WO2019039261A1 - Sleep quality assessment system, sleep quality model creation program, and sleep quality assessment program - Google Patents

Sleep quality assessment system, sleep quality model creation program, and sleep quality assessment program Download PDF

Info

Publication number
WO2019039261A1
WO2019039261A1 PCT/JP2018/029517 JP2018029517W WO2019039261A1 WO 2019039261 A1 WO2019039261 A1 WO 2019039261A1 JP 2018029517 W JP2018029517 W JP 2018029517W WO 2019039261 A1 WO2019039261 A1 WO 2019039261A1
Authority
WO
WIPO (PCT)
Prior art keywords
sleep
model
likelihood
unit
sound
Prior art date
Application number
PCT/JP2018/029517
Other languages
French (fr)
Japanese (ja)
Inventor
福井　健一
ホングル ウ
隆史 加藤
正行 沼尾
Original Assignee
国立大学法人大阪大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国立大学法人大阪大学 filed Critical 国立大学法人大阪大学
Priority to JP2019538050A priority Critical patent/JPWO2019039261A1/en
Publication of WO2019039261A1 publication Critical patent/WO2019039261A1/en

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/16—Devices for psychotechnics; Testing reaction times ; Devices for evaluating the psychological state

Definitions

  • a sleep quality determination system that determines the quality of sleep based on sounds emitted by a living person or animal living in sleep or sounds generated by movement of a living body, a sleep quality model creation program, and a sleep quality determination program .
  • sleep disorders caused by inadequate sleep quality and quantity such as sleeplessness, cause physical, mental, social, and emotional functions to be disturbed during awakening. Therefore, assessing sleep conditions such as sleep quality is important information to restore these functions.
  • sleep state brain waves, breathing and heart rate change.
  • various events such as body movement, snoring, apnea and bruxism also occur during sleep in association with changes in sleep state and stability.
  • sleep stages such as REM / REM sleep are effective for sleep disorder determination and sleep science research, but are inappropriate as information to be presented to general users.
  • the present invention has been made in view of the above problems, and can acquire data even in a non-contact state with a living body, and can easily determine the quality of sleep instead of the sleep stage by machine learning data over a long period of time Quality determination system, a sleep quality model creation program, and a sleep quality determination program.
  • the sleep quality determination system which is one of the present invention is characterized by a sound recording unit which records a sound emitted by a living body during sleep as a sleep sound and a feature of the recorded sleep sound.
  • An extraction unit that extracts a plurality of pieces of feature information that is a part, an analysis unit that calculates each feature vector based on the extracted feature information, a classification unit that classifies the plurality of feature vectors into a plurality of types, sleep quality Processing the classified feature vector of the sleep sound in the first state by using the hidden Markov model, the first model being one of probabilistic models, and the sleep sound in which the sleep state is the second state
  • a probabilistic model generation unit for generating a second model which is one of the probabilistic models by processing the classified feature vectors of the second group using the hidden Markov model, the first model, and the second model
  • the sound recording unit records sleep sound to be determined, the extraction unit extracts feature information from the sleep sound, and the analysis unit extracts a feature vector from the feature information.
  • the classification unit classifies the feature vector, and further applies the classified feature vector of the determination target to the first model and applies the first likelihood as a likelihood and the second model And the second state of the sleep quality of the judgment target on the basis of the first likelihood and the second likelihood thus calculated. And a determination unit that determines whether the
  • a probabilistic model for determining the quality of sleep based on a large number of feature information included in the sleep sound recorded for a long time is generated, and the probabilistic model is applied to the sleep sound to be determined. This makes it possible to accurately determine the sleep state.
  • the classification unit may perform cluster analysis on a plurality of the feature vectors to automatically classify the plurality of feature vectors into a small number of categories, and assign the same label to the feature vectors belonging to each category.
  • classification of feature information can be performed by computer processing, and a probability model based on a large amount of sleep sounds can be easily generated.
  • the probability model it is possible to determine the quality of sleep with high accuracy.
  • the probability model generation unit may generate a first model of N (N is an integer) and a second model in which the number of hidden variables is one in a range of 3 or more and 9 or less.
  • the probability model generation unit generates a first model and a second model in which the number of hidden variables is N ⁇ 1, N, and N + 1, and the likelihood calculation unit generates the generated first model, And calculating the first likelihood and the second likelihood for each of the second models, and the determination unit compares the first likelihood and the second likelihood corresponding to the number of the same number of hidden variables Then, the one with high likelihood may be won and it may be determined by majority decision whether the sleep quality is in the first state or the second state.
  • the probability model generation unit generates a first model and a second model for the number of the plurality of hidden variables, and the likelihood calculation unit generates the generated first model and the second model. For each, the first likelihood and the second likelihood are calculated, and the determination unit determines the calculated likelihood as a likelihood vector and determines whether the sleep quality is in the first state or the second state by pattern identification. You may
  • the pattern identification is a classifier for supervised learning
  • the determination unit includes a likelihood model generation unit that generates a model corresponding to the classifier, and the likelihood is based on the generated likelihood model. It is also possible to determine whether the sleep quality is in the first state or the second state using a vector as an input. According to this, it is possible to improve the accuracy of the determination.
  • the terminal device includes: a probability model generation device including the probability model generation unit; and a plurality of terminal devices connected to the probability model generation device via a network, the terminal device further includes the sound recording unit. You may provide the result presentation part which makes a display apparatus show the determination result by a determination part.
  • the server centrally manages the sleep sounds recorded by the plurality of terminal devices, it is possible to generate a more appropriate probability model.
  • a sleep quality model creating program is a sound recording unit for recording as a sleep sound a sound emitted by a living body during sleep, and the recorded sleep sound.
  • An extraction unit that extracts a plurality of pieces of feature information that are feature portions from among them, an analysis unit that calculates each feature vector based on the extracted feature information, and a classification unit that classifies the plurality of feature vectors into a plurality of types;
  • the state of sleep was in the second state, the first model being one of the probabilistic models by processing the classified feature vectors of the sleep sound in which the quality of sleep was the first state using the hidden Markov model
  • a probability model generation unit for generating a second model which is one of the probability models by processing the classified feature vectors of the sleep sound using a hidden Markov model, the first model, and And a storage unit for storing a second model, to realize the above-mentioned respective units by executing the respective processes in the computer.
  • the sleep quality determination program is a sound recording unit for recording a sound emitted by a living body during sleep as a sleep sound, and in the recorded sleep sound.
  • An extracting unit that extracts a plurality of pieces of feature information that are feature portions from the image, an analyzing unit that calculates each feature vector based on the extracted feature information, a classifying unit that classifies a plurality of the feature vectors into a plurality of types;
  • a likelihood calculating unit that calculates a second likelihood that is a likelihood by applying to the second model and a first likelihood that is a degree, and the first likelihood and a second likelihood that are calculated Quality of sleep to be judged
  • a determination section for determining a first state or a second state, to realize the above-mentioned respective units by executing the respective processes in the computer.
  • kernel function used in the claims and the specification etc. may not be strictly included in the kernel function because it does not satisfy positive definiteness, but the claims and the specification Etc. are described as being included in the kernel function.
  • FIG. 1 is a block diagram showing a functional configuration of a sleep quality determination system.
  • the sleep quality determination system 100 is a system that determines sleep quality based on the generated probability model, and includes a sound recording unit 110, an extraction unit 120, an analysis unit 130, a classification unit 140, and a probability model generation unit. 150, a storage unit 160, a likelihood calculation unit 170, and a determination unit 180. In the case of the present embodiment, the sleep quality determination system 100 further includes a result presentation unit 190.
  • the sound recording unit 110 is a processing unit that records the sound emitted by the living body 200 during sleep in the storage device 119 as the sleep sound 210.
  • the sleep sound 210 also includes environmental sound around the living body 200.
  • the sound recording unit 110 is connected to the microphone 111 disposed in the vicinity of the living body 200 wirelessly or by wire, and the microphone 111 converts the sound emitted by the living body 200 and the surrounding sound.
  • the sound recording unit 110 converts an analog electrical signal into a digital signal and stores the digital signal in the storage device 119.
  • the sound recording unit 110 is provided in each of the plurality of terminal devices 101 connected to the probability model generation device 199 via the network 300 or the like, and the sound recording unit 110 is the entire sleep quality determination system 100. There are multiple.
  • the method of storing the sleep sound 210 of the sound recording unit 110 is not particularly limited.
  • a semiconductor storage device such as a flash memory and a magnetic storage device such as a hard disk
  • the sleep sound 210 is recorded.
  • the sleep sound 210 recorded by the sound recording unit 110 may be continuously recorded from when the living body 200 sleeps until it wakes up, or may be intermittently recorded at a predetermined interval.
  • the position to which the microphone 111 is attached is not particularly limited as long as it can acquire the sound emitted by the living body 200, and may be attached to the ceiling, floor or wall of a room where the living body 200 is sleeping. In the case of the present embodiment, the microphone 111 is attached to the top of the bed of the bed where the living body 200 is sleeping. In addition, it is preferable to arrange
  • the living body 200 may be not only human but also animals other than human.
  • the sound emitted by the living body 200 is a sound generated based on movement (body movement) of the living body 200 such as noise from the mouth or throat of the living body 200, such as bruises, snoring, sleeping sounds, rubbing noise of a futon or squeaking noise of a bed.
  • the sound around the living body 200 is a sound emitted from equipment around the living body 200 such as a cooler or a humidifier, or a car sound reaching around the living body 200 through a window or the like.
  • FIG. 2 is a diagram showing a part of the recorded sleep sound.
  • the extraction unit 120 is a processing unit that extracts a plurality of pieces of feature information 201 that are characteristic portions from the sleep sound 210 recorded by the sound recording unit 110.
  • the method for the extraction unit 120 to extract the feature information 201 is not particularly limited, and the feature information 201 may be extracted by a statistical method, and a predetermined method based on the detection time of a sound wave having an amplitude of a predetermined threshold or more.
  • the sound window included in the time interval of (1) may be extracted as the feature information 201.
  • the extraction unit 120 extracts a continuous period of burst level 1 or more as one feature information 201 by using Kleinberg's burst extraction method among statistical methods, and the noise recorded regularly. Was eliminated.
  • the number of pieces of feature information 201 extracted by the extraction unit 120 from the sleep sound 210 included in a predetermined time was n.
  • the analysis unit 130 is a processing unit that calculates feature vectors respectively indicating the plurality of pieces of feature information 201 extracted by the extraction unit 120.
  • the method for the analysis unit 130 to calculate the feature vector is not particularly limited, in the case of the present embodiment, the analysis unit 130 can obtain the feature information 201 by performing Fourier transform, in particular, fast Fourier transform on the extracted feature information 201. Discrete points of the frequency spectrum are calculated as feature vectors.
  • the classification unit 140 is a processing unit that classifies the plurality of feature vectors calculated by the analysis unit 130 into a plurality of types.
  • the classification method of feature vectors is not particularly limited, but in the case of the present embodiment, the classification unit 140 receives a set of feature vectors as input, and uses kernel self-organizing map, hierarchical clustering, and each feature value according to silhouette values.
  • the information 201 is automatically classified into a small number of similar event categories, and feature vectors belonging to the same category are labeled the same. These clusters roughly correspond to the categories of sounds and environmental sounds produced by biological activity during sleep.
  • the probability model generation unit 150 is a processing unit that generates a probability model by processing using a hidden Markov model.
  • the probability model generation unit 150 corresponds to the feature information 201 corresponding to the feature vectors classified for each of a plurality of sleep sounds 210 included in the first state such as the sleep sound 210 determined to be in the good state (first state) by the questionnaire.
  • the first model which is one of the probabilistic models, is generated by processing using a hidden Markov model based on the time of occurrence of.
  • the probability model generation unit 150 classifies each of a plurality of sleep sounds 210 included in the second state such as the sleep sound 210 which is considered to be a bad state (the second state) by a questionnaire. Based on the time at which the feature information 201 corresponding to the feature vector is generated, processing using a hidden Markov model is performed to generate a second model which is one of the probability models.
  • the first model and the second model having the number N of hidden variables of 2 to 6 were separately generated. Then, the sleep sound 210 known to be in the first state and the sleep sound 210 known to be in the second state were respectively applied to the first model and the second model to calculate the likelihood.
  • the result is shown in FIG. As shown in the figure, it is understood that the number 5 of hidden variables is suitable when judging the good state, and the number 3 of the half of the hidden is suitable when judging the bad state. Therefore, when determining the quality of sleep based on the sleep sound 210 whose state is unknown, it can be concluded that the number of hidden variables is in the range of 3, 4, and 5.
  • the probability model generation unit 150 generates probability models in which the number N of hidden variables is 3, 4 and 5 for the first state and the second state, respectively. That is, the first model (3) of the number 3 of the hidden variable, the first model (4) of the number 4 of the hidden variable, and the first model (5) of the hidden variable 5) is generated. Also, as the second model reflecting the second state, the second model (3) of the number 3 of the hidden variable, the second model (4) of the number 4 of the hidden variable, and the second model (5) of the hidden variable 5) is generated.
  • the storage unit 160 is a processing unit that causes the storage device 159 to store the first model generated by the probability model generation unit 150 and the second model.
  • the storage unit 160 may update the probability model stored in the storage device 159 when a probability model is newly generated by a set of newly added sleep sounds 210.
  • the extraction unit 120 extracts feature information for the sleep sound 210 whose condition to be determined is unknown by the sound recording unit 110, like the sleep sound whose state is known, and the analysis unit 130
  • the feature vector is calculated from the feature information, and the feature vector labeled by the classification unit 140 is applied to the first model stored in advance in the storage device 159 to calculate the first likelihood, which is the likelihood, and the second model It is a processing unit that calculates the second likelihood by applying.
  • the probability models are the first model (3), the first model (4), the first model (5), the second model (3), the second model (4), the second model (4) Because there are six types of 5), the likelihood calculation unit 170 calculates the first likelihood (3), the first likelihood (4), the first likelihood (5), and the second likelihood for one sleep sound 210. Degree (3), second likelihood (4), second likelihood (5) Six likelihoods are calculated.
  • the determination unit 180 is a processing unit that determines whether the quality of sleep to be determined is the first state or the second state based on the first likelihood and the second likelihood calculated by the likelihood calculation unit 170. In the case of the present embodiment, the determination unit 180 compares the first likelihood and the second likelihood corresponding to the number of the same number of hidden variables and wins the one with high likelihood, and the sleep quality is determined by majority decision. Determines whether the first state or the second state.
  • FIG. 4 is a table showing the likelihood calculated by the likelihood calculating unit and the determination result. As shown in the figure, comparing the likelihood in the case of the number 3 of hidden variables, the second likelihood is high, the likelihood in the case of the number 4 of hidden variables is high in the second likelihood, and the number of hidden variables is 5 The first likelihood is high. Therefore, the second likelihood is two wins, and the majority of the majority of the decisions determines that the quality of sleep to be determined is the second state, that is, the bad state.
  • the result presentation unit 190 is a processing unit that causes the display device 112 to present the determination result by the determination unit 180.
  • the result presentation unit 190 and the display device 112 are provided in the terminal device 101.
  • FIG. 5 is a flow chart showing the operation flow of the sleep quality determination system in the learning phase.
  • the probability model generation device 199 uses the network 300 or the like to generate the sleep sound 210 recorded by the sound recording unit 110 such as the terminal device 101 personally owned such as a smartphone and the terminal device 101 installed in a facility such as a hospital. collect.
  • the sound recording unit 110 such as the terminal device 101 personally owned such as a smartphone and the terminal device 101 installed in a facility such as a hospital. collect.
  • information such as whether the sleep sound 210 belongs to the first state or the second state is also collected in a linked state, and stored in the storage device 159 of the probability model generation device 199. It is stored (S101).
  • the feature information 201 is extracted by the extraction unit 120 of the probability model generation device 199 for each state such as the first state and the second state (S102), and the feature vector is calculated by the analysis unit 130 (S103) And the classified feature vector is classified into a plurality of categories by machine learning in the classification unit 140, and a label indicating which category it belongs to is assigned to each feature vector (S104).
  • the probabilistic model generation unit 150 generates a probabilistic model using the hidden Markov model for the feature vector labeled for each state (S105), and the obtained result is stored in the storage unit 159 by the storage unit 160. (S106).
  • the probability model generation device 199 waits until collecting a new sleeping sound 210 whose state is known (S108), and when collecting a new sleeping sound 210 whose state is known (S107: Yes), a new probability A model is generated and the probability model stored in the storage unit 159 is updated.
  • FIG. 6 is a flowchart showing the flow of the operation of the sleep quality determination system in the determination phase.
  • the sleep sound 210 recorded by the sound recording unit 110 realized as a program in the terminal device 101 such as a smartphone is transmitted to the probabilistic model generation device 199 via the network 300 and the like (S201).
  • the transmitted sleep sound 210 is stored in the storage unit 159 of the probability model generation device 199.
  • the feature information 201 is extracted by the extraction unit 120 (S202), the feature vector is calculated by the analysis unit 130 (S203), and the calculated feature vector is processed by the machine learning in the classification unit 140.
  • a label that is classified into a category and indicates which category it belongs to is assigned to each feature vector (S204).
  • the likelihood calculation unit 170 calculates the likelihood by applying the feature vector labeled for each state to each of the probability model corresponding to the first state and the probability model corresponding to the second state. (S205).
  • the determination unit 180 determines whether the sleep sound 210 to be determined is in the first state or the second state based on the obtained likelihood (S206).
  • the determination result is transmitted to the terminal device 101, and is displayed on the display device 112 of the terminal device 101 by the result presentation unit 190 (S207).
  • a probability model in which the sleep patterns of at least the first state and the plurality of states including the second state are reflected is hidden Markov by the number of different hidden variables.
  • the likelihood of the sleep sound 210 to be determined is calculated using a plurality of probability models generated by the model, and the quality of the sleep to be determined is determined by majority decision. It is possible to determine the quality of sleep.
  • the present invention is not limited to the above embodiment.
  • another embodiment realized by arbitrarily combining the components described herein and excluding some of the components may be used as an embodiment of the present invention.
  • the present invention also includes modifications obtained by applying various modifications to those skilled in the art without departing from the spirit of the present invention, that is, the meaning described in the claims with respect to the above embodiment.
  • the sleep sound 210 for generating a probability model is not only collected from an unspecified number but also stores the sleep sound 210 in the first state and the sleep sound 210 in the second state for each individual. It is also possible to generate a probability model for each individual based on the sleep sound 210 of.
  • the likelihood by the probability model obtained by the number of one hidden variable may be made in degrees.
  • the determination unit 180 determines the likelihood calculated based on the probability model obtained by the number of different hidden variables as the likelihood vector, and determines whether the sleep quality is the first state or the second state by pattern identification. It does not matter.
  • the pattern identification is, for example, a support vector machine (SVM), and can be exemplified by multilayer perceptron, random forest, logistic identification, and the like.
  • the determination unit 180 includes a likelihood model generation unit 181.
  • the likelihood model generation unit 181 is configured based on the feature vectors classified respectively for the plurality of sleep sounds 210 included in the first state and the plurality of sleep sounds 210 included in the second state. Using the first model and the second model generated by the probability model generation unit 150 to calculate the likelihood by the likelihood calculation unit 170, and using the obtained likelihood vector as an input, the first state and the second state And generate a model of the discriminator as a label.
  • the pattern identification classifier is not particularly limited as described above, but a neural network is suitable when the number of data points (number of feature vectors) for generating the likelihood model is relatively large. When the number is relatively small, it is considered that a support vector machine (SVM) or the like is suitable.
  • SVM support vector machine
  • Other discriminators include AdaBoost, naive Bayes, decision trees, and the like.
  • the determination unit 180 applies the first likelihood and the second likelihood calculated by the likelihood calculation unit 170 based on the feature vector of the sleep sound 210 whose state is unknown to the model generated by the likelihood model generation unit 181. Then, determine whether the quality of sleep is in the first state or the second state.
  • the sleep sound 210 for creating the likelihood model may be one used when the probability model generation unit 150 generates a model, or another sleep sound 210 may be used. Moreover, you may mix these.
  • each processing unit may be realized by one device.
  • the sleep quality determination system 100 is configured by a plurality of devices, which device each processing unit is provided with is optional, and a plurality of devices may be provided with the same processing unit. .
  • the sleep quality determination system 100 includes a sound storage unit 110, an extraction unit 120, an analysis unit 130, a classification unit 140, a probability model generation unit 150, a storage unit 160, a likelihood calculation unit 170, and a smart phone or a tablet terminal.
  • a smartphone or the like may function as the sleep quality determination system 100 by installing a program called an application that can realize each processing unit such as the unit 180 and the result presentation unit 190.
  • the present invention can be used to determine and present sleep states of humans and animals, and can be used to manage or improve health and lifestyle habits.

Abstract

Provided is a sleep quality assessment system (100) comprising: a sound recording unit (110) for recording sounds during sleep as sleep sounds; an extraction unit (120) for extracting from the sleep sounds a plurality of instances of feature information; an analysis unit (130) for computing feature vectors from the feature information; a classification unit (140) for classifying the feature vectors into a plurality of types; a probability model generation unit (150) for generating a first model reflecting a sleep sound in a first state and a second model reflecting a sleep sound in a second state by subjecting the feature vectors to a hidden Markov model-based process; a likelihood computation unit (170) for computing a first likelihood obtained by applying the feature vectors classified based on a sleep sound (210) to be assessed to the first model and a second likelihood obtained by applying the feature vectors classified based on the sleep sound (210) to be assessed to the second model; and an assessment unit (180) for assessing the quality of sleep on the basis of the computed likelihoods.

Description

睡眠の質判定システム、睡眠の質モデル作成プログラム、および、睡眠の質判定プログラムSleep quality determination system, sleep quality model creation program, and sleep quality determination program
　睡眠中の人や動物などの生体が発する音や生体の動きにより発生する音などに基づき睡眠の質を判定する睡眠の質判定システム、睡眠の質モデル作成プログラム、および、睡眠の質判定プログラムに関する。 A sleep quality determination system that determines the quality of sleep based on sounds emitted by a living person or animal living in sleep or sounds generated by movement of a living body, a sleep quality model creation program, and a sleep quality determination program .
　例えば人にとって、不眠など睡眠の質や量が不十分なことによって生じる睡眠障害は、覚醒時における身体的、精神的、社会的、感情的な機能を妨げる原因になることは知られている。従って、睡眠の質などの睡眠状態を評価することはこれらの機能を回復させるための重要な情報となる。睡眠状態に応じて、脳波や呼吸、心拍が変化する。さらに、睡眠状態の変化や安定性と関連して、体動、いびき、無呼吸、歯ぎしりなど、様々なイベントも睡眠中に生じている。 For example, it is known that sleep disorders caused by inadequate sleep quality and quantity, such as sleeplessness, cause physical, mental, social, and emotional functions to be disturbed during awakening. Therefore, assessing sleep conditions such as sleep quality is important information to restore these functions. Depending on your sleep state, brain waves, breathing and heart rate change. In addition, various events such as body movement, snoring, apnea and bruxism also occur during sleep in association with changes in sleep state and stability.
　そこで従来は、特許文献１に記載の技術のように、睡眠中の重さ（圧力）の変動を、圧力センサを用いて測定したり、特許文献２に記載の技術のように、睡眠中の脳波を測定したり、特許文献３に記載の技術のように、睡眠中の脈波を測定したりすることによりレム睡眠・ノンレム睡眠などの睡眠段階を評価している。 Therefore, conventionally, as in the technique described in Patent Document 1, fluctuations in weight (pressure) during sleep are measured using a pressure sensor, and as in the technique described in Patent Document 2, sleep during sleep. Sleep stages such as REM sleep and non-REM sleep are evaluated by measuring an electroencephalogram or measuring a pulse wave during sleep as in the technique described in Patent Document 3.
特開２０１１－１１５１８８号公報JP 2011-115188 A 特開２０１１－０８３３９３号公報JP, 2011-083393, A 特開２０１５－１３６３８０号公報JP, 2015-136380, A
　ところが、前記特許文献に記載の技術のように、睡眠環境を計測し、睡眠段階を判定する技術はいくつか存在するものの、体に何らかのセンサなど専用のデバイスを取り付けなければならない。さらに、特殊な装置を備えた特殊な環境下において電線に接続された状態で睡眠させ、データを取得しなければならない。従って、被験体はストレスを受けながら睡眠しなければならず、通常の睡眠状態におけるデータを取得することは困難である。また、このような手法を動物に適用することは困難である。 However, although there are several techniques for measuring the sleep environment and determining the sleep stage as in the technique described in the patent document, it is necessary to attach a dedicated device such as some sensor to the body. Furthermore, it is necessary to sleep and acquire data while connected to the electric wire in a special environment equipped with a special device. Thus, the subject must sleep while being stressed, and it is difficult to obtain data in a normal sleep state. Moreover, it is difficult to apply such an approach to animals.
　また、１回の測定で５～８時間程度の膨大なデータとなり、そこから意味のあるデータを抽出して睡眠段階を評価するにはかなりの労力が必要となる。さらに、レム・ノンレム睡眠などの睡眠段階は、睡眠障害の判定や睡眠科学の研究には有効であるが、一般ユーザに提示する情報としては不適切である。 In addition, it takes 5 to 8 hours of huge data in one measurement, and it takes considerable effort to extract meaningful data from it and evaluate the sleep stage. Furthermore, sleep stages such as REM / REM sleep are effective for sleep disorder determination and sleep science research, but are inappropriate as information to be presented to general users.
　本願発明は、上記課題に鑑みなされたものであり、生体と非接触の状態でもデータを取得することができ、長時間にわたるデータを機械学習によって睡眠段階ではなく睡眠の質を簡単に判定できる睡眠の質判定システム、睡眠の質モデル作成プログラム、および、睡眠の質判定プログラムの提供を目的とする。 The present invention has been made in view of the above problems, and can acquire data even in a non-contact state with a living body, and can easily determine the quality of sleep instead of the sleep stage by machine learning data over a long period of time Quality determination system, a sleep quality model creation program, and a sleep quality determination program.
　上記目的を達成するために、本願発明の１つである睡眠の質判定システムは、睡眠中における生体が発する音を睡眠音として記録する音記録部と、記録された前記睡眠音の中から特徴部分である特徴情報を複数個抽出する抽出部と、抽出された特徴情報に基づきそれぞれの特徴ベクトルを算出する解析部と、複数の前記特徴ベクトルを複数種類に分類する分類部と、睡眠の質が第一状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第一モデルと、睡眠の状態が第二状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第二モデルを生成する確率モデル生成部と、前記第一モデル、および、前記第二モデルを記憶させる記憶部とを備え、前記音記録部は、判定対象の睡眠音を記録し、前記抽出部は、前記睡眠音から特徴情報を抽出し、前記解析部は、前記特徴情報から特徴ベクトルを算出し、前記分類部は、前記特徴ベクトルを分類し、さらに、前記判定対象の分類された特徴ベクトルを前記第一モデルに適用して尤度である第一尤度と前記第二モデルに適用して尤度である第二尤度を算出する尤度算出部と、算出された前記第一尤度と第二尤度とに基づき前記判定対象の睡眠の質が第一状態か第二状態かを判定する判定部とを備える。 In order to achieve the above object, the sleep quality determination system which is one of the present invention is characterized by a sound recording unit which records a sound emitted by a living body during sleep as a sleep sound and a feature of the recorded sleep sound. An extraction unit that extracts a plurality of pieces of feature information that is a part, an analysis unit that calculates each feature vector based on the extracted feature information, a classification unit that classifies the plurality of feature vectors into a plurality of types, sleep quality Processing the classified feature vector of the sleep sound in the first state by using the hidden Markov model, the first model being one of probabilistic models, and the sleep sound in which the sleep state is the second state A probabilistic model generation unit for generating a second model which is one of the probabilistic models by processing the classified feature vectors of the second group using the hidden Markov model, the first model, and the second model The sound recording unit records sleep sound to be determined, the extraction unit extracts feature information from the sleep sound, and the analysis unit extracts a feature vector from the feature information. The classification unit classifies the feature vector, and further applies the classified feature vector of the determination target to the first model and applies the first likelihood as a likelihood and the second model And the second state of the sleep quality of the judgment target on the basis of the first likelihood and the second likelihood thus calculated. And a determination unit that determines whether the
　これによれば、長時間にわたり記録された睡眠音の中に含まれる多数の特徴情報に基づく睡眠の質を判定するための確率モデルを生成し、当該確率モデルを判定対象の睡眠音に適用することで、正確に睡眠状態の判定を行う事が可能となる。 According to this, a probabilistic model for determining the quality of sleep based on a large number of feature information included in the sleep sound recorded for a long time is generated, and the probabilistic model is applied to the sleep sound to be determined. This makes it possible to accurately determine the sleep state.
　また、前記分類部は、複数の前記特徴ベクトルに対しクラスタ分析を行って少数のカテゴリに自動分類し、各カテゴリに属する前記特徴ベクトルに同一のラベルを付与してもよい。 Further, the classification unit may perform cluster analysis on a plurality of the feature vectors to automatically classify the plurality of feature vectors into a small number of categories, and assign the same label to the feature vectors belonging to each category.
　これにより、特徴情報の分類をコンピュータの処理により実行することができ、大量の睡眠音に基づく確率モデルを容易に生成することが可能となる。また、当該確率モデルを用いることにより、高い精度で睡眠の質を判定することが可能となる。 As a result, classification of feature information can be performed by computer processing, and a probability model based on a large amount of sleep sounds can be easily generated. In addition, by using the probability model, it is possible to determine the quality of sleep with high accuracy.
　また、前記確率モデル生成部は、隠れ変数の数が３以上、９以下の範囲の１つであるＮ（Ｎは整数）の第一モデル、および、第二モデルを生成してもよい。 Further, the probability model generation unit may generate a first model of N (N is an integer) and a second model in which the number of hidden variables is one in a range of 3 or more and 9 or less.
　当該数の隠れ変数の数を用いることで、掲載の処理数を抑制することができ、かつ、睡眠の質が適切に反映した確率モデルを生成することができる。 By using the number of hidden variables of the number, it is possible to suppress the number of processings of posting, and to generate a probabilistic model in which the quality of sleep is properly reflected.
　また、前記確率モデル生成部は、隠れ変数の数がＮ－１、Ｎ、Ｎ＋１の第一モデル、および、第二モデルをそれぞれ生成し、前記尤度算出部は、生成された第一モデル、および、第二モデルのそれぞれについて、第一尤度と第二尤度とを算出し、前記判定部は、同一数の隠れ変数の数に対応する第一尤度と第二尤度とを比較して尤度の高いものを勝ちとし、多数決により睡眠の質が第一状態か第二状態かを判定してもよい。 Further, the probability model generation unit generates a first model and a second model in which the number of hidden variables is N−1, N, and N + 1, and the likelihood calculation unit generates the generated first model, And calculating the first likelihood and the second likelihood for each of the second models, and the determination unit compares the first likelihood and the second likelihood corresponding to the number of the same number of hidden variables Then, the one with high likelihood may be won and it may be determined by majority decision whether the sleep quality is in the first state or the second state.
　これによれば、特徴情報の比較的多い睡眠と特徴情報の比較的少ない睡眠とを広く対応付けて判定することができ、睡眠の質の判定精度を向上させることが可能となる。 According to this, it is possible to determine the sleep with a relatively large amount of feature information and the sleep with a relatively small amount of feature information in a broad manner, and it is possible to improve the determination accuracy of sleep quality.
　また、前記確率モデル生成部は、複数の隠れ変数の数について第一モデル、および、第二モデルをそれぞれ生成し、前記尤度算出部は、生成された第一モデル、および、第二モデルのそれぞれについて、第一尤度と第二尤度とを算出し、前記判定部は、算出された尤度を尤度ベクトルとし、パターン識別により睡眠の質が第一状態か第二状態かを判定してもよい。 Further, the probability model generation unit generates a first model and a second model for the number of the plurality of hidden variables, and the likelihood calculation unit generates the generated first model and the second model. For each, the first likelihood and the second likelihood are calculated, and the determination unit determines the calculated likelihood as a likelihood vector and determines whether the sleep quality is in the first state or the second state by pattern identification. You may
　また、前記パターン識別は、教師有り学習の識別器であり、前記判定部は、前記識別器に対応するモデルを生成する尤度モデル生成部を備え、生成された尤度モデルに基づき前記尤度ベクトルを入力として睡眠の質が第一状態か第二状態かを判定してもよい。これによれば、判定の精度を向上させることが可能となる。 The pattern identification is a classifier for supervised learning, and the determination unit includes a likelihood model generation unit that generates a model corresponding to the classifier, and the likelihood is based on the generated likelihood model. It is also possible to determine whether the sleep quality is in the first state or the second state using a vector as an input. According to this, it is possible to improve the accuracy of the determination.
　また、前記確率モデル生成部を備える確率モデル生成装置と、前記確率モデル生成装置にネットワークを介して接続される複数の端末装置とを備え、前記端末装置は、前記音記録部と、さらに、前記判定部による判定結果を表示装置に提示させる結果提示部を備えてもよい。 Further, the terminal device includes: a probability model generation device including the probability model generation unit; and a plurality of terminal devices connected to the probability model generation device via a network, the terminal device further includes the sound recording unit. You may provide the result presentation part which makes a display apparatus show the determination result by a determination part.
　これによれば、端末装置によって判定対象の睡眠音を録音し、睡眠の質の判定結果を端末装置で簡単に確認することが可能となる。また、複数の端末装置で録音された睡眠音をサーバーが集中して管理することで、より適切な確率モデルを生成することが可能となる。 According to this, it is possible to record the sleep sound to be determined by the terminal device and to easily confirm the determination result of the quality of sleep by the terminal device. In addition, since the server centrally manages the sleep sounds recorded by the plurality of terminal devices, it is possible to generate a more appropriate probability model.
　また、上記目的を達成するために、本願発明の１つである睡眠の質モデル作成プログラムは、睡眠中における生体が発する音を睡眠音として記録する音記録部と、記録された前記睡眠音の中から特徴部分である特徴情報を複数個抽出する抽出部と、抽出された特徴情報に基づきそれぞれの特徴ベクトルを算出する解析部と、複数の前記特徴ベクトルを複数種類に分類する分類部と、睡眠の質が第一状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第一モデルと、睡眠の状態が第二状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第二モデルを生成する確率モデル生成部と、前記第一モデル、および、前記第二モデルを記憶させる記憶部とを含み、前記各処理をコンピュータに実行させることにより上記各部を実現する。 In addition, in order to achieve the above object, a sleep quality model creating program according to the present invention is a sound recording unit for recording as a sleep sound a sound emitted by a living body during sleep, and the recorded sleep sound. An extraction unit that extracts a plurality of pieces of feature information that are feature portions from among them, an analysis unit that calculates each feature vector based on the extracted feature information, and a classification unit that classifies the plurality of feature vectors into a plurality of types; The state of sleep was in the second state, the first model being one of the probabilistic models by processing the classified feature vectors of the sleep sound in which the quality of sleep was the first state using the hidden Markov model A probability model generation unit for generating a second model which is one of the probability models by processing the classified feature vectors of the sleep sound using a hidden Markov model, the first model, and And a storage unit for storing a second model, to realize the above-mentioned respective units by executing the respective processes in the computer.
　これによれば、睡眠の質を判定するための確率モデルを生成することが可能となる。 According to this, it is possible to generate a probabilistic model for determining the quality of sleep.
　また、上記目的を達成するために、本願発明の１つである睡眠の質判定プログラムは、睡眠中における生体が発する音を睡眠音として記録する音記録部と、記録された前記睡眠音の中から特徴部分である特徴情報を複数個抽出する抽出部と、抽出された特徴情報に基づきそれぞれの特徴ベクトルを算出する解析部と、複数の前記特徴ベクトルを複数種類に分類する分類部と、隠れマルコフモデルを用いて生成された異なる睡眠の質にそれぞれ対応する第一モデル、および、前記第二モデルを記憶させる記憶部とを備え、分類された特徴ベクトルを前記第一モデルに適用して尤度である第一尤度と前記第二モデルに適用して尤度である第二尤度を算出する尤度算出部と、算出された前記第一尤度と第二尤度とに基づき前記判定対象の睡眠の質が第一状態か第二状態かを判定する判定部とを含み、前記各処理をコンピュータに実行させることにより上記各部を実現する。 In addition, in order to achieve the above object, the sleep quality determination program according to the present invention is a sound recording unit for recording a sound emitted by a living body during sleep as a sleep sound, and in the recorded sleep sound. An extracting unit that extracts a plurality of pieces of feature information that are feature portions from the image, an analyzing unit that calculates each feature vector based on the extracted feature information, a classifying unit that classifies a plurality of the feature vectors into a plurality of types; A first model corresponding to different sleep qualities generated using a Markov model, and a storage unit for storing the second model, and applying classified feature vectors to the first model; A likelihood calculating unit that calculates a second likelihood that is a likelihood by applying to the second model and a first likelihood that is a degree, and the first likelihood and a second likelihood that are calculated Quality of sleep to be judged And a determination section for determining a first state or a second state, to realize the above-mentioned respective units by executing the respective processes in the computer.
　これによれば、睡眠の質を判定するための生成された確率モデルを用いて睡眠の質を適切に判断することが可能となる。 According to this, it is possible to appropriately determine the sleep quality using the generated probability model for determining the sleep quality.
　本願発明によれば、専用デバイスを必要とせず、汎用の録音デバイスを用いて簡便に睡眠の質をコンピュータに判定させることができる。 According to the present invention, it is possible to cause a computer to easily determine the quality of sleep using a general-purpose recording device without requiring a dedicated device.
睡眠の質判定システムの機能構成を示すブロック図である。It is a block diagram showing functional composition of a sleep quality judging system. 記録された睡眠音の一部を示す図である。It is a figure which shows a part of sleep sound recorded. 複数の隠れ変数の数の確率変数に睡眠音を適用した結果の尤度を示す表である。It is a table | surface which shows the likelihood of the result of having applied a sleep sound to the random variable of the number of several hidden variables. 尤度算出部が算出した尤度と判定結果とを示す表である。It is a table | surface which shows the likelihood and determination result which the likelihood calculation part calculated. 学習フェーズにおける睡眠の質判定システムの動作の流れを示すフローチャートである。It is a flowchart which shows the flow of operation | movement of the sleep quality determination system in a learning phase. 判定フェーズにおける睡眠の質判定システムの動作の流れを示すフローチャートである。It is a flowchart which shows the flow of operation | movement of the sleep quality determination system in a determination phase. 判定部の別例を示すブロック図である。It is a block diagram showing another example of a judgment part.
　次に、本願発明に係る睡眠の質判定システム、睡眠の質モデル作成プログラム、および、睡眠の質判定プログラムの実施の形態について、図面を参照しつつ説明する。なお、以下の実施の形態は、本願発明に係る睡眠の質判定システム、睡眠の質モデル作成プログラム、および、睡眠の質判定プログラムの一例を示したものに過ぎない。従って本願発明は、以下の実施の形態を参考に請求の範囲の文言によって範囲が画定されるものであり、以下の実施の形態のみに限定されるものではない。よって、以下の実施の形態における構成要素のうち、本発明の最上位概念を示す独立請求項に記載されていない構成要素については、本発明の課題を達成するのに必ずしも必要ではないが、より好ましい形態を構成するものとして説明される。 Next, an embodiment of a sleep quality determination system, a sleep quality model creation program, and a sleep quality determination program according to the present invention will be described with reference to the drawings. The following embodiments are merely examples of the sleep quality determination system, the sleep quality model creation program, and the sleep quality determination program according to the present invention. Accordingly, the scope of the present invention is defined by the wording of the claims with reference to the following embodiments, and is not limited to only the following embodiments. Therefore, among the components in the following embodiments, components that are not described in the independent claim showing the highest concept of the present invention are not necessarily required to achieve the object of the present invention, It is described as constituting a preferred embodiment.
　また、図面は、本願発明を示すために適宜強調や省略、比率の調整を行った模式的な図となっており、実際の形状や位置関係、比率とは異なる場合がある。 In addition, the drawings are schematic diagrams in which emphasis, omission, and adjustment of ratios are appropriately performed to illustrate the present invention, and may differ from actual shapes, positional relationships, and ratios.
　また、請求の範囲、および、明細書等で用いているカーネル関数は、正定値性を満たさないため、厳密にはカーネル関数に含まれない恐れがあるが、本請求の範囲、および、明細書等ではカーネル関数に含まれるものとして記載している。 Also, the kernel function used in the claims and the specification etc. may not be strictly included in the kernel function because it does not satisfy positive definiteness, but the claims and the specification Etc. are described as being included in the kernel function.
　図１は、睡眠の質判定システムの機能構成を示すブロック図である。 FIG. 1 is a block diagram showing a functional configuration of a sleep quality determination system.
　睡眠の質判定システム１００は、生成した確率モデルに基づき睡眠の質を判定するシステムであって、音記録部１１０と、抽出部１２０と、解析部１３０と、分類部１４０と、確率モデル生成部１５０と、記憶部１６０と、尤度算出部１７０と、判定部１８０とを備える。本実施の形態の場合、睡眠の質判定システム１００は、さらに結果提示部１９０を備える。 The sleep quality determination system 100 is a system that determines sleep quality based on the generated probability model, and includes a sound recording unit 110, an extraction unit 120, an analysis unit 130, a classification unit 140, and a probability model generation unit. 150, a storage unit 160, a likelihood calculation unit 170, and a determination unit 180. In the case of the present embodiment, the sleep quality determination system 100 further includes a result presentation unit 190.
　音記録部１１０は、睡眠中における生体２００が発する音を睡眠音２１０として記憶装置１１９に記録する処理部である。なお、睡眠音２１０には、生体２００の周辺の環境音も含まれる。本実施の形態の場合、音記録部１１０は、生体２００の近傍に配置されるマイクロフォン１１１と無線や有線によって接続されており、生体２００が発する音、および、周辺の音をマイクロフォン１１１が変換したアナログの電気信号を音記録部１１０がデジタル信号に変換して記憶装置１１９に保存している。また、音記録部１１０は、確率モデル生成装置１９９とネットワーク３００等を介して接続される複数の端末装置１０１にそれぞれ備えられるものであり、睡眠の質判定システム１００全体としては、音記録部１１０は、複数備えられている。 The sound recording unit 110 is a processing unit that records the sound emitted by the living body 200 during sleep in the storage device 119 as the sleep sound 210. The sleep sound 210 also includes environmental sound around the living body 200. In the case of the present embodiment, the sound recording unit 110 is connected to the microphone 111 disposed in the vicinity of the living body 200 wirelessly or by wire, and the microphone 111 converts the sound emitted by the living body 200 and the surrounding sound. The sound recording unit 110 converts an analog electrical signal into a digital signal and stores the digital signal in the storage device 119. In addition, the sound recording unit 110 is provided in each of the plurality of terminal devices 101 connected to the probability model generation device 199 via the network 300 or the like, and the sound recording unit 110 is the entire sleep quality determination system 100. There are multiple.
　音記録部１１０の睡眠音２１０の保存方法は、特に限定されるものではなく、例えば、フラッシュメモリなどの半導体記憶装置や、ハードディスクなどの磁気記憶装置のほか、光などに基づき記憶する装置などに睡眠音２１０を記録する。音記録部１１０が記録する睡眠音２１０は、生体２００が入眠してから覚醒するまでを連続的に記録してもよく、また、所定の間隔で断続的に記録しても構わない。 The method of storing the sleep sound 210 of the sound recording unit 110 is not particularly limited. For example, in addition to a semiconductor storage device such as a flash memory and a magnetic storage device such as a hard disk, The sleep sound 210 is recorded. The sleep sound 210 recorded by the sound recording unit 110 may be continuously recorded from when the living body 200 sleeps until it wakes up, or may be intermittently recorded at a predetermined interval.
　マイクロフォン１１１が取り付けられる位置は生体２００が発する音を取得できる位置であれば特に限定されるものではなく、生体２００が睡眠している部屋の天井や床や壁に取り付けられるものでもよい。本実施形態の場合、生体２００が睡眠しているベッドの宮の頂上部分にマイクロフォン１１１は取り付けられている。なお、定常的にノイズを発生させる物（例えばクーラー）からは遠い位置に配置することが好ましい。 The position to which the microphone 111 is attached is not particularly limited as long as it can acquire the sound emitted by the living body 200, and may be attached to the ceiling, floor or wall of a room where the living body 200 is sleeping. In the case of the present embodiment, the microphone 111 is attached to the top of the bed of the bed where the living body 200 is sleeping. In addition, it is preferable to arrange | position from the thing (for example, cooler) which generate | occur | produces noise regularly.
　ここで、生体２００とは、人間ばかりでなく人間以外の動物であっても構わない。また、生体２００が発する音とは、歯ぎしりやいびきや寝言など生体２００の口や喉から出る音や、布団の擦れる音やベッドのきしむ音など生体２００の動き（体動）に基づき発生する音などである。生体２００の周辺の音とは、クーラーや加湿器など生体２００の周囲にある備品から発せられる音や、窓などを介して生体２００の周辺に達する車の音などである。 Here, the living body 200 may be not only human but also animals other than human. Further, the sound emitted by the living body 200 is a sound generated based on movement (body movement) of the living body 200 such as noise from the mouth or throat of the living body 200, such as bruises, snoring, sleeping sounds, rubbing noise of a futon or squeaking noise of a bed. Etc. The sound around the living body 200 is a sound emitted from equipment around the living body 200 such as a cooler or a humidifier, or a car sound reaching around the living body 200 through a window or the like.
　図２は、記録された睡眠音の一部を示す図である。 FIG. 2 is a diagram showing a part of the recorded sleep sound.
　同図に示すように、抽出部１２０は、音記録部１１０により記録された睡眠音２１０の中から特徴的な部分である特徴情報２０１を複数個抽出する処理部である。抽出部１２０が特徴情報２０１を抽出する方法は特に限定されるものではなく、統計的手法により特徴情報２０１を抽出してもよく、所定の閾値以上の振幅の音波の検出時刻を基準にした所定の時間間隔に含まれる音波を特徴情報２０１とする時間窓を設けて抽出してもよい。本実施形態の場合、抽出部１２０は、統計的手法の中でもＫｌｅｉｎｂｅｒｇのバースト抽出法を用い、バーストレベル１以上の連続する期間をひとつの特徴情報２０１として抽出し、定常的に記録されているノイズを排除した。所定の時間に含まれる睡眠音２１０から抽出部１２０が抽出した特徴情報２０１の数はｎ個であった。 As shown in the figure, the extraction unit 120 is a processing unit that extracts a plurality of pieces of feature information 201 that are characteristic portions from the sleep sound 210 recorded by the sound recording unit 110. The method for the extraction unit 120 to extract the feature information 201 is not particularly limited, and the feature information 201 may be extracted by a statistical method, and a predetermined method based on the detection time of a sound wave having an amplitude of a predetermined threshold or more. The sound window included in the time interval of (1) may be extracted as the feature information 201. In the case of the present embodiment, the extraction unit 120 extracts a continuous period of burst level 1 or more as one feature information 201 by using Kleinberg's burst extraction method among statistical methods, and the noise recorded regularly. Was eliminated. The number of pieces of feature information 201 extracted by the extraction unit 120 from the sleep sound 210 included in a predetermined time was n.
　解析部１３０は、抽出部１２０により抽出された複数の特徴情報２０１をそれぞれ示す特徴ベクトルを算出する処理部である。解析部１３０が特徴ベクトルを算出する方法は特に限定されるものではないが、本実施形態の場合、解析部１３０は、抽出された特徴情報２０１をフーリエ変換、特に高速フーリエ変換することにより得られる周波数スペクトルの離散点を特徴ベクトルとして算出している。 The analysis unit 130 is a processing unit that calculates feature vectors respectively indicating the plurality of pieces of feature information 201 extracted by the extraction unit 120. Although the method for the analysis unit 130 to calculate the feature vector is not particularly limited, in the case of the present embodiment, the analysis unit 130 can obtain the feature information 201 by performing Fourier transform, in particular, fast Fourier transform on the extracted feature information 201. Discrete points of the frequency spectrum are calculated as feature vectors.
　分類部１４０は、解析部１３０により算出された複数の特徴ベクトルを複数種類に分類する処理部である。特徴ベクトルの分類方法は特に限定されるものでは無いが、本実施の形態の場合、分類部１４０は、特徴ベクトルの集合を入力として、カーネル自己組織化マップ、階層型クラスタリング、シルエット値により各特徴情報２０１を少数の類似事象カテゴリに自動分類し、同じカテゴリに属する特徴ベクトルに同じラベルを付している。これらのクラスタは、睡眠中の生体活動によって生じる音や環境音のカテゴリに大まかに対応する。 The classification unit 140 is a processing unit that classifies the plurality of feature vectors calculated by the analysis unit 130 into a plurality of types. The classification method of feature vectors is not particularly limited, but in the case of the present embodiment, the classification unit 140 receives a set of feature vectors as input, and uses kernel self-organizing map, hierarchical clustering, and each feature value according to silhouette values. The information 201 is automatically classified into a small number of similar event categories, and feature vectors belonging to the same category are labeled the same. These clusters roughly correspond to the categories of sounds and environmental sounds produced by biological activity during sleep.
　確率モデル生成部１５０は、隠れマルコフモデルを用いた処理により確率モデルを生成する処理部である。確率モデル生成部１５０は、アンケートによって良状態（第一状態）であったとされる睡眠音２１０など第一状態に含まれる複数の睡眠音２１０について、それぞれ分類された特徴ベクトルを対応する特徴情報２０１が発生した時刻に基づき隠れマルコフモデルを用いた処理により確率モデルの１つである第一モデルを生成する。また、確率モデル生成部１５０は、第一状態と同様に、アンケートによって悪状態（第二状態）であったとされる睡眠音２１０など第二状態に含まれる複数の睡眠音２１０について、それぞれ分類された特徴ベクトルを対応する特徴情報２０１が発生した時刻に基づき隠れマルコフモデルを用いた処理により確率モデルの１つである第二モデルを生成する。 The probability model generation unit 150 is a processing unit that generates a probability model by processing using a hidden Markov model. The probability model generation unit 150 corresponds to the feature information 201 corresponding to the feature vectors classified for each of a plurality of sleep sounds 210 included in the first state such as the sleep sound 210 determined to be in the good state (first state) by the questionnaire. The first model, which is one of the probabilistic models, is generated by processing using a hidden Markov model based on the time of occurrence of. Also, as in the first state, the probability model generation unit 150 classifies each of a plurality of sleep sounds 210 included in the second state such as the sleep sound 210 which is considered to be a bad state (the second state) by a questionnaire. Based on the time at which the feature information 201 corresponding to the feature vector is generated, processing using a hidden Markov model is performed to generate a second model which is one of the probability models.
　ここで、睡眠の質を判定するのに適した確率モデルを生成するために、隠れ変数の数Ｎが２～６の第一モデルと第二モデルとを別途生成した。そして、第一状態であることが既知の睡眠音２１０と第二状態であることが既知の睡眠音２１０をそれぞれ第一モデルと第二モデルに適用し尤度を算出した。その結果が図３である。同図に示すように、良状態を判断する場合は隠れ変数の数５が適しており、悪状態を判断する場合は、隠れ半数の数３が適していることがわかる。従って、状態が未知の睡眠音２１０に基づき睡眠の質を判定する場合、隠れ変数の数が３、４、５の範囲が適切であると結論付けることができる。 Here, in order to generate a probabilistic model suitable for determining the quality of sleep, the first model and the second model having the number N of hidden variables of 2 to 6 were separately generated. Then, the sleep sound 210 known to be in the first state and the sleep sound 210 known to be in the second state were respectively applied to the first model and the second model to calculate the likelihood. The result is shown in FIG. As shown in the figure, it is understood that the number 5 of hidden variables is suitable when judging the good state, and the number 3 of the half of the hidden is suitable when judging the bad state. Therefore, when determining the quality of sleep based on the sleep sound 210 whose state is unknown, it can be concluded that the number of hidden variables is in the range of 3, 4, and 5.
　そこで本実施の形態の場合、確率モデル生成部１５０は、第一状態、および、第二状態についてそれぞれ隠れ変数の数Ｎが３と４と５の確率モデルを生成している。つまり、第一状態が反映された第一モデルとして隠れ変数の数３の第一モデル（３）、隠れ変数の数４の第一モデル（４）と、隠れ変数の数５の第一モデル（５）が生成される。また、第二状態が反映された第二モデルとして隠れ変数の数３の第二モデル（３）、隠れ変数の数４の第二モデル（４）と、隠れ変数の数５の第二モデル（５）が生成される。 Therefore, in the case of the present embodiment, the probability model generation unit 150 generates probability models in which the number N of hidden variables is 3, 4 and 5 for the first state and the second state, respectively. That is, the first model (3) of the number 3 of the hidden variable, the first model (4) of the number 4 of the hidden variable, and the first model (5) of the hidden variable 5) is generated. Also, as the second model reflecting the second state, the second model (3) of the number 3 of the hidden variable, the second model (4) of the number 4 of the hidden variable, and the second model (5) of the hidden variable 5) is generated.
　なお、本実施の形態では、第一状態に対応した第一モデルと第二状態に対応した第二モデルの作成を例示したが、３以上の睡眠状態を反映させた確率モデルを複数個構築してもかまわない。 In this embodiment, although the creation of the first model corresponding to the first state and the second model corresponding to the second state has been exemplified, a plurality of probability models reflecting three or more sleep states are constructed. It does not matter.
　記憶部１６０は、確率モデル生成部１５０により生成された第一モデル、および、第二モデルを記憶装置１５９に記憶させる処理部である。記憶部１６０は、新しく追加された睡眠音２１０の集合により新しく確率モデルが生成された場合、記憶装置１５９に記憶されている確率モデルを更新してもかまわない。 The storage unit 160 is a processing unit that causes the storage device 159 to store the first model generated by the probability model generation unit 150 and the second model. The storage unit 160 may update the probability model stored in the storage device 159 when a probability model is newly generated by a set of newly added sleep sounds 210.
　尤度算出部１７０は、判定対象が音記録部１１０により録音した状態が未知の睡眠音２１０について、状態が既知の睡眠音と同様に、抽出部１２０が特徴情報を抽出し、解析部１３０が特徴情報から特徴ベクトルを算出し、分類部１４０がラベリングした特徴ベクトルを事前に記憶装置１５９に保存された第一モデルに適用して尤度である第一尤度を算出し、第二モデルに適用して第二尤度を算出する処理部である。 In the likelihood calculation unit 170, the extraction unit 120 extracts feature information for the sleep sound 210 whose condition to be determined is unknown by the sound recording unit 110, like the sleep sound whose state is known, and the analysis unit 130 The feature vector is calculated from the feature information, and the feature vector labeled by the classification unit 140 is applied to the first model stored in advance in the storage device 159 to calculate the first likelihood, which is the likelihood, and the second model It is a processing unit that calculates the second likelihood by applying.
　本実施の形態の場合、確率モデルは、第一モデル（３）、第一モデル（４）、第一モデル（５）、第二モデル（３）、第二モデル（４）、第二モデル（５）の６種類が存在するため、尤度算出部１７０は、１つの睡眠音２１０について第一尤度（３）、第一尤度（４）、第一尤度（５）、第二尤度（３）、第二尤度（４）、第二尤度（５）６個の尤度を算出する。 In the case of the present embodiment, the probability models are the first model (3), the first model (4), the first model (5), the second model (3), the second model (4), the second model (4) Because there are six types of 5), the likelihood calculation unit 170 calculates the first likelihood (3), the first likelihood (4), the first likelihood (5), and the second likelihood for one sleep sound 210. Degree (3), second likelihood (4), second likelihood (5) Six likelihoods are calculated.
　判定部１８０は、尤度算出部１７０で算出された第一尤度と第二尤度とに基づき判定対象の睡眠の質が第一状態か第二状態かを判定する処理部である。本実施の形態の場合、判定部１８０は、同一数の隠れ変数の数に対応する第一尤度と第二尤度とを比較して尤度の高いものを勝ちとし、多数決により睡眠の質が第一状態か第二状態かを判定する。 The determination unit 180 is a processing unit that determines whether the quality of sleep to be determined is the first state or the second state based on the first likelihood and the second likelihood calculated by the likelihood calculation unit 170. In the case of the present embodiment, the determination unit 180 compares the first likelihood and the second likelihood corresponding to the number of the same number of hidden variables and wins the one with high likelihood, and the sleep quality is determined by majority decision. Determines whether the first state or the second state.
　図４は、尤度算出部が算出した尤度と判定結果とを示す表である。同図に示すように、隠れ変数の数３の場合の尤度を比較すると第二尤度が高く、隠れ変数の数４の場合の尤度は第二尤度が高く、隠れ変数の数５の尤度は第一尤度が高い。従って、第二尤度が２勝であり、多数決により判定対象の睡眠の質は、第二状態、即ち悪い状態であると判定される。 FIG. 4 is a table showing the likelihood calculated by the likelihood calculating unit and the determination result. As shown in the figure, comparing the likelihood in the case of the number 3 of hidden variables, the second likelihood is high, the likelihood in the case of the number 4 of hidden variables is high in the second likelihood, and the number of hidden variables is 5 The first likelihood is high. Therefore, the second likelihood is two wins, and the majority of the majority of the decisions determines that the quality of sleep to be determined is the second state, that is, the bad state.
　結果提示部１９０は、判定部１８０による判定結果を表示装置１１２に提示させる処理部である。本実施の形態の場合、結果提示部１９０、および、表示装置１１２は、端末装置１０１に備えられている。 The result presentation unit 190 is a processing unit that causes the display device 112 to present the determination result by the determination unit 180. In the case of the present embodiment, the result presentation unit 190 and the display device 112 are provided in the terminal device 101.
　次に、学習フェーズにおける睡眠の質判定システム１００の動作の例を説明する。図５は、学習フェーズにおける睡眠の質判定システムの動作の流れを示すフローチャートである。 Next, an example of the operation of the sleep quality determination system 100 in the learning phase will be described. FIG. 5 is a flow chart showing the operation flow of the sleep quality determination system in the learning phase.
　スマートフォンなど個人的に所有される端末装置１０１、病院などの施設に設置される端末装置１０１などの音記録部１１０で記録された睡眠音２１０を、ネットワーク３００等を介して確率モデル生成装置１９９が収集する。睡眠音２１０を収集する際には、睡眠音２１０が第一状態に属するか、第二状態に属するかなどの情報も紐付けられた状態で収集され、確率モデル生成装置１９９の記憶装置１５９に記憶される（Ｓ１０１）。 The probability model generation device 199 uses the network 300 or the like to generate the sleep sound 210 recorded by the sound recording unit 110 such as the terminal device 101 personally owned such as a smartphone and the terminal device 101 installed in a facility such as a hospital. collect. When collecting the sleep sound 210, information such as whether the sleep sound 210 belongs to the first state or the second state is also collected in a linked state, and stored in the storage device 159 of the probability model generation device 199. It is stored (S101).
　収集した睡眠音２１０は、第一状態、第二状態などの状態毎に確率モデル生成装置１９９の抽出部１２０によって特徴情報２０１が抽出され（Ｓ１０２）、解析部１３０によって特徴ベクトルが算出され（Ｓ１０３）、算出された特徴ベクトルが分類部１４０における機械学習によって複数のカテゴリに分類され、どこのカテゴリに属するかを示すラベルが各特徴ベクトルに付与される（Ｓ１０４）。 In the collected sleep sound 210, the feature information 201 is extracted by the extraction unit 120 of the probability model generation device 199 for each state such as the first state and the second state (S102), and the feature vector is calculated by the analysis unit 130 (S103) And the classified feature vector is classified into a plurality of categories by machine learning in the classification unit 140, and a label indicating which category it belongs to is assigned to each feature vector (S104).
　次に、確率モデル生成部１５０は、状態毎にラベルが附された特徴ベクトルを隠れマルコフモデルを用いて確率モデルを生成し（Ｓ１０５）、得られた結果は記憶部１６０により記憶装置１５９に保存される（Ｓ１０６）。 Next, the probabilistic model generation unit 150 generates a probabilistic model using the hidden Markov model for the feature vector labeled for each state (S105), and the obtained result is stored in the storage unit 159 by the storage unit 160. (S106).
　また、確率モデル生成装置１９９は、状態が既知の新たな睡眠音２１０を収集するまで待機し（Ｓ１０８）、状態が既知の新たな睡眠音２１０を収集した場合（Ｓ１０７：Ｙｅｓ）は、新しく確率モデルを生成し、記憶装置１５９に保存される確率モデルを更新する。 In addition, the probability model generation device 199 waits until collecting a new sleeping sound 210 whose state is known (S108), and when collecting a new sleeping sound 210 whose state is known (S107: Yes), a new probability A model is generated and the probability model stored in the storage unit 159 is updated.
　次に、判定フェーズにおける睡眠の質判定システム１００の動作の例を説明する。図６は、判定フェーズにおける睡眠の質判定システムの動作の流れを示すフローチャートである。 Next, an example of the operation of the sleep quality determination system 100 in the determination phase will be described. FIG. 6 is a flowchart showing the flow of the operation of the sleep quality determination system in the determination phase.
　スマートフォンなど端末装置１０１においてプログラムとして実現される音記録部１１０で記録された睡眠音２１０を、ネットワーク３００等を介して確率モデル生成装置１９９に送信する（Ｓ２０１）。送信された睡眠音２１０は、確率モデル生成装置１９９の記憶装置１５９に記憶される。 The sleep sound 210 recorded by the sound recording unit 110 realized as a program in the terminal device 101 such as a smartphone is transmitted to the probabilistic model generation device 199 via the network 300 and the like (S201). The transmitted sleep sound 210 is stored in the storage unit 159 of the probability model generation device 199.
　記憶された睡眠音２１０は、抽出部１２０によって特徴情報２０１が抽出され（Ｓ２０２）、解析部１３０によって特徴ベクトルが算出され（Ｓ２０３）、算出された特徴ベクトルが分類部１４０における機械学習によって複数のカテゴリに分類され、どこのカテゴリに属するかを示すラベルが各特徴ベクトルに付与される（Ｓ２０４）。 The feature information 201 is extracted by the extraction unit 120 (S202), the feature vector is calculated by the analysis unit 130 (S203), and the calculated feature vector is processed by the machine learning in the classification unit 140. A label that is classified into a category and indicates which category it belongs to is assigned to each feature vector (S204).
　次に、尤度算出部１７０は、状態毎にラベルが附された特徴ベクトルを第一状態に対応した確率モデル、および、第二状態に対応した確率モデルのそれぞれに適用して尤度を算出する（Ｓ２０５）。判定部１８０は、得られた尤度に基づき判定対象の睡眠音２１０が第一状態か第二状態を判定する（Ｓ２０６）。判定結果は端末装置１０１に送信され、結果提示部１９０により端末装置１０１の表示装置１１２に表示される（Ｓ２０７）。 Next, the likelihood calculation unit 170 calculates the likelihood by applying the feature vector labeled for each state to each of the probability model corresponding to the first state and the probability model corresponding to the second state. (S205). The determination unit 180 determines whether the sleep sound 210 to be determined is in the first state or the second state based on the obtained likelihood (S206). The determination result is transmitted to the terminal device 101, and is displayed on the display device 112 of the terminal device 101 by the result presentation unit 190 (S207).
　本実施の形態の睡眠の質判定システム１００によれば、少なくとも第一状態、および、第二状態を含む複数状態の睡眠パターンをそれぞれ反映させた確率モデルを複数の異なる隠れ変数の数で隠れマルコフモデルにより生成し、得られた複数の確率モデルを用いて判定対象の睡眠音２１０の尤度を算出し、多数決により判定対象の睡眠の質を判定することにより、音の情報のみから高い精度で睡眠の質を判定することが可能となる。 According to the sleep quality determination system 100 of the present embodiment, a probability model in which the sleep patterns of at least the first state and the plurality of states including the second state are reflected is hidden Markov by the number of different hidden variables. The likelihood of the sleep sound 210 to be determined is calculated using a plurality of probability models generated by the model, and the quality of the sleep to be determined is determined by majority decision. It is possible to determine the quality of sleep.
　従って、大がかりな睡眠ポリグラフや専用デバイスを必要とせず、汎用の端末装置を用いて簡便に録音された睡眠音２１０から睡眠の質を判定することができる。 Therefore, it is possible to determine the quality of sleep from the sleep sound 210 recorded easily using a general-purpose terminal device without requiring a large-scale sleep polygraph or a dedicated device.
　なお、本願発明は、上記実施の形態に限定されるものではない。例えば、本明細書において記載した構成要素を任意に組み合わせて、また、構成要素のいくつかを除外して実現される別の実施の形態を本願発明の実施の形態としてもよい。また、上記実施の形態に対して本願発明の主旨、すなわち、請求の範囲に記載される文言が示す意味を逸脱しない範囲で当業者が思いつく各種変形を施して得られる変形例も本願発明に含まれる。 The present invention is not limited to the above embodiment. For example, another embodiment realized by arbitrarily combining the components described herein and excluding some of the components may be used as an embodiment of the present invention. The present invention also includes modifications obtained by applying various modifications to those skilled in the art without departing from the spirit of the present invention, that is, the meaning described in the claims with respect to the above embodiment. Be
　例えば、確率モデルを生成するための睡眠音２１０は、不特定多数から収集する場合ばかりで無く、個体毎に第一状態の睡眠音２１０、および、第二状態の睡眠音２１０を蓄積し、過去の睡眠音２１０に基づき個体毎の確率モデルを生成してもかまわない。 For example, the sleep sound 210 for generating a probability model is not only collected from an unspecified number but also stores the sleep sound 210 in the first state and the sleep sound 210 in the second state for each individual. It is also possible to generate a probability model for each individual based on the sleep sound 210 of.
　また、複数の異なる隠れ変数の数により得られた確率モデルに基づく尤度によって多数決により第一状態か、第二状態かを判定したが、１つの隠れ変数の数により得られた確率モデルによる尤度で判定を行ってもかまわない。 Moreover, although it was determined whether it is the first state or the second state by majority decision by the likelihood based on the probability model obtained by the number of different hidden variables, the likelihood by the probability model obtained by the number of one hidden variable The determination may be made in degrees.
　また、判定部１８０は、複数の異なる隠れ変数の数により得られた確率モデルに基づき算出された尤度を尤度ベクトルとし、パターン識別により睡眠の質が第一状態か第二状態かを判定してもかまわない。パターン識別としては、例えばサポートベクトルマシン（ＳＶＭ）であり、多層パーセプトロン、Ｒａｎｄｏｍ　Ｆｏｒｅｓｔ、ロジスティック識別などを例示することができる。 Further, the determination unit 180 determines the likelihood calculated based on the probability model obtained by the number of different hidden variables as the likelihood vector, and determines whether the sleep quality is the first state or the second state by pattern identification. It does not matter. The pattern identification is, for example, a support vector machine (SVM), and can be exemplified by multilayer perceptron, random forest, logistic identification, and the like.
　具体的に例えば、判定部１８０は、図７に示すように、尤度モデル生成部１８１を備えている。尤度モデル生成部１８１は、第一状態に含まれる複数の睡眠音２１０について、それぞれ分類された特徴ベクトル、および第二状態に含まれる複数の睡眠音２１０について、それぞれ分類された特徴ベクトルに基づき、確率モデル生成部１５０で生成された第一モデル、および第二モデルを用いて尤度算出部１７０に尤度を算出させ、得られた尤度ベクトルを入力とし、第一状態と第二状態とをラベルとして識別器のモデルを生成する。 Specifically, for example, as illustrated in FIG. 7, the determination unit 180 includes a likelihood model generation unit 181. The likelihood model generation unit 181 is configured based on the feature vectors classified respectively for the plurality of sleep sounds 210 included in the first state and the plurality of sleep sounds 210 included in the second state. Using the first model and the second model generated by the probability model generation unit 150 to calculate the likelihood by the likelihood calculation unit 170, and using the obtained likelihood vector as an input, the first state and the second state And generate a model of the discriminator as a label.
　ここで、パターン識別の識別器としては、上記の通り特に限定されるものではないが、尤度モデルを生成するためのデータ点数（特徴ベクトルの数）が比較的多い場合はニューラルネットワークなどが適し、比較的少ない場合はサポートベクトルマシン（ＳＶＭ）などが適していると考えられる。その他の識別器としてはＡｄａＢｏｏｓｔ、ナイーブベイズ、決定木などを例示できる。 Here, the pattern identification classifier is not particularly limited as described above, but a neural network is suitable when the number of data points (number of feature vectors) for generating the likelihood model is relatively large. When the number is relatively small, it is considered that a support vector machine (SVM) or the like is suitable. Other discriminators include AdaBoost, naive Bayes, decision trees, and the like.
　次に、判定部１８０は、状態が未知の睡眠音２１０による特徴ベクトルに基づき尤度算出部１７０が算出した第一尤度、第二尤度を尤度モデル生成部１８１が生成したモデルに適用して睡眠の質が第一状態か第二状態かを判定する。 Next, the determination unit 180 applies the first likelihood and the second likelihood calculated by the likelihood calculation unit 170 based on the feature vector of the sleep sound 210 whose state is unknown to the model generated by the likelihood model generation unit 181. Then, determine whether the quality of sleep is in the first state or the second state.
　なお、尤度モデルを作成するための睡眠音２１０は、確率モデル生成部１５０がモデルを生成する際に用いられるものでもよく、別の睡眠音２１０でも構わない。またこれらを混在させても構わない。 The sleep sound 210 for creating the likelihood model may be one used when the probability model generation unit 150 generates a model, or another sleep sound 210 may be used. Moreover, you may mix these.
　また、睡眠の質判定システム１００が、ネットワーク３００で接続された端末装置１０１と確率モデル生成装置１９９とを備える場合を説明したが、各処理部が１つの装置により実現されるものでもかまわない。また、複数の装置により睡眠の質判定システム１００が構築されている場合、各処理部がいずれの装置に備えられるかは任意であり、複数の装置が同じ処理部をそれぞれ備えていてもかまわない。 Although the case where the sleep quality determination system 100 includes the terminal device 101 and the probabilistic model generation device 199 connected by the network 300 has been described, each processing unit may be realized by one device. In addition, when the sleep quality determination system 100 is configured by a plurality of devices, which device each processing unit is provided with is optional, and a plurality of devices may be provided with the same processing unit. .
　また、睡眠の質判定システム１００は、スマートフォンやタブレット端末などに音記録部１１０、抽出部１２０、解析部１３０、分類部１４０、確率モデル生成部１５０、記憶部１６０、尤度算出部１７０、判定部１８０、結果提示部１９０などの各処理部を実現することのできるアプリと称されるプログラムをインストールすることにより、スマートフォンなどを睡眠の質判定システム１００として機能させてもかまわない。 In addition, the sleep quality determination system 100 includes a sound storage unit 110, an extraction unit 120, an analysis unit 130, a classification unit 140, a probability model generation unit 150, a storage unit 160, a likelihood calculation unit 170, and a smart phone or a tablet terminal. A smartphone or the like may function as the sleep quality determination system 100 by installing a program called an application that can realize each processing unit such as the unit 180 and the result presentation unit 190.
　本願発明は、人間および動物の睡眠状態を判定して提示し、健康・生活習慣の管理や改善などに利用できる。 INDUSTRIAL APPLICABILITY The present invention can be used to determine and present sleep states of humans and animals, and can be used to manage or improve health and lifestyle habits.
１００　睡眠の質判定システム
１０１　端末装置
１１０　音記録部
１１１　マイクロフォン
１１２　表示装置
１１９　記憶装置
１２０　抽出部
１３０　解析部
１４０　分類部
１５０　確率モデル生成部
１５９　記憶装置
１６０　記憶部
１７０　尤度算出部
１８０　判定部
１９０　結果提示部
１９９　確率モデル生成装置
２００　生体
２０１　特徴情報
２１０　睡眠音
３００　ネットワーク 100 sleep quality determination system 101 terminal device 110 sound recording unit 111 microphone 112 display device 119 storage device 120 extraction unit 130 analysis unit 140 classification unit 150 probability model generation unit 159 storage unit 160 storage unit 170 likelihood calculation unit 180 determination unit 190 Result presentation part 199 probability model generation device 200 living body 201 feature information 210 sleep sound 300 network

Claims (9)

  1. 　睡眠中における生体が発する音を睡眠音として記録する音記録部と、
    　記録された前記睡眠音の中から特徴部分である特徴情報を複数個抽出する抽出部と、
    　抽出された特徴情報に基づきそれぞれの特徴ベクトルを算出する解析部と、
    　複数の前記特徴ベクトルを複数種類に分類する分類部と、
    　睡眠の質が第一状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第一モデルと、睡眠の状態が第二状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第二モデルとを生成する確率モデル生成部と、
    　前記第一モデル、および、前記第二モデルを記憶させる記憶部とを備え、
    　前記音記録部は、
    　判定対象の睡眠音を記録し、
    　前記抽出部は、
    　前記睡眠音から特徴情報を抽出し、
    　前記解析部は、
    　前記特徴情報から特徴ベクトルを算出し、
    　前記分類部は、
    　前記特徴ベクトルを分類し、
    　さらに、
    　前記判定対象の分類された特徴ベクトルを前記第一モデルに適用して尤度である第一尤度と前記第二モデルに適用して尤度である第二尤度を算出する尤度算出部と、
    　算出された前記第一尤度と第二尤度とに基づき前記判定対象の睡眠の質が第一状態か第二状態かを判定する判定部とを備える
    睡眠の質判定システム。     A sound recording unit that records the sound emitted by the living body during sleep as sleep sound;
    An extraction unit that extracts a plurality of pieces of feature information that is a feature portion from the recorded sleep sound;
    An analysis unit that calculates each feature vector based on the extracted feature information;
    A classification unit that classifies a plurality of feature vectors into a plurality of types;
    The state of sleep was in the second state, the first model being one of the probabilistic models by processing the classified feature vectors of the sleep sound in which the quality of sleep was the first state using the hidden Markov model A probability model generation unit configured to generate the second feature, which is one of the probability models, by processing the classified feature vectors of the sleep sound using a hidden Markov model;
    A storage unit configured to store the first model and the second model;
    The sound recording unit is
    Record the sleep sound of the judgment target,
    The extraction unit
    Extracting feature information from the sleep sound;
    The analysis unit
    Calculate a feature vector from the feature information;
    The classification unit
    Classify the feature vectors,
    further,
    A likelihood calculation unit that calculates a second likelihood that is a likelihood by applying the classified feature vector of the determination target to the first model and applying the likelihood to the first likelihood and the second model When,
    A sleep quality determination system comprising: a determination unit that determines whether the sleep quality of the determination target is the first state or the second state based on the calculated first likelihood and second likelihood.
  2. 　前記分類部は、
    　複数の前記特徴ベクトルに対しクラスタ分析を行って少数のカテゴリに自動分類し、各カテゴリに属する前記特徴ベクトルに同一のラベルを付与する
    請求項１に記載の睡眠の質判定システム。     The classification unit
    The sleep quality determination system according to claim 1, wherein cluster analysis is performed on a plurality of the feature vectors to automatically classify them into a small number of categories, and the same label is attached to the feature vectors belonging to each category.
  3. 　前記確率モデル生成部は、
    　隠れ変数の数が３以上、９以下の範囲の１つであるＮ（Ｎは整数）の第一モデル、および、第二モデルを生成する
    請求項１または２に記載の睡眠の質判定システム。     The probabilistic model generation unit
    The system according to claim 1 or 2, wherein a first model of N (N is an integer) whose number of hidden variables is one of 3 or more and 9 or less and N is an integer is generated.
  4. 　前記確率モデル生成部は、
    　隠れ変数の数がＮ－１、Ｎ、Ｎ＋１の第一モデル、および、第二モデルをそれぞれ生成し、
    　前記尤度算出部は、
    　生成された第一モデル、および、第二モデルのそれぞれについて、第一尤度と第二尤度とを算出し、
    　前記判定部は、
    　同一数の隠れ変数の数に対応する第一尤度と第二尤度とを比較して尤度の高いものを勝ちとし、多数決により睡眠の質が第一状態か第二状態かを判定する
    請求項３に記載の睡眠の質判定システム。     The probabilistic model generation unit
    Generate a first model and a second model with N−1, N, N + 1 number of hidden variables, respectively
    The likelihood calculation unit
    Calculate first likelihood and second likelihood for each of the generated first model and second model,
    The determination unit is
    The first likelihood and the second likelihood corresponding to the same number of hidden variables are compared and the one with the highest likelihood is determined as the winner, and the majority decision determines whether the sleep quality is the first state or the second state The sleep quality determination system according to claim 3.
  5. 　前記確率モデル生成部は、
    　複数の隠れ変数の数について第一モデル、および、第二モデルをそれぞれ生成し、
    　前記尤度算出部は、
    　生成された第一モデル、および、第二モデルのそれぞれについて、第一尤度と第二尤度とを算出し、
    　前記判定部は、
    　算出された尤度を尤度ベクトルとし、パターン識別により睡眠の質が第一状態か第二状態かを判定する
    請求項３に記載の睡眠の質判定システム。     The probabilistic model generation unit
    Generate a first model and a second model for the number of multiple hidden variables,
    The likelihood calculation unit
    Calculate first likelihood and second likelihood for each of the generated first model and second model,
    The determination unit is
    The sleep quality determination system according to claim 3, wherein the calculated likelihood is a likelihood vector, and it is determined whether the sleep quality is the first state or the second state by pattern identification.
  6. 　前記パターン識別は、教師有り学習の識別器であり、
    　前記判定部は、
    　前記識別器に対応するモデルを生成する尤度モデル生成部を備え、生成された尤度モデルに基づき前記尤度ベクトルを入力として睡眠の質が第一状態か第二状態かを判定する
    請求項５に記載の睡眠の質判定システム。     The pattern identification is a classifier for supervised learning,
    The determination unit is
    A likelihood model generation unit that generates a model corresponding to the classifier, and based on the generated likelihood model, receives the likelihood vector and determines whether the sleep quality is in the first state or the second state. The sleep quality determination system according to 5.
  7. 　前記確率モデル生成部を備える確率モデル生成装置と、前記確率モデル生成装置にネットワークを介して接続される複数の端末装置とを備え、
    　前記端末装置は、
    　前記音記録部と、
    　さらに、前記判定部による判定結果を表示装置に提示させる結果提示部を備える
    請求項１から６のいずれか一項に記載の睡眠の質判定システム。     A probability model generation device including the probability model generation unit; and a plurality of terminal devices connected to the probability model generation device via a network;
    The terminal device is
    The sound recording unit;
    The sleep quality determination system according to any one of claims 1 to 6, further comprising a result presentation unit configured to present a determination result by the determination unit on a display device.
  8. 　睡眠中における生体が発する音を睡眠音として記録する音記録部と、
    　記録された前記睡眠音の中から特徴部分である特徴情報を複数個抽出する抽出部と、
    　抽出された特徴情報に基づきそれぞれの特徴ベクトルを算出する解析部と、
    　複数の前記特徴ベクトルを複数種類に分類する分類部と、
    　睡眠の質が第一状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第一モデルと、睡眠の状態が第二状態であった前記睡眠音の分類された特徴ベクトルを隠れマルコフモデルを用いた処理により確率モデルの１つである第二モデルとを生成する確率モデル生成部と、
    　前記第一モデル、および、前記第二モデルを記憶させる記憶部とを含み、
    前記各処理をコンピュータに実行させることにより上記各部を実現する睡眠の質モデル作成プログラム。     A sound recording unit that records the sound emitted by the living body during sleep as sleep sound;
    An extraction unit that extracts a plurality of pieces of feature information that is a feature portion from the recorded sleep sound;
    An analysis unit that calculates each feature vector based on the extracted feature information;
    A classification unit that classifies a plurality of feature vectors into a plurality of types;
    The state of sleep was in the second state, the first model being one of the probabilistic models by processing the classified feature vectors of the sleep sound in which the quality of sleep was the first state using the hidden Markov model A probability model generation unit configured to generate the second feature, which is one of the probability models, by processing the classified feature vectors of the sleep sound using a hidden Markov model;
    A storage unit configured to store the first model and the second model;
    A sleep quality model creation program for realizing the above-mentioned sections by causing a computer to execute the above-mentioned respective processing.
  9. 　睡眠中における生体が発する音を睡眠音として記録する音記録部と、
    　記録された前記睡眠音の中から特徴部分である特徴情報を複数個抽出する抽出部と、
    　抽出された特徴情報に基づきそれぞれの特徴ベクトルを算出する解析部と、
    　複数の前記特徴ベクトルを複数種類に分類する分類部と、
    　隠れマルコフモデルを用いて生成された異なる睡眠の質にそれぞれ対応する第一モデル、および、前記第二モデルを記憶させる記憶部とを備と、
    　分類された特徴ベクトルを前記第一モデルに適用して尤度である第一尤度と前記第二モデルに適用して尤度である第二尤度を算出する尤度算出部と、
    　算出された前記第一尤度と第二尤度とに基づき前記判定対象の睡眠の質が第一状態か第二状態かを判定する判定部とを含み、
    前記各処理をコンピュータに実行させることにより上記各部を実現する睡眠の質判定プログラム。     A sound recording unit that records the sound emitted by the living body during sleep as sleep sound;
    An extraction unit that extracts a plurality of pieces of feature information that is a feature portion from the recorded sleep sound;
    An analysis unit that calculates each feature vector based on the extracted feature information;
    A classification unit that classifies a plurality of feature vectors into a plurality of types;
    A first model respectively corresponding to different sleep qualities generated using a Hidden Markov Model, and a storage unit for storing the second model;
    A likelihood calculator that applies the classified feature vector to the first model to apply a first likelihood as a likelihood and the second model to calculate a second likelihood as a likelihood;
    A determination unit that determines whether the quality of sleep to be determined is the first state or the second state based on the calculated first likelihood and second likelihood;
    A sleep quality determination program that implements the above-described units by causing a computer to execute the processes.
PCT/JP2018/029517 2017-08-22 2018-08-07 Sleep quality assessment system, sleep quality model creation program, and sleep quality assessment program WO2019039261A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2019538050A JPWO2019039261A1 (en) 2017-08-22 2018-08-07 Sleep quality assessment system, sleep quality modeling program, and sleep quality assessment program

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2017158957 2017-08-22
JP2017-158957 2017-08-22

Publications (1)

Publication Number Publication Date
WO2019039261A1 true WO2019039261A1 (en) 2019-02-28

Family

ID=65438736

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2018/029517 WO2019039261A1 (en) 2017-08-22 2018-08-07 Sleep quality assessment system, sleep quality model creation program, and sleep quality assessment program

Country Status (2)

Country Link
JP (1) JPWO2019039261A1 (en)
WO (1) WO2019039261A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110464303A (en) * 2019-08-15 2019-11-19 深圳和而泰家居在线网络科技有限公司 Sleep quality appraisal procedure and device

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011155048A1 (en) * 2010-06-10 2011-12-15 富士通株式会社 Audio processing device, and breathing detection device
JP2013094340A (en) * 2011-10-31 2013-05-20 Omron Healthcare Co Ltd Sleep evaluation device and sleep evaluation program
JP2013165828A (en) * 2012-02-15 2013-08-29 Tanita Corp Sleep evaluation device and sleep evaluation system
US20140188006A1 (en) * 2011-05-17 2014-07-03 University Health Network Breathing disorder identification, characterization and diagnosis methods, devices and systems
JP2015217130A (en) * 2014-05-16 2015-12-07 トヨタ自動車株式会社 Sleep stage estimation device
JP2016209327A (en) * 2015-05-11 2016-12-15 沖電気工業株式会社 Sleep depth estimation device, sleep depth estimation method and program
JP2016209311A (en) * 2015-05-08 2016-12-15 花王株式会社 Sleep feeling determination method and device

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011155048A1 (en) * 2010-06-10 2011-12-15 富士通株式会社 Audio processing device, and breathing detection device
US20140188006A1 (en) * 2011-05-17 2014-07-03 University Health Network Breathing disorder identification, characterization and diagnosis methods, devices and systems
JP2013094340A (en) * 2011-10-31 2013-05-20 Omron Healthcare Co Ltd Sleep evaluation device and sleep evaluation program
JP2013165828A (en) * 2012-02-15 2013-08-29 Tanita Corp Sleep evaluation device and sleep evaluation system
JP2015217130A (en) * 2014-05-16 2015-12-07 トヨタ自動車株式会社 Sleep stage estimation device
JP2016209311A (en) * 2015-05-08 2016-12-15 花王株式会社 Sleep feeling determination method and device
JP2016209327A (en) * 2015-05-11 2016-12-15 沖電気工業株式会社 Sleep depth estimation device, sleep depth estimation method and program

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110464303A (en) * 2019-08-15 2019-11-19 深圳和而泰家居在线网络科技有限公司 Sleep quality appraisal procedure and device
CN110464303B (en) * 2019-08-15 2022-04-01 深圳数联天下智能科技有限公司 Sleep quality evaluation method and device

Also Published As

Publication number Publication date
JPWO2019039261A1 (en) 2020-09-24

Similar Documents

Publication Publication Date Title
Bardou et al. Lung sounds classification using convolutional neural networks
Zhang et al. Pdvocal: Towards privacy-preserving parkinson's disease detection using non-speech body sounds
Xie et al. Wavelet-based sparse functional linear model with applications to EEGs seizure detection and epilepsy diagnosis
Alqudah Towards classifying non-segmented heart sound records using instantaneous frequency based features
Zeinali et al. Heart sound classification using signal processing and machine learning algorithms
Taran et al. Sleep apnea detection using artificial bee colony optimize hermite basis functions for EEG signals
Emanet et al. A comparative analysis of machine learning methods for classification type decision problems in healthcare
Lima et al. Kernel machines for epilepsy diagnosis via EEG signal classification: A comparative study
Shahnaz et al. Emotion recognition based on wavelet analysis of Empirical Mode Decomposed EEG signals responsive to music videos
WO2017075601A1 (en) Programmable electronic stethoscope devices, algorithms, systems, and methods
Arnau-González et al. ES1D: A deep network for EEG-based subject identification
Nalwaya et al. Automated emotion identification using Fourier–Bessel domain-based entropies
Tucker et al. A data mining methodology for predicting early stage Parkinson's disease using non-invasive, high-dimensional gait sensor data
Singer et al. Classification of severity of trachea stenosis from EEG signals using ordinal decision-tree based algorithms and ensemble-based ordinal and non-ordinal algorithms
Yang et al. Hybrid machine learning scheme for classification of BECTS and TLE patients using EEG brain signals
JP2023089729A (en) Computer system and emotion estimation method
Wang et al. A multi-channel UNet framework based on SNMF-DCNN for robust heart-lung-sound separation
Ardani et al. A new approach to signal filtering method using K-means clustering and distance-based Kalman filtering
WO2019039261A1 (en) Sleep quality assessment system, sleep quality model creation program, and sleep quality assessment program
Wu et al. Statistical sleep pattern modelling for sleep quality assessment based on sound events
Shervegar et al. Heart sound classification using Gaussian mixture model
Wu et al. Personal sleep pattern visualization using sequence-based kernel self-organizing map on sound data
Hsu Embedded grey relation theory in Hopfield neural network: application to motor imagery EEG recognition
Liu et al. Human behavior sensing: challenges and approaches
Nehary et al. A deep convolutional neural network classification of heart sounds using fractional fourier transform

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18847692

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2019538050

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18847692

Country of ref document: EP

Kind code of ref document: A1