Model generation device, model generation method, state prediction device, state prediction method, and recording medium

Abstract

A model generation device (1) comprises: an acquiring means (111) for acquiring learning data (122) including a first index value (S_D_sample) indicating a state of a sample patient at a first point in time (t_D) and a second index value (S_F_sample, S_Y_sample) indicating a state of the sample patient at a second point in time (t_F, t_Y); and a learning means (112) for learning a prediction model (M) for predicting, on the basis of the state of a subject patient at the first point in time, the state of the subject patient at the second point in time. The learning means learns the prediction model by updating the prediction model on the basis of a third index value (S_F_predict, S_Y_predict) indicating a state of the sample patient at the second point in time that the prediction model has predicted on the basis of the first index value, and change information (C) representing a tendency of change in patient state over time.

Description

[Technical field]

[0001] The present invention relates to the technical fields of a model generation device, a model generation method, and a recording medium capable of generating a prediction model for predicting a patient's condition, as well as a condition prediction device, a condition prediction method, and a recording medium capable of predicting a patient's condition using a prediction model. [Background technology]

[0002] There are known devices that predict the condition of patients in facilities such as hospitals. For example, Patent Literature 1 describes a device that predicts the degree of recovery of a patient's condition in order to match the patient with a facility that will accept the patient.

[0003] Other prior art documents related to the present invention include Non-Patent Document 1. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. 2020 / 071540 Brochure [Non-patent literature]

[0005] [Non-Patent Document 1] Tetsuo Koyama, "Functional Prognosis Prediction for Stroke Patients and Regional Collaboration Pathways," 3rd Rehabilitation Medicine Specialist Association Academic Conference, JpnJ Rehabil Med 2009, pp108-117, 2008 Summary of the Invention [Problem to be solved by the invention]

[0006] In order to predict the state of a patient at a certain time point, a prediction model that predicts the state of a patient at a certain time point based on the state of the patient at another time point different from the certain time point may be used. When such a prediction model is used, it is preferable to perform a learning operation to update the prediction model (for example, update the parameters of the prediction model) using learning data including the actual state of the patient at a certain time point and the actual state of the patient at another time point in order to improve the prediction accuracy of the prediction model. On the other hand, in order to sufficiently improve the prediction accuracy of the prediction model, a certain amount of learning data is required. However, in order to generate the learning data, it is necessary to actually measure the state of the patient. For this reason, a sufficient amount of learning data is not always available. As a result, the prediction accuracy of the prediction model cannot be appropriately improved, and as a result, there is a possibility that the state of the patient cannot be appropriately predicted.

[0007] The present invention aims to provide a model generation device, a model generation method, a state prediction device, a state prediction method, and a recording medium that can solve the above-mentioned technical problems. As an example, the present invention aims to provide a model generation device, a model generation method, and a recording medium that can appropriately improve the prediction accuracy of a prediction model, and a state prediction device, a state prediction method, and a recording medium that can appropriately predict the state of a patient. [Means for solving the problem]

[0008] One embodiment of the model generating device comprises an acquisition means for acquiring learning data including a first index value indicating the condition of a sample patient at a first time point and a second index value indicating the condition of the sample patient at a second time point after the first time point, and a learning means for learning a prediction model that predicts the condition of the target patient at the second time point based on the condition of the target patient at the first time point, wherein the learning means learns the prediction model based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and change information indicating the trend of change in the condition over time.

[0009] One aspect of the model generation method includes obtaining training data including a first index value indicating the condition of a sample patient at a first time point and a second index value indicating the condition of the sample patient at a second time point later than the first time point, and training a prediction model that predicts the condition of the target patient at the second time point based on the condition of the target patient at the first time point, wherein in training the prediction model, the prediction model is trained based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating the tendency of the condition to change over time.

[0010] A first aspect of the recording medium is a recording medium having recorded thereon a computer program that causes a computer to execute a model generation method, comprising: acquiring learning data including a first index value indicating the condition of a sample patient at a first time point and a second index value indicating the condition of the sample patient at a second time point after the first time point; learning a predictive model that predicts the condition of the target patient at the second time point based on the condition of the target patient at the first time point; and learning the predictive model based on a third index value indicating the condition of the sample patient at the second time point predicted by the predictive model based on the first index value and change information indicating the tendency of the condition to change over time.

[0011] One aspect of the condition prediction device is a condition prediction device comprising: a condition acquisition means for acquiring condition information regarding the condition of a target patient at a first time point; and a prediction means for predicting the condition of the target patient at a second time point after the first time point using the condition information and a prediction model for predicting the condition of the target patient at the second time point based on the condition of the target patient at the first time point, wherein the prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating the tendency of the condition to change over time.

[0012] One aspect of the condition prediction method includes obtaining condition information regarding the condition of a target patient at a first time point, and predicting the condition of the target patient at a second time point after the first time point using the condition information and a prediction model that predicts the condition of the target patient at a second time point after the first time point based on the condition of the target patient at the first time point, wherein the prediction model is a model generated by obtaining learning data including a first index value indicating the condition of a sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and change information indicating the tendency of the condition to change over time.

[0013] A second aspect of the recording medium is a recording medium having recorded thereon a computer program for causing a computer to execute a condition prediction method, the method comprising: acquiring condition information regarding a condition of a target patient at a first time point; predicting the condition of the target patient at a second time point after the first time point using the condition information and a prediction model that predicts the condition of the target patient at the second time point based on the condition of the target patient at the first time point; the prediction model being a model generated by acquiring learning data including a first index value indicating the condition of a sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point; and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and change information indicating the tendency of the condition to change over time. Effect of the Invention

[0014] According to the above-mentioned model generation device, model generation method, and recording medium in a first aspect, it is possible to appropriately improve the prediction accuracy of the prediction model. Also, according to the above-mentioned condition prediction device, condition prediction method, and recording medium in a second aspect, it is possible to appropriately predict the condition of a patient. [Brief description of the drawings]

[0015] [Figure 1] FIG. 1 is a block diagram showing the configuration of a model generating device according to this embodiment. [Diagram 2] FIG. 2 is a data structure diagram showing the data structure of the learning data set in this embodiment. [Diagram 3] FIG. 3 is a flowchart showing the flow of a model generating operation performed by the model generating device in this embodiment. [Figure 4] FIG. 4 is a graph showing an example of a change curve showing the change tendency of the FIM score. [Diagram 5] FIG. 5 is a block diagram showing the configuration of a state prediction device in this embodiment. [Figure 6]FIG. 6 is a flowchart showing the flow of the state prediction operation performed by the state prediction device in this embodiment. [Figure 7] FIG. 7 is a graph showing the prediction accuracy of a prediction model generated by the model generating operation of the present embodiment and the prediction accuracy of a prediction model generated by the model generating operation of a comparative example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, embodiments of a model generation device, a model generation method, a state prediction device, a state prediction method, and a recording medium will be described with reference to the drawings. Below, a model generation device 1 to which the embodiments of the model generation device, the model generation method, and the recording medium are applied, and a state prediction device 2 to which the embodiments of the state prediction device, the state prediction method, and the recording medium are applied will be described in order.

[0017] (1) Model generation device 1 First, the model generating device 1 will be described. The model generating device 1 is a device capable of performing a model generating operation. The model generating operation is an operation for generating a prediction model M. The prediction model M is a model capable of predicting each of a patient's state at a time tB after the time tA and a patient's state at a time tC after the time tB based on the patient's state at the time tA. The prediction model M may separately include a model capable of predicting a patient's state at a time tB after the time tA based on the patient's state at the time tA, and a model capable of predicting a patient's state at a time tC after the time tB based on the patient's state at the time tA. Alternatively, the prediction model M may be a single model capable of predicting each of a patient's state at a time tB after the time tA and a patient's state at a time tC after the time tB based on the patient's state at the time tA. In this embodiment, an example will be described in which the prediction model M is a model capable of outputting an index value indicating the patient's state at a time tB and an index value indicating the patient's state at a time tC when an index value indicating the patient's state at a time tA is input. The prediction model M is a model that can be trained (e.g., a model that can be trained by machine learning). In other words, the prediction model M is a model that can update parameters that define the operation of the prediction model M by a learning operation. An example of such a trainable model is a model that includes a neural network. When the prediction model M is a model that includes a neural network, the parameters that define the operation of the prediction model M may include, for example, at least one of the structure of the neural network (i.e., the connection state of the nodes), the weight of the neural network, and the bias of the neural network.

[0018] The patient's condition in this embodiment may mean, for example, a patient's condition that changes over time. An example of such a patient's condition is the patient's medical condition. This is because the patient's medical condition changes due to various factors. For example, the patient's medical condition may change to improve due to a treatment administered to the patient. Alternatively, for example, the patient's medical condition may change to worsen even if the patient is treated. As an example, if the patient suffers from Alzheimer's or the like, the patient's medical condition may change to worsen even if the patient is treated. Another example of the patient's condition is at least one of the patient's cognitive function and the patient's physical function. This is because at least one of the patient's cognitive function and the patient's physical function improves (i.e., changes) due to the rehabilitation that the patient engages in. In this embodiment, an example will be described in which at least one of the patient's cognitive function and the patient's physical function is used as the patient's condition.

[0019] The index value indicating the state of the patient may mean an index value that changes depending on whether the state of the patient is good or bad. In other words, the index value indicating the state of the patient may mean an index value that quantitatively indicates the degree of the state of the patient. For example, an index value that increases as the state of the patient improves may be used as the index value. For example, an index value that decreases as the state of the patient improves may be used as the index value. For example, an index value that increases as the state of the patient worsens may be used as the index value. For example, an index value that decreases as the state of the patient worsens may be used as the index value. As described above, when the state of the patient changes over time, the index value indicating the state of the patient also changes over time. As described above, in this embodiment, an example in which at least one of the cognitive function of the patient and the physical function of the patient is used as the state of the patient will be described. In this case, an index based on FIM (Functional Independence Measure) (hereinafter referred to as "FIM score") may be used as the index value indicating the state of the patient. Therefore, in this embodiment, an example in which the FIM score is used as the index value indicating the state of the patient will be described.

[0020] The time tA may mean a time between when the patient is admitted to the hospital and when the patient is discharged from the hospital. The time tB may mean a time when a predetermined period pB has passed since the patient is discharged from the hospital. The time tC may mean a time when a predetermined period pC (however, the predetermined period pC is longer than the predetermined period pB) has passed since the patient is discharged from the hospital. In this embodiment, an example will be described in which the time tA is the time t_D when the patient is admitted to the hospital, the time tB is the time t_F when six months have passed since the patient is discharged from the hospital, and the time tC is the time t_Y when one year has passed since the patient is discharged from the hospital. In this case, the prediction model M is a model capable of predicting each of the FIM score S_F indicating the patient's condition at the time t_F and the FIM score S_Y indicating the patient's condition at the time t_Y based on the FIM score S_D indicating the patient's condition at the time t_D. However, the times tA, tB, and tC are not limited to these examples.

[0021] The configuration of the model generation device 1 and the model generation operation performed by the model generation device 1 will be described below in order.

[0022] (1-1) Configuration of model generation device 1 First, the configuration of a model generation device 1 in this embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of a model generation device 1 in this embodiment.

[0023] 1, the model generation device 1 includes a calculation device 11 and a storage device 12. The model generation device 1 may further include an input device 13 and an output device 14. However, the model generation device 1 does not have to include at least one of the input device 13 and the output device 14. The calculation device 11, the storage device 12, the input device 13, and the output device 14 may be connected via a data bus 15.

[0024] The arithmetic device 11 includes at least one of a central processing unit (CPU), a graphical processing unit (GPU), and a field programmable gate array (FPGA). The arithmetic device 11 reads a computer program. For example, the arithmetic device 11 may read a computer program stored in the storage device 12. For example, the arithmetic device 11 may read a computer program stored in a computer-readable and non-transitory recording medium using a recording medium reading device (not shown). The arithmetic device 11 may acquire (i.e., download or read) a computer program from a device (not shown) disposed outside the model generation device 1 via an input device 13 capable of functioning as a communication device. The arithmetic device 11 executes the read computer program. As a result, a logical function block for executing an operation (e.g., the above-mentioned model generation operation) that the model generation device 1 should perform is realized in the arithmetic device 11. In other words, the arithmetic device 11 can function as a controller for realizing a logical function block for executing an operation that the model generation device 1 should perform.

[0025] FIG. 1 shows an example of logical functional blocks realized in the arithmetic device 11 to execute the model generation operation. As shown in FIG. 1, a data acquisition unit 111, which is a specific example of an "acquisition means", and a model generation unit 112, which is a specific example of a "learning means", are realized in the arithmetic device 11. Note that the operations performed by the data acquisition unit 111 and the model generation unit 112 will be described in detail later, but will be outlined below. The data acquisition unit 111 acquires learning data 122 (see FIG. 2) used to generate the prediction model M from a learning dataset 121 stored in the storage device 12. The model generation unit 112 generates (in other words, learns) the prediction model M using the learning data 122 acquired by the data acquisition unit 111.

[0026] The storage device 12 can store desired data. For example, the storage device 12 may temporarily store a computer program executed by the arithmetic device 11. The storage device 12 may temporarily store data that is temporarily used by the arithmetic device 11 when the arithmetic device 11 is executing a computer program. The storage device 12 may store data that is to be stored by the model generating device 1 for a long period of time. The storage device 12 may include at least one of a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device. In other words, the storage device 12 may include a non-temporary recording medium.

[0027] In this embodiment, the storage device 12 stores data used by the model generation device 1 to perform a model generation operation. Fig. 1 illustrates a training data set 121 as an example of data used by the model generation device 1 to perform a model generation operation. That is, Fig. 1 illustrates an example in which the storage device 12 stores the training data set 121.

[0028] An example of the data structure of the training data set 121 is shown in FIG. 2. As shown in FIG. 2, the training data set 121 includes a plurality of training data 122. Each training data 122 includes information on the condition of the patient measured in the past. Specifically, each training data 122 includes at least one of the FIM score S_D obtained as a result of measuring the condition of the patient at time t_D, the FIM score S_F obtained as a result of measuring the condition of the patient at time t_F, and the FIM score S_Y obtained as a result of measuring the condition of the patient at time t_Y, as information on the condition of the patient measured in the past. In this case, the training data set 121 may include training data 122 including all of the FIM score S_D, the FIM score S_F, and the FIM score S_Y (see the training data 122 with a data ID of #1 in FIG. 2). The training data set 121 may include training data 122 that includes the FIM score S_D and the FIM score S_F, but does not include the FIM score S_Y (see the training data 122 with a data ID of #2 in FIG. 2). The training data set 121 may include training data 122 that includes the FIM score S_D and the FIM score S_Y, but does not include the FIM score S_F (see the training data 122 with a data ID of #3 in FIG. 2).

[0029] The patient whose FIM score is included in the learning data 122 is a patient whose FIM score has been measured in the past, and may be different from a patient whose future FIM score is to be estimated using the prediction model M. For this reason, in the following description, the patient whose FIM score is included in the learning data 122 is referred to as a "sample patient" as necessary, while the patient whose future FIM score is to be estimated using the prediction model M is referred to as a "target patient" to conveniently distinguish between the two. The target patient may be the same person as the sample patient. In other words, the learning dataset 121 may include the learning data 122 including the FIM score measured in the past of the target patient whose future FIM score is to be estimated using the prediction model M.

[0030] In the following description, the FIM score S_D, the FIM score S_F, and the FIM score S_Y included in the learning data 122 are written as FIM score S_D_sample, FIM score S_F_sample, and FIM score S_Y_sample, respectively, as necessary, while the FIM score S_F and the FIM score S_Y predicted using the prediction model M are written as FIM score S_F_predict and FIM score S_Y_predict, respectively, to distinguish between the two. In this case, the FIM score S_D_sample corresponds to the actual measurement value of the FIM score indicating the state of the sample patient at the time t_D when the sample patient is admitted to the hospital. Moreover, the FIM score S_F_sample corresponds to the actual measurement value of the FIM score indicating the state of the sample patient at the time t_F when six months have passed since the sample patient was discharged from the hospital. Moreover, the FIM score S_Y_sample corresponds to the actual measurement value of the FIM score indicating the state of the sample patient at the time t_Y when one year has passed since the sample patient was discharged from the hospital. The FIM score S_F_predict corresponds to a predicted value of the FIM score indicating the state of the target patient at time t_F, six months after the target patient is discharged from the hospital, and the FIM score S_Y_predict corresponds to a predicted value of the FIM score indicating the state of the target patient at time t_Y, one year after the target patient is discharged from the hospital.

[0031] The training data set 121 may include a plurality of training data 122 each including information on the condition of a plurality of different sample patients. For example, the training data set 121 includes (i) first training data including at least one of an FIM score S_D_sample obtained by measuring the condition of the first sample patient at time t_D when the first sample patient is admitted to the hospital, an FIM score S_F_sample obtained by measuring the condition of the first sample patient at time t_F when six months have passed since the first sample patient was discharged from the hospital, and an FIM score S_Y_sample obtained by measuring the condition of the first sample patient at time t_Y when one year has passed since the first sample patient was discharged from the hospital; and (ii) ) second learning data including at least one of an FIM score S_D_sample obtained by measuring the condition of a second sample patient different from the first sample patient at time t_D when the second sample patient is admitted to the hospital, and an FIM score S_F_sample obtained by measuring the condition of the second sample patient at time t_F when six months have passed since the second sample patient was discharged from the hospital, and an FIM score S_Y_sample obtained by measuring the condition of the second sample patient at time t_Y when one year has passed since the second sample patient was discharged from the hospital.

[0032] In the case where a sample patient is admitted to a hospital multiple times, the training dataset 121 may include training data 122 containing information on the condition of the sample patient when the sample patient is admitted to the hospital for the Nth time (note that N is a variable indicating an integer equal to or greater than 1) and training data 122 containing information on the condition of the same sample patient when the sample patient is admitted to the hospital for the Mth time (note that M is a variable different from N and indicating an integer equal to or greater than 1).

[0033] 1 again, the input device 13 is a device that accepts input of information to the model generation device 1 from outside the model generation device 1. For example, the input device 13 may include an operation device (e.g., at least one of a keyboard, a mouse, and a touch panel) that can be operated by a user of the model generation device 1. For example, the input device 13 may include a receiving device that can receive information transmitted as data from outside the model generation device 1 via a communication network to the model generation device 1.

[0034] The output device 14 is a device that outputs information. For example, the output device 14 may output information about the prediction model M generated by the model generation operation performed by the model generation device 1. As described above, since the prediction model M is used by the state prediction device 2, the output device 14 may output information about the prediction model M to the state prediction device 2. As a result, the state prediction device 2 can predict the state of the target patient using the prediction model M. An example of such an output device 14 is a transmission device that can transmit information as data via a communication network or a data bus. However, the output device 14 may include, for example, a display (display device) that can output (i.e., display) information as an image. The output device 14 may include, for example, a speaker (audio output device) that can output information as sound. For example, the output device 14 may include, for example, a printer that can output a document on which information is printed.

[0035] (1-2) Model Generation Operation Performed by the Model Generation Device 1 Next, the model generation operation performed by the model generation device 1 will be described with reference to Fig. 3. Fig. 3 is a flowchart showing the flow of the model generation operation performed by the model generation device 1.

[0036] As shown in FIG. 3, first, the data acquisition unit 111 acquires the learning data 122 used to generate the prediction model M from the learning dataset 121 stored in the storage device 12 (step S11). The data acquisition unit 111 acquires all of the multiple learning data 122 included in the learning dataset 121. However, the data acquisition unit 111 may acquire a part of the multiple learning data 122 included in the learning dataset 121, while not acquiring the other part of the multiple learning data 122 included in the learning dataset 121. In the following description, for convenience of explanation, an example will be described in which the data acquisition unit 111 acquires K (note that K is a constant indicating an integer equal to or greater than 2) pieces of learning data 122 (specifically, learning data 122#1 to 122#K).

[0037] Thereafter, the model generation unit 112 generates (i.e., learns) a prediction model M using the multiple learning data 122 acquired in step S11 (step S12). Specifically, when the model generation device 1 generates a prediction model M for the first time, the model generation unit 112 inputs the FIM score S_D_sample included in the learning data 122#k (where k is a variable indicating an integer satisfying 1≦k≦K) acquired in step S11 to a default prediction model M (or a prediction model M in an initial state) prepared in advance. In the following description, the FIM scores S_D_sample, S_F_sample, and S_Y_sample included in the learning data 122#k are referred to as FIM scores S_D_sample#k, S_F_sample#k, and S_Y_sample#k, respectively. Alternatively, if the model generation device 1 has already generated the prediction model M, the model generation unit 112 inputs the FIM score S_D_sample#k included in the learning data 122#k acquired in step S11 to the generated prediction model M. As a result, the prediction model M outputs the FIM score S_F_predict and the FIM score S_Y_predict. That is, the model generation unit 112 acquires the FIM score S_F_predict and the FIM score S_Y_predict. In the following description, the FIM scores S_F_predict and S_Y_predict output by the prediction model M to which the FIM score S_D_sample#k has been input are referred to as the FIM score S_F_predict#k and S_Y_predict#k, respectively.

[0038] The model generation unit 112 repeats the operation of acquiring the FIM score S_F_predict#k and the FIM score S_Y_predict#k by inputting the FIM score S_D_sample#k to the prediction model M, for the K pieces of learning data 122#1 to 122#K acquired in step S11. As a result, the model generation unit 112 acquires the FIM scores S_F_predict#1 to S_F_predict#K and the FIM scores S_Y_predict#1 to S_Y_predict#K.

[0039] Thereafter, the model generation unit 112 updates the default or generated prediction model M based on the FIM scores S_F_predict#1 to S_F_predict#K and the FIM scores S_Y_predict#1 to S_Y_predict#K. That is, the model generation unit 112 updates parameters that define the operation of the prediction model M based on the FIM scores S_F_predict#1 to S_F_predict#K and the FIM scores S_Y_predict#1 to S_Y_predict#K. As a result, an updated (that is, learned) prediction model M is generated. In order to update the prediction model M, the model generation unit 112 uses a loss function (in other words, an objective function) Loss that is determined based on the FIM scores S_F_predict#1 to S_F_predict#K and the FIM scores S_Y_predict#1 to S_Y_predict#K. That is, the model generation unit 112 generates the prediction model M based on the loss function Loss. Specifically, the model generation unit 112 may update the prediction model M so that the loss function Loss is minimized, for example.

[0040] The loss function Loss may include a loss term Loss1 based on the prediction errors of the FIM scores S_F and S_Y. Specifically, the FIM score S_F_predict#k output by the prediction model M is a predicted value of the FIM score S_F of the sample patient at the time t_F. Meanwhile, the actual value of the FIM score S_F of the sample patient at the time t_F is included as the FIM score S_F_sample#k in the learning data 122#k. For this reason, it is estimated that the smaller the difference between the FIM score S_F_predict#k and the FIM score S_F_sample#k (i.e., the prediction error of the FIM score S_F), the better the prediction accuracy of the prediction model M. For the same reason, it is estimated that the smaller the difference between the FIM score S_Y_predict#k and the FIM score S_Y_sample#k (i.e., the prediction error of the FIM score S_Y), the better the prediction accuracy of the prediction model M. Therefore, the loss term Loss1 may be a term that decreases as the difference between the FIM score S_F_predict#k and the FIM score S_F_sample#k decreases and / or as the difference between the FIM score S_Y_predict#k and the FIM score S_Y_sample#k decreases. An example of such a loss term Loss1 is shown in Equations 1 and 2. Equation 1 shows an example in which a squared error is used as the prediction error. Equation 2 shows an example in which an absolute value is used as the prediction error. Note that, considering that the model generating unit 112 generally calculates the loss term Loss1 by performing a matrix operation, it is preferable to use the loss term Loss1 shown in Equation 1.

[0041]

number

[0042]

number

[0043] Particularly in this embodiment, the loss function Loss includes a loss term Loss2 determined based on a change curve C that represents a change tendency of the FIM score over time. In this case, the model generation unit 112 may update the prediction model M so that the loss function Loss including the sum of the loss term Loss1 and the loss term Loss2 is minimized.

[0044] The change curve C is information that represents a typical (in other words, general, standard, average, or ideal) change tendency of the FIM score over time. Specifically, the condition (in this case, the disease state) of a patient suffering from a certain disease is relatively likely to recover in a manner that is generally similar to the tendency of the disease state of other patients suffering from the same disease to recover. In other words, the FIM score of a patient suffering from a certain disease is relatively likely to change (typically increase) in a manner that is generally similar to the tendency of the FIM scores of other patients suffering from the same disease to change (typically increase). In other words, the FIM scores of multiple patients suffering from the same disease are relatively likely to change in a similar manner to each other. For example, the above-mentioned Non-Patent Document 1 shows that the FIM scores of patients suffering from stroke tend to change along a logarithmic curve.

[0045] In this way, the change curve C (or any information) showing a typical change tendency of the FIM score over time can be said to substantially represent a predetermined relationship that is estimated to be established between the FIM scores S_D_sample#k, S_F_sample#k, and S_Y_sample#k included in the learning data 122#k. In this case, the closer the relationship between the FIM score S_D_sample#k input to the prediction model M and the FIM scores S_F_predict#k and S_Y_predict#k output by the prediction model M is to the predetermined relationship represented by the change curve C, the better the prediction accuracy of the prediction model M is estimated to be. In other words, the closer the change between the FIM score S_D_sample#k and the FIM scores S_F_predict#k and S_Y_predict#k is to the change tendency of the FIM score from the time t_D to the time t_F and t_Y represented by the change curve C, the better the prediction accuracy of the prediction model M is estimated to be. For this reason, the loss term Loss2 may be a term that becomes smaller as the relationship between the FIM scores S_D_sample#k, S_F_predict#k, and S_Y_predict#k approaches the predetermined relationship represented by the change curve C. In other words, the loss term Loss2 may be a term that becomes smaller as the change between the FIM score S_D_sample#k and the FIM scores S_F_predict#k and S_Y_predict#k approaches the change trend of the FIM score from time t_D to time t_F and t_Y represented by the change curve C. In this case, the model generation unit 112 generates the prediction model M so that the change between the FIM score S_D_sample#k and the FIM scores S_F_predict#k and S_Y_predict#k approaches the change trend of the FIM score from time t_D to time t_F and t_Y represented by the change curve C.

[0046] An example of the loss term Loss2 will be described with reference to FIG. 4. FIG. 4 shows an example in which the FIM score is defined by a formula a, S_Day1=S_Day2+β×ln(Day1 / Day2). That is, FIG. 4 shows an example in which a change curve C showing the change tendency of the FIM score is expressed by a formula a, S_Day1=S_Day2+β×ln(Day1 / Day2). Note that "Day1" and "Day2" in the formula a indicate the number of days elapsed from a predetermined reference time (for example, the time when the patient developed the disease). However, "Day1" and "Day2" are different from each other. Also, "β" in the formula a means an arbitrary constant. Also, "ln" in the formula a means a logarithm with Napier's constant e as the base (i.e., a natural logarithm). Therefore, formula a indicates that a predetermined relationship defined by formula b, S_Y=S_D+β×ln(Day_Y / Day_D), is established between the FIM scores S_D, S_F, and S_Y. Note that "Day_D" and "Day_Y" in formula b indicate the number of days elapsed from a predetermined reference time to time t_D and the number of days elapsed from a predetermined reference time to time t_Y, respectively. Furthermore, the constant β in formula b is a constant that can be expressed by formula c, β=S_F-S_D / ln(Day_F / Day_D). Note that "Day_F" in formula c indicates the number of days elapsed from a predetermined reference time to time t_F. Formula b may be replaced with formula b', S_F=S_D+β×ln(Day_F / Day_D), in which case the constant β may be expressed as formula c', β=S_Y-S_D / ln(Day_Y / Day_D). Here, by substituting formula c into formula b and rearranging the formula, formula b is converted into formula d, S_Y=S_D×(1-A)+S_F×A. Note that "A" in formula d indicates a constant expressed by the formula A=ln(Day_Y / Day_D) / ln(Day_F / Day_D).

[0047] As a result, it can be said that the change curve C shown in Fig. 4 indicates that a predetermined relationship expressed by the formula d, S_Y = S_D x (1-A) + S_F x A, is established between the FIM scores S_D, S_F, and S_Y. In other words, the change curve C indicates that it is estimated that a predetermined relationship expressed by the formula d, S_Y_sample = S_D_sample x (1-A) + S_F_sample x A, is likely to be established between the FIM scores S_D_sample, S_F_sample, and S_Y_sample. Then, considering that the FIM scores S_F_predict and S_Y_predict are the predicted values ​​of the FIM scores S_F_sample and S_Y_sample, respectively, it is estimated that the prediction accuracy of the prediction model M is better as the relationship between the FIM scores S_D_sample, S_F_predict, and S_Y_predict approaches a predetermined relationship expressed by the formula e of S_Y_predict=S_D_sample×(1-A)+S_F_predict×A. Therefore, the loss term Loss2 may be a term that becomes smaller as the difference between the left side of the formula e (=S_Y_predict) and the right side of the formula e (=S_D_sample×(1-A)+S_F_predict×A) becomes smaller. An example of such a loss term Loss2 is shown in formulas 3 and 4. Formula 3 shows an example in which the squared value of the difference is used. Formula 4 shows an example in which the absolute value of the difference is used. In addition, considering that the model generating unit 112 generally calculates the loss term Loss2 by performing a matrix operation, it is preferable to use the loss term Loss2 shown in Equation 3. In addition, the loss term Loss2 may be used as a so-called regularization term.

[0048]

number

[0049]

number

[0050] The loss term Loss2 may be set in advance based on the change curve C. Alternatively, the model generation unit 112 may set the loss term Loss2 based on the change curve C. In either case, it can be said that the model generation unit 112 updates the prediction model M based on the change curve C. In other words, it can be said that the model generation unit 112 updates the prediction model M using prior knowledge regarding the change curve C.

[0051] (2) State prediction device 2 Next, the state prediction device 2 will be described. The state prediction device 2 performs a state prediction operation of predicting the state of the target patient at a time tB after the time tA and the state of the target patient at a time tC after the time tB, based on the state of the target patient at the time tA, using the prediction model M generated by the model generation device 1. In this embodiment, as described above, an example is described in which the prediction model M is a model capable of predicting the FIM score S_F indicating the state of the target patient at a time t_F and the FIM score S_Y indicating the state of the target patient at a time t_Y after the time t_F, based on the FIM score S_D indicating the state of the target patient at the time t_D. In this case, the state prediction operation is an operation of predicting the FIM score S_F indicating the state of the target patient at a time t_F and the FIM score S_Y indicating the state of the target patient at a time t_Y, based on the FIM score S_D indicating the state of the target patient at the time t_D, using the prediction model M. Hereinafter, the configuration of the state prediction device 2 and the state prediction operation performed by the state prediction device 2 will be described in order.

[0052] (2-1) Configuration of state prediction device 2 First, the configuration of the state prediction device 2 in this embodiment will be described with reference to Fig. 5. Fig. 5 is a block diagram showing the configuration of the state prediction device 2 in this embodiment.

[0053] 5, the state prediction device 2 includes a calculation device 21 and a storage device 22. Furthermore, the state prediction device 2 may include an input device 23 and an output device 24. However, the state prediction device 2 does not have to include at least one of the input device 23 and the output device 24. The calculation device 21, the storage device 22, the input device 23, and the output device 24 may be connected via a data bus 25.

[0054] The arithmetic device 21 includes at least one of a CPU, a GPU, and an FPGA, for example. The arithmetic device 21 reads a computer program. For example, the arithmetic device 21 may read a computer program stored in the storage device 22. For example, the arithmetic device 21 may read a computer program stored in a computer-readable and non-transient recording medium using a recording medium reading device (not shown). The arithmetic device 21 may acquire (i.e., download or read) a computer program from a device (not shown) disposed outside the state prediction device 2 via an input device 23 capable of functioning as a communication device. The arithmetic device 21 executes the read computer program. As a result, a logical function block for executing an operation (for example, the above-mentioned state prediction operation) to be performed by the state prediction device 2 is realized in the arithmetic device 21. In other words, the arithmetic device 21 can function as a controller for realizing a logical function block for executing an operation to be performed by the state prediction device 2.

[0055] FIG. 5 shows an example of a logical functional block realized in the arithmetic device 21 to execute a state prediction operation. As shown in FIG. 5, a data acquisition unit 211, which is a specific example of an "acquisition means", and a state prediction unit 212, which is a specific example of a "prediction means", are realized in the arithmetic device 21. Note that the operations performed by the data acquisition unit 211 and the state prediction unit 212 will be described in detail later, but will be outlined below. The data acquisition unit 211 acquires an FIM score S_D indicating the state of the target patient at time t_D. The state prediction unit 212 predicts an FIM score S_F indicating the state of the target patient at time t_F and an FIM score S_Y indicating the state of the target patient at time t_Y, using the FIM score S_D acquired by the data acquisition unit 211 and a prediction model M.

[0056] The storage device 22 can store desired data. For example, the storage device 22 may temporarily store a computer program executed by the arithmetic device 21. The storage device 22 may temporarily store data that is temporarily used by the arithmetic device 21 when the arithmetic device 21 is executing a computer program. The storage device 22 may store data that is to be stored by the state prediction device 2 for a long period of time. For example, in the example shown in FIG. 5, the storage device 22 stores a prediction model M used by the state prediction device 2. The storage device 22 may include at least one of a RAM, a ROM, a hard disk device, a magneto-optical disk device, an SSD, and a disk array device. In other words, the storage device 22 may include a non-temporary recording medium.

[0057] The input device 23 is a device that accepts information input to the state prediction device 2 from outside the state prediction device 2. For example, the input device 23 may include an operation device (for example, at least one of a keyboard, a mouse, and a touch panel) that can be operated by a user of the state prediction device 2. For example, the input device 23 may include a receiving device that can receive information transmitted as data from outside the state prediction device 2 to the state prediction device 2 via a communication network.

[0058] The output device 24 is a device that outputs information. For example, the output device 24 may output information related to the prediction result of the state prediction device 2 (for example, each of the FIM scores S_F and S_Y). An example of such an output device 24 is a transmission device that can transmit information as data via a communication network or a data bus. However, the output device 24 may include, for example, a display (display device) that can output (i.e., display) information as an image. The output device 24 may include, for example, a speaker (audio output device) that can output information as sound. For example, the output device 24 may include, for example, a printer that can output a document on which information is printed.

[0059] (2-2) State Prediction Operation Performed by State Prediction Device 2 Next, the state prediction operation performed by the state prediction device 2 will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of the state prediction operation performed by the state prediction device 2.

[0060] 6, first, the data acquisition unit 211 acquires the FIM score S_D indicating the condition of the target patient at time t_D (step S21). For example, when the FIM score S_D is stored in the storage device 22, the data acquisition unit 211 may acquire the FIM score S_D from the storage device 22. For example, when the FIM score S_D is input to the condition prediction device 2 via the input device 23, the data acquisition unit 211 may acquire the FIM score S_D using the input device 23.

[0061] Thereafter, the state prediction unit 212 predicts the FIM score S_F indicating the state of the target patient at time t_F and the FIM score S_Y indicating the state of the target patient at time t_Y, respectively, using the FIM score S_D acquired in step S21 and the prediction model M (step S22). Specifically, the state prediction unit 212 inputs the FIM score S_D acquired in step S21 to the prediction model M. As a result, the prediction model M outputs the FIM score S_F_predict and the FIM score S_Y_predict. That is, the state prediction unit 212 acquires the FIM score S_F_predict from the prediction model M as a predicted value of the FIM score S_F indicating the state of the target patient at time t_F. In addition, the state prediction unit 212 acquires the FIM score S_Y_predict from the prediction model M as a predicted value of the FIM score S_Y indicating the state of the target patient at time t_Y.

[0062] (3) Technical effect As described above, the model generation device 1 of the present embodiment generates a prediction model M based on the change curve C (i.e., information representing the change tendency of the patient's condition). That is, the model generation device 1 generates the prediction model M using a loss function Loss including a loss term Loss2 determined based on the change curve C. Therefore, compared to the model generation device of the comparative example, the model generation device 1 can generate a prediction model M capable of outputting a prediction result that is relatively likely to match the change tendency of the patient's condition represented by the change curve C. Note that the model generation device of the comparative example is a device that generates a prediction model M using a loss function that does not include the loss term Loss2 (for example, a loss function that includes the loss term Loss1 but does not include the loss term Loss2). Therefore, the model generation device 1 can appropriately improve the prediction accuracy of the prediction model M.

[0063] In addition, in this embodiment, the model generating device 1 generates a prediction model M that predicts FIM scores S_F and S_Y indicating the state of the target patient at time points t_F and t_Y after the target patient is discharged from the hospital. Here, as described above, when the loss function Loss includes a loss term Loss1 related to the prediction error, the greater the number of samples of the FIM scores S_F_sample and S_Y_sample indicating the state of the sample patient at time points t_F and t_Y after the sample patient is discharged from the hospital, the higher the prediction accuracy of the FIM scores S_F and S_Y by the prediction model M. However, compared to measuring the state of the patient while the patient is hospitalized in the hospital, it is not necessarily easy to measure the state of the patient once the patient has been discharged from the hospital. This is because it is not easy to encourage the discharged patient to visit the hospital to measure the state of the patient. For this reason, it is not necessarily easy to collect a large number of FIM scores S_F_sample and S_Y_sample. On the other hand, the model generation device 1 of the present embodiment generates the prediction model M using a loss function Loss including a loss term Loss2 based on the change curve C in addition to or instead of the loss term Loss1 related to the prediction error. Therefore, the model generation device 1 of the present embodiment can appropriately improve the prediction accuracy of the prediction model M even under circumstances where the number of samples of the FIM scores S_F_sample and S_Y_sample is relatively small, compared to the model generation device of the comparative example. For example, FIG. 7 shows the prediction accuracy of the prediction model M generated by the model generation device 1 of the present embodiment and the prediction accuracy of the prediction model M generated by the model generation device of the comparative example. The horizontal axis of FIG. 7 shows the ratio of the total number of sample patients whose FIM scores S_F_sample and S_Y_sample were measurable at the time points t_F and t_Y to the total number of sample patients whose FIM scores S_D_sample at the time point t_D have been measured.Specifically, "X%" on the horizontal axis of FIG. 7 indicates a situation in which only the FIM scores S_F_sample and S_Y_sample of X×Y sample patients were measurable under the condition that the FIM scores S_D_sample of Y sample patients have been measured (in other words, the FIM scores S_F_sample and S_Y_sample of (1-X)×Y sample patients could not be measured). As shown in FIG. 7, the prediction accuracy of the prediction model M generated by the model generation device of the comparative example decreases as the number of samples of the FIM scores S_F_sample and S_Y_sample decreases (i.e., the ratio of the total number of sample patients whose FIM scores S_F_sample and S_Y_sample were measurable decreases). On the other hand, the prediction accuracy of the prediction model M generated by the model generation device 1 of the present embodiment does not decrease as much as the prediction accuracy of the prediction model M generated by the model generation device of the comparative example, even if the number of samples of the FIM scores S_F_sample and S_Y_sample decreases.

[0064] Moreover, the condition prediction device 2 of this embodiment predicts the condition of the patient using the prediction model M generated by the model generation device 1 of this embodiment. Therefore, the condition prediction device 2 can appropriately predict the condition of the patient.

[0065] (4) Variations In the above description, the prediction model M is a model capable of predicting, based on the patient's condition at time tA (e.g., time t_D when the patient is hospitalized), the patient's condition at time tB (e.g., time t_F when six months have passed since the patient was discharged from the hospital) later than time tA and the patient's condition at time tC (e.g., time t_Y when one year has passed since the patient was discharged from the hospital) later than time tB based on the patient's condition at time tA. However, the prediction model M may be a model that predicts the patient's condition at either time tB or tC based on the patient's condition at time tA, but does not predict the patient's condition at the other of time tB and tC. Even in this case, the model generation device 1 may generate the prediction model M by performing the above-mentioned model generation operation.

[0066] As a modified example, a model generation operation for generating a prediction model M capable of predicting an FIM score S_Y indicating a state of a patient at a time t_Y based on an FIM score S_D indicating a state of the patient at a time t_D will be described. In the modified example, the data acquisition unit 111 also acquires learning data 122 (step S11 in FIG. 3). However, in the modified example, the learning data 122 may not include an FIM score S_F_sample indicating a state of a sample patient at a time t_F. Thereafter, the model generation unit 112 acquires an FIM score S_Y_predict#k by inputting the FIM score S_D_sample#k to the prediction model M. Thereafter, the model generation unit 112 updates the prediction model M based on the FIM scores S_Y_predict#1 to S_Y_predict#K so that the loss function Loss is minimized. However, in the modified example, the loss term Loss1 is a term that becomes smaller as the difference between the FIM score S_Y_predict#k and the FIM score S_Y_sample#k becomes smaller. An example of such a loss term Loss1 is shown in Equations 4 and 5. In a modified example, the loss term Loss2 may be a term that becomes smaller as the relationship between the FIM scores S_D_sample#k and S_Y_predict#k approaches the predetermined relationship represented by the change curve C. An example of such a loss term Loss2 is shown in Equations 6 and 7. However, in a modified example, the constant β is preferably information known to the model generating device 1 (i.e., a value that is set in advance).

[0067]

number

[0068]

number

[0069]

number

[0070]

number

[0071] Furthermore, by replacing the symbol "Y" with the symbol "F" in the description of the model generation operation in the modified example, the above description of the model generation operation becomes a description of a model generation operation for generating a predictive model M capable of predicting an FIM score S_F indicating a patient's condition at time t_F based on an FIM score S_D indicating the patient's condition at time t_D.

[0072] (5) Additional Notes The following supplementary notes are further disclosed regarding the above-described embodiment. [Appendix 1] an acquisition means for acquiring learning data including a first index value indicating a condition of a sample patient at a first time point and a second index value indicating a condition of the sample patient at a second time point after the first time point; a learning means for learning a prediction model that predicts a condition of the target patient at the second time point based on a condition of the target patient at the first time point; Equipped with The learning means learns the prediction model based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and on change information indicating a change tendency of the condition over time. Model generation device. [Appendix 2] The learning means learns the prediction model so that a change between the first and third index values ​​approaches a first change trend, which is a change trend of the state from the first time point to the second time point represented by the change information. 2. The model generating apparatus of claim 1. [Appendix 3] The learning means learns the prediction model based on a loss function including a term that becomes smaller as the change between the first and third index values ​​approaches the first change tendency. 2. The model generating apparatus of claim 1. [Appendix 4] The first time point is a time point between when a patient, including the sample patient and the target patient, is admitted to a hospital and when the patient is discharged from the hospital, The second time point is a time point when a first predetermined period of time has elapsed since the patient was discharged from the hospital. 4. A model generating device according to any one of claims 1 to 3. [Appendix 5] the predictive model further predicts a condition of the subject patient at a third time point that is later than the first time point and different from the second time point; the learning data includes a fourth index value indicating a condition of the sample patient at the third time point; The learning means learns the prediction model based on the third index value, a fifth index value indicating the condition of the sample patient at the third time point predicted by the prediction model based on the first index value, and the change information. 5. A model generating device according to any one of claims 1 to 4. [Appendix 6] The learning means learns the prediction model so that changes among the first index value, the third index value, and the fifth index value approach a second change tendency which is a change tendency of the state from the first time point to the second time point and the third time point represented by the change information. 6. The model generating apparatus of claim 5. [Appendix 7] The learning means learns the prediction model based on a loss function including a term that becomes smaller as the change among the first index value, the third index value, and the fifth index value approaches the second change tendency. 7. The model generating apparatus of claim 6. [Appendix 8] The first time point is a time point between when a patient, including the sample patient and the target patient, is admitted to a hospital and when the patient is discharged from the hospital, The second time point is a time point when a first predetermined period has elapsed since the patient was discharged from the hospital; The third time point is a time point when a second predetermined period different from the first predetermined period has elapsed since the patient was discharged from the hospital. 8. A model generating device according to any one of appendix 5 to 7. [Appendix 9] The first time point is the time point when the patient is admitted to a hospital. 9. The model generating device according to claim 4 or 8. [Appendix 10] acquiring learning data including a first index value indicating a state of a sample patient at a first time point and a second index value indicating a state of the sample patient at a second time point after the first time point; learning a predictive model that predicts a condition of the subject patient at the second time point based on a condition of the subject patient at the first time point; In the learning of the prediction model, the prediction model is learned based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and on change information indicating a change tendency of the condition over time. Model generation method. [Appendix 11] To your computer acquiring learning data including a first index value indicating a state of a sample patient at a first time point and a second index value indicating a state of the sample patient at a second time point after the first time point; learning a predictive model that predicts a condition of the subject patient at the second time point based on a condition of the subject patient at the first time point; A recording medium having a computer program recorded thereon for executing a model generation method in which the predictive model is trained based on a third index value indicating the condition of the sample patient at the second time point predicted by the predictive model based on the first index value, and change information indicating the trend of change in the condition over time. [Appendix 12] On the computer, acquiring learning data including a first index value indicating a state of a sample patient at a first time point and a second index value indicating a state of the sample patient at a second time point after the first time point; learning a predictive model that predicts a condition of the subject patient at the second time point based on a condition of the subject patient at the first time point; A computer program that executes a model generation method in which the predictive model is trained based on a third index value that indicates the condition of the sample patient at the second time point predicted by the predictive model based on the first index value, and change information that indicates the trend of change in the condition over time. [Appendix 13] A status acquisition means for acquiring status information regarding the status of a subject patient at a first time point; a prediction means for predicting a condition of the target patient at the second time point using the condition information and a prediction model for predicting a condition of the target patient at a second time point after the first time point based on the condition of the target patient at the first time point; Equipped with The prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating a change tendency of the condition over time. State prediction device. [Appendix 14] Obtaining status information regarding the status of the subject patient at a first time point; predicting a condition of the target patient at a second time point after the first time point using the condition information and a prediction model that predicts a condition of the target patient at the second time point based on the condition of the target patient at the first time point; The prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating a change tendency of the condition over time. State prediction methods. [Appendix 15] On the computer, Obtaining status information regarding the status of the subject patient at a first time point; predicting a condition of the target patient at a second time point after the first time point using the condition information and a prediction model that predicts a condition of the target patient at the second time point based on the condition of the target patient at the first time point; A recording medium on which a computer program is recorded for executing a condition prediction method, in which the prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating the trend of change in the condition over time. [Appendix 16] On the computer, Obtaining status information regarding the status of the subject patient at a first time point; predicting a condition of the target patient at a second time point after the first time point using the condition information and a prediction model that predicts a condition of the target patient at the second time point based on the condition of the target patient at the first time point; A computer program for executing a condition prediction method, wherein the prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating the trend of change in the condition over time.

[0073] The present invention can be modified as appropriate within the scope that does not contradict the gist or concept of the invention that can be read from the claims and the entire specification, and the parameter determination device, parameter determination method, signal transmission method, and computer program that involve such modifications are also included in the technical concept of the present invention. [Explanation of symbols]

[0074] 1. Model generation device 11 Arithmetic unit 111 Data Acquisition Section 112 Model Generation Unit 12 Storage device 121 Training Dataset 122 Training Data 2. State Prediction Device 21 Arithmetic unit 211 Data Acquisition Department 212 State Prediction Unit M Prediction Model

Claims

1. an acquisition means for acquiring learning data including a first index value indicating a condition of a sample patient at a first time point and a second index value indicating a condition of the sample patient at a second time point after the first time point; a learning means for learning a prediction model that predicts a condition of the target patient at the second time point based on a condition of the target patient at the first time point; Equipped with The learning means learns the prediction model based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and on change information indicating a change trend of the condition over time, so that a change between the first index value and the third index value approaches a first change trend which is a change trend of the condition from the first time point to the second time point represented by the change information. Model generation device.

2. The learning means learns the prediction model based on a loss function including a term that becomes smaller as the change between the first and third index values ​​approaches the first change tendency. The model generating device according to claim 1 .

3. The first time point is a time point between when a patient, including the sample patient and the target patient, is admitted to a hospital and when the patient is discharged from the hospital, The second time point is a time point when a first predetermined period of time has elapsed since the patient was discharged from the hospital.

3. The model generating device according to claim 1 or 2.

4. the predictive model further predicts a condition of the subject patient at a third time point that is later than the first time point and different from the second time point; the learning data includes a fourth index value indicating a condition of the sample patient at the third time point; The learning means learns the prediction model based on the third index value, a fifth index value indicating the condition of the sample patient at the third time point predicted by the prediction model based on the first index value, and the change information. A model generating device according to any one of claims 1 to 3.

5. The learning means learns the prediction model so that changes among the first index value, the third index value, and the fifth index value approach a second change trend, which is a change trend of the state from the first time point to the second time point and the third time point represented by the change information.

5. The model generating device of claim 4.

6. The learning means learns the prediction model based on a loss function including a term that becomes smaller as the change among the first index value, the third index value, and the fifth index value approaches the second change tendency.

6. The model generating device according to claim 5.

7. The first time point is a time point between when a patient, including the sample patient and the target patient, is admitted to a hospital and when the patient is discharged from the hospital, the second time point is a time point when a first predetermined period has elapsed since the patient was discharged from the hospital; The third time point is a time point when a second predetermined period different from the first predetermined period has elapsed since the patient was discharged from the hospital. A model generating device according to any one of claims 4 to 6.

8. The first time point is the time point when the patient is admitted to a hospital.

8. A model generating device according to claim 3 or 7.

9. acquiring learning data including a first index value indicating a state of a sample patient at a first time point and a second index value indicating a state of the sample patient at a second time point after the first time point; learning a prediction model that predicts a condition of the subject patient at the second time point based on a condition of the subject patient at the first time point; In the learning of the prediction model, the prediction model is learned based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and on change information indicating a change trend of the condition over time, so that a change between the first and third index values ​​approaches a first change trend which is a change trend of the condition from the first time point to the second time point indicated by the change information. A method for generating models using computer software.

10. To your computer acquiring learning data including a first index value indicating a state of a sample patient at a first time point and a second index value indicating a state of the sample patient at a second time point after the first time point; learning a prediction model that predicts a condition of the subject patient at the second time point based on a condition of the subject patient at the first time point; A computer program that executes a model generation method in which, in training the prediction model, the prediction model is trained based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value, and change information indicating the trend of change in the condition over time, so that the change between the first and third index values ​​approaches a first change trend, which is the trend of change in the condition from the first time point to the second time point represented by the change information.

11. A status acquisition means for acquiring status information regarding a status of a subject patient at a first time point; a prediction means for predicting a condition of the target patient at the second time point using the condition information and a prediction model for predicting a condition of the target patient at a second time point after the first time point based on the condition of the target patient at the first time point; Equipped with The prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and change information indicating a change trend of the condition over time so that a change between the first and third index values ​​approaches a first change trend, which is a change trend of the condition from the first time point to the second time point represented by the change information. State prediction device.

12. Obtaining status information regarding a status of a subject patient at a first time point; predicting a condition of the target patient at the second time point using the condition information and a prediction model that predicts a condition of the target patient at a second time point after the first time point based on the condition of the target patient at the first time point; The prediction model is a model generated by acquiring learning data including a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and change information indicating a change trend of the condition over time so that a change between the first and third index values ​​approaches a first change trend, which is a change trend of the condition from the first time point to the second time point represented by the change information. A method for predicting conditions using computer software.

13. On the computer, Obtaining status information regarding a status of a subject patient at a first time point; predicting a condition of the target patient at the second time point using the condition information and a prediction model that predicts a condition of the target patient at a second time point after the first time point based on the condition of the target patient at the first time point; a first index value indicating the condition of the sample patient at the first time point and a second index value indicating the condition of the sample patient at the second time point, and learning, based on a third index value indicating the condition of the sample patient at the second time point predicted by the prediction model based on the first index value and change information indicating a change trend of the condition over time, so that the change between the first and third index values ​​approaches a first change trend, which is the change trend of the condition from the first time point to the second time point represented by the change information.

Images