X F1} , X _F2 ..., X _F60 from the range where the spectrum of respiratory effort is significantly included (for example, 0Hz to 1Hz), and f _i = log (1+). | X _Fi | 2), i = 1,2, ..., 60 are the features of the frequency domain according to the respiratory rate of the subject. In order to avoid the instability of the calculation when the value of the power spectrum | X _Fi | 2 becomes 0, the logarithm is taken after adding 1 to _{| X Fi | 2.}

The classifier determines whether or not it is sleeping based on the above features, and if it is actually sleeping data, the correct answer is taken. When the learning of the classifier is completed by the above procedure, the process proceeds to step S203, and the learned classifier is used to determine sleep / wakefulness for the verification data.

Next, among the epochs of the verification data, only the epochs judged to be sleep by the trained classifier are collected, and the average fa1, ... ., fa60 is calculated (step S204).

Next, ipeak = arg max fai is set as the peak frequency for the subject to be verified, and 19 points from _fipeak-9 _{to fipeak + 9 are extracted for all epochs of the verification data.} The extracted value is used as the feature amount of the verification data (step S205).

By the above procedure, 19-dimensional features are acquired from the respiratory data of the subject. As for how many points to include from before and after the peak frequency, parameter search is performed while changing the number m of points taken from the front and the number n of points taken from the back, and m = n as the value with the highest classification performance. = 9 is selected, but the number of points to be included is not limited to 9.

It is not necessary to perform the learning step of the classifier in steps S201 to S202 every time, and normally, the process of steps S203 to S205 is performed on the subject data for which the sleep / wake determination is to be performed, using the trained classifier. Is executed, and the feature amount of the frequency domain corresponding to the respiratory rate of the subject is acquired.

As described above, according to the present embodiment, the breathing data is acquired by the breathing data acquisition device 20 attached to the subject, and the sleep awakening is determined based on the feature amount extracted from the breathing data. As a result, sleep / wake determination can be performed with high accuracy and robustness using data acquired by a device that can be easily used even at home. The breathing data acquisition device 20 can acquire accurate data even in an environment where there are people sleeping nearby, such as a family member, unlike a sheet-type sensor that is laid on a bed and used. Therefore, it is also suitable for screening for sleep apnea syndrome at home.

In addition, since the feature amount in the time domain and the feature amount in the frequency domain of the respiratory data are used as the feature amount, it is possible to make a highly accurate judgment using the feature amount that can be easily extracted from the respiratory data.

In addition, since the feature amount in the frequency domain corresponding to the respiratory rate of the subject is used as the feature amount, it is possible to perform highly accurate sleep / wake determination according to the characteristics of the individual subject.

It should be noted that the present invention is not limited to the above-described embodiment, and can be carried out in various other forms within a range that does not deviate from the gist of the present invention. For this reason, the above embodiments are merely examples in all respects and are not to be construed in a limited manner. For example, the above-mentioned processing steps can be arbitrarily changed in order or executed in parallel within a range that does not cause a contradiction in the processing contents.

1 ... Sleep awakening determination system 10 ... Sleep awakening determination device 11 ... Control device 12 ... Storage device 20 ... Respiratory data acquisition device N ... Communication network

Claims

A respiratory data acquisition device that is attached to the subject and acquires the respiratory time series data of the subject,
A sleep / awakening determination device that determines sleep / awakening using the respiratory time series data is provided.
The sleep / wake determination device is
A feature amount acquisition unit that extracts one or more feature amounts used for sleep / wake determination from the respiratory time series data, and a feature amount acquisition unit.
A sleep / awakening determination system including a sleep / awakening determination unit that determines sleep / awakening based on the extracted feature amount using a machine learning model.
The feature amount acquisition unit is
From the breathing time series data, at least one of the feature amount in the time domain related to the respiratory amplitude, the feature amount in the frequency domain related to the respiratory rate, and the feature amount in the frequency domain corresponding to the respiratory rate of the subject. The sleep-wake determination system according to claim 1, which extracts a feature amount.
The sleep according to claim 2, wherein the feature amount in the time domain includes at least one of the average, the variance, the maximum value of the absolute value, and the number of zero crossings in the time per epoch for determining sleep / wakefulness. Awakening judgment system.
The feature amount in the frequency domain includes at least one of spectral entropy, spectral centroid, spectral flux, spectral roll-off, power in a predetermined frequency band, peak frequency in a predetermined frequency band, and power at the peak frequency. 2. The sleep / wake determination system according to claim 2.
The sleep-wake determination system according to claim 2, wherein the feature amount in the frequency domain corresponding to the respiratory rate of the subject includes a power spectrum around the peak frequency in the epoch determined that the subject is sleeping.
It is a sleep / awakening determination device that determines sleep / awakening.
A feature amount acquisition unit that extracts one or more feature amounts used for sleep / wake determination from the breathing time series data of the subject acquired by the respiratory data acquisition device attached to the subject, and a feature amount acquisition unit.
A sleep / awakening determination device including a sleep / awakening determination unit that determines sleep / awakening based on the extracted feature amount using a machine learning model.
A computer that makes sleep / wake judgments,
A feature amount acquisition unit that extracts one or more feature amounts used for sleep / wake determination from the breathing time series data of the subject acquired by the respiratory data acquisition device attached to the subject, and a feature amount acquisition unit.
A program that uses a machine learning model to function as a sleep / awakening determination unit that determines sleep / awakening based on the extracted features.