State transition estimation device, state transition estimation method, and state transition estimation program

Abstract

We estimate long-term state changes by taking into account variability among individuals. [Solution] The state transition estimation device acquires a first feature vector of a first individual in a first age range, a second feature vector of a second individual in a second age range, and individual data of the first individual. From the individual data, it generates a latent vector representing the variability among the first individuals. Using the latent vector, it performs a mapping to the estimated feature vector when the first individual reaches the second age range. The mapping parameters are updated so that the mapping approaches the optimal transport from the probability distribution of the first feature vector to the probability distribution of the second feature vector, and the probability distribution of the estimated feature vector when an individual in the first age range reaches the second age range is estimated.

Description

Technical Field

[0001] The present disclosure relates to a state transition estimation device, a state transition estimation method, and a state transition estimation program.

Background Art

[0002] Predicting the long-term transition of disease risks and estimating the transition of health states are effective for designing future life plans. By predicting the long-term transition of health states and disease risks from health examination results and daily health activities, the prediction of lifetime costs can be made, and for example, it can be utilized for asset management support.

[0003] In recent years, a large amount of cross-sectional data has been accumulated, and data analysis has become possible even in the fields of healthcare and medicine. For example, Patent Document 1 describes a technique for constructing a health model in which the health state of a subject changes over time.

[0004] In addition, techniques for predicting long-term transitions are also useful in the fields of healthcare and medicine.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Non-Patent Documents

[0006]

Non-Patent Document 1

Summary of the Invention

[0007] However, it can be difficult to obtain sufficient longitudinal data with temporal connections. Based on our own knowledge, the inventors are considering estimating longitudinal data by using technologies such as optimal transport, and estimating the changes between distributions based on cross-sectional data collected for each age group.

[0008] One example of a technology related to optimal transport is the technology described in Non-Patent Document 1, which combines machine learning and optimal transport. According to the technology described in Non-Patent Document 1, a distribution with a probabilistic spread can be assigned to the destination data for a given source data point.

[0009] This allows us to, for example, probabilistically show how test values ​​change as a person ages, based on the test values ​​from a health checkup of a subject in a certain age group. At first glance, this seems to represent a situation where even if the test value is the same at a given point in time, it does not necessarily mean that the value will change to the same value due to individual variability. However, the technology described in Non-Patent Literature 1 generates a spread distribution by introducing noise, and therefore does not reflect actual individual variability.

[0010] This disclosure has been made in view of the above-mentioned problems, and one exemplary purpose is to provide a technique for estimating long-term state changes while taking into account variability among individuals. [Means for solving the problem]

[0011] A state transition estimation device relating to an exemplary aspect of this disclosure includes: a data input unit that acquires a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals; an individual variability generation unit that generates a latent vector representing the variability among the first individuals from the individual data of the first individuals; and a mapping from the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, using the latent vector of the first individual. The system includes: a data mapping unit that performs the following: an update unit that updates the mapping parameters in the data mapping unit based on the first feature vectors and estimated feature vectors of the plurality of first individuals, and the second feature vectors of the plurality of second individuals, so that the mapping in the data mapping unit approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors; and a probability distribution estimation unit that uses the data mapping unit whose mapping parameters have been updated by the update unit to estimate the probability distribution of the estimated feature vectors when an individual reaches the second age range, based on the first feature vectors of individuals in the first age range.

[0012] A state transition estimation method relating to an illustrative aspect of this disclosure includes: a data input process that acquires a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals; an individual variability generation process that generates a latent vector representing the variability among the first individuals from the individual data of the first individuals; and a mapping that uses the latent vector of the first individual to the estimated feature vector when the first individual reaches the second age range. The data mapping process includes: a data mapping process; an update process that updates the mapping parameters applied to the data mapping process so that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vector to the probability distribution of the second feature vector, based on the first feature vector and estimated feature vector of the plurality of first individuals, and the second feature vector of the plurality of second individuals; and a probability distribution estimation process that estimates the probability distribution of the estimated feature vector when an individual reaches the second age range, based on the first feature vector of an individual in the first age range, using the data mapping process whose mapping parameters have been updated by the update process.

[0013] A state transition estimation program relating to an exemplary aspect of this disclosure includes a data input process that obtains from a computer a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals; an individual variability generation process that generates a latent vector representing the variability among the first individuals from the individual data of the first individuals; and a mapping from the first feature vector of the first individual to an estimated feature vector when the first individual reaches the second age range, using the latent vector of the first individual. The system performs a data mapping process to create an image; an update process to update the mapping parameters applied to the data mapping process so that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vector to the probability distribution of the second feature vector, based on the first feature vector and estimated feature vector of the plurality of first individuals and the second feature vector of the plurality of second individuals; and a probability distribution estimation process to estimate the probability distribution of the estimated feature vector when an individual reaches the second age range, based on the first feature vector of an individual in the first age range, using the data mapping process whose mapping parameters have been updated by the update process. [Effects of the Invention]

[0014] One illustrative aspect of this disclosure provides the exemplary effect of offering a technique for estimating long-term state transitions while taking into account variability among individuals. [Brief explanation of the drawing]

[0015] [Figure 1] This is a block diagram showing the configuration of the state transition estimation device related to this disclosure. [Figure 2] This flowchart shows the flow of the state transition estimation method related to this disclosure. [Figure 3] This is a block diagram showing the configuration of the state transition estimation device related to this disclosure. [Figure 4]This figure shows an example of a forecast circle related to this disclosure. [Figure 5] This figure shows an example of a forecast circle related to this disclosure. [Figure 6] This flowchart shows the flow of the state transition estimation method related to this disclosure. [Figure 7] This figure shows an example of a forecast circle related to this disclosure. [Figure 8] This figure shows an example of a forecast circle related to this disclosure. [Figure 9] This is a block diagram showing the configuration of the life plan decision support device related to this disclosure. [Figure 10] This is a block diagram showing the configuration of a computer that functions as a state transition estimation device related to this disclosure. [Modes for carrying out the invention]

[0016] The following are examples of embodiments of the present invention. However, the present invention is not limited to the exemplary embodiments shown below, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining some or all of the technologies (things or methods) employed in each of the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, embodiments obtained by appropriately omitting some of the technologies employed in each of the exemplary embodiments shown below may also be included in the scope of the present invention. In addition, the effects mentioned in each of the exemplary embodiments shown below are examples of effects that can be expected in that exemplary embodiment and do not define the scope of the present invention. That is, embodiments that do not produce the effects mentioned in each of the exemplary embodiments shown below may also be included in the scope of the present invention.

[0017] [First Exemplary Embodiment] A first exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is the basic form for each of the exemplary embodiments described later. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur. Furthermore, each technology shown in the drawings referenced to explain this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur.

[0018] (Configuration of the state transition estimation device) The configuration of the state transition estimation device 100 will be explained with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the state transition estimation device 100. As shown in Figure 1, the state transition estimation device 100 includes a data input unit 110, an individual variation generation unit 120, a data mapping unit 130, an update unit 140, and a probability distribution estimation unit 150.

[0019] The data input unit 110, by accepting external input or by converting external input, obtains a first feature vector representing the state of each of multiple first individuals in a first age range, a second feature vector representing the state of each of multiple second individuals in a second age range that is older than the first age range, and individual data that is different from the first feature vector of each of the multiple first individuals.

[0020] The first and second individuals may be people or equipment such as machines. The first and second individuals do not need to be identical, but they may be partially identical. The first and second age ranges are not particularly limited, but for example, they may correspond to age groups in 10-year increments. For example, if the first and second individuals are people, the first age range may be in their 60s and the second age range may be in their 70s. If the first and second individuals are equipment, the first and second age ranges may be the equipment's age range.

[0021] The states indicated by the first and second feature vectors are not particularly limited, but may be two or more states, or even three or more states. If the first and second individuals are people, each state may be test data such as health checkup values ​​(HbA1c, etc.), disease risk indicators, or life log data (weight, blood pressure, steps, etc.). If the first and second individuals are equipment, each state may be test data such as periodic inspection values, output values, or power consumption.

[0022] Individual data is data that represents the attributes and state of each individual, which are different from the state indicated by the first and second feature vectors. While not particularly limited, if the first individual is a person, it may include attribute data such as gender and age, or it may include test values, disease risk indicators, life log data, etc., which are different from the state indicated by the first and second feature vectors. If the first individual is equipment, the individual data may include attribute data such as type and manufacturer, or it may include test value data such as periodic inspections, output values, power consumption, etc.

[0023] The individual variation generation unit 120 generates a latent vector representing the variation between first individuals from the individual data of the first individual. Representing the variation between first individuals can also be rephrased as showing the individuality of each first individual, meaning that it serves as an indicator of what characteristics each first individual possesses from a perspective other than the state shown by the first feature vector. The method for generating the latent vector is not particularly limited, but any method that converts individual data into a vector is acceptable.

[0024] The data mapping unit 130 uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range. In other words, the data mapping unit 130 maps from the feature space where the first feature vector is distributed to the feature space where the second feature vector is distributed.

[0025] The update unit 140 updates the mapping parameters in the data mapping unit 130 so that the mapping in the data mapping unit 130 approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors, based on the first feature vectors and estimated feature vectors of multiple first individuals, as well as the second feature vectors of multiple second individuals. For example, the update unit 140 approaches the optimal transport by repeatedly calculating the transport cost from the first feature vectors to the estimated feature vectors and updating the mapping parameters to reduce the transport cost.

[0026] The probability distribution estimation unit 150 uses the data mapping unit 130, whose mapping parameters have been updated by the update unit 140, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, based on the first feature vector of an individual in a first age range. This individual may be a different individual from the first individual, or it may be the first individual.

[0027] (Effects of the state transition estimation device) As described above, the state transition estimation device 100 employs a configuration in which a latent vector generated from the individual data of the first individual is used to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range. Therefore, the state transition estimation device 100 has the effect of being able to estimate long-term state transitions while taking into account the variability between individuals.

[0028] (Flowchart of the state transition estimation method) The flow of the state transition estimation method S10 will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the state transition estimation method S10. As shown in Figure 2, the state transition estimation method S10 includes data input processing S11, individual variability generation processing S12, data mapping processing S13, update processing S14, and probability distribution estimation processing S15.

[0029] The data input process S11 obtains a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data that is different from the first feature vector of each of the plurality of first individuals.

[0030] The individual variability generation process S12 generates a latent vector representing the variability among the first individuals from the individual data of the first individual.

[0031] The data mapping process S13 uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range.

[0032] The update process S14 updates the mapping parameters applied to the data mapping process S13, based on the first feature vectors and estimated feature vectors of multiple first individuals, as well as the second feature vectors of multiple second individuals, so that the mapping in the data mapping process S13 approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors.

[0033] The probability distribution estimation process S15 uses the data mapping process S13, whose mapping parameters have been updated by the update process S14, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, based on the first feature vector of an individual in a first age range.

[0034] (Effects of the state transition estimation method) As described above, the state transition estimation method S10 employs a configuration in which a latent vector generated from the individual data of the first individual is used to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range. Therefore, the state transition estimation method S10 has the effect of being able to estimate long-term state transitions while taking into account the variability between individuals.

[0035] [Second exemplary embodiment] A second exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiment are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise.

[0036] In the following, the first and second individuals are individuals, and the first and second feature vectors are described using examples that represent states related to health or disease risk, but this embodiment is not limited to these examples. Examples of states related to health or disease risk include test results data from health checkups (such as HbA1c), disease risk indicators, and life log data (such as weight, blood pressure, and steps taken).

[0037] (Configuration of the state transition estimation device) The configuration of the state transition estimation device 100A will be explained with reference to Figure 3. Figure 3 is a block diagram showing the configuration of the state transition estimation device 100A. The state transition estimation device 100A includes a data input unit 110A, an individual variation generation unit 120A, a data mapping unit 130A, an update unit 140A, a probability distribution estimation unit 150A, and a display unit 160.

[0038] The data input unit 110A acquires and stores the first feature vector 111, the second feature vector 112, and the individual data 113, respectively. The first feature vector 111 is a vector having components corresponding to multiple states of the first individual, and in the following description, x1, ..., x NThis is written as (N is the number of first individuals). The second feature vector 112 is a vector in which each of the second individual has components corresponding to multiple states of the second individual, and in the following explanation, y1, ..., y N This is written as (N is the number of the second individual). Note that the following explanation describes the case where both the number of the first individual and the number of the second individual are N, but this embodiment is not limited to this. Also, individual data 113 is data that shows a state or attribute different from the first feature vector 111 of the first individual, and in the following explanation, z1, ..., z N This is how it is written (N is the number of the first individual).

[0039] The individual variability generation unit 120A uses a mapping model (first machine learning model) 121, which is a machine learning model that takes individual data 113 as input and outputs a latent vector, to generate a latent vector w for each first individual from the individual data 113. The number of components of the latent vector w is, for example, the same as the number of components of the first feature vector 111. However, the number of components of the latent vector w may be different from the number of components of the first feature vector 111, provided that the calculations in the data mapping unit 130A described later are possible. The mapping model 121 can be composed of, for example, a neural network (NN) such as a multi-layered fully connected (FC) network.

[0040] The data mapping unit 130A performs mapping using an encoder model 131 and a decoder model 132 that constitute a Variational Auto Encoder (VAE) type machine learning model. The data mapping unit 130A inputs the first feature vector x of the first individual to the encoder model 131, and the mean μ and variance σ of the estimated feature vector y^ obtained by mapping are obtained. 2 It generates the following. Then, the data mapping unit 130A calculates the variance σ 2 Multiply by the latent vector w (calculate the Hadamard product), and find the mean μ and variance σ 2 The latent vector w is input to the decoder model 132, and the mapped estimated feature vector y^ is output.

[0041] As described above, in the present embodiment, the mean μ and variance σ are generated from the first feature vector x using a machine learning model. 2 On the other hand, in the technique described in Non-Patent Document 1, noise sampled from a normal distribution is added to the first feature vector x to generate a plurality of y^, thereby calculating the variance σ. 2 However, in the present embodiment, since the latent vector w is calculated from the individual data z, only one latent vector w is generated for one first feature vector x. Therefore, it is not possible to calculate the variance σ by generating a plurality of y^ as in Non-Patent Document 1. Thus, in the present embodiment, the mean μ and variance σ are directly generated from the first feature vector x using a machine learning model. 2 As described above, in the present embodiment, the mean μ and variance σ are generated from the first feature vector x using a machine learning model. 2 are generated.

[0042] The encoder model 131 and the decoder model 132 can be configured by a neural network (NN) such as a multi-layer fully connected layer (FC), for example.

[0043] The update unit 140A calculates the transport cost from the first feature vector x, the second feature vector y, and the estimated feature vector y^, and performs an update so as to reduce the transport cost. As the transport cost, the L2 distance may be used as in general optimal transport, or other distances may be used. Also, in order to obtain a distributed distribution, the constraints of optimal transport may be relaxed by adding a regularization term. As the regularization term, an entropy term may be added as a general method, or a variance term may be added. Hereinafter, the case using the variance term will be described. As the variance, the variance σ generated in the data mapping unit 130A is used. 2 is used.

[0044] In one example, the parameters (mapping parameters) of the mapping model 121, encoder model 131, and decoder model 132 are updated to reduce the transport loss, which is a loss function defined to minimize transport costs. This allows the parameters (mapping parameters) of the mapping model 121, encoder model 131, and decoder model 132 to be updated so that the mapping in the data mapping unit 130A approaches the optimal transport from the probability distribution of the first feature vector x to the probability distribution of the second feature vector y.

[0045] In one example, the update unit 140A may further calculate one or more loss functions, such as VAE loss, variance loss, and mean loss, and update the parameters (mapping parameters) of the mapping model 121, encoder model 131, and decoder model 132 to minimize these loss functions.

[0046] The VAE loss is a common VAE function used to minimize the squared error between the VAE's input and output. The variance loss is the variance σ output by encoder model 131. 2 This is a loss function that minimizes the squared error between the mean and the variance of the estimated feature vector y^ output by decoder model 132 across all data. The mean loss is a loss function that minimizes the squared error between the mean μ output by encoder model 131 and the estimated feature vector y^ output by decoder model 132. Note that the condition may be relaxed so that there is no penalty if the estimated feature vector y^ output by decoder model 132 is within 1σ of the mean μ output by encoder model 131.

[0047] The update unit 140A may further calculate individual losses and update the parameters of the individualized model 151 to minimize the loss function. Details will be described later.

[0048] Next, the probability distribution estimation unit 150A will be described. As shown in Figure 4, the probability distribution estimation unit 150A uses the mean μ and variance σ generated by the data mapping unit 130A. 2A first probability distribution P(y|x) of the Gaussian type is estimated from this. The first probability distribution represents the range of the estimated feature vector y^ with a variability of 1σ when the individuality of the first individual is not considered. In this specification, this range is also called the forecast circle C1.

[0049] As shown in the upper part of Figure 5, the probability distribution estimation unit 150A also calculates the variance σ that reflects the individuality of the first individual based on the personal data z. p 2 A second probability distribution P(y|z,z) having the following characteristics may be estimated. The second probability distribution exhibits a more limited range (forecast circle C2) than the first probability distribution.

[0050] First, the probability distribution estimation unit 150A generates the difference vector μ-y^ from the individual data z using an individualization model 151, which is a machine learning model that takes individual data z as input and outputs a difference vector μ-y^. The difference vector μ-y^ is the difference vector between the estimated feature vector y^ obtained by the data mapping unit 130A and the mean μ of the estimated feature vector y^.

[0051] The personalized model 151 may consist of a neural network (NN), such as a multi-layered fully connected (FC) network. The update unit 140A can machine-learn the personalized model 151 by updating its parameters to minimize the squared error between the output of the personalized model 151 and μ-y^.

[0052] As shown in the upper part of Figure 5, if the first feature vector x is different (in the upper part of Figure 5, x and x' are shown as first feature vectors that are different from each other), the size of the 1σ forecast circle C1 changes. Therefore, as shown in the lower part of Figure 5, the update unit 140A can absorb the differences and perform machine learning by normalizing so that the radius is 1.

[0053] Then, the probability distribution estimation unit 150A uses the machine learning-developed individualized model 151 to calculate the variance σ based on the squared error of the difference vector μ-y^ generated from the individual data z, according to the ratio of the radius σ of the prediction circle. p2 Calculate.

[0054] Here the variance σ p 2 This is defined as the prediction variance, which is the average of the squared errors obtained during normalized learning, being normalized again by multiplying it by the radius σ of the forecast circle.

[0055] As a result, the probability distribution estimation unit 150A uses the mapped estimated feature vector y^ as the mean, and σ p 2 A second probability distribution P(y|z,z) of the Gaussian type with variance can be estimated. The second probability distribution represents the range of the estimated feature vector y^ that reflects the individuality of the first individual. In this specification, this range is also called the forecast circle C2.

[0056] The display unit 160 is comprised of a display device and the like. The display unit 160 displays a forecast circle C1 showing the first probability distribution estimated by the probability distribution estimation unit 150A and a forecast circle C2 showing the second probability distribution.

[0057] (Flowchart of the state transition estimation method) The flow of the state transition estimation method S10A will be explained with reference to Figure 6. Figure 6 is a flowchart showing the flow of the state transition estimation method S10A. As shown in Figure 6, the state transition estimation method S10A includes data input processing S11A, individual variability generation processing S12A, data mapping processing S13A, update processing S14A, probability distribution estimation processing S15A, and display processing S16.

[0058] In step S11A, the data input unit 110A receives the first feature vector x1, ..., x N Obtain the first feature vector x k The data (1≦k≦N) is sequentially provided to the data mapping unit 130A and the update unit 140A. The data input unit 110A also provides the second feature vectors y1, ..., y N Obtain the second feature vector y k (1≦k≦N) is provided to the update unit 140A. The data input unit 110A receives individual data z1, ..., zN Obtain the individual data z k (1≦k≦N) is sequentially supplied to the individual variation generation unit 120A.

[0059] In step S12A, the individual variation generation unit 120A uses the mapping model 121 to process the individual data z provided from the data input unit 110A. k A latent vector w is generated from this and provided to the data mapping unit 130A.

[0060] In step 13A, the data mapping unit 130A receives the first feature vector x from the data input unit 110A. k The average μ is input to encoder model 131. φ and variance σ φ 2 The data mapping unit 130A generates the variance σ. φ 2 The average μ is obtained by multiplying it by the w provided from the individual variation generation unit 120A (calculating the Hadamard product). φ By adding these, we obtain the following value (1).

[0061]

number

[0062] The data mapping unit 130A inputs the obtained values ​​into the decoder model 132 and estimates the feature vector y^ k The data mapping unit 130A calculates the average μ. φ , variance σ φ 2 and estimated feature vector y^ k Provide to the updated section 140A.

[0063] In step S14A, the update unit 140A calculates the transport loss, VAE loss, variance loss, mean loss, and individual loss, and updates the parameters of each model.

[0064] First, let's explain the transport loss. The update unit 140A receives the first feature vector x from the data input unit 110A.k And the estimated feature vector y^ provided by the data mapping unit 130A k and variance σ φ 2 From this, we calculate the following value (2).

[0065]

number

[0066] The update unit 140A calculates this for k=1 to N and averages it to obtain the following value (3).

[0067]

number

[0068] Furthermore, the update unit 140A receives the estimated feature vector y^ from the data mapping unit 130A. k Input this into the function f, and obtain the resulting f(y^ k The following values ​​(4) are obtained by calculating the values ​​for k=1 to N and averaging them.

[0069]

number

[0070] Then, the update unit 140A calculates the following transport loss L1 from values ​​(3) and (4).

[0071]

number

[0072] Furthermore, the update unit 140A receives the second feature vector y from the data input unit 110A. k Input this into the function f, and obtain the resulting f(y k The following values ​​(5) are obtained by calculating () for each of k=1 to N and averaging them.

[0073]

number

[0074] Then, the update unit 140A calculates the following transport loss L2 from values ​​(4) and (5).

[0075]

number

[0076] The update unit 140A updates the parameters (mapping parameters) of the mapping model, encoder model, and decoder model so that the transport loss L1 decreases and the transport loss L2 increases.

[0077] Next, we will explain the VAE loss. The update unit 140A uses the second feature vector y provided from the data input unit 110A as the VAE loss. k and the estimated feature vector y^ provided from the data mapping unit 130A k The squared error is calculated, and the parameters (mapping parameters) of the mapping model, encoder model, and decoder model are updated to minimize the VAE loss.

[0078] Next, let's explain the variance loss. The update unit 140A uses the variance σ provided by the data mapping unit 130A over k=1 to N as the variance loss. 2 The squared error between the mean and the variance of the estimated feature vector y^ provided by the data mapping unit 130A is calculated, and the parameters (mapping parameters) of the mapping model, encoder model, and decoder model are updated to minimize the variance loss.

[0079] Next, let's explain the average loss. The update unit 140A uses the average μ provided by the data mapping unit 130A over the range k=1 to N as the average loss. φ The mean and the estimated feature vector y^ provided by the data mapping unit 130A. k The squared error with respect to the mean is calculated. Note that the estimated feature vector y^ k is the mean μ φfrom 1σ φ The conditions may be relaxed so that there is no penalty if it falls within that limit.

[0080] Next, we will explain the individual loss. The update unit 140A calculates the individual loss using the estimated value of the difference vector μ-y^ generated by the probability distribution estimation unit 150A using the individualization model 151, and the actual mean μ φ and estimated feature vector y^ k The squared error between the difference vector and the result is calculated, and the parameters of the individualized model 151 are updated to minimize the individual loss.

[0081] As a result, the update unit 140A updates the parameters of each model (trains each model). In one example, each model may be optimized by repeating steps S12A to S14A.

[0082] In step S15A, the probability distribution estimation unit 150A receives the mean μ from the data mapping unit 130A. φ and variance σ φ 2 A first probability distribution P(y|x) of the Gaussian type is estimated from this. In step S16, the display unit 160 shows the range 1σ corresponding to the first probability distribution. φ The forecast circle C1 is shown in the diagram.

[0083] Furthermore, in step S15A, the probability distribution estimation unit 150A uses the individualized model 151 to estimate the individual data z k and the first feature vector x k From the variance σ that reflects individuality p 2 Calculate the estimated feature vector y^ k σ p 2 A second probability distribution P(y|x,z) of the Gaussian type with variance is estimated. In step S16, the display unit 160 displays the range 11σ corresponding to the second probability distribution. p A more limited forecast circle C2 is shown.

[0084] [Third Exemplary Embodiment] A third exemplary embodiment, which is an example of an embodiment of the present invention, will now be described in detail. Components having the same function as those described in the above-described exemplary embodiments will be denoted by the same reference numerals, and their descriptions will be omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs.

[0085] In this embodiment, at least some of the first individuals correspond to at least some of the second individuals. The correspondence between the first and second individuals means that the second individual is an older version of the first individual, or, if the first and second individuals are individuals, that they are the same individual.

[0086] At this time, the update unit 140A calculates the transport loss, VAE loss, variance loss, mean loss, and individual loss using the estimated feature vector y^, and in addition to calculating the various losses using the estimated feature vector y^, it also uses the second feature vector y of the second individual corresponding to the first individual, y ~ By calculating an additional loss function using this method, the accuracy can be further improved.

[0087] Furthermore, the update unit 140A maps the estimated feature vector y^ obtained from the first feature vector x of the first individual for which a corresponding second individual exists, and the second feature vector y of the second individual corresponding to the first individual. ~ The parameters (mapping parameters) of the mapping model 121, encoder model 131, and decoder model 132 may be updated to bring them closer together.

[0088] For example, the update unit 140A considers the mapping loss as the estimated feature vector y^ that the data mapping unit 130A mapped from the first feature vector x of the first individual to which the corresponding second individual exists, and the second feature vector y of the second individual corresponding to the first individual. ~ The difference (for example, the squared error) may be calculated, and the parameters (mapping parameters) of the mapping model 121, encoder model 131, and decoder model 132 may be updated to minimize the mapping loss.

[0089] For example, the update unit 140A maps the estimated feature vector y^ obtained by the data mapping unit 130A from the first feature vector x of the first individual to which the corresponding second individual exists, and the second feature vector y of the second individual corresponding to the first individual. ~ The update unit 140A may have a machine learning model that learns the relationship between the two. The update unit 140A may then calculate an R2 loss, which is a loss function for reducing the R2 score of the machine learning model, and update the parameters (mapping parameters) of the mapping model 121, encoder model 131, and decoder model 132 so that the R2 loss is reduced.

[0090] [Fourth exemplary embodiment] A fourth exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiments are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs.

[0091] The state transition prediction device 100A is not limited to one stage, but may predict two or more stages. For example, if the first individual is in their 60s, feature vectors for three or more stages, such as their 70s and 80s, may be estimated. Note that it is not limited to 10-year units. The first feature vector of the first individual is denoted as x, the second feature vector of the second individual as y1, and the third feature vector of the third individual as y2.

[0092] The probability distribution estimation unit 150A first calculates the mean μ1 and variance σ1 of the estimated feature vector for the age range corresponding to the second individual, using the first feature vector x of the first individual. 2 The probability distribution P(y1|x) is estimated. Furthermore, the probability distribution estimation unit 150A calculates the mean μ2 and variance σ2 of the estimated feature vector in the age range corresponding to the third individual from the second feature vector y1 of the second individual. 2 And estimate the probability distribution P(y2|y1).

[0093] As shown in Figure 7, the probability distribution estimation unit 150A calculates the mean μ1 and variance σ1 when x is given to the probability distribution P(y1|x). 2 The display unit 160 may then calculate the mean μ2 and variance σ2 when μ1 is given to the probability distribution P(y2|y1). 2 The display unit 160 may then calculate the value and display the corresponding forecast circle. As in the example above, if x is the value at age 60, then μ1 can be interpreted as a trend forecast for 10 years later, and μ2 as a trend forecast for 20 years later.

[0094] The probability distribution estimation unit 150A further calculates, when the personal data z includes daily changing life log data such as the average number of steps per week or average weight, first using the individual data z collected at the same time as the first feature vector x, the mean is y^1 and the variance is σ p 2 The probability distribution P(y1|x,z) is estimated. The display unit 160 shows y^1 and σ p 2The forecast circle is displayed using [this method]. Furthermore, the probability distribution estimation unit 150A estimates P(y2|y1) and calculates the mean μ2 and variance σ2 when y^1 is given. 2 The display unit 160 calculates the value and displays the corresponding forecast circle.

[0095] Furthermore, the probability distribution estimation unit 150A adjusts the mean y^1(z') and variance σ in response to changes in the individual data z. p 2 Given (z') and y^1(z'), the mean μ'² and variance σ 2’ 2 The display unit 160 calculates the value and displays the corresponding forecast circle.

[0096] At this time, the display unit 160 displays the change in the forecast circle, such as by animating the new forecast circle, so that the change from the previous state can be seen. In this way, if the individual data includes data that can change on the time axis, the probability distribution estimation unit 150A may re-estimate the probability distribution of the estimated feature vector in accordance with the change in the individual data, and the display unit 160 may display the change in the probability distribution of the estimated feature vector estimated by the probability distribution estimation unit 150A. By performing timing detection and updates automatically and in real time, it is expected that the effect of encouraging health behaviors related to the life log can be achieved.

[0097] [Fifth Exemplary Embodiment] A fourth exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiments are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical hindrance occurs.

[0098] As shown in Figure 9, the life plan decision support device 1 comprises a state transition estimation device 100, an input unit 200, and a presentation unit 300. When a user inputs their health or disease risk-related states and other personal data via the input unit 200, the state transition estimation device 100 predicts the user's future health or disease risk-related states, and the presentation unit 300 may present the user with asset management suggestions based on lifetime cost predictions. This can support the user in making decisions about their life plan.

[0099] [Examples of implementation using software] Some or all of the functions of the state transition estimation devices 100 and 100A (hereinafter also referred to as "the above devices") may be implemented by hardware such as integrated circuits (IC chips) or by software.

[0100] In the latter case, each of the above devices is implemented, for example, by a computer that executes instructions for a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in Figure 10. Figure 10 is a block diagram showing the hardware configuration of computer C, which functions as each of the above devices.

[0101] Computer C comprises at least one processor C1 and at least one memory C2. Memory C2 stores a program P that causes computer C to operate as each of the above-mentioned devices. In computer C, processor C1 reads program P from memory C2 and executes it, thereby realizing each of the above-mentioned devices.

[0102] For processor C1, for example, a CPU (Central Processing Unit), GPU (Graphic Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof can be used. For memory C2, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination thereof can be used.

[0103] Computer C may also be equipped with RAM (Random Access Memory) for loading program P at runtime and for temporarily storing various data. Furthermore, computer C may be equipped with communication interfaces for sending and receiving data with other devices. Additionally, computer C may be equipped with input / output interfaces for connecting input / output devices such as keyboards, mice, displays, and printers.

[0104] Furthermore, program P can be recorded on a non-temporary, tangible recording medium M that is readable by computer C. Such a recording medium M could be, for example, tape, disk, card, semiconductor memory, or programmable logic circuitry. Computer C can acquire program P via such a recording medium M. Program P can also be transmitted via a transmission medium. Such a transmission medium could be, for example, a communication network or broadcast waves. Computer C can also acquire program P via such a transmission medium.

[0105] Furthermore, each of the above functions of each of the above devices may be implemented by a single processor in a single computer, by multiple processors in a single computer working together, or by multiple processors in each of multiple computers working together. In addition, the programs for implementing each of the above functions in each of the above devices may be stored in a single memory in a single computer, distributed and stored in multiple memories in a single computer, or distributed and stored in multiple memories in each of multiple computers.

[0106] [Additional Note 1] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0107] (Note 1) A data input unit that acquires a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals. An individual variation generation unit generates a latent vector representing the variation between the first individuals from the individual data of the first individual, A data mapping unit that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update unit updates the mapping parameters in the data mapping unit so that the mapping in the data mapping unit approaches the optimal transport from the probability distribution of the first feature vector to the probability distribution of the second feature vector, based on the first feature vector and estimated feature vector of the plurality of first individuals, and the second feature vector of the plurality of second individuals. A state transition estimation device comprising: a probability distribution estimation unit that uses the data mapping unit, whose mapping parameters have been updated by the update unit, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, from the first feature vector of an individual in a first age range.

[0108] (Note 2) The state transition estimation device described in Appendix 1, wherein the data mapping unit uses a Variational Auto Encoder (VAE) type machine learning model to output the mean and variance of the estimated feature vector obtained by the mapping.

[0109] (Note 3) The state transition estimation device described in Appendix 2, wherein the probability distribution estimation unit takes the individual data as input and uses a machine learning model that outputs a difference vector between the estimated feature vector obtained by the data mapping unit and the mean of the estimated feature vector to estimate the probability distribution of the estimated feature vector when an individual reaches the second age range, based on the individual data and first feature vector of an individual in the first age range.

[0110] (Note 4) The probability distribution estimation unit estimates a first probability distribution of the estimated feature vector from the first feature vector of individuals within the first age range, and estimates a second probability distribution of the estimated feature vector from the individual data and the first feature vector of individuals within the first age range. The state transition estimation device according to Appendix 3, further comprising a display unit for displaying the first probability distribution and the second probability distribution.

[0111] (Note 5) At least some of the first individuals and at least some of the second individuals correspond to each other. The state transition estimation device according to Appendix 1, wherein the update unit updates the mapping parameters in the data mapping unit so that the estimated feature vector mapped by the data mapping unit from the first feature vector of at least some of the first individuals approaches the second feature vector of at least some of the second individuals.

[0112] (Note 6) The aforementioned individual data includes data that may change over time. The probability distribution estimation unit re-estimates the probability distribution of the estimated feature vector in response to the change in the individual data. The state transition estimation device according to Appendix 1, further comprising a display unit that displays the change in the probability distribution of the estimated feature vector estimated by the probability distribution estimation unit.

[0113] (Note 7) The first individual, the second individual, and the first individual are individuals, The first feature vector, the second feature vector, and the estimated feature vector indicate a state related to health or disease risk, as described in any one of the appendices 1 to 7 of the state transition estimation device.

[0114] (Note 8) A life plan decision support device equipped with a state transition estimation device as described in Appendix 7.

[0115] (Note 9) A data input process that acquires a first feature vector representing the state of each of multiple first individuals in a first age range, a second feature vector representing the state of each of multiple second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector for each of the multiple first individuals. A method for generating individual variability from the individual data of the first individual to generate a latent vector representing the variability between the first individuals, A data mapping process that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update process updates the mapping parameters applied to the data mapping process, based on the first feature vectors and estimated feature vectors of the plurality of first individuals, and the second feature vectors of the plurality of second individuals, such that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors. A state transition estimation method comprising: a probability distribution estimation process that uses the data mapping process, in which the mapping parameters have been updated by the update process, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, from the first feature vector of an individual in a first age range.

[0116] (Note 10) On the computer, A data input process that acquires a first feature vector representing the state of each of multiple first individuals in a first age range, a second feature vector representing the state of each of multiple second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector for each of the multiple first individuals. A method for generating individual variability from the individual data of the first individual to generate a latent vector representing the variability between the first individuals, A data mapping process that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update process updates the mapping parameters applied to the data mapping process, based on the first feature vectors and estimated feature vectors of the plurality of first individuals, and the second feature vectors of the plurality of second individuals, such that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors. A state transition estimation program that uses the data mapping process, in which the mapping parameters have been updated by the update process, to perform a probability distribution estimation process that estimates the probability distribution of the estimated feature vector when an individual reaches a second age range, from the first feature vector of an individual in a first age range.

[0117] [Additional Note 2] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0118] (Note 1) It comprises at least one processor, and the at least one processor is A data input process that acquires a first feature vector representing the state of each of multiple first individuals in a first age range, a second feature vector representing the state of each of multiple second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector for each of the multiple first individuals. A method for generating individual variability from the individual data of the first individual to generate a latent vector representing the variability between the first individuals, A data mapping process that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update process updates the mapping parameters applied to the data mapping process, based on the first feature vectors and estimated feature vectors of the plurality of first individuals, and the second feature vectors of the plurality of second individuals, such that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors. A state transition estimation device that performs a probability distribution estimation process using the data mapping process in which the mapping parameters have been updated by the update process, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, from the first feature vector of an individual in a first age range.

[0119] The state transition estimation device may further include a memory. The memory may also store a program for causing at least one processor to execute each of the aforementioned processes. [Explanation of Symbols]

[0120] 100, 100A ···State transition estimation device 110, 110A ···Data input section 120, 120A ···Individual variation generation section 130, 130A ···Data mapping section 140,140A...Update section 150, 150A ···Probability distribution estimation unit 160...Display section

Claims

1. A data input unit that acquires a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals. An individual variation generation unit generates a latent vector representing the variation between the first individuals from the individual data of the first individual, A data mapping unit that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update unit updates the mapping parameters in the data mapping unit so that the mapping in the data mapping unit approaches the optimal transport from the probability distribution of the first feature vector to the probability distribution of the second feature vector, based on the first feature vector and estimated feature vector of the plurality of first individuals, and the second feature vector of the plurality of second individuals. A state transition estimation device comprising: a probability distribution estimation unit that uses the data mapping unit, whose mapping parameters have been updated by the update unit, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, from the first feature vector of an individual in a first age range.

2. The state transition estimation device according to claim 1, wherein the data mapping unit outputs the mean and variance of the estimated feature vector obtained by the mapping using a Variational Auto Encoder (VAE) type machine learning model.

3. The state transition estimation device according to claim 2, wherein the probability distribution estimation unit takes the individual data as input and uses a machine learning model that outputs a difference vector between the estimated feature vector obtained by the data mapping unit and the mean of the estimated feature vector to estimate the probability distribution of the estimated feature vector when an individual reaches the second age range, based on the individual data and first feature vector of an individual in the first age range.

4. The probability distribution estimation unit estimates a first probability distribution of the estimated feature vector from the first feature vector of individuals within the first age range, and estimates a second probability distribution of the estimated feature vector from the individual data and the first feature vector of individuals within the first age range. The state transition estimation device according to claim 3, further comprising a display unit for displaying the first probability distribution and the second probability distribution.

5. At least some of the first individuals correspond to at least some of the second individuals. The state transition estimation device according to claim 1, wherein the update unit updates the mapping parameters in the data mapping unit so that the estimated feature vector mapped by the data mapping unit from the first feature vector of at least some of the first individuals approaches the second feature vector of at least some of the second individuals.

6. The aforementioned individual data includes data that may change over time. The probability distribution estimation unit re-estimates the probability distribution of the estimated feature vector in response to the change in the individual data. The state transition estimation device according to claim 1, further comprising a display unit that displays the change in the probability distribution of the estimated feature vector estimated by the probability distribution estimation unit.

7. The first individual, the second individual, and the first individual are individuals, The state transition estimation device according to any one of claims 1 to 6, wherein the first feature vector, the second feature vector, and the estimated feature vector indicate a state related to health or disease risk.

8. A life plan decision support device comprising the state transition estimation device described in claim 7.

9. A data input process that acquires a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals. A method for generating individual variability from the individual data of the first individual, which generates a latent vector representing the variability between the first individuals, A data mapping process that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update process to update the mapping parameters applied to the data mapping process, based on the first feature vectors and estimated feature vectors of the plurality of first individuals, and the second feature vectors of the plurality of second individuals, such that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors. A state transition estimation method comprising: a probability distribution estimation process that uses the data mapping process, in which the mapping parameters have been updated by the update process, to estimate the probability distribution of the estimated feature vector when an individual reaches a second age range, from the first feature vector of an individual in a first age range.

10. On the computer, A data input process that acquires a first feature vector representing the state of each of a plurality of first individuals in a first age range, a second feature vector representing the state of each of a plurality of second individuals in a second age range that is older than the first age range, and individual data different from the first feature vector of each of the plurality of first individuals. A method for generating individual variability from the individual data of the first individual, which generates a latent vector representing the variability between the first individuals, A data mapping process that uses the latent vector of the first individual to map the first feature vector of the first individual to the estimated feature vector when the first individual reaches the second age range, An update process to update the mapping parameters applied to the data mapping process, based on the first feature vectors and estimated feature vectors of the plurality of first individuals, and the second feature vectors of the plurality of second individuals, such that the mapping in the data mapping process approaches the optimal transport from the probability distribution of the first feature vectors to the probability distribution of the second feature vectors. A state transition estimation program that uses the data mapping process, in which the mapping parameters have been updated by the update process, to perform a probability distribution estimation process that estimates the probability distribution of the estimated feature vector when an individual enters a second age range, from the first feature vector of an individual in the first age range.

Images