Model generation device, model generation method, and program

Abstract

The goal is to enable the training of a model for estimating health status using data representing the health status of individuals belonging to a specific group at a single point in time, and by incorporating characteristics that can be considered as factors related to the temporal changes in health status values ​​into the model's training. [Solution] Using two datasets that have a temporal relationship as a dataset of health status data values, the starting point is taken from one of the two datasets and the ending point is taken from the other dataset. The mean of the Gaussian distribution that the health status data values ​​follow during the temporal change from the starting point to the ending point is set as a continuous function that weights the health status data values ​​such that the higher the data density estimated by kernel density estimation, the greater the weight of the health status data values. The model is then trained using conditional flow matching.

Description

Technical Field

[0001] The present invention relates to a model generation device, a model generation method, and a program.

Background Art

[0002] It is conceivable to use time-series data of an individual's health state for learning a model for estimating a health state, such as aggregating the time-series data of an individual's health state to predict the health state. For example, Patent Document 1 describes a method for predicting the future health rank of a prediction target person using a hidden Markov model for each age and gender learned using time-series data of medical examination results for 6 years for 5,000 people each by age and gender. In the method described in Patent Document 1, the hidden Markov model has 6 states each, and the 6 states are classified into 3 health ranks of "healthy", "attention required", and "careful examination required (onset)". And in the method described in Patent Document 1, the health rank of the prediction target person is predicted using the hidden Markov model selected according to the age and gender of the prediction target person and the time-series data belonging to the population of the same age and gender as the prediction target person.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Time-series data of an individual's health state may be difficult to collect, for example, because it takes time to collect the data and for reasons such as privacy. Data indicating the state of a person belonging to a specific group at each one point (that is, data that is not time-series data), such as measurement data in a single medical examination of a person of a specific age, may be easier to collect than time-series data of an individual's health state. It is preferable to train a model for estimating health status using data that shows the state of each individual belonging to a specific group at a single point in time. In this case, if characteristics related to the temporal changes in health status values ​​can be reflected in the model's training, it is expected that the model will be able to estimate health status with higher accuracy.

[0005] One example of the purpose of this disclosure is to provide a model generation device, a model generation method, and a program that can solve the problems described above. [Means for solving the problem]

[0006] According to a first aspect of this disclosure, the model generation device includes an input processing means for acquiring a first dataset, which is a dataset of first health status data values, which are health status data values ​​for each person belonging to a first group, and a second dataset, which is a dataset of second health status data values, which are health status data values ​​for each person belonging to a second group, which is a group that has a temporal relationship with the first group, and for each combination of the first health status data value and the second health status data value, the path of the temporal change of the mean of a Gaussian distribution, which is set as the probability distribution that the health status data value follows during the temporal change from the first health status data value to the second health status data value, by connecting the first health status data value and the second health status data value, and Carne The system includes a flow setting means that sets a continuous function for weighting health status data values ​​such that the higher the data density estimated by data density estimation, the greater the weight given to the health status data values, and sets a flow that shows the direction vector of the change in health status data values ​​over time, with the first health status data value as the starting point and the second health status data value as the ending point, so that the health status data values ​​follow a Gaussian distribution whose mean is shown by the continuous function, and an optimization means that optimizes the parameter values ​​of the model that calculates the direction vector of the change in health status data values ​​over time, so that for all combinations of the first health status data value and the second health status data value, the direction vector calculated using the model approaches the direction vector shown by the flow.

[0007] According to a second aspect of this disclosure, the model generation method involves a computer obtaining a first dataset, which is a dataset of first health status data values, which are health status data values ​​for each person belonging to a first group, and a second dataset, which is a dataset of second health status data values, which are health status data values ​​for each person belonging to a second group, which is a group that has a temporal relationship with the first group. For each combination of the first health status data values ​​and the second health status data values, the computer determines the path of the temporal change of the mean of a Gaussian distribution, which is set as the probability distribution that the health status data values ​​follow during the temporal change from the first health status data values ​​to the second health status data values, and the first health status data values ​​and the second health status data values. This includes connecting the two values, setting a continuous function that weights the health status data values ​​such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data values; setting a flow that shows the direction vector of the change in the health status data values ​​over time, with the first health status data value as the starting point and the second health status data value as the ending point, such that the health status data values ​​follow a Gaussian distribution whose mean is shown by the continuous function; and performing an optimization calculation of the parameter values ​​of the model that calculates the direction vector of the change in the health status data values ​​over time, such that the direction vector calculated using the model for all combinations of the first health status data value and the second health status data value approaches the direction vector shown by the flow.

[0008] According to a third aspect of this disclosure, the program obtains from a computer a first dataset, which is a dataset of first health status data values, which are health status data values ​​for each person belonging to a first group, and a second dataset, which is a dataset of second health status data values, which are health status data values ​​for each person belonging to a second group, which is a group that has a temporal relationship with the first group, and for each combination of the first health status data value and the second health status data value, the program calculates the path of the temporal change of the mean of a Gaussian distribution, which is set as the probability distribution that the health status data value follows during the temporal change from the first health status data value to the second health status data value, by connecting the first health status data value and the second health status data value. This program performs the following actions: setting a continuous function for weighting health status data values ​​so that the higher the data density estimated by Nell density estimation, the greater the weight given to the health status data values; setting a flow that shows the direction vector of the change in health status data values ​​over time, with the first health status data value as the starting point and the second health status data value as the ending point, so that the health status data values ​​follow a Gaussian distribution whose mean is shown by the continuous function; and optimizing the parameter values ​​of the model that calculates the direction vector of the change in health status data values ​​over time, for all combinations of the first health status data value and the second health status data value, so that the direction vector calculated using the model approaches the direction vector shown in the flow. [Effects of the Invention]

[0009] According to one aspect of this disclosure, a model for estimating health status can be trained using data that shows the health status of each person belonging to a specific group at a single point in time, and in such a way that characteristics that are considered to be characteristics related to the change over time of values ​​related to health status are reflected in the model's training. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows an example of the configuration of a model generation device according to at least one embodiment. [Figure 2] This figure shows an example of data input and output in a model generation device according to at least one embodiment. [Figure 3] This figure shows an example of the processing procedure performed by a model generation device according to at least one embodiment. [Figure 4] This figure shows an example of the configuration of a model generation device according to at least one embodiment. [Figure 5] This figure shows an example of the processing steps in a model generation method according to at least one embodiment. [Figure 6] This figure shows an example of a computer configuration according to at least one embodiment. [Modes for carrying out the invention]

[0011] The embodiments will be described below with reference to the drawings.

[0012] <First Embodiment> Figure 1 shows an example of the configuration of a model generation device according to at least one embodiment. In the configuration shown in Figure 1, the model generation device 100 includes a communication unit 110, a display unit 120, an operation input unit 130, a storage unit 180, and a processing unit 190. The processing unit 190 includes an input processing unit 191, a flow setting unit 192, an optimization unit 193, and an output processing unit 194.

[0013] The model generation device 100 generates a model for estimating health status. For example, with the growing interest in healthcare, the model generation device 100 can be used to predict health status. The model generator 100 uses the first dataset and the second dataset to train a model that estimates the changes in health status values ​​over time. The first dataset is a dataset of first health status data values, which are values ​​related to the health status of each person belonging to the first group. The second dataset is a dataset of second health status data values, which are values ​​related to the health status of each person belonging to the second group, which is a group that has a temporal relationship with the first group. The data included in the first dataset is also referred to as first health status data. The data included in the second dataset is also referred to as second health status data. Data relating to health status is also called health status data. Values ​​relating to health status (values ​​of health status data) are also called health status data values. A model that estimates the changes in health status values ​​over time is also called a state estimation model. Health status data may also be vector data showing values ​​for multiple items related to health status.

[0014] In this context, "health status" refers to either a physical or mental state. In this context, the relationship between the first and second groups over time means that the health status data values ​​are associated with time, and that the first health status data values ​​and the second health status data values ​​are associated with different times. For example, the first and second groups may be groups of people from different age groups.

[0015] The second group may be associated with a later time than the first group. In this case, the model generation device 100 trains the state estimation model so that it predicts health status data values ​​that are further in the future than the health status data values ​​that are input to the state estimation model. For example, if the first group is a group of people in their 20s and the second group is a group of people in their 30s, the model generator 100 trains the state transition model so that the state estimation model receives input of health status data values ​​for people in their 20s and estimates the changes in health status data values ​​over time until those people reach their 30s.

[0016] Alternatively, the second group may be associated with a more past time period than the first group. In this case, the model generation device 100 trains the state estimation model so that it estimates health status data values ​​that are past the time period of the input health status data values. For example, if the first group is a group of people in their 30s and the second group is a group of people in their 20s, the model generator 100 trains the state transition model so that the state estimation model receives input of health status data values ​​for people in their 30s and estimates the changes in health status data values ​​over time, going back to their 20s.

[0017] Furthermore, shortening the time interval for each group may be used to improve the accuracy of the trained model. For example, the first pattern involves using a dataset of health status data for people aged 20 to under 30 as the first dataset, and a dataset of health status data for people aged 30 to under 40 as the second dataset. The second pattern involves using a dataset of health status data for people aged 20 to under 25 as the first dataset, and a dataset of health status data for people aged 30 to under 35 as the second dataset. The second pattern has less temporal variability within the dataset than the first pattern, and as a result, it is expected that the trained model can estimate the changes in health status data values ​​over time with relatively high accuracy.

[0018] The health status data used by the model generation device 100 to train the state estimation model does not need to be time-series data. The people belonging to the first group and the people belonging to the second group may be different people. Time-series data on an individual's health status can be difficult to collect, for example, due to the time required for data collection and privacy concerns. Data showing the status of individuals belonging to a specific group at a single point in time (i.e., data that is not time-series data) may be easier to collect than time-series data on an individual's health status. According to the model generation device 100, it is expected that the data used to train the state estimation model will be relatively easy to obtain. A dataset of time-series data can also be called a longitudinal dataset. A dataset of data associated with the same time period can also be called a cross-sectional dataset.

[0019] In this context, model learning refers to adjusting the model's parameter values. Model learning can also be called model training. Model learning can be understood as optimizing the model's parameter values, or simply optimizing the model itself. Furthermore, model learning can also be understood as generating a pre-trained model. Learning could, for example, be machine learning. The model generation device 100 may be configured using a computer.

[0020] The communication unit 110 communicates with other devices. For example, the communication unit 110 may communicate with a database device that stores anonymized measurement data from periodic health checkups, categorized by the age group of the person being examined, and receive age-specific datasets.

[0021] The display unit 120 includes a display screen such as a liquid crystal panel or an LED (Light Emitting Diode) panel, and displays various images. For example, the display unit 120 may display information related to the learning of the state estimation model by the model generation device 100, such as the progress of the state estimation model's learning.

[0022] The operation input unit 130 includes, for example, input devices such as a keyboard and a mouse, and accepts user operations. For example, the operation input unit 130 may accept user operations that perform settings related to the learning of the state estimation model by the model generation device 100, such as setting the variance of the Gaussian distribution assumed to be the distribution of health status data values.

[0023] The memory unit 180 stores various types of data. For example, the memory unit 180 may store the state estimation model to be trained. Alternatively, the memory unit 180 may store various types of data acquired by the model generation device 100 that are used to train the state estimation model. The memory unit 180 is configured using the memory devices provided by the model generation device 100.

[0024] The processing unit 190 controls various parts of the model generation device 100 to perform various processes. The functions of the processing unit 190 are performed, for example, by the CPU (Central Processing Unit) of the model generation device 100 reading a program from the storage unit 180 and executing it.

[0025] The input processing unit 191 acquires various data used for training the state estimation model. For example, the input processing unit 191 may acquire a dataset of health status data from another device via the communication unit 110. Alternatively, the input processing unit 191 may acquire parameter values ​​specified by the user for training the state estimation model, such as the variance of a Gaussian distribution assumed to represent the distribution of values ​​related to health status, via the operation input unit 130. The input processing unit 191 is an example of an input processing means.

[0026] The flow setting unit 192 performs various settings for training the state estimation model. The flow setting unit 192 is an example of a flow processing means. The optimization unit 193 learns the state estimation model using the settings made by the flow setting unit 192. The optimization unit 193 is an example of an optimization means. The output processing unit 194 outputs the state estimation model that has been trained by the optimization unit 193.

[0027] The combination of the flow setting unit 192 and the optimization unit 193 may be used to learn the parameter values of the state estimation model by means of the Conditional Flow Matching (CFM) method. Conditional Flow Matching is a method for optimizing the model parameter values in Continuous Normalizing Flow (CNF) that models the change in distribution by a differential equation. Here, for the flow u: [0, 1] × R t →R d with an initial value of x ∈ R d the solution of the differential equation of formula (1) is denoted as φ t (x).

[0028]

Equation

[0029] R represents the real number space. d is an integer where d ≥ 1, and R d represents the d-dimensional real number space. t is a variable that takes values of 0 ≤ t ≤ 1. In the model generation device 100, t is treated as a variable indicating the progress of time. t is also referred to as time t. In Continuous Normalizing Flow, when the probability density functions p0 and p1 on R d are given, for X ~ p0, learning is performed on the model v t of the flow u t (x; θ) such that φ1(X) ~ p1 holds. θ represents the parameters of the model v t .

[0030] After learning, the model v t receives an input of x ∈ R d and approximately calculates (d / dt)φ t . That is, the learned model v t approximately calculates the differential coefficient value indicating the path from x0 ~ p0 to x1 ~ p1. In the model generation device 100, the model v tThe derivative value calculated by this process can be considered as the direction vector of the time-dependent change in health status data values. In the model generation device 100, model v t (x;θ) corresponds to the state estimation model. In the model generation device 100, point x represents the health status data value. The health status data is assumed to be represented by a d-dimensional real vector.

[0031] In conditional flow matching, the loss function L is given by equation (2). CFM The value of the parameter θ is optimized to minimize the value of (θ).

[0032]

number

[0033] In conditional flow matching, the flow u is determined when the start point x0 and end point x1 are fixed. t Consider (x|x0,x1), and flow u t (x|x0,x1) and model v t Squared error with (x;θ) ||u t (x|x0,x1)-v t (x;θ)|| 2 We consider minimizing ||·||. ||·|| represents the norm. Flow u when the starting point x0 and ending point x1 are fixed t (x|x0,x1) is also called a conditional flow.

[0034] E represents the expected value. The expression (x0,x1)~q shows that the combination of the starting point x0 and the ending point x1 follows a probability distribution q. Here, the probability distribution q is the probability distribution resulting from the combination of probability distributions p0 and p1. That is, the probability distribution q is the probability distribution of the Cartesian product space R d ×R d The distribution shown above is the first component of the direct product space (R d Assume that the marginal distribution of the first component is p0, and the marginal distribution of the second component is p1. p t(·|x0,x1) is defined as a case where the pair (x0,x1) is fixed, the initial value is x0, and the flow is u t The value of φ at time t, given that (·|x0,x1) is determined by equation (1). t The distribution is shown. x~p t (·|x0,x1) is a value of x that corresponds to the probability distribution p t This shows that the result is given by a random number following the pattern (x0, x1).

[0035] Here, the probability distribution p for 0 ≤ t ≤ 1 t Let (x0,x1) follow a Gaussian distribution. This can be expressed as shown in equation (3).

[0036]

number

[0037] Thus, assuming that point x follows a Gaussian distribution for 0 ≤ t ≤ 1, the conditional flow u t It is known that this is determined as shown in equation (4).

[0038]

number

[0039] σ t (x0,x1) is p t Each component of (|x0,x1) (p at each time t) t This shows the standard deviation of (x0,x1). σ' t (x0,x1) is σ t We will show the derivative of (x0,x1) with respect to t. μ t (x0,x1) is p t This shows the average of each component of (x0,x1). μ' t (x0, x1) is μ t We will show the derivative of (x0,x1) with respect to t. Average μ t(x0,x1) can be set as shown in equation (5), for example.

[0040]

number

[0041] Average μ t Given that (x0, x1) is set as in equation (5), the probability distribution p follows point x. t (x|x0,x1) can be expressed as shown in equation (6).

[0042]

number

[0043] Here, the probability distribution p follows point x. t If the (x|x0,x1) coordinate system can reflect the characteristics of the distribution of values ​​related to a person's health status (health status data values), it is expected that the state estimation model generated by the model generation device 100 will be able to estimate the health status with greater accuracy. Values ​​related to a person's health status have standard values ​​or standard ranges, and it is conceivable that data tends to concentrate around these standard values ​​or ranges. Therefore, it is conceivable that values ​​related to a person's health status tend to change through areas of high data density. In order to reflect in conditional flow matching that values ​​related to a person's health status tend to change as they pass through areas with high data density, the flow setting unit 192 sets the mean μ t Alternatively, (x0, x1) can be set as shown in equation (7).

[0044]

number

[0045] Here, taking the expected value can be understood as taking the sample mean of the values ​​in the expression within the brackets for x, which follows the probability distribution indicated by the subscript in E. K represents the kernel function. The subscripts "t||x0-x1||" and "(1-t)||x0-x1||" after K both indicate the bandwidth in kernel density estimation. Furthermore, the bandwidth h can be expressed as shown in equation (8).

[0046]

number

[0047] When using a Gaussian kernel (Radius Basis Function; RBF) as the kernel function, it can be expressed as shown in equation (9).

[0048]

number

[0049] E x~p0 [K t||x0-x1|| (x-x0) can be understood as an estimate of the probability density of point x, obtained by kernel density estimation that focuses on the area around the starting point x0, and the larger the value of t, the wider the range of focus. E x~p1 [K (1-t)||x0-x1|| (x-x1) can be understood as an estimate of the probability density of point x, obtained by kernel density estimation that focuses on the area around the endpoint x1, and the larger the value of 1-t, the larger the range of focus. Equation (7) shows that when point x moves from the starting point x0 to the ending point x1, near t=0 it passes through a region with high data density near the starting point x0, and near t=1 it passes through a region with high data density near the starting point x1, with an average μ t This can be interpreted as an expression that sets a value.

[0050] Alternatively, set the bandwidth to "t α ||x0-x1||”, (1-t) α It can also be written as "||x0-x1||". In this case, the mean μ t This is shown in equation (10).

[0051] [Number]

[0052] α is a real constant set as a parameter of the bandwidth. When 0 < t < 1, the larger the value of α, the narrower (smaller) the bandwidths “t α ||x0 - x1||” and “(1 - t) α ||x0 - x1||”. Therefore, it can be understood that the larger the value of α, the easier it is for the model generation device 100 to focus on the region close to the starting point x₀ and the ending point x₁, and the smaller the value of α, the easier it is for the model generation device 100 to focus on the region far from the starting point x₀ and the ending point x₁. For example, the operation input unit 130 may receive a user operation specifying the value of the parameter α, and the flow setting unit 192 may set the specified value as the parameter α of the average μ t . Alternatively, as the formula for setting the bandwidth, it is not limited to those shown in formula (7) or formula (10), and a formula represented by various functions f(t) satisfying f(0) = 0 can be used.

[0053] Variance σ t (x0, x1) may be set as in formula (11).

[0054] [Number]

[0055] σ on the right side is a constant. Alternatively, the variance σ t (x0, x1) may be set as in formula (12).

[0056] [Number]

[0057] The probability distribution p that point x follows t (x|x0,x1) is the mean μ t (x0,x1) and variance σ t Using (x0, x1), it can be shown as in equation (3) above.

[0058] Note that the mean μ shown in equation (10) t The derivative μ' of (x0,x1) with respect to t t (x0,x1) can be shown as in equation (13).

[0059]

number

[0060] The function f(t,y,g) is given by equation (14).

[0061]

number

[0062] The function "id" in equation (13) represents the identity map. It can be expressed as id(x) = x. The function "1" in equation (13) represents a constant function that maps any argument to the constant 1. It can be expressed as 1(x)=1. The derivative of the function f(t,y,g) with respect to t, f'(t,y,g), is given by equation (15).

[0063]

number

[0064] Also, σ t Let =σ√(t(1-t)), then σ' t / σ t This is shown in equation (16).

[0065]

number

[0066] Flow u t (x|x0,x1) can be expressed as shown in equation (17).

[0067]

number

[0068] The optimization unit 193 can optimize the model parameter θ using equation (17). The mean μ shown in equation (10) t When using (x0,x1), the derivative with respect to t is μ' t (x0, x1) can be calculated, and the flow u t (x|x0,x1) can be calculated. The optimization unit 193 calculates the flow u t The model parameter θ can be optimized using (x|x0,x1).

[0069] Figure 2 shows an example of data input and output in the model generation device 100. In the example shown in Figure 2, the input processing unit 191 acquires various data used to train the state estimation model, such as the dataset for the first group, the dataset for the second group, and the variance of the Gaussian distribution followed by point x. The input processing unit 191 then outputs the acquired data to the flow setting unit 192 and the optimization unit 193.

[0070] Regarding the acquisition of variance, for example, the operation input unit 130 may accept a user operation to select one of several variance options, such as equations (11) and (12). The input processing unit 191 may then output the selected variance to the flow setting unit 192 and the optimization unit 193.

[0071] Alternatively, the variance used by the model generation device 100 may be fixed to a specific value. In this case, the storage unit 180 may store the variance. The input processing unit 191 may then read the variance from the storage unit 180 and output it to the flow setting unit 192 and the optimization unit 193. Alternatively, the flow setting unit 192 may read the variance from the storage unit 180.

[0072] The flow setting unit 192 sets a conditional flow based on the setting of the mean and variance of the Gaussian distribution that point x follows. For example, the flow setting unit 192 sets the mean μ shown in equation (10). t The settings for (x0,x1) and the variance σ t Based on the setting = σ√(t(1-t)), conditional flow u t You may also set (x|x0,x1) as shown in equation (17). The setting of a conditional flow by the flow setting unit 192 can also be understood as defining a conditional flow. The flow setting unit 192 outputs the set conditional flow to the optimization unit 193.

[0073] The optimization unit 193 performs optimization calculations to optimize the parameter values ​​of the state estimation model using the conditional flow set by the flow setting unit 192. Specifically, the optimization unit 193 optimizes the parameter values ​​of the state estimation model so that, for each combination of the first health state data and the second health state data, the path to point x indicated by the state estimation model approaches the path to point x indicated by a conditional flow that starts with the first health state data value as the starting point x0 and ends with the second health state as the ending point x1. The optimization unit 193 repeatedly performs optimization calculations to optimize the parameter values ​​of the state estimation model for all combinations of the first health status data and the second health status data.

[0074] In the optimization calculation, the optimization unit 193 searches for parameter values ​​that will make the evaluation shown by the evaluation function in equation (2) as good as possible. Specifically, the optimization unit 193 searches for parameter values ​​that will make the evaluation shown by the evaluation function in equation (2) as good as possible.t Output of (x;θ) and conditional flow u t We search for a parameter θ that minimizes the integral of the squared error between the output (x|x0,x1) and the output, over the path from the starting point x0 to the ending point x1.

[0075] However, the optimization method used by the optimization unit 193 is not limited to a specific method. For example, the optimization unit 193 may perform optimization calculations using gradient methods such as the steepest descent method, but is not limited to this.

[0076] The optimization unit 193 outputs the trained state estimation model to the output processing unit 194. The output processing unit 194 outputs the trained state estimation model to the outside of the model generation device 100. For example, the output processing unit 194 may transmit the trained state estimation model to another device via the communication unit 110.

[0077] Alternatively, the model generation device 100 may, after training the state estimation model, use the trained state estimation model to estimate the changes in health status data values ​​over time. In this case, the model generation device 100 does not need to output the trained state estimation model externally. Therefore, the processing unit 190 may be configured without an output processing unit 194.

[0078] Figure 3 shows an example of the processing steps performed by the model generation device. In the process shown in Figure 3, the input processing unit 191 acquires various data used to train the state estimation model (step S101).

[0079] Next, the flow setting unit 192 sets the average and conditional flows (step S102). In setting the average, the flow setting unit 192 sets values ​​for the parameters of the expression that represents the average, such as parameter α in expression (10). If the expression that represents the average does not include parameters, the flow setting unit 192 may set the expression as is. For example, the flow setting unit 192 may read the expression that represents the average from the storage unit 180 and use the read expression as is for setting the conditional flow. In setting a conditional flow, the flow setting unit 192 sets an expression that represents the conditional flow, for example, expression (10), based on the set mean and variance.

[0080] Next, the optimization unit 193 starts a loop L11 that processes each combination of the first health status data value and the second health status data value (step S103). The first health status data value and the second health status data value that are processed in loop L11 are used as the start and end points of the conditional flow. The first health status data value and the second health status data value that are processed in loop L11 are also referred to as the fixed start point and fixed end point.

[0081] In the processing of loop L11, the optimization unit 193 performs optimization calculations for the parameter values ​​of the state estimation model so that the path of health state data values ​​shown by the state estimation model approaches as closely as possible the path shown by the conditional flow with fixed start and end points (step S104).

[0082] For example, the optimization unit 193 samples combinations of starting point x0 and ending point x1 according to the probability distribution q based on equation (2), and then the model v t (x;θ) and conditional flow u t The squared error with (·|x0,x1) is calculated. Then, the optimization unit 193 calculates the average of the squared errors when the starting point x0 and the ending point x1 are moved as the evaluation function value, and optimizes the model v such that the evaluation function value becomes smaller. t We search for the value of the parameter θ in (x;θ).

[0083] Next, the optimization unit 193 performs termination processing for loop L11. Specifically, the optimization unit 193 determines whether or not the processing of loop L11 has been performed for all combinations (all possible combinations) of the first health status data and the second health status data. If the optimization unit 193 determines that there are combinations that have not yet been processed in loop L11, it continues to process the unprocessed combinations in loop L11. On the other hand, if the optimization unit 193 determines that it has performed the processing of loop L11 for all combinations, it terminates loop L11.

[0084] After loop L11, the output processing unit 194 outputs the trained state estimation model (step S106). After step S106, the model generation device 100 completes the process shown in Figure 3.

[0085] As described above, the input processing unit 191 acquires the first dataset and the second dataset. The first dataset is a dataset of first health status data values, which are health status data values ​​for each person belonging to the first group. Health status data values ​​are values ​​related to health status. The second dataset is a dataset of second health status data values, which are health status data values ​​for each person belonging to the second group, which is a group that has a temporal relationship with the first group.

[0086] The flow setting unit 192 sets the mean of a Gaussian distribution for each combination of the first health status data value and the second health status data value, and sets the flow. Regarding the setting of the mean of the Gaussian distribution, the flow setting unit 192 sets the path of the time-dependent change of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time-dependent change from the first health status data value to the second health status data value. The flow setting unit 192 sets this path of time-dependent change of the mean of the Gaussian distribution as a continuous function that connects the first health status data value and the second health status data value, and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value. Regarding the flow settings, the flow setting unit 192 sets a flow that shows the direction vector of the change in health status data values ​​over time, with the first health status data value as the starting point and the second health status data value as the ending point, so that the health status data values ​​follow a Gaussian distribution whose mean is shown by the continuous function described above.

[0087] The optimization unit 193 performs optimization calculations for the parameter values ​​of the model that calculates the direction vector of the change in health status data values ​​over time. For all combinations of the first health status data value and the second health status data value, the optimization unit 193 performs optimization calculations so that the direction vector calculated using the model approaches the direction vector shown in the flow chart.

[0088] According to the model generation device 100, there is no need to use time-series data for model training, and in this respect, it is expected that the data used for model training can be obtained relatively easily. Furthermore, the model generation device 100 allows for the incorporation of a characteristic of the time-dependent changes in health status values, such as the tendency for changes to occur more easily through areas of high data density, into the model's learning process. In this respect, the model generation device 100 is expected to be able to estimate the time-dependent changes in human health values ​​with relatively high accuracy.

[0089] Here, one might consider incorporating the characteristics of each health-related item, such as height, weight, and blood glucose level, into the model's training as a way to reflect the changes in health-related values ​​over time. However, in this case, it would be necessary to change the formula to reflect the characteristics each time the health-related item changes, which would be burdensome for the person setting up the formula. Furthermore, for items whose characteristics are unknown, it would be impossible to reflect the characteristics of the changes in health-related values ​​over time into the model's training. Moreover, if there are many health-related items, incorporating the characteristics of each item into the formula would be particularly burdensome for the person setting up the formula, making it impractical.

[0090] In contrast, the model generation device 100 incorporates a common characteristic of health status items—that values ​​related to health status tend to change through areas of high data density—into the model's learning process. As a result, the model generation device 100 only needs to pre-incorporate characteristics that are considered to be the characteristics of the time-dependent changes in health status values ​​into its formulas; it does not need to incorporate item-specific characteristics into the formulas. In this respect, the model generation device 100 can relatively reduce the burden on the operator setting up the formulas.

[0091] Furthermore, the continuous function set as the path of the time-dependent change in the mean of the Gaussian distribution includes parameters that adjust the bandwidth in kernel density estimation. According to the model generation device 100, by adjusting the parameter values ​​to adjust the bandwidth, it is possible to adjust which parts of the data density are given more importance.

[0092] For example, if you want the model generation device 100 to prioritize the area around the starting point x0 and the area around the ending point x1, you can set a relatively large value for the parameter α in equation (10). On the other hand, if you want the model generation device 100 to also prioritize areas relatively far from the starting point x0 and the ending point x1, you can set a relatively small value for the parameter α in equation (10).

[0093] Furthermore, the model generation device 100 adjusts the model parameter values ​​using machine learning. According to the model generation device 100, known machine learning techniques can be used in part of the process of adjusting the values ​​of the model parameters. In this respect, it is expected that the design of the model generation device 100 can be carried out relatively easily.

[0094] <Second Embodiment> Figure 4 shows an example of the configuration of a model generation device according to at least one embodiment. In the configuration shown in Figure 4, the model generation device 610 comprises an input processing unit 611, a flow setting unit 612, and an optimization unit 613.

[0095] In this configuration, the input processing unit 611 acquires a first data set, which is a dataset of first health status data values, which are health status data values ​​for each person belonging to the first group, and a second data set, which is a dataset of second health status data values, which are health status data values ​​for each person belonging to the second group, which is a group that has a temporal relationship with the first group.

[0096] The flow setting unit 612 sets the path of the time-dependent change of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time-dependent change from the first health status data value to the second health status data value, for each combination of the first health status data value and the second health status data value. This path is set as a continuous function that connects the first health status data value and the second health status data value, and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value. The flow setting unit 612 sets the flow, which shows the direction vector of the time-dependent change of the health status data value, starting from the first health status data value and ending from the second health status data value, so that the health status data value follows a Gaussian distribution whose mean is shown by the aforementioned continuous function.

[0097] The optimization unit 613 optimizes the parameter values ​​of the model that calculates the direction vector of the change in health status data values ​​over time, so that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data value and the second health status data value. The input processing unit 611 is an example of an input processing means. The flow setting unit 612 is an example of a flow setting means. The optimization unit 613 is an example of an optimization means.

[0098] According to the model generation device 610, there is no need to use time-series data for model training, and in this respect, it is expected that the data used for model training can be obtained relatively easily. Furthermore, the model generation device 610 allows for the incorporation of a characteristic of the time-dependent changes in health status values, such as the tendency for changes to occur more easily through areas of high data density, into the model's learning process. In this respect, the model generation device 610 is expected to enable the learning-based model to estimate the time-dependent changes in human health values ​​with relatively high accuracy.

[0099] Here, one might consider incorporating the characteristics of each health-related item, such as height, weight, and blood glucose level, into the model's training as a way to reflect the changes in health-related values ​​over time. However, in this case, it would be necessary to change the formula to reflect the characteristics each time the health-related item changes, which would be burdensome for the person setting up the formula. Furthermore, for items whose characteristics are unknown, it would be impossible to reflect the characteristics of the changes in health-related values ​​over time into the model's training. Moreover, if there are many health-related items, incorporating the characteristics of each item into the formula would be particularly burdensome for the person setting up the formula, making it impractical.

[0100] In contrast, the model generation device 610 incorporates a common characteristic of health status items—that values ​​related to health status tend to change through areas of high data density—into the model's learning process. As a result, the model generation device 610 only needs to pre-incorporate characteristics that are considered to be the characteristics of the time-dependent changes in health status values ​​into its formulas; it does not need to incorporate item-specific characteristics into the formulas. In this respect, the model generation device 610 can relatively reduce the burden on the operator setting up the formulas.

[0101] <Third Embodiment> Figure 5 shows an example of the processing steps in a model generation method according to at least one embodiment. The processing shown in Figure 5 includes acquiring data (step S611), setting up the flow (step S612), and performing optimization calculations (step S613).

[0102] In acquiring data (step S611), the computer acquires a first dataset, which is a dataset of first health status data values, which are values ​​related to health status, for each person belonging to the first group, and a second dataset, which is a dataset of second health status data values, which are each person belonging to the second group, which is a group that has a temporal relationship with the first group.

[0103] In setting up the flow (step S612), the computer sets up a continuous function that connects the first and second health status data values, and assigns weights to the health status data values ​​such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data values. The flow shows the direction vector of the change in health status data values ​​over time, with the first health status data value as the starting point and the second health status data value as the ending point, so that the health status data values ​​follow a Gaussian distribution whose mean is shown by the aforementioned continuous function.

[0104] In the optimization calculation (step S613), the computer optimizes the parameter values ​​of the model that calculates the direction vector of the change over time of the health status data values, so that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data values ​​and the second health status data values.

[0105] According to the process shown in Figure 5, it is not necessary to use time-series data for model training, and in this respect, it is expected that the data used for model training can be obtained relatively easily. Furthermore, the process shown in Figure 5 allows us to reflect in the model's learning the characteristic that values ​​related to health status tend to change more easily when passing through areas with high data density, which is considered a characteristic of the time course of health status values. In this respect, the model obtained through the process shown in Figure 5 is expected to be able to estimate the time course of values ​​related to a person's health with relatively high accuracy.

[0106] Here, one might consider incorporating the characteristics of each health-related item, such as height, weight, and blood glucose level, into the model's training as a way to reflect the changes in health-related values ​​over time. However, in this case, it would be necessary to change the formula to reflect the characteristics each time the health-related item changes, which would be burdensome for the person setting up the formula. Furthermore, for items whose characteristics are unknown, it would be impossible to reflect the characteristics of the changes in health-related values ​​over time into the model's training. Moreover, if there are many health-related items, incorporating the characteristics of each item into the formula would be particularly burdensome for the person setting up the formula, making it impractical.

[0107] In contrast, the process shown in Figure 5 incorporates a common characteristic of health status items—that values ​​related to health status tend to change through areas of high data density—into the model's learning. As a result, in the process shown in Figure 5, it is sufficient to pre-incorporate characteristics that are considered to be characteristics of the time-dependent changes in health status values ​​into the formula, and there is no need to incorporate item-specific characteristics into the formula. In this respect, the process shown in Figure 5 can relatively reduce the burden on the person setting up the formula.

[0108] Figure 6 shows an example of a computer configuration according to at least one embodiment. In the configuration shown in Figure 6, the computer 700 comprises a CPU 710, a main memory 720, an auxiliary memory 730, an interface 740, and a non-volatile recording medium 750.

[0109] One or more of the above-described model generation devices 100 and 610, or parts thereof, may be implemented in the computer 700. In that case, the operation of each processing unit described above is stored in the auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, expands it in the main memory 720, and executes the above-described processing according to the program. The CPU 710 also allocates memory areas in the main memory 720 corresponding to each of the above-described memory units according to the program. Communication between each device and other devices is performed by the interface 740 having a communication function and communicating according to the control of the CPU 710. The interface 740 also has a port for the non-volatile recording medium 750 and reads information from the non-volatile recording medium 750 and writes information to the non-volatile recording medium 750.

[0110] When the model generation device 100 is implemented in the computer 700, the operation of the processing unit 190 and each of its parts is stored in auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, expands it into the main memory 720, and executes the above processing according to the program.

[0111] Furthermore, the CPU 710 reserves a memory area for the memory unit 180 in the main memory 720 according to the program. Communication with other devices by the communication unit 110 is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Display of images by the display unit 120 is performed by the interface 740 having a display device and displaying various images under the control of the CPU 710. Acceptance of user operations by the operation input unit 130 is performed by the interface 740 having an input device and accepting user operations under the control of the CPU 710.

[0112] When the model generation device 610 is implemented in the computer 700, the operations of the input processing unit 611, the flow setting unit 612, and the optimization unit 613 are stored in auxiliary storage device 730 in the form of a program. The CPU 710 reads the program from the auxiliary storage device 730, expands it into the main memory 720, and executes the above processes according to the program.

[0113] Furthermore, the CPU 710 reserves memory in the main memory 720 for the model generation device 610 to process according to the program. Communication between the model generation device 610 and other devices is performed by the interface 740 having a communication function and operating under the control of the CPU 710. Interaction between the model generation device 610 and the user is performed by the interface 740 having input and output devices, presenting information to the user via the output device and accepting user operations via the input device under the control of the CPU 710.

[0114] One or more of the above-mentioned programs may be recorded on the non-volatile recording medium 750. In this case, the interface 740 may read the program from the non-volatile recording medium 750. The CPU 710 may then either directly execute the program read by the interface 740, or temporarily save it in the main memory 720 or auxiliary memory 730 before executing it.

[0115] Alternatively, a program for executing all or part of the processing performed by the model generation device 100 and the model generation device 610 may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be loaded into a computer system and executed to perform the processing of each part. The term "computer system" here includes hardware such as an operating system (OS) and peripheral devices. Furthermore, "computer-readable recording media" refers to portable media such as flexible disks, magneto-optical disks, ROMs (Read Only Memory), CD-ROMs (Compact Disc Read Only Memory), and storage devices such as hard disks built into computer systems. The above-mentioned program may be intended to implement only a part of the functions described above, and may also be able to implement the above-mentioned functions in combination with programs already recorded in the computer system.

[0116] Although the present disclosure has been described above with reference to embodiments, the present disclosure is not limited to the embodiments described above. Various modifications to the structure and details of the present disclosure can be made as understandable to those skilled in the art within the scope of the present disclosure. Furthermore, the embodiments described above may be combined with other embodiments as appropriate.

[0117] Some or all of the above embodiments may also be described as follows, but are not limited to these.

[0118] (Note 1) An input processing means for acquiring a first dataset, which is a dataset of first health status data values, which are values ​​related to health status, for each person belonging to the first group, and a second dataset, which is a dataset of second health status data values, which is a dataset of each person belonging to the second group, which is a group that has a temporal relationship with the first group. For each combination of the first health status data value and the second health status data value, the path of the time-dependent change of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time-dependent change from the first health status data value to the second health status data value, is set as a continuous function that connects the first health status data value and the second health status data value and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value; and the flow, which shows the direction vector of the time-dependent change of the health status data value with the first health status data value as the starting point and the second health status data value as the ending point, is set as a flow setting means that sets the health status data value so that its mean follows a Gaussian distribution shown by the continuous function; An optimization means for optimizing the parameter values ​​of a model that calculates the direction vector of the change over time of the health status data values, such that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data values ​​and the second health status data values, A model generation device equipped with the following features.

[0119] (Note 2) The continuous function includes a parameter that adjusts the bandwidth in the kernel density estimation. The model generation device described in Appendix 1.

[0120] (Note 3) The model generation device adjusts the values ​​of the model parameters using machine learning. A model generation device as described in Appendix 1 or Appendix 2.

[0121] (Note 4) Computers We obtain two datasets: a first dataset, which is a dataset of first health status data values, which are values ​​related to health status, and a second dataset, which is a dataset of second health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, and a second dataset, which is second health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are values ​​related to health status, and For each combination of the first health status data value and the second health status data value, the path of the time course of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time course from the first health status data value to the second health status data value, is set as a continuous function that connects the first health status data value and the second health status data value and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value. The flow showing the direction vector of the time course of the health status data value, with the first health status data value as the starting point and the second health status data value as the ending point, is set such that the health status data value follows a Gaussian distribution whose mean is shown by the continuous function. The optimization calculation of the parameter values ​​of the model that calculates the direction vector of the change over time of the health status data values ​​is performed so that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data values ​​and the second health status data values. A model generation method that includes the following.

[0122] (Note 5) The continuous function includes a parameter that adjusts the bandwidth in the kernel density estimation. The model generation method according to claim 2.

[0123] (Note 6) The computer adjusts the parameter values ​​of the model using machine learning. The model generation method described in Appendix 4 or Appendix 5.

[0124] (Note 7) On the computer, This involves obtaining two datasets: a first dataset, which is a dataset of first health status data values, consisting of health status data values ​​for each individual belonging to the first group, and a second dataset, which is a dataset of second health status data values, consisting of health status data values ​​for each individual belonging to the second group, which is a group that has a temporal relationship with the first group. For each combination of the first health status data value and the second health status data value, the path of the time-dependent change of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time-dependent change from the first health status data value to the second health status data value, is set as a continuous function that connects the first health status data value and the second health status data value and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value. The flow that shows the direction vector of the time-dependent change of the health status data value, with the first health status data value as the starting point and the second health status data value as the ending point, is set such that the health status data value follows a Gaussian distribution whose mean is shown by the continuous function. The optimization calculation of the parameter values ​​of the model that calculates the direction vector of the change over time of the health status data values ​​is performed so that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data values ​​and the second health status data values. A program that executes the command.

[0125] (Note 8) The continuous function includes a parameter that adjusts the bandwidth in the kernel density estimation. The program according to claim 5.

[0126] (Note 9) The computer is instructed to adjust the parameter values ​​of the model using machine learning. The program described in Appendix 7 or Appendix 8. [Explanation of Symbols]

[0127] 100, 610 Model Generator 110 Communications Department 120 Display section 130 Operation Input Section 180 Storage section 190 Processing Unit 191, 611 Input Processing Unit 192, 612 Flow setting section 193, 613 Optimization Unit 194 Output Processing Unit

Claims

1. An input processing means for acquiring a first dataset, which is a dataset of first health status data values, which are values ​​related to health status, and which is a dataset of first health status data values, which are the health status data values ​​for each person belonging to the first group, and a second dataset, which is a dataset of second health status data values, which are the health status data values ​​for each person belonging to the second group, which is a group that has a temporal relationship with the first group. For each combination of the first health status data value and the second health status data value, the path of the time-dependent change of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time-dependent change from the first health status data value to the second health status data value, is set as a continuous function that connects the first health status data value and the second health status data value and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value; and a flow setting means that sets the direction vector of the time-dependent change of the health status data value, with the first health status data value as the starting point and the second health status data value as the ending point, such that the health status data value follows a Gaussian distribution whose mean is shown by the continuous function. An optimization means for optimizing the parameter values ​​of a model that calculates the direction vector of the change in the health status data value over time, such that the direction vector calculated using the model for all combinations of the first health status data value and the second health status data value approaches the direction vector shown in the flow chart. A model generation device equipped with the following features.

2. The continuous function includes a parameter that adjusts the bandwidth in the kernel density estimation. The model generation apparatus according to claim 1.

3. The model generation device adjusts the values ​​of the model parameters using machine learning. A model generation apparatus according to claim 1 or claim 2.

4. Computers We obtain two datasets: a first dataset, which is a dataset of first health status data values, which are values ​​related to health status, and a second dataset, which is a dataset of second health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, and a second dataset, which is second health status data values, which are a dataset of first health status data values, which are a dataset of first health status data values, which are values ​​related to health status, and For each combination of the first health status data value and the second health status data value, the path of the time course of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time course from the first health status data value to the second health status data value, is set as a continuous function that connects the first health status data value and the second health status data value and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value. The flow showing the direction vector of the time course of the health status data value, with the first health status data value as the starting point and the second health status data value as the ending point, is set such that the health status data value follows a Gaussian distribution whose mean is shown by the continuous function. The optimization calculation of the parameter values ​​of the model that calculates the direction vector of the change over time of the health status data values ​​is performed so that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data values ​​and the second health status data values. A model generation method that includes the following.

5. On the computer, This involves obtaining a first dataset, which is a dataset of first health status data values, consisting of health status data values ​​for each person belonging to the first group, and a second dataset, which is a dataset of second health status data values, consisting of health status data values ​​for each person belonging to the second group, which is a group that has a temporal relationship with the first group. For each combination of the first health status data value and the second health status data value, the path of the time-dependent change of the mean of the Gaussian distribution, which is set as the probability distribution that the health status data value follows during the time-dependent change from the first health status data value to the second health status data value, is set as a continuous function that connects the first health status data value and the second health status data value and weights the health status data value such that the higher the data density estimated by kernel density estimation, the greater the weight given to the health status data value. The flow that shows the direction vector of the time-dependent change of the health status data value, with the first health status data value as the starting point and the second health status data value as the ending point, is set such that the health status data value follows a Gaussian distribution whose mean is shown by the continuous function. The optimization calculation of the parameter values ​​of the model that calculates the direction vector of the change over time of the health status data values ​​is performed so that the direction vector calculated using the model approaches the direction vector shown in the flow chart for all combinations of the first health status data values ​​and the second health status data values. A program that executes the command.

Images