Method for generating body condition variation prediction model, body condition variation prediction system, body condition variation prediction method, body condition variation prediction model, and method for generating regularity identification model

Abstract

To provide a method for generating a body condition variation prediction model capable of generating a body condition variation prediction model having improved prediction accuracy.SOLUTION: A method for generating a body condition variation prediction model for predicting variation in a body condition of a subject on a bed comprises the steps of: acquiring biological information on the subject for teacher data by sensors; determining a clinical state of the subject for the teacher data on the basis of the biological information on the subject for the teacher data; acquiring a state trend which is a trend per unit time of the clinical state; identifying regularity included in temporal fluctuation in the state trend; and generating a body condition variation prediction model trained with a correlation between the regularity and body condition variation information on the subject for the teacher data by supervised machine learning using the teacher data in which the regularity and the body condition variation information indicating variation in the body condition of the subject for the teacher data are associated with each other.SELECTED DRAWING: Figure 11

Description

[Technical field]

[0001] The present invention relates to a method for creating a physical condition variation prediction model, a physical condition variation prediction system, a physical condition variation prediction method, a physical condition variation prediction model, and a method for creating a regularity identification model. [Background technology]

[0002] 2. Description of the Related Art In medical and nursing care settings, it has been proposed to obtain data on a subject lying in bed using a sensor and predict changes in the subject's physical condition based on the obtained data.

[0003] Patent document 1 discloses a health condition prediction method that acquires body movement data of a target person, constantly determines the target person's sleeping state based on the acquired body movement data, and predicts changes in the target person's health condition based on the determined sleeping state. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-76689 A Summary of the Invention [Problem to be solved by the invention]

[0005] Regarding techniques for predicting changes in a subject's physical condition, including the physical condition prediction method disclosed in Patent Document 1, further improvements in prediction accuracy are desired.

[0006] An object of the present invention is to provide a method for creating a physical condition variation prediction model, which is capable of creating a physical condition variation prediction model with improved prediction accuracy.

[0007] An object of the present invention is to provide a health condition variation prediction system, a health condition variation prediction method, and a health condition variation prediction model with improved prediction accuracy.

[0008] An object of the present invention is to provide a method for creating a regularity specifying model for use in a health condition variation prediction system and a health condition variation prediction method with improved prediction accuracy. [Means for solving the problem]

[0009] According to a first aspect of the present invention, A method for creating a physical condition variation prediction model for predicting a variation in the physical condition of a subject in bed, comprising: acquiring biological information of a subject for use as teacher data by a sensor; determining a lying state of the subject for the teacher data based on biological information of the subject for the teacher data; Obtaining a state trend which is a trend of the lying state per unit time; Identifying a pattern contained in the time variation of the condition trend; creating a health condition variation prediction model that learns the correlation between the regularity and the health condition variation information of the subject for the teacher data by supervised machine learning using teacher data that corresponds the regularity to health condition variation information indicating variations in the health condition of the subject for the teacher data.

[0010] According to a second aspect of the present invention, A physical condition variation prediction system for predicting a variation in a physical condition of a subject in bed, a bed state determination unit that determines a bed state of the subject based on biological information of the subject acquired by a sensor; a state tendency acquisition unit that acquires a state tendency that is a tendency of the lying state per unit time; a pattern identifying unit that identifies a pattern contained in the time variation of the state tendency; A physical condition variation prediction unit inputs the regularity into a physical condition variation prediction model to predict a variation in the physical condition of the subject, The health condition variation prediction model is a trained model that has learned the correlation between the regularity identified for the subjects for the training data and the health condition variation information of the subjects for the training data, through supervised machine learning using training data that matches the regularity identified for the subjects for the training data with health condition variation information that indicates variations in the health of the subjects for the training data.

[0011] According to a third aspect of the present invention, A method for predicting a change in a physical condition of a subject in bed, comprising: a lying state determining step of determining a lying state of the subject based on biological information of the subject acquired by a sensor; A state tendency acquisition step of acquiring a state tendency which is a tendency of the lying state per unit time; a pattern identifying step of identifying a pattern contained in the time variation of the state trend; and a physical condition variation prediction step of inputting the regularity into a physical condition variation prediction model to predict a variation in the physical condition of the subject, The health condition variation prediction model is a trained model that has been trained to learn the correlation between the regularity identified for the subjects for the teacher data and the health condition variation information of the subjects for the teacher data, by supervised machine learning using teacher data that matches the regularity identified for the subjects for the teacher data with health condition variation information that indicates variations in the health of the subjects for the teacher data.

[0012] According to a fourth aspect of the present invention, A physical condition variation prediction model for predicting a variation in the physical condition of a subject in bed, acquiring biological information of a subject for use as teacher data by a sensor; determining a lying state of the subject for the teacher data based on biological information of the subject for the teacher data; Obtaining a state trend which is a trend of the lying state per unit time; Identifying a pattern contained in the time variation of the condition trend; Corresponding the regularity to physical condition variation information indicating a variation in the physical condition of the subject for use as the teacher data; A correlation between the regularity and the subject's physical condition variation information for the teacher data has been learned by supervised machine learning using the teacher data created by the above method, A physical condition variation prediction model is provided that predicts variations in the physical condition of the subject based on the regularity.

[0013] According to a fifth aspect of the present invention, acquiring biological information of a subject on a bed using a sensor; determining a lying state of the subject based on biological information of the subject; Obtaining a state trend which is a trend of the lying state per unit time; and creating a regularity identification model that identifies a predetermined regularity contained in the temporal fluctuation of the state trend by machine learning using the state trend as training data. Effect of the Invention

[0014] According to the method for creating a physical condition variation prediction model of the present invention, it is possible to create a physical condition variation prediction model with improved prediction accuracy.

[0015] According to the physical condition variation prediction system, physical condition variation prediction method, and physical condition variation prediction model of the present invention, variations in a subject's physical condition can be predicted with improved prediction accuracy.

[0016] According to the method for creating a regularity specific model of the present invention, it is possible to create a regularity specific model for use in a health condition variation prediction system and health condition variation prediction method with improved prediction accuracy. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a block diagram showing the configuration of a physical condition variation prediction system. [Diagram 2] FIG. 2 is an explanatory diagram showing the arrangement of the load sensors with respect to the bed. [Diagram 3]FIG. 3 is a flowchart showing the steps of a method for predicting a change in physical condition executed by the system for predicting a change in physical condition. [Figure 4] FIG. 4 is a diagram conceptually showing the flow of processing in the method for predicting a change in physical condition executed by the prediction system for predicting a change in physical condition. [Diagram 5] FIG. 5 is a graph showing an example of a frequency spectrum obtained based on a load signal. [Figure 6] FIG. 6 is a graph showing the subject's lying state over time. [Figure 7] FIG. 7 is a graph showing the subject's condition trends over time. [Figure 8] FIG. 8 is a flowchart showing steps of a method for creating a lying state determination model. [Figure 9] FIG. 9 is a graph showing an example of the temporal fluctuation of four load signals detected by four load sensors. [Figure 10] FIG. 10 is a flow chart showing the steps of a method for creating a regularity identification model. [Figure 11] FIG. 11 is a flowchart showing the steps of a method for creating a physical condition variation prediction model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] <Embodiment> A health condition change prediction system 100 and a health condition change prediction method according to an embodiment of the present invention will be described with reference to FIGS. 1 to 7. FIG.

[0019] [Health condition mutation prediction system] 1, the physical condition variation prediction system 100 mainly includes a load measuring unit 10, an information processing device 20, and a terminal device 30. The load measuring unit 10, the information processing device 20, and the terminal device 30 are connected to each other via a network NW such as the Internet. At least two of the load measuring unit 10, the information processing device 20, and the terminal device 30 may be connected by wiring.

[0020] The load measuring unit 10 mainly includes four load sensors LS1, LS2, LS3, and LS4. Each of the four load sensors LS1 to LS4 measures the load using, for example, a beam-shaped load cell. Such load sensors are described in, for example, Japanese Patent No. 4829020 and Japanese Patent No. 4002905.

[0021] As shown in FIG. 2, the four load sensors LS1 to LS4 are disposed under casters C1, C2, C3, and C4 attached to the lower ends of legs BL1, BL2, BL3, and BL4 at the four corners of a bed BD used by a subject S, respectively.

[0022] The information processing device 20 mainly includes a control unit 21 and a storage unit 22. The control unit 21 includes a preprocessing unit 211, a bed state determination unit 212, a trend acquisition unit 213, a regularity specification unit 214, a physical condition variation prediction unit 215, a teacher data creation unit 216, and a model creation unit 217 (described in detail later). The storage unit 22 includes a bed state determination model M STA , the regularity specific model M REG , Health condition mutation prediction model M PRE is preserved (see below for details).

[0023] The information processing device 20, and the control unit 21 and memory unit 22 therein, may be configured within a single physical server, or may be configured within multiple physical servers (i.e., cloud servers) that are distributed and connected via the cloud.

[0024] The terminal device 30 mainly includes a display unit 31 and an input unit 32 .

[0025] The display unit 31 receives the prediction of the change in the physical condition of the subject S output by the physical condition change prediction unit 215, and displays the prediction to a user (such as a doctor, nurse, caregiver, or caregiver at home) of the physical condition change prediction system 100. The display unit 31 has, for example, an image display unit such as a liquid crystal display and / or an audio display unit such as a speaker.

[0026] The input unit 32 is an interface for performing various inputs to the physical condition variation prediction system 100. The input unit 32 may be, for example, a keyboard and / or a mouse, or may be a touch panel integrated with the liquid crystal display of the display unit 31.

[0027] [Method of predicting changes in physical condition] Each step of the method for predicting a change in physical condition executed by the system 100 for predicting a change in physical condition will be described with reference to the flow chart of FIG. 3 and the conceptual diagram of FIG.

[0028] The health condition variation prediction method executed by the health condition variation prediction system 100 includes a load measurement process S11, a preprocessing process S12, a bed state determination process S13, a trend acquisition process S14, a pattern identification process S15, a health condition variation prediction process S16, and a display process S17, as shown in the flowchart of Figure 3.

[0029] [Load measurement process S11] In the load measurement step S11, the load of the subject S on the bed BD (i.e., the load applied by the subject S to the bed BD) is measured using the load sensors LS1 to LS4 of the load measurement unit 10. The load of the subject S is distributed and applied to the load sensors LS1 to LS4 arranged under the legs BL1 to BL4 at the four corners of the bed BD, and is distributed and measured by these.

[0030] Each of the load sensors LS1 to LS4 detects the load of the subject S as an analog signal. Hereinafter, the analog signals (that is, raw data signals) detected by the load sensors LS1, LS2, LS3, and LS4 will be referred to as analog signals s1, s2, s3, and s4, respectively.

[0031] The analog signals s1 to s4 are sent to the information processing device 20 via a communication interface (not shown) of the load measuring unit 10 and the network NW.

[0032] [Pretreatment step S12] In the pre-processing step S12, the pre-processing unit 211 of the information processing device 20 calculates various parameters based on the load values ​​(analog signals s1, s2, s3, s4) of the subject S received from the load measuring unit 10 via the network NW.

[0033] First, the pre-processing unit 211 receives the analog signals s1 to s4 sent from the load measuring unit 10 via the network NW and a communication interface (not shown) of the information processing device 20. Then, the pre-processing unit 211 digitally converts the analog signals s1 to s4 with a sampling period of, for example, 5 milliseconds to obtain digital signals. Hereinafter, the digital signals obtained by digitally converting the analog signals s1, s2, s3, and s4, respectively, are referred to as load signals ss1, ss2, ss3, and ss4.

[0034] In this embodiment, the pre-processing unit 211 calculates at least one of the following parameters: The pre-processing unit 211 stores the calculated parameters in the storage unit 22.

[0035] (1) Load signals ss1 to ss4 The load signals ss1 to ss4 are used as parameters as they are.

[0036] (2) Spectral amplitude SAn for each frequency band X-Y The weight signals ss1 to ss4 are sequentially subjected to a short-time Fourier transform to sequentially generate frequency spectra (an example is shown in FIG. 5). The length of the window function of the short-time Fourier transform may be, for example, about 10 [s]. The short-time Fourier transform may be performed every 5 [s] to 15 [s], for example. The window functions of two consecutive short-time Fourier transforms may partially overlap each other.

[0037] It is known that the length of the window function in the short-time Fourier transform correlates with the time resolution and frequency resolution. In other words, if the time length of the window function is increased, the time resolution becomes lower and the frequency resolution becomes higher, and if it is decreased, the time resolution becomes higher and the frequency resolution becomes lower. One method known for improving these characteristics is to suppress side lobes by introducing a window function. Other known methods include Generalized Harmonic Analysis (GHA) and Multi-Windows STFT, which uses multiple window lengths. These methods can be used to analyze the frequency spectrum and determine explanatory variables.

[0038] For each of the generated frequency spectra, the frequency is divided into 10 bands, and the average value of the spectral amplitude in each band is calculated. In the present embodiment, as an example, the average value of the spectral amplitude in each band is calculated for 0.2 to 0.5 [Hz], 0.6 to 1.0 [Hz], 1.1 to 1.5 [Hz], 1.6 to 2.0 [Hz], 2.1 to 2.5 [Hz], 2.6 to 3.0 [Hz], 3.1 to 3.5 [Hz], 3.6 to 4.0 [Hz], 4.1 to 4.5 [Hz], and 4.6 to 5.0 [Hz].

[0039] In the following, the load signal ss n The average spectral amplitude of the frequency band from X to Y [Hz] of the frequency spectrum generated based on (n = 1, 2, 3, 4) is the spectral amplitude SAn X―Y Write (n=1, 2, 3, 4).

[0040] (3) Center of gravity position The position of the center of gravity of the subject S on the bed BD is calculated based on the load signals ss1 to ss4. Any known method may be used to calculate the position of the center of gravity. An example of the method of calculating the position of the center of gravity is described in the specification of Japanese Patent No. 6321719.

[0041] (4)Respiration waveform The center of gravity G of the subject S vibrates in a direction along the body axis (spine) of the subject S in response to the subject S's breathing (hereinafter, this vibration will be referred to as "respiratory vibration"). The respiratory waveform shows the state of respiratory vibration, and is represented by the horizontal axis representing time and the vertical axis representing the amplitude of respiratory vibration. An example of a method for drawing a respiratory waveform is described in the specification of Japanese Patent No. 6661173.

[0042] (5)Respiration rate The respiratory rate of the subject S is calculated, for example, based on the respiratory waveform. Since one cycle of the respiratory waveform corresponds to one breath (exhalation and inhalation), if the distance between peaks of the respiratory waveform is T, then the respiratory rate B per minute is estimated as B=60 / T.

[0043] (6) Respiration amplitude The respiration rate of the subject S is calculated, for example, based on a respiration waveform. The amplitude of the respiration waveform can be estimated as a respiration amplitude value.

[0044] (7) Heart rate waveform As an example, a waveform obtained by separating components included in the frequency band of the heartbeat (approximately 0.5 Hz to approximately 3.3 Hz) from each of the load signals ss1 to ss4 can be used as the heartbeat waveform.

[0045] (8) ACI (Activity Index) ACI is an index indicating the activity level of a subject. For example, ACI is calculated as a time integral value of a simple average value of standard deviations σ1 to σ4 over a predetermined period (e.g., 20 seconds) with standard deviations of time fluctuations of load signals ss1, ss2, ss3, and ss4 as σ1, σ2, σ3, and σ4, respectively. Instead of standard deviations σ1 to σ4, values ​​obtained by dividing standard deviations σ1 to σ4 by the amplitude of the respiratory waveform may be used. Further details of ACI are described in the specification of Japanese Patent No. 6661173.

[0046] (9) Total load It can be calculated as the sum of the load values ​​indicated by the load signals ss1 to ss4.

[0047] (10)Other In addition, depending on the state to be classified, a predetermined parameter that is presumed to be related to the state can be created and used.

[0048] [Bed state determination process S13] In the lying-down state determination step S13, the lying-down state determination unit 212 calculates at least one of the various parameters calculated in the pre-processing step S12 into a lying-down state determination model M STA to determine the bedside state of subject S.

[0049] Bedside state judgment model M STA is a trained model created to determine the lying state of the subject S on the bed BD based on predetermined explanatory variables (the creation method will be described later). STA For example, the spectral amplitude SAn in the frequency band of 0.2 to 0.5 Hz 0.2-0.5 (n=1, 2, 3, 4) and the spectral amplitude SAn in the frequency band of 4.6 to 5.0 [Hz] 4.6-5.0 When (n=1, 2, 3, 4) are input as features (explanatory variables), the lying state of subject S may be classified into cough, body movement, abnormal breathing, or normal. Abnormal breathing here includes one or more of hypopnea, hyperpnea, and other abnormal breathing states. In the following explanation, a case will be described in which "hypopnea" is determined as one example of abnormal breathing to be one of the lying states.

[0050] The subject S's lying state being "coughing" means that the subject S is coughing on the bed BD. The subject S's lying state being "body movement" means that the subject S is exhibiting body movement on the bed BD. Body movement may include large body movement accompanied by movement of the torso, small body movement accompanied by movement of the head and / or limbs without movement of the torso, etc. The subject S's lying state being "hypopnea" means that the breathing rate of the subject S on the bed BD is equal to or lower than a predetermined value. The subject S is in a "hypopnea" state, for example, when the subject S temporarily reaches a state of apnea or respiratory arrest due to a respiratory disorder or the like, or when the frequency of breathing decreases due to a decline in physical strength or the like. The subject S's lying state being "normal" means that the subject S on the bed BD is not in any of the coughing state, body movement state, and hypopnea state.

[0051] In addition, the bed state judgment model M STA may be capable of identifying and classifying, for example, hyperventilation as the lying-down state of the subject S in addition to the above-mentioned cough, body movement, hypoventilation, and normal. The lying-down state of the subject S being "hyperventilation" means that the breathing rate of the subject S on the bed BD is equal to or higher than a predetermined value (however, a value different from the predetermined value when determining whether or not the subject S is "hypopnea"). In addition, the lying-down state determination model M STA may identify and classify the aforementioned hypopnea and hyperpnea, as well as other abnormal respiratory conditions, as one category (eg, "breathing caution") as the lying-down state of the subject S.

[0052] The lying-down state determination unit 212 sequentially calculates the parameters calculated in the pre-processing step S12 into the lying-down state determination model M at a predetermined period (for example, 5 [s] to 15 [s]). STA The bed state determination model M STA The lying-in-bed state determination unit 212 receives the input and outputs a determination result. STA The output (i.e., the determination result) is stored in the storage unit 22 one by one.

[0053] During the period when the subject S is out of bed BD, the load signals ss1 to ss4 are not obtained and the various parameters are not calculated in the pre-processing step S12, so the lying state determination unit 212 cannot obtain the parameter values. When the lying state determination unit 212 cannot obtain the parameter values ​​during a certain period, it determines that the lying state of the subject S during that period is "out of bed".

[0054] Fig. 6 is an example of a graph displaying, in chronological order, the lying-in-bed state of the subject S determined by the lying-in-bed state determination unit 212. From the graph in Fig. 6, it can be seen that, although the subject S is generally in a "normal" state, he or she occasionally enters a "coughing" state, a "body movement" state, a "low breathing" state, and an "out of bed" state.

[0055] [Trend acquisition process S14] In the trend acquisition step S14, the trend acquisition section 213 calculates (acquires) the trend per unit time of the lying state of the subject S determined in the lying state determination step S13.

[0056] In this embodiment, the tendency to be in bed per unit time means the proportion of the lying state within the unit time. The unit time can be set arbitrarily, and examples include one hour and one day. Hereinafter, the tendency of the subject S to be in bed per unit time is referred to as the "state tendency." The unit time referred to here is synonymous with the calculation span of the tendency to be in bed.

[0057] The trend acquisition unit 213 calculates a state trend indicating the proportion of the tendency of lying in bed during the elapsed period from the time series data of the lying in bed state during the elapsed period each time a unit time passes. For convenience of explanation, such state trend data generated by one calculation by the trend acquisition unit 213 will be counted as "one record (of state trend)". However, the actual state trend data does not have to be in record format. The trend acquisition unit 213 saves the calculated state trend of one record in the storage unit 22. As a result, records of state trends are generated at any time at unit time intervals and are saved in the storage unit 22 in chronological order.

[0058] FIG. 7 is a graph in which the condition trends of the subject S are arranged as bar graphs for each record. In the example of FIG. 7, the unit time is one day, and 30 records of the condition trends (i.e., 30 days) are arranged. Note that the graph shown in FIG. 7 is merely a schematic graph for explanation. That is, the information processing device 20 or the terminal device 30 according to the present disclosure may or may not create a graph as shown in FIG. 7. In FIG. 7, one bar graph shows the condition trend of one record (unit time=one day). For example, the leftmost bar graph shows the condition trend for 24 hours on November 1st, and the condition trends are about 25% for getting out of bed, about 60% for normal, about 2% for body movement, and about 13% for hypopnea. Similarly, the second graph from the left shows the condition trend for 24 hours on November 2nd, and the condition trends are about 20% for getting out of bed, about 59% for normal, about 1% for body movement, and about 20% for hypopnea.

[0059] [Regularity identification step S15] In the pattern identification step S15, the pattern identification unit 214 calculates the pattern of the state tendency calculated in the tendency acquisition step S14 using a pattern identification model M REG Identify using:

[0060] The regularity of the state tendency means an arbitrary pattern, periodicity, etc. of the state tendency derived when the state tendency is captured in a time series. The regularity of the state tendency includes, as non-limiting examples, an increase (rapid increase, gradual increase, etc.) or decrease (sudden decrease, gradual decrease, etc.) of any one state, an increase (sudden increase, gradual increase, etc.) or decrease (sudden decrease, gradual decrease, etc.) of any two or more states, an increase (sudden increase, gradual increase, etc.) of any one state and a decrease (sudden decrease, gradual decrease, etc.) of any one or more other states, and an increase (sudden increase, gradual increase, etc.) of any two or more states and a decrease (sudden decrease, gradual decrease, etc.) of any two or more other states.

[0061] Regularity specific model M REG is a trained model that, when two or more records of state trends are input, outputs regularity information that includes information indicating regularities in the state trends derived from the input records. REGOne example of the regularity identification model M is a trained model that is generated to identify regularities in state trends by pattern mining. REG The method for creating this will be described later.

[0062] The regularity identification unit 214 periodically (for example, the least common multiple of the records of two or more status trends) calculates a regularity identification model M REG and identifies a regularity derived from the time variation of the condition tendency of the subject S. The period for identifying the regularity may be different from the collection range of the records used to identify the regularity. For example, when identifying a regularity on a daily basis, the records collected for identifying the regularity may be those from up to 7 days prior to the time of identifying the regularity. When a regularity is identified, the regularity identifying unit 214 generates regularity information that associates the content of the identified regularity with the date and time when the regularity was indicated, and stores the information in the storage unit 22.

[0063] Here, the "date and time when the regularity was indicated" refers to, for example, (i) the date and time when the regularity information was generated, (ii) the regularity identification model M REG (iii) the date and time of the most recent record of the state trend entered in the regularity identification model M REG The "date and time" may be the date and time of the oldest record among the records of the status trends input in the "Records" field. Also, the "date and time" may indicate a period such as a date, month, or year other than time.

[0064] In this embodiment, when the subject S is receiving care, the regularity identification unit 214 uses the regularity identification model M REGThe correlation between the presence or absence of "getting out of bed" and the biological condition of the subject S is relatively small, so by identifying the pattern while excluding the state tendency with a high rate of "getting out of bed," the correlation between the identified pattern and the fluctuation of the biological condition of the subject S can be further increased. In this way, increasing the correlation between the pattern and the fluctuation of the biological condition of the subject S can further increase the physical condition variation prediction model M PRE This will lead to improved prediction accuracy.

[0065] [Health condition change prediction process S16] In the health condition variation prediction step S16, the health condition variation prediction unit 215 uses the regularity information generated in the regularity identification step S15 as a health condition variation prediction model M STA By inputting the data into the health condition mutation prediction model M STA As an output from the , information indicating the predicted result of the change in physical condition of the subject S (hereinafter, referred to as predicted information) is obtained.

[0066] Health condition mutation prediction model M PRE is a trained model created to predict changes in the physical condition of subject S on bed BD using regularity information as explanatory variables (the creation method will be described later). In other words, the physical condition change prediction model M PRE is a trained model that outputs prediction information when regularity information is input.

[0067] Health condition mutation prediction model M PRE The variation in the physical condition of the subject S predicted by the physical condition variation prediction model M can be in various forms. PRE may (i) predict whether or not a change in the physical condition of the subject S will occur, (ii) predict the probability that a change in the physical condition of the subject S will occur, or (iii) predict the content of a change in the physical condition that may occur in the subject S. In addition, the physical condition change prediction model M PRE In addition to at least one of (i) to (iii), the timing of when a change in physical condition will occur may be predicted.

[0068] The physical condition variation prediction model M PRESpecific examples of changes in physical condition predicted by the method include the occurrence of attacks of various diseases such as fever, asthma, epilepsy, and heart disease, the occurrence of a specific disease (pneumonia, etc.), high blood pressure, swallowing disorders, and aspiration, and the subject's entering a specific state, as well as improvement in physical condition.

[0069] The physical condition variation prediction unit 215 uses a physical condition variation prediction model M PRE The output (i.e., prediction information) is sequentially stored in the storage unit 22.

[0070] As mentioned above, the health mutation prediction model M PRE predicts the changes in the physical condition of the subject S. In other words, the physical condition change prediction model M PRE can be said to be a model that estimates a change in physical condition of subject S at a specific time or period in the future. Here, the "specific time" or "specific period" may be automatically determined based on the date and time included in the regularity information, for example. In other words, the time of the change in physical condition to be predicted may be determined based on the date and time included in the input regularity information.

[0071] Or, the health mutation prediction model M PRE (i) may be a model that predicts a pinpoint change in physical condition at a timing specified by date and time information, using regularity information and information specifying a date and / or time (hereinafter, for convenience, referred to as "date and time information") as explanatory variables. In addition, the physical condition change prediction model M PRE (ii) may be a model that predicts changes in physical condition for a period indicated by the period information from the timing specified by the date and time information, using regularity information, date and time information, and information specifying a period (hereinafter referred to as period information) as explanatory variables. Specific examples of (i) include "predicting changes in physical condition three days after the prediction time" and "predicting changes in physical condition on April 1, 2023". Specific examples of (ii) include "predicting changes in physical condition that may occur within one month from the prediction time".

[0072] The timing indicated by the date and time information may be specified or identified by the information processing device 20, or may be specified by the user of the physical condition variation prediction system 100 via the input unit 32 of the terminal device 30. In the latter case, the terminal device 30 transmits information indicating the timing (e.g., date) specified by the user to the information processing device 20, and the information processing device 20 receives and uses the information.

[0073] The period indicated by the period information may also be specified or identified by the information processing device 20, or may be specified by the user of the physical condition variation prediction system 100 via the input unit 32 of the terminal device 30. In this case, as in the case of the date and time information, information indicating the predetermined period is transmitted from the terminal device 30 to the information processing device 20.

[0074] In the health condition variation prediction step S16, the health condition variation prediction unit 215 uses a health condition variation prediction model M PRE By inputting the required input data, the health mutation prediction model M PRE As an output from

[0075] [Display process S17] In the display step S17, the control unit 21 of the information processing device 20 sequentially transmits the prediction information to the terminal device 30, thereby causing the terminal device 30 to execute a predetermined display related to the prediction of a physical condition change. The terminal device 30 performs a predetermined display on the display unit 31 based on the prediction result received from the physical condition change prediction unit 215.

[0076] Specifically, for example, the display unit 31 displays, by image and / or sound, the presence or absence of a change in physical condition of the subject S indicated by the prediction result, the details of the change in physical condition, the predicted time when the change in physical condition is to occur, etc. Furthermore, the terminal device 30 may display at least one of the graphs shown in Figs. 5 to 7 on the display unit 31.

[0077] In the regularity identification step S15, the regularity identification model M REG If there is no particular pattern derived from the state trend records input to the REGAlternatively, the information processing device 20 may output regularity information indicating "no regularity." In this case, the information processing device 20 may not execute the health condition change prediction step S16, but may notify the terminal device 30 of "unable to predict health condition change" as prediction information. Then, the terminal device 30 may display on the display unit 31, in the display step S17, information indicating that a health condition change could not be predicted.

[0078] The method of predicting a change in physical condition executed by the system 100 for predicting a change in physical condition can be summarized as follows, with reference to the conceptual diagram of Fig. 4. Note that the diagrams and data shown in each row of Fig. 4 are merely schematic representations of the data for the purpose of understanding, and the actual data does not have to be in this format.

[0079] First, as shown in the first and second rows of Fig. 4, the load measuring unit 10 measures analog loads s1 to s4 using the load sensors LS1 to LS4 and sends them to the information processing device 20 (load measuring step S11). Next, as shown in the second and third rows of Fig. 4, the pre-processing unit 211 converts the analog loads s1 to s4 into digital loads ss1 to ss4 and calculates various parameters (pre-processing step S12).

[0080] Next, as shown in the third and fourth rows of FIG. 4, the lying state determination unit 212 calculates various parameters based on the digital loads ss1 to ss4 for a predetermined period into a state determination model M STA 4, and the lying state of the subject S is sequentially determined at a predetermined cycle (lying state determination step S13). After that, as shown in the fourth and fifth rows of FIG. 4, the trend acquisition unit 213 acquires the state trend of the lying state per unit time (trend acquisition step S14).

[0081] Next, as shown in the fifth and sixth rows of Fig. 4, the regularity identifying unit 214 inputs the condition trends (at least two or more condition trend records, four records in the example of Fig. 4) for a predetermined period into the regularity identifying model to identify regularities (regularity identifying step S15). Then, as shown in the sixth and seventh rows of Fig. 4, the health condition variation identifying unit 215 inputs the identified regularities (and, if necessary, date and time information and / or period information) into the health condition variation prediction model M PREBy inputting the information into the above, it is possible to predict whether or not subject S will experience a change in his / her physical condition, the probability that a change in his / her physical condition will occur, the details of the change in his / her physical condition, and / or the time when the change in his / her physical condition is likely to occur.

[0082] [How to create a bedside state judgment model] Next, the bed state determination model M used in the physical condition change prediction system 100 of the above embodiment STA This explains how to create it.

[0083] Bedside state judgment model M used in the physical condition variation prediction system 100 STA is created according to the content of the lying state to be determined.

[0084] As shown in the flowchart of FIG. 8, the method for creating the lying state determination model includes a teacher data creation step S21, an explanatory variable determination step S22, and a model creation step S23.

[0085] In the teacher data creation step S21, teacher data to be used in the model creation step S22 is created, for example, according to the following procedure.

[0086] First, with the subject S lying on the bed BD, the load of the subject S is measured by the load sensors LS1 to LS4 and the subject S on the bed BD is photographed in parallel for a predetermined period of time, and the load signals ss1 to ss4 and the image (moving image) of the subject S are stored. An example of the stored load signals ss1 to ss4 is shown in FIG. At this time, the voice of the subject S may also be recorded.

[0087] Next, the stored load signals ss1 to ss4 are labeled (tagged) based on the stored image of the subject S. The labeling is performed by linking the load signals ss1 to ss4 with the type of state (motion) of the subject S and the start and end times of the state.

[0088] Examples of the types of conditions to be labeled are "hypopnea," "cough," "normal," and "body movement." In the example shown in FIG. 9, the period from time t1 to time t2 is labeled "hypopnea," the period from time t3 to time t4 and the period from time t7 to time t8 is labeled "body movement," and the period from time t5 to time t6 and the period from time t9 to time t 10 The period was labeled "cough" and the other periods were labeled "normal."

[0089] In addition, the types of states to be labeled may include, instead of or in addition to "hypopnea," "cough," "normal," and "body movement," at least one of the following: "getting into bed" (subject moving from under the bed to on top of the bed), "getting out of bed" (subject moving from on top of the bed to under the bed), "sitting on the edge of the bed" (subject sitting on the edge of the bed), "sitting position" (subject sitting on the bed), "change in reclining bed angle" (rising or lowering), dangerous movements in bed (jumping out over the bed rails, etc.), "eating," "lying on one's back," "lying on one's right side," "lying on one's left side," "lying on one's face," etc.

[0090] Next, a plurality of feature quantities are created based on the stored weight signals ss1 to ss4. The feature quantities used in this embodiment are the above-mentioned (1) weight signals ss1 to ss4, (2) spectrum amplitudes SAn of each frequency band, X-Y , (3) center of gravity position, (4) respiratory waveform, (5) respiratory rate, (6) respiratory amplitude value, (7) heart rate waveform, (8) ACI (Activity Index), and (9) total load.

[0091] As a result, a data set is created as teacher data, in which a combination of multiple features obtained based on the weight signals ss1 to ss4 (hereinafter referred to as the "feature set FS") is used as input, and the states of the subject S labeled with the weight signals ss1 to ss4 are used as output. Since each feature is created based on the weight signals ss1 to ss4, each feature is associated with the state of the subject S labeled with the weight signals ss1 to ss4.

[0092] In the explanatory variable determination step S22, at least one feature is selected from the feature set FS included in the training data created in the training data creation step S21, and is determined as an explanatory variable to be used in the model creation step S23 described below.

[0093] The selection of features (determination of explanatory variables) can be performed, for example, by using a gradient boosting method based on a decision tree algorithm (specifically, for example, XGBoost, LightGBM, CatBoost, etc.). Note that logistic regression may be used instead of the gradient boosting method.

[0094] Specifically, using the teacher data created in the teacher data creation step S21, training data is created in which the feature set FS is input and a predetermined state of the subject S (for example, "cough", "body movement", "low breathing", and "normal") is output. The predetermined state includes a state to be classified by the model to be created, and at least one is selected from a plurality of labeled states. Then, using the training data, a trained model is created by a gradient boosting method, which uses the feature set FS as input and classifies the state of the subject S into the predetermined state. After that, the contribution of each feature in the created trained model is calculated based on the importance of the decision tree algorithm, and one or more features with a high contribution are determined as explanatory variables to be used in the model creation step S23.

[0095] In the model creation step S23, a bed state determination model M is created that classifies the state of the subject S into a predetermined state (for example, "cough", "body movement", "low breathing", and "normal") using the feature quantities selected in the explanatory variable determination step S22 as explanatory variables. STA is generated by supervised machine learning. Specifically, for example, a naive Bayes classifier, a decision tree, a support vector machine (SVM), a neural network, K nearest neighbors, gradient boosting, logistic regression, etc. can be used for the supervised machine learning. The teacher data can be created, for example, by extracting data related to the feature selected in the explanatory variable determination step S22 from the feature set FS.

[0096] [How to create a regularity identification model] Next, the regularity specification model M used in the physical condition change prediction system 100 of the above embodiment REG Here, a method for creating the regularity specification model M is explained using the health condition variation prediction system 100. REG An example of creating a new .

[0097] Regularity specific model M REG As shown in the flowchart of FIG. 10, the method of creating the model includes a load measurement step S31, a lying state determination step S32, a tendency acquisition step S33, and a model creation step S34.

[0098] The load measuring step S31 is the same as the load measuring step S11 in the physical condition variation prediction method. That is, in the load measuring step S31, the load measuring unit 10 detects the load of the subject S on the bed BD using the load sensors LS1 to LS4, and transmits the detected loads to the information processing device 20 as analog signals s1 to s4.

[0099] The lying-in-bed state determination step S32 is the same as the pre-processing step S12 and the lying-in-bed state determination step S13 in the method for predicting a change in physical condition. That is, in the lying-in-bed state determination step S32, the pre-processing unit 211 converts the analog signals s1 to s4 into digital signals ss1 to ss4, and calculates various parameters based on the digital signals. Then, the lying-in-bed state determination unit 212 converts the calculated various parameters into a lying-in-bed state determination model M STA to determine the bedside state of subject S.

[0100] The trend acquisition step S33 is the same as the trend acquisition step S14 in the physical condition change prediction method. That is, in the trend acquisition step S33, the trend acquisition unit 213 calculates the trend per unit time of the lying state of the subject S determined by the lying state determination unit 212.

[0101] In the trend acquisition step S33, records with different unit times may be generated. For example, in the trend acquisition step S33, a record of the state trend may be generated every hour (i.e., the unit time is one hour), and a record of the state trend may be generated every day with the unit time being one day. In addition, the state trend may be calculated with the unit time being an arbitrary period other than one hour or one day.

[0102] In the model creation step S34, the model creation unit 217 creates a regularity identification model M that identifies a predetermined regularity based on one or more lying states by unsupervised machine learning using the tendency of the lying state acquired in the tendency acquisition step S33 as training data. REG Specifically, unsupervised machine learning can use, for example, random forests, clustering, principal component analysis, association analysis, generative adversarial networks, normal mixture models (GMM), local outlier factor methods (LOF), isolation forests, kernel density estimation, etc.

[0103] In the model creation step S34, it is more preferable to use a plurality of records of the state tendency acquired in the tendency acquisition step S33 as training data. This allows the regularity identification model M to be able to identify regularities with higher accuracy. REG The amount and content of the training data used in the model creation step S34 can be determined by the regularity identification model M REG This may be determined appropriately as long as the generalization performance of the method is not impaired.

[0104] When using records of multiple state trends as training data, each record may be a record group having the same specific unit time of the state trend (i.e., a specific span of the state trend), or the unit time of some or all of the records may be different. When using a record group having the same unit time, it is possible to improve the accuracy of identifying regularities when records of the unit time are input to the model. On the other hand, when using a record group having different unit times, it is possible to create a model with high generalization performance.

[0105] Furthermore, in this embodiment, among the condition trends acquired in the trend acquisition step S33, condition trends in which the proportion of "getting out of bed" is equal to or greater than a predetermined value (for example, about 50 to 70%) may not be adopted as training data. Since the correlation between the presence or absence of "getting out of bed" and the biological condition of the subject S is relatively small, by using training data that excludes condition trends in which the proportion of "getting out of bed" is high, the correlation between the identified regularity and the fluctuation in the biological condition of the subject S can be further increased. In this way, increasing the correlation between the identified regularity and the fluctuation in the biological condition of the subject S will in turn improve the physical condition variation prediction model M. PRE This could lead to improved prediction accuracy.

[0106] Here, the regularity specific model M REG Specifically, the regularity of the state tendency to be identified by the above is, for example, the following regularity. (1) The proportion of people experiencing “cough” has been increasing day by day for two or more days. (2) The rate of coughing increases at a specific time (for example, every day after 5:00 p.m.) for three or more days. (3) The proportion of “hypopnea” patients decreases for three or more consecutive days, and the proportion of “normal” patients increases for three or more consecutive days. (4) The rate of “hypopnea” has been increasing day by day for 25 days or more. (5) The percentage of “hypoventilation” decreases day by day and the percentage of “body movement” increases day by day for more than one week.

[0107] These regularities are examples of regularities that are correlated with changes in the subject's physical condition, and regularity information indicating these regularities is included in the training data as explanatory variables in the method for creating a physical condition change prediction model described below.

[0108] [How to create a model for predicting health condition variations] Next, the physical condition change prediction model M used in the physical condition change prediction system 100 of the above embodiment PRE Here, the method of creating the health change prediction model M is described. PRE The following will be explained by taking the case of creating a

[0109] Health condition mutation prediction model M PRE As shown in the flowchart of FIG. 11, the method of creating the health information includes a physical condition information creation process S41, a data input process S42, a load measurement process S43, a bed state determination process S44, a trend acquisition process S45, a regularity identification process S46, a teacher data creation process S47, and a model creation process S48.

[0110] In the health information creation step S41, the information processing device 20 creates or acquires health information showing the health of the subject S in a chronological order. Specifically, the health information is created by, for example, a data creator such as a doctor, nurse, or caregiver visually observing the health of the subject S and recording the observation results as electronic data in the information processing device 20, the terminal device 30, or another device. Examples of health information include nursing records created by nurses in hospitals, and care records created by caregivers in care facilities. The information processing device 20 creates the health information based on the observation results input to its own device. Alternatively, the information processing device 20 obtains the health information by acquiring the health information from the terminal device 30 or another device periodically or at a predetermined timing.

[0111] Instead of or in addition to visual data creation by the data creator, various sensors may be used to measure the biological information of the subject S, and the measured biological information may be recorded together with the date and time of measurement in the information processing device 20, the terminal device 30, or other devices, and the recorded data may be used as physical condition information. Examples of various sensors include a thermometer, a pulse meter, etc. that measure vital data such as pulse, blood pressure, respiration, and body temperature.

[0112] The created physical condition information is data that links the physical condition of the subject S with time. Therefore, based on the physical condition information, it is possible to understand when and what state the physical condition of the subject S was. Note that "time" here does not only refer to the time of day, but is not particularly limited to any division that indicates a period such as date, month, or year.

[0113] The physical condition of the subject S that may be included in the physical condition information specifically includes, for example, heart rate, respiratory rate, blood pressure, body temperature, blood oxygen concentration, and the like.

[0114] In this embodiment, the behavioral information of the subject S, such as eating, cleaning, etc., is also included in the physical condition information. This behavioral information may be included in the nursing record and / or care record described above. In the model creation step S48 described later, learning is performed taking into account the influence of the load fluctuation according to the behavior of the subject S, so that the physical condition variation prediction model M PRE This can improve the prediction accuracy.

[0115] In the data input step S42, the physical condition information created in the physical condition information creation step S41 is input into the information processing device 20. The physical condition information can be input into the information processing device 20, for example, via the input unit 32 of the terminal device 30. Alternatively, the physical condition information may be input by transmitting it to the information processing device 20 from an external device such as a PC. The information processing device 20 stores the received physical condition information in the memory unit 22.

[0116] The load measuring step S43 is the same as the load measuring step S11 in the physical condition variation prediction method. That is, in the load measuring step S43, the load measuring unit 10 detects the load of the subject S on the bed BD using the load sensors LS1 to LS4, and sends the detected loads to the information processing device 20 as analog signals s1 to s4.

[0117] The lying-in-bed state determination step S44 is the same as the pre-processing step S12 and the lying-in-bed state determination step S13 in the method for predicting a change in physical condition. That is, in the lying-in-bed state determination step S44, the pre-processing unit 211 converts the analog signals s1 to s4 into load signals ss1 to ss4, and calculates various parameters based on the load signals. Then, the lying-in-bed state determination unit 212 converts the calculated various parameters into a lying-in-bed state determination model M STA to determine the bedside state of subject S.

[0118] The trend acquisition step S45 is the same as the trend acquisition step S14 in the physical condition change prediction method. That is, in the trend acquisition step S45, the trend acquisition unit 213 acquires the trend per unit time of the lying state of the subject S determined by the lying state determination unit 212.

[0119] The regularity identifying step S46 is the same as the regularity identifying step S15 in the physical condition variation prediction method. That is, in the regularity identifying step S46, the regularity identifying unit 214 identifies the regularity included in the time-dependent fluctuation of the state tendency by using the regularity identifying model M REG When a regularity is identified, the regularity identification unit 214 generates regularity information and stores it in the storage unit 22. In addition, in the regularity identification step S46, the regularity identification model M REG The regularity information is information that associates the contents of the regularity identified by the model with the date and time when the regularity was indicated. The "date and time when the regularity was indicated" is, for example, (i) the date and time when the regularity information was generated, (ii) the regularity identification model M REG (iii) the date and time of the most recent record of the state trend entered in the regularity identification model M REGThe "date and time" may be the date and time of the oldest record among the records of the status trends input in the "Records" field. Also, the "date and time" may indicate a period such as a date, month, or year other than time.

[0120] In the teacher data creation step S47, the teacher data creation unit 217 creates teacher data using the physical condition information stored in the memory unit 22 in the data input step S42 and the regularity information stored in the memory unit 22 in the regularity identification step S46.

[0121] The created teacher data is a data set that inputs the regularity of the condition tendency of the subject S and outputs the subject S's physical condition information. Since both the regularity information and the physical condition information are linked to the date and time, for example, by combining certain regularity information with physical condition information indicating a date and time within a predetermined period from the date and time indicated by the certain regularity information, a set of teacher data can be created that indicates "how the actual physical condition has changed when the condition tendency up to a specific date and time has a certain regularity." Note that, if there are multiple pieces of physical condition information indicating a date and time within a predetermined period from the date and time indicated by the certain regularity information, multiple pieces of teacher data may be created from the certain regularity information. In other words, teacher data that combines certain regularity information with first physical condition information and teacher data that combines the same regularity information with second physical condition information may be created. In the created teacher data, the regularity of the condition tendency and the physical condition of the subject S are associated via the date and time.

[0122] In the model creation step S48, the model creation unit 218 makes the prediction model learn the correlation between the regularity of the condition tendency of the subject S and the physical condition information of the subject S by supervised machine learning using the training data created in the training data creation step S47. As a result, a physical condition variation prediction model M that predicts the physical condition variation of the subject S on the bed BD based on the regularity of the condition tendency is created. PRE Specifically, supervised machine learning can use, for example, a naive Bayes classifier, a decision tree, a support vector machine (SVM), a neural network, K nearest neighbors, gradient boosting, logistic regression, and the like.

[0123] A prediction model M is created by learning the correlation between the pattern of the condition trend of subject S and the physical condition information of subject S. PRE The reasons why it is possible to create

[0124] According to the findings of the inventors of the present invention, before any change occurs in the physical condition of the subject S, some regular fluctuations in the state of the subject S lying down often occur. As a specific example, the following relationship has been observed. Note that the number of days in the following example is explained based on the acquisition date of the record of the condition trend that is the starting point for identifying the regularity.

[0125] (1) If the rate of coughing increases day by day for two or more days, fever will develop after four days. (2) If the rate of coughing increases at a specific time (for example, 5 p.m.) for three or more days, fever will develop three days and 12 hours later. (3) In subjects with a high rate of “hypopnea,” as the rate of “hypopnea” decreases and the rate of “normal” increases, they subsequently develop a fever. (4) In subjects who do not have a high rate of “hypopnea,” if the rate of “hypopnea” increases daily for 25 days or more, the subject will die within one month. (5) In subjects with a high proportion of “normal” subjects, if the proportion of “hypoventilation” decreases day by day and the proportion of “body movement” increases day by day for more than one week, the subject’s physical condition will improve from the eighth day.

[0126] In this way, there is a correlation between the fluctuation of the subject S's lying state and the fluctuation of the subject's physical condition. Therefore, by having the prediction model learn the correlation between the fluctuation of the subject S's lying state and the fluctuation of the subject's physical condition, a prediction model that predicts the fluctuation of the subject's physical condition based on the fluctuation of the subject's lying state can be created. In addition, by using the regularity of the time fluctuation of the state tendency of the subject S's lying state, rather than the time fluctuation of the subject S's lying state itself, as an input to the prediction model, it is possible to perform a more accurate prediction in which the influence of irregular fluctuations in the subject S's lying state is suppressed.

[0127] As described above, the health condition change prediction system 100 and the health condition change prediction method of the above embodiment use a health condition change prediction model M PRE Using this, a change in physical condition of the subject S is predicted. Therefore, the influence of irregular changes in the state of the subject S in bed can be suppressed and changes in the physical condition of the subject S can be predicted with high accuracy. Furthermore, the method of creating a physical condition change prediction model of the above embodiment allows the prediction model to learn the correlation between the regularity of the state tendency of the subject S and the physical condition information of the subject S, thereby making it possible to create a highly accurate physical condition change prediction model in which the influence of irregular changes in the state of the subject S in bed is suppressed.

[0128] In the above embodiment, the following modifications may be used.

[0129] In the above embodiment, various data on a certain subject (subject S) are used to generate a bed state determination model M STA , the regularity specific model M REG , and the health mutation prediction model M PRE That is, in the above embodiment, a subject (subject S) whose physical condition change is to be predicted and a bed state determination model M STA and the health mutation prediction model M PRE The case where the subjects for the training data are the same as those for the regularity identification model M REG The record of the state tendency input when creating the lying-in-bed state determination model M STA , the regularity specific model M REG , and the health mutation prediction model M PRE Each of them was a trained model that learned the autocorrelation of multiple parameters for subject S.

[0130] However, the bed state determination model M STA , the regularity specific model M REG , and the health mutation prediction model M PREThe method of creating the models is not limited to this. For example, one or more of these three types of models may be created from data on subjects (or subject groups) other than the subject (subject S) whose physical condition variation is to be predicted.

[0131] That is, the bed state judgment model M STA , the regularity specific model M REG , and the health mutation prediction model M PRE may be trained models that have learned correlations between multiple parameters obtained from the other subjects (or subject groups). In other words, the subjects from which data (e.g., load data and physical condition information) required for creating various trained models are collected may be different from the subjects whose physical condition changes are actually predicted.

[0132] In addition, the bed state judgment model M STA , the regularity specific model M REG , and the health mutation prediction model M PRE may be created using both data on the subject (subject S) whose physical condition change is to be predicted and data obtained from other subjects (or subject groups). That is, the bed state determination model M STA , the regularity specific model M REG , and the health mutation prediction model M PRE may each be a trained model that has learned correlations between parameters for multiple humans, including subject S.

[0133] In the above embodiment, the lying state determination model M STA Creation of regularity identification model M REG Creation of a health mutation prediction model M PRE The creation of the bed state determination model M and the prediction of the health change were all performed by the health change prediction system 100 shown in FIG. STA , the regularity specific model M REG , and the health mutation prediction model M PREIn this case, the information processing device 20 of the health condition variation prediction system 100 does not need to include the teacher data creation unit 216 and the model creation unit 217.

[0134] In addition, various trained models (bed state determination model M) can be used in systems and devices other than the physical condition variation prediction system 100. STA , the regularity specific model M REG , and the health mutation prediction model M PRE ), a configuration similar to that of the physical condition variation prediction system 100 is not necessarily required.

[0135] For example, the bed state judgment model M STA When creating the lying state determination model M, the load measuring unit 10 and a model creating device such as a computer that can communicate with the load measuring unit 10 can be used. STA In this case, the control unit of the model creation device corresponds to the control unit 21, and the control unit of the model creation device includes at least a preprocessing unit 211 and a lying-down state determination unit 212. In addition, the lying-down state determination model M can be created in the storage unit of the computer or an external storage device that can communicate with the computer. STA An untrained model serving as a template may be stored.

[0136] For example, the regularity specific model M REG When creating the lying state determination model, a load measuring unit 10, a model creating device such as a computer capable of communicating with the load measuring unit 10, and a created lying state determination model M STA Using this, the regularity identification model M REG In this case, the control unit of the model creating device corresponds to the control unit 21, and the control unit of the model creating device includes at least a preprocessing unit 211, a lying-in-bed state determination unit 212, a tendency acquisition unit 213, and a regularity specification unit 214. In addition, the created lying-in-bed state determination model M can be stored in the storage unit of the computer or an external storage device that can communicate with the computer. STA The control unit of the model creating device uses this to determine the state of lying down.REG The processor 10 may store an untrained model that serves as a template for the above.

[0137] For example, the health mutation prediction model M PRE When creating the bed state determination model, the load measuring unit 10, the terminal device 30 or other device from which physical condition information is obtained, a model creation device such as a computer that can communicate with these two devices, and the created bed state determination model M STA and the regularity identification model M REG Using this, the health mutation prediction model M PRE In this case, the control unit of the model creating device corresponds to the control unit 21, and the control unit of the model creating device includes at least a preprocessing unit 211, a bed state determination unit 212, a tendency acquisition unit 213, a regularity specification unit 214, and a physical condition change prediction unit 215. In addition, the created bed state determination model M can be stored in the storage unit of the computer or an external storage device that can communicate with the computer. STA and the created regularity identification model M REG The control unit of the model creation device uses this to determine the state of lying down and to identify patterns. PRE The processor 10 may store an untrained model that serves as a template for the above.

[0138] The creation of the three trained models described above and the prediction of changes in physical condition may be performed in two or more systems together. For example, the bed state determination model M STA and the regularity specific model M REG may be created in advance by the same model creation device, and the created model may be stored in the storage unit 22 of the information processing device 20 of the health variation prediction system 100. Then, in the health variation prediction system 100, the health variation prediction model M PRE It is also possible to perform the creation of a health condition profile and a prediction of changes in health condition.

[0139] In the method for predicting a change in physical condition according to the above embodiment, the display step S17 may be omitted. In this case, the user of the system 100 for predicting a change in physical condition may periodically check whether or not information indicating the predicted result of a change in physical condition (prediction information) is stored in the storage unit 22 of the information processing device 20. If information indicating the predicted result of a change in physical condition (prediction information) is stored, the user may provide the subject S with preventive measures according to the contents of the information.

[0140] In the teacher data creation step S47 of the physical condition variation model creation method of the above embodiment, data that associates the regularity of the condition tendency of the subject S with environmental information such as the temperature and humidity of the hospital room in which the bed BD is installed over time may be included as teacher data. Then, the teacher data may be used in the model creation step S48. The surrounding environment of the subject S directly or indirectly affects the physical condition (body temperature, respiratory rate, etc.) of the subject S. Therefore, by having the prediction model learn the correlation between the lying state of the subject S and the surrounding environment of the subject S, the prediction accuracy of the created physical condition variation prediction model can be further improved.

[0141] In the above embodiment of the method for creating a physical condition variation model, the case where teacher data is created and used using regularity in the physical condition information and condition tendency of the subject S himself who is the subject of the prediction of a physical condition variation has been described. However, the regularity information and / or physical condition information used as teacher data is not limited to the information of the subject S himself. For example, in addition to or instead of teacher data based on the subject S himself who is to be predicted of a physical condition variation, one or more teacher data based on one or more subjects other than the subject S who is to be predicted of a physical condition variation may be created and used.

[0142] In the method for creating a physical condition variation prediction model according to the above embodiment, the physical condition information of the subject S includes behavioral information of the subject S, such as eating and cleaning, but is not limited to this. The physical condition information does not have to include behavioral information.

[0143] In addition, the health condition included in the health condition information is used to create a health condition variation prediction model M PREIt is also possible to use only the physical condition variation to be predicted by the training data. Even with such training data, the prediction model can learn the correlation between the regularity of the condition tendency of the subject S and the condition variation of the subject S to be predicted. In the present invention, the data included in the physical condition information and indicating the correspondence between the physical condition variation to be predicted and time is called "physical condition variation information".

[0144] In the above-described embodiment of the system 100 for predicting a change in physical condition, the method for predicting a change in physical condition, and the method for creating a prediction model for a change in physical condition, the pattern of the time variation of the state tendency of the subject S is specified using a learned pattern specification model M REG However, the method of identifying the pattern of the state tendency of the subject S is not limited to this. For example, the identification of the pattern can be performed by any method of identifying a pattern in data, such as pattern matching.

[0145] In the physical condition variation prediction system 100 of the above embodiment, the load measurement section 10 may measure the load of the subject S using a plurality of pressure-sensitive sensors (pressure sensors) arranged in a matrix under the sheet, instead of the load sensors LS1 to LS4.

[0146] The physical condition variation prediction system 100 of the above embodiment does not necessarily have to include all of the load sensors LS1 to LS4, and may include only one of them. Also, the load detectors do not necessarily have to be placed at the four corners of the bed, and can be placed at any position so as to detect the load of the subject on the bed and its fluctuation. Also, the load sensors LS1 to LS4 are not limited to load sensors using beam-type load cells, and for example, force sensors can be used.

[0147] In the system 100 for predicting a change in physical condition according to the above embodiment, a camera, an acceleration sensor, or the like for acquiring biometric information of the subject on the bed BD may be used instead of or in addition to the load measuring unit 10. In the case where image information of the subject S is acquired as biometric information using a camera, or where the body movement of the subject S is acquired as biometric information using an acceleration sensor attached to the subject S, a change in physical condition of the subject S can be predicted in a manner substantially similar to that in the case where the load of the subject S is acquired as biometric information.

[0148] For example, there are various conventional technologies that can identify what action a person in an image is performing by analyzing an image captured by a camera. By utilizing such known technologies, for example, it is possible to capture an image of the bed BD (and the subject S) with a camera and analyze the image to identify the action of the subject S on the bed BD (i.e., the lying-in state described in this disclosure). Note that when the bed BD is captured by a camera and the subject S is not captured in the image, it is sufficient to determine that the subject S is in the "out of bed" state. The processing after identifying the lying-in state is similar to that described in the above embodiment.

[0149] Also, for example, by having the subject S wear a wearable device equipped with an acceleration sensor, it is possible to identify the lying state of the subject S lying on the bed BD. The processing after identifying the lying state is similar to that described in the above embodiment.

[0150] The physical condition change prediction system 100 of the above embodiment may not include the terminal device 30. In addition, the information processing device 20 is not limited to being connected to the load measuring unit 10 via a network, and may be disposed near the load measuring unit 10 and the bed BD (for example, in the same room) and connected to the load measuring unit 10 by wiring.

[0151] The bed state determination model M STA , the regularity specific model M REG , Health condition mutation prediction model M PREThe bed state determination model M may be recorded in a non-transitory computer-readable recording medium. By reading the recording medium with a computer such as the information processing device 20, the computer can be made to execute the bed state determination step S13, the pattern specification step S15, and the physical condition change prediction step S16. STA , the regularity specific model M REG , Health condition mutation prediction model M PRE The trained models such as the above may be recorded as a program on a recording medium. In addition, if there are untrained models that serve as templates for these trained models, the untrained models may also be recorded on a non-transitory computer-readable recording medium.

[0152] As long as the characteristics of the present invention are maintained, the present invention is not limited to the above-described embodiments, and other forms conceivable within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention. [Explanation of symbols]

[0153] 10 load measuring unit; 20 information processing device; 211 preprocessing unit; 212 bed state determining unit; 213 trend acquiring unit; 214 pattern identifying unit; 215 physical condition change predicting unit; 22 memory unit; 30 terminal device; BD bed; LS1 to LS4 load sensors; M PRE Health condition mutation prediction model; M REG Regularity specific model; M STA Bedside state judgment model; S Subject

Claims

1. A method for creating a health condition change prediction model that predicts changes in the health condition of a subject in bed, The sensors are used to acquire the subject's biometric information for training data, The lying-down state of the subject for the training data is determined based on the subject's biological information for the training data, To obtain the state trend, which is the trend per unit time of the aforementioned lying-down state, To identify the regularity contained in the temporal variation of the aforementioned state trend, A method comprising creating a health condition change prediction model that learns the correlation between the regularity and the health condition change information of subjects for training data by supervised machine learning using training data that associates the regularity with health condition change information of subjects for training data.

2. The method according to claim 1, wherein the subject and the subject for the training data are the same.

3. The method according to claim 1, wherein the subject and the subject for the training data are different.

4. The method according to any one of claims 1 to 3, wherein the information on changes in physical condition includes whether or not there have been changes in the physical condition of the subject for the training data, and / or the content of the changes in the physical condition of the subject for the training data.

5. The method according to any one of claims 1 to 3, wherein the supine state includes a coughing state, a moving state, a normal state, and a respiratory abnormality state.

6. The aforementioned lying-down state includes a state of being out of bed, The method according to any one of claims 1 to 3, which identifies the regularity included in the temporal variation of the state trend, excluding the state trend in which the proportion of the state where the bed is out of bed is greater than or equal to a predetermined value.

7. The method according to any one of claims 1 to 3, wherein the training data is data that further associates the regularity with behavioral information indicating the behavior of a subject for the training data and / or environmental information indicating the environment around the subject for the training data.

8. The method according to any one of claims 1 to 3, wherein the identification of regularities included in the temporal variation of the state trend is performed by a trained regularity identification model that is trained to output the regularity when the state trend is input.

9. The aforementioned sensor is a load sensor placed on the bed, The method according to any one of claims 1 to 3, wherein the biological information is the load of the subject for the training data.

10. The load sensors consist of four sensors, each positioned at one of the four corners of the bed. The method according to claim 9, wherein each of the four sensors includes a strain generating body that generates strain in accordance with the load of a subject for training data, and a strain sensor attached to the strain generating body.

11. A system for predicting changes in the physical condition of a subject lying in bed, A lying-down state determination unit that determines the lying-down state of the subject based on the subject's biological information acquired by the sensor, A state trend acquisition unit acquires the state trend, which is the trend per unit time of the aforementioned lying-down state, A regularity identification unit that identifies regularities included in the temporal variation of the aforementioned state trend, The system includes a health condition change prediction unit that inputs the aforementioned regularity into a health condition change prediction model to predict changes in the subject's health condition, The aforementioned physical condition change prediction model is a trained model that learns the correlation between the aforementioned regularity and the physical condition change information of the subjects in the training data by supervised machine learning using training data which associates the aforementioned regularity identified for subjects in the training data with physical condition change information indicating changes in the physical condition of the subjects in the training data.

12. The physical condition change prediction system according to claim 11, wherein the subject and the subject for the training data are the same.

13. The physical condition change prediction system according to claim 11, wherein the subject and the subject for the training data are different.

14. The physical condition change prediction system according to any one of claims 11 to 13, wherein the physical condition change information includes whether or not there has been a change in the physical condition of the subject for the training data, and / or the content of the change in the physical condition of the subject for the training data.

15. A system for predicting changes in physical condition according to any one of claims 11 to 13, wherein the aforementioned lying-down state includes a coughing state, a movement state, a normal state, and a respiratory abnormality state.

16. The aforementioned lying-down state includes a state of being out of bed, A physical condition change prediction system according to any one of claims 11 to 13, wherein, in identifying the regularity included in the temporal variation of the state trend, the regularity included in the temporal variation of the state trend is identified by excluding the state trend in which the proportion of the state of being out of bed is greater than or equal to a predetermined value.

17. The physical condition change prediction system according to any one of claims 11 to 13, wherein the training data is data that further associates the regularity with behavioral information indicating the behavior of a subject for the training data and / or environmental information indicating the environment around the subject for the training data.

18. The physical condition change prediction system according to any one of claims 11 to 13, wherein the regularity identification unit identifies regularities included in the temporal fluctuation of the state trend using a trained regularity identification model that has been trained to output the regularity when the state trend is input.

19. The aforementioned sensor is a load sensor placed on the bed, The physical condition change prediction system according to any one of claims 11 to 13, wherein the biological information is the weight of the subject.

20. The load sensors consist of four sensors, each positioned at one of the four corners of the bed. The physical condition change prediction system according to claim 19, wherein each of the four sensors includes a strain generating body that generates strain in accordance with the load of the subject, and a strain sensor attached to the strain generating body.

21. A method for predicting changes in the physical condition of a subject lying in bed, A lying-down state determination step, which determines the lying-down state of the subject based on the subject's biological information acquired by the sensor, A state trend acquisition step to acquire the state trend, which is the trend per unit time of the aforementioned lying-down state, A regularity identification step to identify the regularity included in the temporal variation of the aforementioned state trend, The process includes inputting the aforementioned regularity into a health condition change prediction model to predict changes in the subject's health condition, The physical condition change prediction model is a trained model that learns the correlation between the regularity identified for subjects used as training data and the physical condition change information of subjects used as training data by supervised machine learning using training data which associates the regularity identified for subjects used as training data with physical condition change information indicating changes in the physical condition of the subjects used as training data.

22. The method for predicting changes in physical condition according to claim 21, wherein the subject and the subject for training data are the same.

23. The method for predicting changes in physical condition according to claim 21, wherein the subject and the subject for the training data are different.

24. The method for predicting changes in physical condition according to any one of claims 21 to 23, wherein the information on changes in physical condition includes whether or not there has been a change in the physical condition of the subject for the training data, and / or the content of the change in the physical condition of the subject for the training data.

25. A method for predicting changes in physical condition according to any one of claims 21 to 23, wherein the aforementioned lying-down state includes a coughing state, a movement state, a normal state, and a respiratory abnormality state.

26. The aforementioned lying-down state includes a state of being out of bed, A method for predicting changes in physical condition according to any one of claims 21 to 23, wherein, in identifying the regularity included in the temporal variation of the state trend, the regularity included in the temporal variation of the state trend is identified by excluding the state trend in which the proportion of the state of being out of bed is greater than or equal to a predetermined value.

27. The method for predicting changes in physical condition according to any one of claims 21 to 23, wherein the training data is data that further associates the regularity with behavioral information indicating the behavior of a subject for the training data and / or environmental information indicating the environment around the subject for the training data.

28. In the aforementioned regularity identification step, the identification of regularity included in the temporal variation of the state trend is A method for predicting changes in physical condition according to any one of claims 21 to 23, which is performed using a trained regularity identification model that has been trained to output the regularity when the aforementioned state trend is input.

29. The aforementioned sensor is a load sensor placed on the bed, The method for predicting changes in physical condition according to any one of claims 21 to 23, wherein the biological information is the weight of the subject.

30. The load sensors consist of four sensors, each positioned at one of the four corners of the bed. The method for predicting changes in physical condition according to claim 29, wherein each of the four sensors includes a strain generating body that generates strain in accordance with the load of the subject, and a strain sensor attached to the strain generating body.

31. A health condition change prediction model that predicts changes in the health condition of a subject in bed, The sensors are used to acquire the subject's biometric information for training data, The lying-down state of the subject for the training data is determined based on the subject's biological information for the training data, To obtain the state trend, which is the trend per unit time of the aforementioned lying-down state, To identify the regularity contained in the temporal variation of the aforementioned state trend, The regularity described above is associated with the physical condition variation information of the subject for the training data, Supervised machine learning using training data created by has already learned the correlation between the aforementioned regularity and the subject's physical condition change information for the training data. A health condition change prediction model that predicts changes in the subject's health condition based on the aforementioned regularity.

32. The physical condition change prediction model according to claim 31, wherein the subject and the subject for the training data are the same.

33. The physical condition change prediction model according to claim 31, wherein the subject and the subject for the training data are different.

34. The physical condition change prediction model according to any one of claims 31 to 33, wherein the physical condition change information includes whether or not there has been a change in the physical condition of the subject for the training data, and / or the content of the change in the physical condition of the subject for the training data.

35. The physical condition change prediction model according to any one of claims 31 to 33, wherein the aforementioned lying-down state includes a coughing state, a moving state, a normal state, and a respiratory abnormality state.

36. The aforementioned lying-down state includes a state of being out of bed, A physical condition change prediction model according to any one of claims 31 to 33, wherein, in identifying the regularity included in the temporal variation of the state trend, the regularity included in the temporal variation of the state trend is identified by excluding the state trend in which the proportion of the state of being out of bed is greater than or equal to a predetermined value.

37. The physical condition change prediction model according to any one of claims 31 to 33, wherein the training data is data that further associates the regularity with behavioral information indicating the behavior of a subject for the training data and / or environmental information indicating the environment around the subject for the training data.

38. A physical condition change prediction model according to any one of claims 31 to 33, wherein the identification of regularities included in the temporal variation of the state trend is performed by a trained regularity identification model that is trained to output the regularity when the state trend is input.

39. The aforementioned sensor is a load sensor placed on the bed, The physical condition change prediction model according to any one of claims 31 to 33, wherein the biological information is the weight of the subject used for the training data.

40. The load sensors consist of four sensors, each positioned at one of the four corners of the bed. The physical condition change prediction model according to claim 39, wherein each of the four sensors includes a strain generating body that generates strain in accordance with the load of a subject for training data, and a strain sensor attached to the strain generating body.

41. A non-temporary and computer-readable recording medium that records a physical condition change prediction model according to any one of claims 31 to 33.

42. The sensors acquire the subject's biometric information while they are in bed, Determining the supine state of the subject based on the subject's biological information, To obtain the state trend, which is the trend per unit time of the aforementioned lying-down state, A method for creating a regularity identification model, comprising: creating a regularity identification model that identifies a predetermined regularity included in the temporal variation of the state trend by machine learning using the state trend as training data.

Images