JP2015217130A - Sleep stage estimation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device that makes a determination of a sleep stage more accurately and more clearly in response to a change in a state of a subject.SOLUTION: One selected from determination model data previously obtained under a subject model condition different to each other is selected, and a group of a first feature value extracted from respiratory waveform data is integrated by using information of the selected determination model data, and thereby a group of a second feature value is generated. The group of the second feature value is created so that the distribution of equivalent points to the group of the second feature value may be separated substantially to each other by a sleep stage in a determination space where the group of the second feature value of the selected determination model data becomes the variable and prepared. The determination space of the selected determination model data is divided into two or more determination regions by the sleep stage, and a determination is made for the sleep stage corresponding to the region belonging to the group of the second feature value.

Description

  The present invention relates to an apparatus for estimating a sleep stage, and more particularly, to an apparatus for estimating a sleep stage of a person to be measured based on measurement data of a respiratory waveform of the person to be measured.

  It is known that the sleep state of a person can be classified into stages (sleep stages) such as stages I to IV in REM sleep and non-REM sleep, based on the electroencephalogram and the pattern of eye movement. In actual sleep, the above five sleep stages are passed in various orders, and the history of the sleep stages is important information for human health management. Various techniques have been proposed for estimating or determining the sleep stage. In this regard, the sleep stage is determined by measuring brain waves and eye movements based on international criteria. In that case, large-scale equipment is required, for small-scale facilities, homes or vehicles. It is not realistic to introduce equipment for measuring electroencephalogram and eye movement as a device for use. On the other hand, it has been found that the sleep stage is reflected in a person's respiratory state. In view of this, various techniques have been proposed for monitoring the respiratory state of a person during sleep and estimating the sleep stage from the monitored respiratory state. For example, in Patent Document 1, a respiration signal (a signal representing human vibration due to respiration) is extracted from pressure fluctuations of the measurement subject lying on the bedding, and the amplitude and frequency magnitude and fluctuation of the extracted respiration signal are extracted. Deep sleep and light sleep after judging the state of coarse body movement, thin body movement, and inbody movement, wakefulness, sleep falling, unstable and stable , A device for determining a sleep stage such as REM sleep for each unit time (epoch) is described (Patent Document 1 further based on a history of sleep stage determination results obtained during a single sleep) A configuration for evaluating sleep quality is disclosed.) As a sensor technology that can be used for estimating a sleep stage based on a respiratory state, in Patent Document 2, a heart signal is separated from data of an accelerometer attached to a chest of a measurement subject. An example of a technique for deriving a respiratory signal from a cardiac signal is disclosed.

JP2008-301951 Special table 2013-511349

"Easy-to-understand pattern recognition" Kenichiro Ishii and three others, Ohmsha, 1998, pages 123-133

  As already mentioned, it has been found that a person's sleep stage is reflected in the person's breathing state, and the difference in sleep stage is called the breathing waveform or breathing signal of the person's chest associated with breathing exercises. Or, it appears in a change in the aspect of the movement of the abdomen (more specifically, for example, a change in potential of a piezoelectric sensor or the like attached to the chest or abdomen). However, the correspondence between changes in the respiratory waveform and the sleep stage is complex, and the state of the person whose respiration is measured, such as age, gender, physical characteristics (height, weight, body shape, etc.) There is a possibility that the correspondence between the change in the respiratory waveform and the sleep stage changes depending on the physical condition and the like. For example, values measured directly from the signal waveform, such as the amplitude or frequency magnitude of the potential signal of the respiratory waveform, do not directly correspond to each stage of the sleep stage (thus usually from the signal waveform). It is difficult to determine the sleep stage based only on the directly measured value.) The difference in the sleep stage is due to various state quantities obtained from the potential signal of the respiratory waveform, for example, the average value of the peak-to-peak distance of the waveform, Differences in multiple parameters such as standard deviation, average distance between bottoms, standard deviation, average value between peak and bottom, standard deviation, etc. appear in various ways, and differences in these multiple parameters The appearance of can vary depending on the condition of the person whose respiration is measured even in the same person. Therefore, when trying to estimate the sleep stage from the respiratory state, there are differences in the sleep stage in any state quantity in the respiratory waveform while taking into account the various states assumed in the person whose respiration is measured. It is necessary to generate a feature value from the respiration waveform after accurately searching for appearance, that is, a search for a feature value representing the sleep stage in the respiration waveform. In addition, when estimating or determining the sleep stage based on the feature amount representing the sleep stage in the respiratory waveform thus generated, it is preferable that the sleep stage can be determined more accurately and more clearly. Furthermore, it is preferable that the sleep stage can be determined with the highest possible accuracy even if there is a change in the state of a person who measures the same respiration.

  Thus, one object of the present invention is to measure respiration using a characteristic amount searched for as an amount representing a sleep stage in a respiration waveform in an apparatus for estimating a sleep stage from a person's respiration state. The determination of the sleep stage is achieved more accurately and more clearly in response to changes in the person's condition. Note that, as already described, the determination of the sleep stage is performed through measurement of brain waves and eye movements based on the international determination standard, so the determination of the sleep stage based on the respiratory state is an estimation of the sleep stage. Therefore, in the description of the apparatus of the present invention, “determination of sleep stage” is synonymous with “estimation of sleep stage” unless otherwise specified.

  According to this invention, said subject is achieved by the following sleep stage estimation apparatus. The sleep stage estimation device of the present invention, as a basic configuration, extracts a respiratory waveform data acquisition unit that acquires respiratory waveform data of a measurement subject, a first feature amount group from the acquired respiratory waveform data, and A feature value generating unit that generates a second feature value set generated by combining the first feature value group, and respiratory waveform data in which sleep stages are labeled in advance under different measurement target model conditions A determination model data storage unit that stores a plurality of determination model data obtained based on the determination model data, and is used for the measurement subject among the plurality of determination model data stored in the determination model data storage unit. A determination model selection unit that selects determination model data to be determined, and a sleep stage determination unit that determines a sleep stage corresponding to the second set of feature quantities. In the above device, the “sleep stage” is a classification of sleep states according to the depth of sleep, and in principle, as already mentioned, based on the behavior of the brain wave and eye movement patterns. It may be the same as the sleep state classified according to the stage. In the present invention, the sleep state does not necessarily have to be classified into the five sleep stages and the wakefulness stage as described above, and is classified into three stages such as wakefulness, light sleep and deep sleep, for example. May be.

  In the above basic configuration, first, the “respiration waveform data” in the respiration waveform data acquisition unit may be any amount that changes in conjunction with the respiration movement of the measurement subject. As a typical example, the movement of the chest or abdomen of a person corresponding to a breathing exercise, more specifically, a piezoelectric sensor attached to the chest or abdomen reflecting the movement of the subject's chest or abdomen The potential data corresponding to the respiratory movement measured using the above may be employed.

  In the feature value generation unit, the “first feature value group” means amplitude, frequency, peak-to-peak, bottom-to-bottom, peak-to-bottom in respiratory waveform data measured along time. A group of various state quantities and / or their rate of change, typically measured directly from the respiratory waveform data, such as length of time, rate of change, etc. The “second feature value set” is an amount generated from the “first feature value group” in the form described later, and is directly used in the sleep stage determination process. The amount will be.

  In the determination model data storage unit, as described above, determination model data obtained in advance based on respiratory waveform data measured in a state in which the sleep stage is specified under a plurality of different measurement subject model conditions are stored. Remembered. Here, the “measured person model” is a person to be measured who measures the respiratory waveform data during sleep for collecting determination model data. In addition, the “measured person model conditions different from each other” are conditions of a person to be measured in various states, for example, a plurality of persons to be measured having different age, sex, and physical characteristics. In the measurer model, it is information obtained from respiratory waveform data measured in a state where the sleep stage is specified (for example, a state where the sleep state is determined by monitoring brain waves and eye movements), and the sleep stage Data used or referred to for determination. That is, the determination model data storage unit prepares determination model data for a plurality of various measurement subject model conditions.

Specifically, the individual determination model data stored in the determination model data storage unit includes the following data. :
(1) A first feature amount corresponding to a first feature amount group in respiration waveform data measured in advance at various sleep stages under one measured person model condition among different measured person model conditions Data of at least a part of the group of model features.
(2) Each of the first model feature values when the second set of model feature values obtained by linearly combining the first model feature value group of the one measured person model condition is generated. A group of weighting factor data. Here, the second set of model feature values has a distribution of point groups corresponding to the second set of model feature values in the discriminant space in which the second set of model feature values becomes a variable. A set of quantities generated to be substantially separated from each other by a sleep stage associated with each of the second set of model feature quantities.
(3) Discriminant boundary function data that provides a discriminant boundary that divides the discriminant space into regions corresponding to each of the sleep stages associated with each of the second set of model feature values.

  In the above, first, the “first model feature group” in (1) is the first extracted from the measured respiratory waveform data of the measurement subject model under a certain measurement subject model condition. This is a group of feature amounts corresponding to a group of feature amounts. A part of the “group of first model feature values” is prepared for selecting determination model data used for a sleep stage determination process described later.

  An amount obtained by linearly combining the “first group of model feature values” is the “second set of model feature values” referred to in (2) above. More specifically, the “second set of model feature values” is such that, when there are a plurality of sets of second model feature values, the second set of model feature values becomes a variable as described above. The sleep stage in which the distribution of the point cloud corresponding to the second set of model feature values is associated with each of the second set of model feature values in the extended space (referred to as “discriminant space”) So as to be substantially separated from each other. That is, since a sleep stage is specified for each set of second model feature values, a space around each of the second set of model feature values is considered, When points corresponding to a plurality of sets of second model feature values generated from the first model feature value group are plotted, a sleep stage distribution is configured in the space. In that case, in the space, the distribution of the sleep stages is substantially separated from each other for each stage, that is, the distribution areas of the points exist in different spatial areas for each sleep stage. A set of model feature quantities is generated. As described above, the “group of weighting coefficient data” in the above (2) is the first model feature value group from the first model feature quantity group so that the sleep stage distributions are substantially separated from each other. This is a group of weighting coefficients to be multiplied to each of the first group of model feature values when the set of second model feature values is generated. Further, the discriminant boundary function data of (3) above is the boundary of the region where the point cloud corresponding to each sleep stage is distributed in the discriminant space formed by the set of the second model feature values prepared as described above. This is data for specifying a function that defines (discrimination boundary).

  The generation of the second model feature quantity set from the first model feature quantity group can be achieved using, for example, a canonical discrimination method or a linear discrimination method (Non-Patent Document 1). . The function that defines the discriminant boundary uses, for example, a canonical discriminant method (Non-patent Document 1) using a group of second model feature amounts in which the sleep stage in the discriminant space is specified. Can be achieved. It should be understood that in the process of generating the second model feature set from the first model feature set, as described above, the second model feature set is In the discriminant space spanned as a variable, that is, as an axis, the point cloud corresponding to the plurality of sets of second model feature values is substantially separated from each other by the sleep stage associated with each point. That is, the areas distributed in the areas and associated with the sleep stages are in a state of being divided by the discrimination boundary. That is, the second set of model feature values obtained from a first group of model feature values is that the region to which the discriminant space belongs can be clearly identified by the associated sleep stage. Become. Therefore, it is obtained from respiration waveform data measured by an arbitrary measurement subject under conditions close to or equivalent to the measurement subject model condition of certain discriminant model data (corresponding to the first model feature amount group). The second feature value group (corresponding to the second model feature value group) obtained by linearly combining the first feature value group using the weighting coefficient group of the discriminant model data. The corresponding point of is included in the sleep stage region corresponding to the sleep stage of the person being measured in the discrimination space, and this can be clearly understood from the positional relationship between the point and the discrimination boundary. It becomes. The “first model feature group” (and “first feature group”) may be a group of more than three variables (typically, from several tens of variables). The “second model feature set” and the “second feature set” typically include the “first feature so as to form a three-dimensional space. It is a set of three variables obtained by reducing the "quantity" dimension. In that case, the discrimination space and the area corresponding to each sleep stage defined in the space can be visually recognized, and the area corresponding to the sleep stage to which each point belongs can be grasped more clearly.

  In the determination model selection unit in the apparatus of the present invention described above, more specifically, in order to determine the sleep stage, selection of determination model data used for each person to be measured is performed by operating the apparatus. Or a first feature amount corresponding to at least a part of the first model feature amount group and at least a part of the first model feature amount group of each of the plurality of determination model data. It is performed based on a comparison with a part of the group. That is, the determination model selection unit is configured to select determination model data used for each individual person to be measured by one of two processes. Specifically, in the first process, determination model data in which an operator of the apparatus (which may be the same person as the person being measured or a different person) is expected to be suitable for the person being measured; That is, the determination model data of the measurement subject model condition that matches or is close to the measurement subject condition is selected. In another process, a part of the group of first feature values extracted from the respiratory waveform data actually obtained by the measurement subject is converted into the first model feature values corresponding to the plurality of determination model data. Compared with a part of the group, the measurement target model condition expected to be suitable for the measurement target is determined by the determination model selection unit of the apparatus, and the determination model data of the measurement target model condition is selected It becomes. In this regard, the method of determining a suitable measurement target model condition by comparing a part of the first feature quantity group and a part of the first model feature quantity group is sufficiently accurate. In order to achieve this, it has been found necessary to use a part of the first feature quantity group obtained over a certain period of time. Therefore, typically, at the start of use of the apparatus, the determination model data is selected by the operator of the apparatus, and after a certain period (or a predetermined period) thereafter, the first feature amount group is selected. The selection model data may be selected based on a comparison between a part of the first model feature quantity group and a part of the first model feature quantity group.

  More specifically, the feature quantity generation unit in the above-described apparatus according to the present invention linearly combines the first feature quantity group by using the weighting coefficient data group of the selected determination model data as a weighting coefficient. A set of feature quantities is generated. In the sleep stage determination unit, more specifically, in the areas corresponding to the plurality of sleep stages divided by the determination boundary defined by the determination boundary function data in the determination space of the selected determination model data, The sleep stage corresponding to the region to which the point corresponding to the second feature quantity set belongs is determined as the sleep stage corresponding to the (current) second feature quantity set. Here, the current second feature quantity set is generated from the first feature quantity group using the weighting coefficient data group of the selected determination model data as the weighting coefficient. The second feature set was selected and expected to be generated in the discriminant space so that the distribution regions are substantially separated from each other by the corresponding sleep stages. Since the discriminant boundary in the discriminant space of the judgment model data is defined by the discriminant boundary function data, it is determined which sleep stage corresponds to the point corresponding to the second feature amount set. As a result, the sleep stage can be determined more accurately and clearly. In the sleep stage determination unit, the sleep stage determined by referring to the transition of the temporarily determined sleep stage after temporarily determining the sleep stage corresponding to the second feature amount set for each unit time. The stage may be corrected.

  In the operation of the above-described apparatus of the present invention, typically, when the determination of the sleep stage is executed for the first time for a certain person to be measured, a determination model that is expected to be suitable for the person to be measured first. Data selection is made by the operator. Then, the determination model selection unit passes the weighting coefficient data group of the selected determination model data to the feature amount generation unit, and passes the determination boundary function data to the sleep stage determination unit. After that, when the respiration waveform data of the measurement subject is acquired, the feature quantity generation unit extracts the first feature quantity group and the weighting coefficient of the first feature quantity group and the selected determination model data. A second set of feature values is generated using the data group, and the second set of feature values is passed to the sleep stage determination unit. The sleep stage determination unit refers to the position of the discriminant boundary defined by the discriminant boundary function data in the discriminant space and the position of the point of the second feature amount pair, and The region to which the point belongs is determined, and the sleep stage corresponding to the region is determined as the second feature amount set, that is, the current sleep stage of the measurement subject.

  For the above-mentioned measured person, the sleep stage is determined by the apparatus of the present invention for a certain period, and after the first feature amount group for the measured person is collected over the period. When the sleep stage is determined, the determination model selection unit first stores a part of the first feature amount group obtained so far of the measurement subject and the determination model data storage unit. The first model feature close to a part of the first feature amount group obtained so far by comparing a part of the first model feature amount group of the plurality of determination model data obtained The determination model data having a part of the quantity group is selected, and the determination of the sleep stage by the above-described series of processes is executed using the selected determination model data.

  One of the features of the device of the present invention described above is that a plurality of state quantities are first extracted from the respiratory waveform data as a first feature quantity, and further a second feature quantity obtained by using them is extracted. The pair is generated so that the distribution of points corresponding to the second feature quantity set is substantially separated by the sleep stage in the feature quantity space in which the second feature quantity set is a variable. It is a point that has been made. That is, in the apparatus of the present invention, a large number of first feature quantities are extracted as quantities reflecting the difference between a plurality of types of sleep stages, and are used for the subsequent sleep stage determination. Therefore, since more information related to the sleep stage is collected from the respiratory waveform data and reflected in the determination of the sleep stage, improvement in the accuracy of the sleep stage determination is expected. In addition, the sleep stage determination process is performed using a second feature value set generated from a plurality of first feature values. As described above, the second feature value set is as described above. In the discriminant space in which the second set of feature values is set as a variable, the distribution of points corresponding to the second set of feature values is substantially separated by the sleep stage, that is, in the discriminant space. The second feature set is generated such that the distribution of the point cloud of the second feature set at each sleep stage is substantially separated from each other, and in the discrimination space Since the point cloud region of the second feature value set is separated for each sleep stage by the discrimination boundary, it becomes possible to more clearly determine the sleep stage indicated by the second feature value set. .

  In addition, regarding selection of determination model data used for determination processing of a measurement subject's sleep stage, preferably, the determination model selection unit is predetermined after selecting one determination model data for one measurement target person. The selected determination model data is maintained over a period of time, and the second feature is referred to after the lapse of the predetermined period with reference to the set of second feature values in the predetermined period. The determination model data to which the set of the second feature amount obtained by the amount generation unit belongs may be selected again. That is, in this case, the determination model data to be used is not changed every moment, but the determination model data once selected is used for at least a predetermined period. According to such a configuration, it is expected that the sleep stage can be stably determined for one person to be measured.

  Thus, according to the present invention described above, it is expected that the determination of the sleep stage can be achieved more accurately and clearly by collecting more information from the respiratory waveform data and aggregating the information. The In addition, by preparing determination model data obtained in advance under many different conditions for determining the sleep stage, it is possible to determine the sleep stage corresponding to the state of a plurality of various subjects. Is expected to be achieved appropriately. Since the apparatus of the present invention can determine the sleep stage using data such as a piezoelectric sensor attached to the chest or abdomen of the person to be measured, it does not require large-scale equipment, and is for home use or in-vehicle use. It can be used as a device.

  Other objects and advantages of the present invention will become apparent from the following description of preferred embodiments of the present invention.

FIG. 1A is a diagram schematically illustrating a sleep stage estimation apparatus according to the present invention. FIGS. 1B and 1C are diagrams for collecting respiratory waveform data worn by a measurement subject. It is the figure which represented the sensor apparatus and its internal structure typically. FIG. 1D is a block diagram showing the configuration of the processing unit inside the sleep stage estimation apparatus according to the present invention. In the figure, arrows indicate the flow of information. FIG. 2A is a diagram for explaining a partial definition of a plurality of waveform feature quantities (first feature quantities) acquired by the feature quantity processing unit of the sleep stage estimation apparatus according to the present invention. 2 (B) is a diagram schematically depicting a space (feature amount space) spanned by waveform feature amounts, and FIG. 2 (C) explains a distribution region of sleep stages appearing in the feature amount space. FIG. The feature amount space is a multi-dimensional space having the number of waveform feature amounts, but in FIG. 2C, as an example, two of the waveform feature amounts having xi and xj as axes. The distribution region of the sleep stage when viewed in a dimensional space is shown. FIG. 3A shows a configuration of a set of discriminating feature amounts (second feature amounts) from a feature amount space formed by a group of waveform feature amounts (first feature amounts) in the sleep stage estimation apparatus according to the present invention. It is a figure explaining the reduction process to the discrimination space to perform. FIG. 3B is a diagram schematically illustrating the distribution of sleep stages in a space (discriminant space) spanned by a set of discriminant feature values generated in the sleep stage estimating apparatus according to the present invention. FIG. 4A is a diagram illustrating a process of smoothing the respiratory waveform data acquired from the sensor device and setting the vertex / bottom point in the filter processing unit of the sleep stage estimation device according to the present invention. . FIG. 4B is a diagram for explaining a mode of smoothing of respiratory waveform data in the filter processing unit. FIG. 4C is a diagram illustrating processing for specifying a portion not used for sleep stage determination in the respiratory waveform data in the filter processing unit. FIG. 5 is a diagram illustrating a configuration of a plurality of determination model data stored in the determination model data storage unit of the sleep stage estimation apparatus according to the present invention. FIG. 6 is a diagram illustrating a process of selecting determination model data to be used for a measurement subject in the determination model selection unit of the sleep stage estimation apparatus according to the present invention. FIGS. 7A and 7B show sleep determined based on data (second feature value set) obtained from the measurement subject in the sleep stage determination unit of the sleep stage estimation apparatus according to the present invention. The example of the determination result of the step is shown. (A) is an example of temporary determination (determination for each epoch), and (B) is an example of positive determination (determination that corrects temporary determination).

DESCRIPTION OF SYMBOLS 1 ... Computer body 2 ... Computer terminal 3 ... Monitor 4 ... Keyboard, mouse (input device)
5. Respiratory waveform sensor

Outline of Apparatus The sleep stage estimation apparatus according to the present invention may be realized by the operation of a program in the computer apparatus 1 of the type normally used in this field (FIG. 1 (A)). The computer apparatus 1 is equipped with a CPU, a storage device, and an input / output device (I / O) connected to each other by a bidirectional common bus in a normal manner, and the storage device is used in the calculation of the present invention. A memory storing each program for executing the arithmetic processing, a work memory used during the arithmetic operation, and a data memory. In addition, the display and output of instructions, calculation results, and other information to the computer device 1 by the practitioner are performed through the computer terminal device 2 connected to the computer device 1. The computer terminal device 2 is provided with a monitor 3 and an input device 4 such as a keyboard and a mouse in a normal manner. When each of the programs is started, the practitioner follows the procedure of each program. In accordance with the above display, various instructions and inputs are made to the computer device 1 using the input device 4, and the calculation state, calculation result, determination result, etc. from the computer device 1 are visually confirmed on the monitor 3. Is possible. Note that the computer device 1 and the computer terminal device 2 may be equipped with other peripheral devices (not shown) (printer for outputting results, storage device for inputting / outputting calculation conditions and calculation result information, etc.). Should be understood.

  Furthermore, the sleep stage estimation apparatus according to the present invention, as schematically illustrated in FIG. 1 (B), is a respiratory waveform that is a detected value from the sensor device 5 attached to the chest or abdomen of the measurement subject P. An input / output device (communication I / O) for data communication for acquiring data is provided. The sensor device 5 typically has a sensor unit composed of a piezoelectric element, the operation control of the sensor unit, and the sensor output to the computer device 1 by wireless T, as schematically illustrated in FIG. It has a control device for transmitting and communication I / O. The sensor device 5 is attached to the chest or abdomen of the measurement subject P by, for example, a support or fixing device such as a wearing belt, and the output of the sensor unit is typically the breathing of the measurement subject P. The movement of the chest or abdomen accompanying the exercise is received as a pressure change. The pressure change is converted into a potential change and transmitted to the computer apparatus 1.

  Further, the sensor device 5 may be provided with a stimulus generating means for giving a stimulus to the measurement subject. Such stimulation may be, for example, voice or vibration given to the measurement subject P in a timely manner according to the determination when the sleep stage is determined. More specifically, the vibration stimulus may be used for awakening by applying vibration after a predetermined time after the person to be measured becomes a sleep state by the sleep stage estimation device. Alternatively, as another method for giving a stimulus to the person to be measured P, a system is constructed in which the stimulus generating means 6 such as a music player, a smartphone, an air conditioner or the like can receive a control command from the computer device 1, and the sleep stage estimating device According to the instruction from the computer device 1 in accordance with the sleep stage determined in Step 1, various operations such as generation of voice and vibration, air conditioning control, etc. (not necessarily stimulating for awakening the measurement subject P during sleep) May not be the purpose). The sensor device 5 may incorporate a CPU and a memory. In this case, the sensor device 5 may have a data storage function.

  In the configuration for executing each process of the sleep stage estimation device according to the present invention, specifically, as shown in FIG. 1 (D), an input unit that captures respiratory waveform data from the sensor device 5; From a filter processing unit that executes processing such as smoothing of respiratory waveform data, a feature amount generation unit that generates feature amounts in respiratory waveform data including a feature amount processing unit, and a measurement subject model under various conditions A determination model database storing a plurality of determination model data obtained in advance, a determination model selection unit for selecting a determination model to be used when determining the sleep stage of the measurement subject, and a feature generated by a feature amount generation unit A sleep stage determination unit that preferentially determines a sleep stage with reference to the quantity and data of the determination model, and a result in the sleep stage determination unit is recorded or output to be used for controlling the stimulus generation means. Output section and It is included. In addition, the operator may input the state of the person to be measured, such as age and sex, and select a model. Generally speaking, regarding the data flow in the above-described configuration, first, the determination model selection unit includes the weighting coefficient data group aji and the determination boundary function data Fk included in the determination model data selected in the manner described below. Each is passed to the feature amount processing unit of the feature amount generation unit and the provisional determination unit of the sleep stage determination unit. The feature amount processing unit performs a smoothing process or the like of the respiration waveform data, and then uses the weighting coefficient data group aji to calculate the waveform feature amount (first feature amount) xj extracted from the processed respiration waveform data. The discriminating feature amount Ai is generated by linear combination, and the discriminating feature amount Ai is passed to the temporary determination unit of the sleep stage determination unit. Then, in the temporary determination unit of the sleep stage determination unit, using the determination boundary function data Fk and the determination feature quantity Ai, a determination is made of the determination feature quantity Ai, that is, the sleep stage in the current respiratory waveform data. In the correct determination unit, the temporarily determined sleep stage is corrected.

  Note that each of the above-described units illustrated in FIG. 1D includes programs necessary for various processes or calculations in the computer apparatus 1 when the sleep stage estimation apparatus performs the sleep stage estimation (determination). It should be understood that it is implemented and realized by arithmetic processing according to the program.

  Hereinafter, the principle of the sleep stage determination process of the sleep stage estimation apparatus and the configuration and operation of each unit will be described in detail.

The principle of the sleep stage determination process in the sleep stage estimation device As described in the “Background” column and the “Summary of the Invention” column, it is determined by the brain wave and eye movement patterns of a sleeping person. The sleep stage is reflected in the respiratory waveform data indicating the movement of the person's respiratory movement, specifically the movement of the person's chest or abdomen. Therefore, the sleep stage is determined by monitoring the respiratory waveform data. Is possible. For example, when the respiratory waveform data during sleep of a person as illustrated in FIG. 2 (A) or FIGS. 4 (A) and 4 (C) is obtained, various state quantities (waveforms) obtained from the respiratory waveform data. Feature quantity), for example, the average value of the distance between the peaks of the waveform, the standard deviation, the average value of the distance between the bottom, the standard deviation, the average value between the peak and the bottom, the standard deviation, etc., depending on the sleep stage Change. However, there are various appearances of changes in the plurality of parameters depending on the sleep stage, and the correspondence between these changes and the sleep stage is complicated. Therefore, in order to determine the sleep stage based on the respiratory waveform data, first, in order to clarify the correspondence between the waveform feature quantity and the sleep stage, the respiratory waveform data of the person during sleep in various sleep stages. Is measured in a state where the sleep stage is specified, a plurality of waveform feature amounts are acquired from the respiratory waveform data, and the waveform feature amount schematically illustrated in FIG. Assuming the feature amount space, the region where the waveform feature amount in each sleep stage is distributed in the feature amount space is grasped.

  However, in reality, it is not clear which of the various state quantities that can be extracted from the respiratory waveform data has a strong correlation and which has a weak correlation with the sleep stage. Therefore, when trying to accurately determine the sleep stage, it may be possible to extract many state quantities that are expected to be correlated with the sleep stage from the respiratory waveform data as waveform feature quantities. As illustrated in FIG. 2B, the feature amount space constituting the waveform feature amount is a multi-dimensional space (four or more dimensions), and points of the waveform feature amount corresponding to each sleep stage in the space. It is not easy to grasp the group distribution area. In addition, as illustrated in FIG. 2C, generally, the distribution areas of the waveform feature amount point cloud corresponding to each sleep stage in the feature amount space formed by the waveform feature amount overlap each other. Therefore, when a certain waveform feature amount is obtained, it is often difficult to clearly determine the sleep stage corresponding thereto.

  Therefore, in the present invention, the correspondence between the state quantity obtained from the respiratory waveform data and the sleep stage can be grasped clearly, that is, without reducing the number of waveform feature quantities to be used. In order to do this, a new approach is used.

Specifically, first, as described above, the respiratory waveform data of a sleeping person at various sleep stages is measured in a state where the sleep stage is specified, and the waveform feature value is changed from the respiratory waveform data to the sleep stage. Many state quantities that are expected to be correlated are extracted. Thereafter, the obtained waveform feature groups are subjected to weighted linear combination of waveform features according to the procedure of the canonical discriminant method (also referred to as linear discriminant method) to reduce the waveform feature groups. Dimensionalization is performed (see FIG. 3A, Non-Patent Document 1). That is, for the waveform feature quantity group (x1, x2, x3,..., Xj,...), Reduced dimension feature quantities A1, A2, and A3 (discriminating feature quantities: second feature quantities) are generated as follows. Is done.
A1 = a11 ・ x1 + a12 ・ x2 + a13 ・ x3… + a1j ・ xj…
A2 = a21 · x1 + a22 · x2 + a23 · x3 ... + a2j · xj ... (1)
A3 = a31 ・ x1 + a32 ・ x2 + a33 ・ x3… + a3j ・ xj…
Here, a11, a12,... Aji are weighting coefficients. In short, A1, A2, and A3 classify the obtained waveform feature group (xj) for each sleep stage and classify the maximum of the interclass variance matrix Σ1 of the waveform feature group (xj). Is a quantity expressed by an eigenvector corresponding to each eigenvalue when the larger one of the eigenvalues of the matrix Σ1 ⁻¹ Σ2 is selected. aji is a coefficient for converting a point of the waveform feature amount space into a point of the space spanned by such eigenvectors.

  When a set of discriminant features is generated by reducing the waveform features by linear combination of the groups of the waveform features as described above, a newly generated discriminant feature is generated as shown in FIG. A reduced-order space (discriminant space) having a set of Then, in this discriminant space, when plotting the points of the discriminant feature amount set, as schematically illustrated in FIG. 3B, the distribution region of the discriminant feature amount point group corresponding to each sleep stage. Are substantially separated from each other. (This effect is obtained by maximizing the inter-class variance matrix Σ1 and minimizing the intra-class variance matrix Σ2 of the group (xj) of waveform feature amounts. In short, the discriminating feature amount is obtained by linear combination of the waveform feature amounts. , The weighting coefficient aji increases the variance between classes (sleep stages) as much as possible by maximizing the interclass variance matrix Σ1, while minimizing the intraclass variance matrix Σ2 to generate the same class (sleep stage). (By calculating the variation of point cloud as much as possible.)

  Then, as described above, in the discriminant space in the state where the dimension reduction from the waveform feature group to the discriminant feature group is performed, the distribution region of the discriminant feature point group corresponding to each sleep stage is expressed as follows. A dividing boundary (discrimination boundary) is determined. Such discriminant boundaries are also determined according to the procedure of the canonical discriminant method (in short, the surface where the variation in the distance to the point cloud distribution region of the discriminant feature value corresponding to each sleep stage is the maximum is the discriminant boundary. Become.). Specifically, the discriminant boundary is given by a functional expression Fk (discriminant boundary function data) that gives it in the discriminant space. Here, it should be noted that the set of discriminant feature values is three variables, and therefore the discriminant space is three-dimensional. Since the discriminant space is three-dimensional, the distribution region of the point cloud of the discriminant feature amount corresponding to each sleep stage can be easily grasped visually. Therefore, the state quantity obtained from the respiratory waveform data and the sleep stage It will be possible to grasp the relationship between and clearly. In addition, since a group of waveform features obtained from respiratory waveform data is considered in the set of discriminating features, it can be obtained from respiratory waveform data accurately without reducing the number of waveform features used. It is expected that the correspondence between the state quantity and the sleep stage can be grasped.

  Thus, as described above, through the waveform feature amount from the respiratory waveform data of a person under certain conditions during sleep measured in the state where the sleep stages in various sleep stages are specified, the discrimination space as described above. When a discriminant feature quantity in which distribution regions corresponding to each sleep stage are substantially separated from each other is generated, a weighting coefficient aji for generating the discriminant feature quantity is obtained, and further, a discriminant boundary in the discriminant space is obtained. A functional expression Fk that gives Then, when the respiratory waveform data of the person who measured the respiratory waveform data or the person with the same condition as that person is obtained (with the sleep stage unknown), the waveform feature amount from the respiratory waveform data at that time and A discriminant feature value Ai is generated using the weighting coefficient aji, and a new discriminant discriminant in the discriminant space is obtained by referring to the positional relationship between the discriminant feature amount Ai and the discriminant boundary given by the function formula Fk By determining which sleep stage region the feature quantity Ai belongs to, it is possible to determine the sleep stage of the discrimination waveform quantity Ai, that is, the respiratory waveform data whose sleep stage is unknown.

  By the way, as already described, the correspondence between the respiratory waveform data and the sleep stage varies depending on the state of the person whose respiration is measured, for example, age, sex, physical characteristics (height, weight, body shape, etc.), physical condition, etc. there is a possibility. Therefore, in the apparatus of the present invention, when the sleep stage determination is performed using the respiration waveform data of an arbitrary measured person, a person (measured person model) having the same or similar condition as the measured person The weighting coefficient aji calculated from the respiratory waveform data during sleep in the various sleep stages obtained in step 1) and the function formula Fk that gives the discrimination boundary must be prepared in advance. Therefore, in the apparatus of the present invention, as already mentioned, the weighting coefficient aji (group of weighting coefficient data) obtained from the respiratory waveform data for various measurement subject models, and the functional expression Fk (discriminant boundary function) Data (discriminant model data) including data is prepared in advance. Then, when performing the determination of the sleep stage in a certain person to be measured, the determination model data of the person model to be measured that is expected to be suitable for the person to be measured is selected in the manner described later, and the determination is made. Using the weighting coefficient data group aji of the model data and the discriminant boundary function data Fk, the discriminant feature value is generated and the discriminant boundary is set.

The following configuration and operation of in each section to sleep stage estimation device will be described operation with the detailed structure of each part of the sleep stage estimation device.
(1) Filter Processing Unit In the filter processing unit in the feature value generation unit, first, smoothing of respiratory waveform data that is potential data transmitted from the sensor device 5 is executed. As schematically depicted in the left diagram of FIG. 4A, in the respiratory waveform data obtained from the sensor device 5, generally, a vibration having a higher frequency than the amplitude representing the respiratory state is superimposed. Yes. Therefore, first, in the respiratory waveform data obtained from the sensor device 5, for example, smoothing processing by a moving average method is executed. In such smoothing processing, typically, a plurality of respiration waveform data X obtained from the sensor device 5 and its time change data ΔX / Δt (X change amount ΔX per sampling time Δt) are plural. A moving average process may be performed. In the moving average process, the number of data points (window frame) averaged at a time and the number of moving average processes are within a range in which the amplitude v of the respiration waveform data X and the time change data ΔX / Δt after smoothing is assumed. May be adjusted in advance. In the case of this embodiment, typically, two moving average processes may be performed. Specifically, as illustrated in FIG. 4B, after the first moving average process is performed on the respiration waveform data X and its time change data ΔX / Δt in the illustrated window frame. The second moving average process is executed using the smoothed points after the first moving average process.

  In the filter processing unit, the respiration waveform data X smoothed as described above and the time change data ΔX / Δt of the respiration waveform data X before the smoothing process and the time change data ΔX / Restoration of vertices and base points at Δt is performed. As can be seen by comparing the left diagram of FIG. 4A and the middle diagram, when smoothing processing is performed on the respiratory waveform data X and its time change data ΔX / Δt, the amplitude a decreases. It becomes. Therefore, in order to more accurately obtain the true amplitude of the respiration waveform data X and its time change data ΔX / Δt, the top and bottom of the respiration waveform data X and its time change data ΔX / Δt before the smoothing process are obtained. An estimate of the point height is made. Specifically, for any data, the amplitude reduction ratio by the above moving average process can be known in advance (for example, it is known that the illustrated example reduces by 20%). For example, in the respiration waveform data X after smoothing and its time change data ΔX / Δt, the vertex and the bottom point are detected, and the height of the detected vertex and the bottom point is reduced by moving average processing. A value obtained by multiplying the amplitude ratio (0.2 v) may be added to calculate the heights of the apex and bottom points of the respiration waveform data X before the smoothing process and its time change data ΔX / Δt. In addition, when the vertex and bottom point are restored, the respiration waveform data X after smoothing and the entire time change data ΔX / Δt are corrected so as to match the height of the restored vertex and bottom point. May be.

  Further, in the filter processing unit, the difference in height between the top and bottom of the smoothed respiration waveform data X and its time change data ΔX / Δt or the height of the restored top and bottom is calculated. With reference to the difference, in the respiration waveform data X and its time change data ΔX / Δt, the section in which the height difference between the apex and the bottom exceeds the normally assumed range is specified. A section where the difference in height between the apex and the bottom exceeds the normal range is likely to be an area where there is body movement such as turning over, and in such a section, the sleep stage is correct. Judgment becomes difficult. Therefore, such a section where the difference in height between the apex and the base is significant may be specified as an unnecessary section for determining the sleep stage. Specifically, referring to FIG. 4C, first, in the respiratory waveform data X or its time change data ΔX / Δt, it is assumed that the height difference between the apex and the bottom follows a Gaussian distribution. The Then, every determination unit time (referred to as “epoch”, typically set to 20 to 30 seconds), the peak and bottom of the normal respiration waveform data X or its time change data ΔX / Δt. An epoch in which the number of amplitudes v (i) exceeding three times the standard deviation σ of the height difference from the point exceeds a predetermined value (for example, 4 times) may be specified as an unnecessary section. For example, in the illustrated example, the epoch e3 is specified as an unnecessary section because the amplitude v (i) exceeding 3σ exceeds a predetermined value (four times). And the determination of the sleep stage mentioned later is not performed in the area specified as an unnecessary area. In specifying the unnecessary section, when referring to the difference in height between the restored vertex and the bottom point, it is advantageous in that the section exceeding the normal amplitude can be specified with high accuracy. In addition, by specifying an unnecessary section and not performing the determination of the sleep stage in such a section, it is possible to accurately grasp the tendency of the sleep stage transition.

(2) Feature Quantity Processing Unit In the feature quantity processing unit in the feature quantity generating unit, the “waveform feature” already described in the respiratory waveform data X and its time change data ΔX / Δt processed by the filter processing unit. The “quantity” group is extracted, and the “discriminating feature quantity” is generated using the waveform feature quantity group.

  In the extraction of the waveform feature amount group, for each epoch, various state amounts are extracted as a waveform feature amount group with reference to the respiration waveform data X and its time change data ΔX / Δt. As such various state quantities, quantities that are known to some extent to be correlated with the sleep stage are selected. Specifically, first, in the respiratory waveform data X and its time change data ΔX / Δt, as illustrated in FIG. 2A, the peak-to-peak difference RTP and the bottom-bottom difference RTB. , Bottom-peak difference IV, peak-bottom difference EV, bottom-peak time difference IT, peak-bottom time difference ET, peak-peak time difference, bottom-bottom time difference, bottom-peak Various state quantities (basic feature quantities) such as a time change amount IV / IT in between are detected. Then, a moving average process is performed on each detected value, and an statistic such as an average value or standard deviation of the values is calculated for each epoch. In addition, the amount of time variation (local difference feature amount) of various detected state quantities is calculated, and their moving average processing is executed. For each epoch, statistics such as the average value and standard deviation of those values are obtained. It may be calculated. Further, a difference between epochs (2 to 3 epochs) of statistics such as an average value and a standard deviation of the basic feature amount and the local difference feature amount may be calculated (difference feature amount between epochs). Thus, all of the above average values of basic feature values, statistics such as standard deviation, average values of local difference feature values, statistics such as standard deviation, and inter-epoch difference feature values are extracted as waveform feature values. A group of quantities xj is constructed.

  Thus, as described above, when a group of waveform features is configured for each epoch, the weighting coefficient data group of the selected determination model data passed from the determination model selection unit in a manner described later for each epoch. Aji is used to generate a set of discriminant feature quantities Ai by linearly combining groups of waveform feature quantities in the manner of Equation (1) described in the above “Principle of Sleep Stage Determination Processing”. The set of discriminating feature values Ai that has been performed is passed to the temporary determination unit of the sleep stage determination unit.

(3) Determination Model Database In the determination model database, as already described, a plurality of determination model data obtained in advance from the measurement subject model under various conditions is stored (see FIG. 5). Each judgment model data is a respiration waveform measured in a state where a sleep stage is specified in a measurement subject under a certain condition with respect to one measurement target model condition, that is, age, sex, physical characteristics, etc. Extraction of a group (x1, x2,... Xj...) Of the waveform feature amounts (first model feature amounts) from the data, and a set (A1, A2, A3) of discriminating feature amounts (second model feature amounts) Weighting coefficient data group aji, discriminating boundary function data Fk, and part of the waveform feature amount (described later) Used to select a determination model.). The measured person model of the determination model data stored in the determination model database may be a group of measured person models (model family) under any conditions for which the apparatus of the present invention is assumed to be used. Classification such as male, young female, senior male, senior female, or classification such as young male / female, senior male / female, 10-20 male, 10-20 female, 30 male, 30 female, ... 60 male, 60 female A typical example of a person-to-be-measured model may be employed for a classification of female. One determination model data may be data obtained from a plurality of person models of the same category.

(4) Judgment Model Selection Unit As already mentioned, the judgment model selection unit selects and selects a measurement subject model that meets or is close to the condition of the measurement subject from the judgment model database. The weighting coefficient data group aji in the determination model data of the measured person model is passed to the feature amount processing unit, and the discrimination boundary function data Fk is passed to the temporary determination unit. Therefore, it should be understood that the selection of the decision model data is executed at the beginning of a series of processes executed in the apparatus.

  Specifically, the judgment model data is selected by the method in which the person to be measured or the operator of the apparatus selects the person model to be measured that is expected to be suitable for the person to be measured, and the judgment model selection unit. The measurement person may be configured to be able to execute two methods of selecting a measurement subject model based on information obtained from respiratory waveform data.

  In the case of a method in which a person to be measured or an operator of the apparatus selects a person model to be measured that is expected to be suitable for the person to be measured, the input device 4 of the computer terminal device 2 is used to start measurement. A measurer model may be selected. Note that the sensor device 5 may be provided with a function for inputting the measurement subject model. Thus, when an input for selecting the measurement subject model is made, as described above, the determination model selection unit transmits the data related to the measurement target model selected from the determination model database to the corresponding processing unit. To do.

  On the other hand, in the method of selecting the measurement subject model in the judgment model selection unit, specifically, a group of waveform feature values obtained from the respiratory waveform data measured by the measurement subject. Is compared with a part of the waveform feature quantity of each judgment model data stored in the judgment model database, and among the part of the waveform feature quantity of each judgment model data, a group of waveform feature quantities of the measurement subject Is selected as determination model data used for determination of the sleep stage. Comparison between a part of the waveform feature of the measurement subject and a part of the waveform feature of each determination model data is typically performed according to a cluster analysis technique as illustrated in FIG. Can be executed in a manner in which determination model data having a part of the waveform feature amount close to a part of the waveform feature amount of the measurement subject is determined (generally, a dendrogram (Ward method) of In other words, in the feature amount space having the waveform feature amount as a variable, determination model data having a waveform feature amount point group that is close to the waveform feature amount point group of the measurement subject is selected. In this regard, when determining the determination model data with sufficient accuracy using the above dendrogram, the data amount of the waveform feature amount is a certain period (for example, about several hours to one day). It has been found that an amount of data obtained from respiratory waveform data measured over time is required. Therefore, in the method for selecting the measurement subject model in the determination model selection unit, the acquisition of the respiratory waveform data is already executed for the certain measurement subject over the certain period by the apparatus of the present invention. It may be used when there is. In addition, according to the method for selecting the measurement subject model in the determination model selection unit, the measurement subject model is selected by using the measurement subject's actual respiration waveform data. The measured person model is expected to be an optimum determination model.

  Thus, when the measurement subject model is selected, as described above, the determination model selection unit transmits data related to the measurement target model selected from the determination model database to the corresponding processing unit.

  By the way, it is considered that the correspondence between the waveform feature amount in the respiratory waveform data and the sleep stage in a certain measurement subject hardly changes in a short period. Therefore, the determination model once selected may be maintained for a predetermined period that may be arbitrarily set, for example, for several days to several weeks. Then, after the predetermined period has elapsed, the method for selecting the measurement subject model in the determination model selection unit may be executed again, and the determination model may be reselected. According to such a configuration, it is possible to stably grasp the trend of the sleep stage of the person being measured, and appropriately change the person model to be measured when the state of the person to be measured has changed. It is expected that sleep stage determination is often achieved.

(5) Temporary Judgment Unit and Positive Judgment Unit In the temporary judgment unit, a point corresponding to a set of discriminating feature values Ai for each epoch generated by the feature value processing unit in the discriminant space is a judgment model. It is determined which of the areas defined by the discriminant boundary function data Fk passed from the selection unit belongs, and for each epoch, the sleep stage corresponding to the area to which the set of points of the discriminant feature Ai belongs is the current sleep stage. Is determined. In that case, typically, as illustrated in FIG. 7A, the determination result of the sleep stage may change unlikely in the actual sleep stage. Therefore, in the apparatus of the present invention, the positive determination unit may further correct the determination result of the sleep stage for each epoch in the temporary determination unit. Specifically, for example, in the illustrated example, for one epoch in the tentative judgment unit, the number of epochs of “deep sleep” epochs appearing and appearing in the range from the epoch to a predetermined epoch in the past Referring to the frequency, the probability of occurrence of “deep sleep” determination and “slow sleep” determination is calculated, and in the determination of one epoch, the sleep stage with the higher probability is adopted as the determination result . Similarly, for one epoch in the tentative judgment unit, refer to the number of consecutive epochs of the “wakefulness” judgment and the appearance frequency in the range from the epoch to a predetermined epoch in the past. "Probability" and "hypnosis" determination are calculated, and in the determination of one epoch, the sleep stage with the higher probability is adopted as the determination result. Then, as illustrated in FIG. 7B, a determination result assumed in the transition of the actual sleep stage is obtained. In addition, the correction of the sleep stage determination in the positive determination unit is not based on the above method, considering the transition of the sleep stage determination in the temporary determination unit and the expected change in the sleep stage, You may correct | amend suitably.

Operation and operation of sleep stage estimation apparatus In the operation of the sleep estimation apparatus of the present invention, first, in the case of a person to be measured who uses the apparatus for the first time, a determination is made when an input of a person to be measured is made. The model selection unit calls the input determination model data of the measurement subject model from the determination model database, passes the weighted coefficient data group aj of the called determination model data to the feature amount processing unit, and determines the boundary function data Fk Is passed to the provisional determination unit. Then, when an instruction to start the determination of the sleep stage is given to the computer device 1, the measurement of the respiratory waveform data from the sensor device 5 attached to the measurement subject is started, and in real time, in the filter processing unit. Respiration waveform data processing, feature feature processing unit extraction of waveform feature quantity groups and generation of discriminating feature quantities are performed, and discriminating feature quantities are passed to a provisional judgment section to perform provisional judgment and correct judgment of sleep stages. The sleep stage correction in the determination unit is executed, and the sleep stage determination result is output.

  In the above state, after the sleep stage is determined for a predetermined period (typically several hours to one day), the same person to be measured uses the apparatus again. When the instruction to start the determination of the sleep stage is given to the computer device 1, the determination model selection unit uses the part of the group of waveform feature values of the measured person up to that point to the measured person The determination model data of the suitable measurement subject model is called from the determination model database, and the determination process of the sleep stage is executed in the same manner as described above using the called determination model data.

  Further, in the above state, after the sleep stage is determined over a predetermined period (typically several days to several weeks), the same person to be measured uses this apparatus. The determination model selection unit executes reselection of determination model data of a measurement subject model that matches the measurement target (the same determination model data may be selected), and the determination model selected here. Using the data, the sleep stage determination process is executed in the same manner as described above.

  Thus, as described above, the determination result of the sleep stage may be used for stimulation of the measurement target, adjustment of the sleep environment of the measurement target, or the like.

  Although the above description has been made in relation to the embodiment of the present invention, many modifications and changes can be easily made by those skilled in the art, and the present invention is limited to the embodiment exemplified above. It will be apparent that the invention is not limited and applies to various devices without departing from the inventive concept.

  For example, in the illustrated embodiment, the determination stage of the sleep stage is three stages of wakefulness, light sleep, and deep sleep, but more detailed determination stages (stages I to IV in REM sleep and non-REM sleep, wakefulness) ). In generating a set of discriminating feature values from a group of waveform feature values, the set of discriminating feature values is three-dimensional, but may be two-dimensional or one-dimensional (discrimination from a group of waveform feature values). The dimension of the discriminating feature quantity can be selected according to the number of eigenvalues selected when generating the feature quantity set.)

Claims (2)

A sleep stage estimation device comprising:
A respiratory waveform data acquisition unit for acquiring respiratory waveform data measured by the measurement subject;
Extracting a first feature quantity group from the acquired respiratory waveform data, and generating a second feature quantity set generated by combining the first feature quantity group; and
A determination model data storage unit that stores a plurality of determination model data obtained based on respiratory waveform data measured in advance in a state where the sleep stage is specified under different measurement subject model conditions;
A determination model selection unit that selects determination model data to be used for the measurement subject among the plurality of determination model data stored in the determination model data storage unit;
A sleep stage determination unit that determines a sleep stage corresponding to the set of the second feature amount,
Each of the plurality of determination model data is
A first model corresponding to the first feature amount group in the respiratory waveform data measured in advance in various sleep stages under one of the different measurement subject model conditions. At least part of the group of features,
Weighting coefficient for each of the first model feature values when generating a set of second model feature values obtained by linearly combining the first model feature value group of the one measured person model condition As a group of data, the second set of model feature values corresponds to the second set of model feature values in a discriminant space spanned by the second set of model feature values as a variable. A group of weighting factor data that is a set of quantities generated such that a distribution of point groups is substantially separated from each other by a sleep stage associated with each of the second set of model feature quantities;
Discriminant boundary function data that provides a discriminant boundary that divides the discriminant space into regions corresponding to each of the sleep stages associated with each of the second model feature amount sets,
The determination model selection unit is configured to select at least a part of the first model feature amount group and the first model feature amount group of each of the plurality of determination model data based on selection by an operator of the device. Selecting determination model data used for the person to be measured based on comparison with a part of the group of the first feature amount corresponding to at least a part of
The feature quantity generation unit generates the second set of feature quantities by linearly combining the first feature quantity group with the weight coefficient coefficient group of the selected determination model data as a weighting coefficient;
The sleep stage determination unit includes the first of the regions corresponding to a plurality of sleep stages delimited by the determination boundary defined by the determination boundary function data in the determination space of the selected determination model data. An apparatus for determining a sleep stage corresponding to an area to which a point corresponding to a second feature quantity set belongs as a sleep stage corresponding to the second feature quantity set.   The apparatus according to claim 1, wherein the determination model selection unit selects the selected determination model data over a predetermined period after selecting one determination model data used for one person to be measured. And after the elapse of the predetermined period, select judgment model data to be used for the one person to be measured with reference to the second set of feature values in the predetermined period. Repairing device.

Images