Statistical data acquisition device, contribution degree calculation device, treatment behavior exploration device, treatment target exploration device, statistical data acquisition program, contribution degree calculation program, treatment behavior exploration program, and treatment target exploration program

Abstract

A learner (26) is trained so as to predict and output a result of a treatment action on a treatment object from attribute information representing an attribute of the treatment object and treatment information representing the content of the treatment action on the treatment object, on the basis of a training data group (16). A statistical data acquisition unit (32) inputs to the trained learner (26) an input data group (38) that includes a plurality of input data (40) including input attribute information (42) representing an attribute of the treatment object and input treatment information (44) representing the content of a treatment action on the treatment object, the input attribute information (42) being mutually different and the input treatment information (44) being mutually the same in the input data group (38), and acquires an output data group (46) which are statistical data of the predicted result of the treatment action indicated by the input treatment information (44).

Description

Technical Field

[0001] This specification discloses a statistical data acquisition device, a contribution degree calculation device, a treatment behavior exploration device, a treatment target exploration device, a statistical data acquisition program, a contribution degree calculation program, a treatment behavior exploration program, and a treatment target exploration program.

Background Art

[0002] Conventionally, the use of learning devices has been carried out in the medical field. For example, Non-Patent Document 1 discloses a system that outputs, as an objective score, the drug to be administered to a patient and its drug efficacy based on patient information such as cancer, genes, and mutations. Non-Patent Document 2 discloses a method of having AI (Artificial Intelligence) propose suspected disease names, recommended treatment methods, etc. based on examination data and symptom data.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Non-Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] By the way, in order to put a certain treatment into practical use, statistical data consisting of a vast amount of data indicating the results of the treatment is required. Here, since obtaining a vast amount of data requires a great deal of cost, labor, or time, it has been a bottleneck in the practical implementation of the treatment. Note that the treatment in this specification is an act performed based on medicine for the treatment, diagnosis, or prevention of injuries or illnesses of humans or animals. In addition, since the treatment also includes administering pharmaceuticals to humans or animals, the results of the treatment may indicate the effects and safety of the pharmaceuticals. In particular, if the treatment aimed at practical implementation is a treatment for humans, the statistical data will be statistical data for clinical trials, and if the treatment aimed at practical implementation is a treatment using pharmaceuticals or medical devices, the statistical data may be statistical data for clinical trials.

[0005] Also, when the attribute information representing the attributes of the treatment target (for example, physical characteristics or health status) or the treatment information representing the content of the treatment has a plurality of data items, it may be beneficial to grasp to what extent each data item affects the result of the treatment.

[0006] Furthermore, it may be beneficial to identify a treatment suitable for a certain treatment target.

[0007] An object of the present disclosure is to reduce the cost, labor, or time for obtaining statistical data indicating the results of a treatment. Alternatively, an object of the present disclosure is to make it possible to easily grasp to what extent each of the plurality of data items of the attribute information of the treatment target and the plurality of data items of the treatment information affects the result of the treatment. Alternatively, an object of the present disclosure is to propose a treatment suitable for a certain treatment target.

Means for Solving the Problem

[0008] The statistical data acquisition device disclosed in this specification uses learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, and is trained to predict and output the results of the treatment actions for a treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment actions for the treatment target. An input data group including a plurality of input data including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment actions for the treatment target, wherein the input attribute information of each of the input data is different from each other, and the input treatment information of each of the input data is the same as each other. By inputting the input data group, a statistical data acquisition unit that acquires statistical data of the prediction results of the treatment actions indicated by the input treatment information is provided.

[0009] The value of the input attribute information may follow a predetermined distribution.

[0010] A distribution adjustment unit that adjusts the distribution based on a plurality of distributions for determining the value of the input attribute information, the statistical data that is the output data of the learned learning device when the input attribute information according to each distribution is included in the input data, and the target statistical data desired by the user, so that the statistical data output by the learned learning device approaches the target statistical data may be further provided.

[0011] The statistical data acquisition unit may acquire an output error distribution that is a distribution of output errors of the learned learning device, and correct the results of the treatment actions indicated by the input treatment information based on the output error distribution.

[0012] The learning device may be learned using virtual learning data created based on statistical information on past treatment actions.

[0013] The learning device may be trained using virtual learning data created to follow the distribution of the treatment actions indicated by the statistical information on the treatment actions performed in the past.

[0014] After being trained using the first learning data group, the learning device may be retrained using a second learning data group different from the first learning data group.

[0015] Further, the contribution degree calculation device disclosed in this specification uses learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, and inputs first input data including first input attribute information representing the attributes of a treatment target and first input treatment information representing the content of a treatment action for the treatment target into a learning device trained to predict and output the results of the treatment action for the treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment action for the treatment target, thereby obtaining a first prediction result, and inputs second input data including second input attribute information and second input treatment information obtained by changing one of a plurality of data items included in the first input attribute information and the first input treatment information into the learning device, thereby obtaining a second prediction result, and includes a contribution degree calculation unit that calculates the contribution degree of the data item regarding the output of the learning device based on the difference between the first prediction result and the second prediction result.

[0016] In addition, the treatment behavior exploration device disclosed in this specification uses learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment behaviors for the past treatment targets, and learning treatment results representing the results of the treatment behaviors for the past treatment targets. Based on attribute information representing the attributes of a treatment target and treatment information representing the content of a treatment behavior for the treatment target, a learning device learned to predict and output the result of the treatment behavior for the treatment target is provided with a plurality of input data each including input attribute information representing the attributes of a treatment target and input treatment information representing the content of a treatment behavior for the treatment target, and the input treatment information of each of the input data is different from each other. Based on the prediction results of a plurality of different treatment behaviors for a predetermined treatment target obtained by inputting the plurality of input data, a treatment behavior exploration unit that explores a treatment behavior suitable for the predetermined treatment target is provided.

[0017] In addition, the treatment target exploration device disclosed in this specification uses learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment behaviors for the past treatment targets, and learning treatment results representing the results of the treatment behaviors for the past treatment targets. Based on attribute information representing the attributes of a treatment target and treatment information representing the content of a treatment behavior for the treatment target, a learning device learned to predict and output the result of the treatment behavior for the treatment target is provided with a plurality of input data each including input attribute information representing the attributes of a treatment target and input treatment information representing the content of a treatment behavior for the treatment target, and the input attribute information of each of the input data is different from each other. Based on the prediction results of a predetermined treatment behavior for a plurality of different treatment targets obtained by inputting the plurality of input data, a treatment target exploration unit that explores a treatment target suitable for the predetermined treatment behavior is provided.

[0018] In addition, the statistical data acquisition program disclosed in this specification causes a computer to use learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions performed on the past treatment targets, and learning treatment results representing the results of the treatment actions performed on the past treatment targets, and to use the learning device learned to predict and output the results of the treatment actions for a treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment actions for the treatment target. An input data group including a plurality of input data including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment actions for the treatment target, wherein the input attribute information of each of the input data is different from each other, and the input treatment information of each of the input data is the same as each other. By inputting the input data group, it functions as a statistical data acquisition unit that acquires statistical data on the prediction results of the treatment actions indicated by the input treatment information.

[0019] In addition, the contribution degree calculation program disclosed in this specification causes a computer to use learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions performed on the past treatment targets, and learning treatment results representing the results of the treatment actions performed on the past treatment targets, and to use the learning device learned to predict and output the results of the treatment actions for a treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment actions for the treatment target. By inputting first input data including first input attribute information representing the attributes of the treatment target and first input treatment information representing the content of the treatment actions for the treatment target, a first prediction result is obtained. By inputting second input data including second input attribute information and second input treatment information obtained by changing one of a plurality of data items included in the first input attribute information and the first input treatment information into the learning device, a second prediction result is obtained. Based on the difference between the first prediction result and the second prediction result, it functions as a contribution degree calculation unit that calculates the contribution degree of the data item regarding the output of the learning device.

[0020] In addition, the treatment behavior exploration program disclosed in this specification causes a computer to use learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment behaviors for the past treatment targets, and learning treatment results representing the results of the treatment behaviors for the past treatment targets, and to use the learning device learned to predict and output the results of the treatment behaviors for the treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment behaviors for the treatment target. Each of a plurality of input data including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment behaviors for the treatment target, and the input treatment information of each of the input data is different from each other. Based on the prediction results of a plurality of different treatment behaviors for a predetermined treatment target obtained by inputting the plurality of input data, a treatment behavior exploration unit that explores a treatment behavior suitable for the predetermined treatment target is made to function. This is a characteristic.

[0021] In addition, the treatment target exploration program disclosed in this specification causes a computer to use learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment behaviors for the past treatment targets, and learning treatment results representing the results of the treatment behaviors for the past treatment targets, and to use the learning device learned to predict and output the results of the treatment behaviors for the treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment behaviors for the treatment target. Each of a plurality of input data including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment behaviors for the treatment target, and the input attribute information of each of the input data is different from each other. Based on the prediction results of a predetermined treatment behavior for a plurality of different treatment targets obtained by inputting the plurality of input data, a treatment target exploration unit that explores a treatment target suitable for the predetermined treatment behavior is made to function. This is a characteristic.

Advantages of the Invention

[0022] According to the present disclosure, it is possible to reduce the cost, labor, or time for obtaining statistical data indicating the results of a treatment action. Alternatively, according to the present disclosure, it is possible to easily grasp the degree to which each of a plurality of data items of attribute information of a treatment target and a plurality of data items of treatment information affects the result of a treatment action. Alternatively, according to the present disclosure, it is possible to propose a treatment action suitable for a certain treatment target.

Brief Description of the Drawings

[0023]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Embodiments for Carrying Out the Invention

[0024] FIG. 1 is a schematic configuration diagram of server 10 as a statistical data acquisition device, contribution degree calculation device, treatment behavior exploration device, or treatment target exploration device according to the present embodiment. Note that as long as the functions described below can be exhibited, the statistical data acquisition device, contribution degree calculation device, treatment behavior exploration device, or treatment target exploration device may be realized by a computer other than server 10. Further, each of the functions described below may be realized by the cooperation of a plurality of computers.

[0025] The communication interface 12 is composed of, for example, a NIC (Network Interface Card) or a short-range wireless communication adapter. The communication interface 12 exhibits a function of communicating with other computers. For example, the communication interface 12 can receive various data necessary for the processes described below from other computers. Further, the communication interface 12 can transmit information indicating the result of the process to other computers.

[0026] The memory 14 is composed of, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), an eMMC (embedded Multi Media Card), a ROM (Read Only Memory), or a RAM (Random Access Memory). A computer program as a statistical data acquisition program, contribution degree calculation program, treatment behavior exploration program, or treatment target exploration program for operating each part of server 10 is stored in the memory 14. Note that the computer program can also be stored in a computer-readable non-transitory storage medium such as a USB (Universal Serial Bus) memory or a CD-ROM. Server 10 can read and execute the computer program from such a storage medium.

[0027] As shown in FIG. 1, a learning data group 16 is stored in the memory 14. The learning data group 16 is a data group for training a learning device 26 described later. FIG. 2 is a diagram showing an example of the learning data group 16. The learning data group 16 is composed of a plurality of learning data 18. Each learning data 18 includes learning attribute information 20 representing the attributes of a past treatment target, learning treatment information 22 representing the content of the treatment action for the treatment target, and learning treatment result 24 representing the result of the treatment action for the treatment target. Here, the treatment target includes humans or animals. Further, the attributes of the treatment target include the physical characteristics of the treatment target (for example, gender, age, height, weight, blood type, underlying disease, allergy, etc.) and the health state (including the mental state). Further, the treatment action is an action performed based on medicine for treating, diagnosing, or preventing the injury or illness of the treatment target, and the treatment action also includes administering a medicine to a human or an animal. Among the learning data 18, the learning treatment result 24 serves as teacher data.

[0028] In the present embodiment, the learning attribute information 20 includes a plurality of data items 20a (for example, each parameter indicating physical characteristics or health state, etc.). Further, in the present embodiment, the learning treatment information 22 also includes a plurality of data items 22a (for example, each parameter representing the treatment content, etc.). Furthermore, in the present embodiment, the learning treatment result 24 also includes a plurality of data items 24a (for example, each parameter related to the treatment target after treatment, etc.). Since the learning data 18 is data input to the learning device 26 or used for comparison with the output data of the learning device 26, the values of each data item 20a, 22a, 24a are data obtained by converting data indicating the attributes of the treatment target, the content of the treatment action, or the result of the treatment action into numerical values.

[0029] The learning data group 16 is prepared in advance by a user of the server 10 or the like and stored in the memory 14. Further, the learning data group 16 may be stored in the memory of another device accessible from the server 10 instead of the memory 14.

[0030] Returning to FIG. 1, a learning device 26 is further stored in the memory 14. The learning device 26 is a program that performs machine learning based on the learning data group 16. As the structure of the learning device 26, typically a neural network, the structure of the learning device 26 is not limited thereto as long as it can exhibit the functions described below. The learning device 26 may also be stored in the memory of another device accessible from the server 10 instead of the memory 14. The learning method and processing content of the learning device 26 will be described together with the processor 28 described later.

[0031] The processor 28 is configured to include at least one of a general-purpose processing device (such as a CPU (Central Processing Unit)) and a dedicated processing device (such as an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a programmable logic device). The processor 28 may be configured not by a single processing device but by the cooperation of a plurality of physically separated processing devices. As shown in FIG. 1, the processor 28 functions as a learning processing unit 30, a statistical data acquisition unit 32, a contribution degree calculation unit 34, and a search unit 36 according to the computer program stored in the memory 14.

[0032] The learning processing unit 30 executes a learning process for training the learning device 26 using the learning data group 16. Specifically, the learning processing unit 30 inputs the learning attribute information 20 and the learning treatment information 22 among the learning data 18 (see FIG. 2) into the learning device 26. The learning device 26 predicts the result of the treatment action indicated by the input learning treatment information 22 for the treatment target indicated by the input learning attribute information 20, and outputs the prediction result as output data. The learning processing unit 30 calculates the difference between the output data and the learning treatment result 24 which is the teacher data among the learning data 18, and adjusts the parameters of the learning device 26 (for example, if the learning device 26 is a neural network, the weights and biases of each neuron) so that the difference becomes smaller. The learning processing unit 30 trains the learning device 26 by repeating the said process.

[0033] The learning processing unit 30 may perform preprocessing for correcting data loss, imbalance, outliers, etc. on each learning data 18 included in the learning data group 16, and then train the learning device 26 using the preprocessed learning data group 16. For example, when there is a loss in the data item 20a, data item 22a, or data item 24a of the learning data 18, the learning processing unit 30 can complement the missing data based on known literature information. By performing preprocessing on the learning data 18, the learning efficiency of the learning device 26 can be improved.

[0034] Also, the learning processing unit 30 may correct the data item 20a, data item 22a, or data item 24a of the learning data 18 using a correction model that estimates the result of the treatment action for a treatment target other than the said treatment target from the result of the treatment action for a certain treatment target based on known literature information. Thereby, for example, based on the results of clinical trials conducted in other countries, learning data 18 indicating the results of clinical trials if they were conducted in one's own country can be obtained.

[0035] In addition, for example, in documents and the like, statistical information regarding treatment actions performed in the past may be shown. The statistical information regarding treatment actions performed in the past includes statistical information regarding the attributes of the treatment targets of the treatment actions and statistical information regarding the treatment results of the treatment actions. The statistical information regarding the attributes of the treatment targets includes, for example, the proportion of treatment targets belonging to each class (for example, when the attribute is age, the class of 0 to 10 years old, the class of 11 to 20 years old, etc.) for each attribute of the treatment target (such as age, gender, disease, etc.). Similarly, the statistical information regarding the treatment results includes, for example, the proportion of treatment results belonging to each class for each attribute of the treatment results (such as effect, variation in effect, etc.). The learning processing unit 30 may generate virtual learning data 18 based on such statistical information regarding treatment actions performed in the past. Then, the learning processing unit 30 may use the virtual learning data 18 to train the learner 26.

[0036] In particular, the statistical information regarding treatment actions performed in the past can be regarded as showing the distribution regarding the treatment actions. The distribution regarding the treatment actions means the distribution of the treatment targets of the treatment actions and the distribution of the treatment results of the treatment actions. Therefore, the learning processing unit 30 may create virtual learning data 18 so as to follow the distribution regarding treatment actions performed in the past. Creating virtual learning data 18 so as to follow the distribution regarding treatment actions performed in the past means that the values of each learning attribute information 20 (specifically, each data item 20a) included in the virtual learning data 18 follow the statistical information regarding the treatment targets of the treatment actions performed in the past, and the values of each learning treatment information 22 (specifically, each data item 22a) included in the virtual learning data 18 follow the statistical information regarding the treatment results of the treatment actions performed in the past, to generate the virtual learning data 18.

[0037] Further, after the learning processing unit 30 has learned the learning device 26 using the first learning data group 16, the learning device 26 may be relearned using a second learning data group 16 different from the first learning data group 16. For example, the learning processing unit 30 learns the learning device 26 using the first learning data group 16 indicating the results of treatment acts performed in other countries, and then, using the second learning data group 16 indicating the results of treatment acts performed in its own country, can further learn the learning device 26 that has already been learned using the first learning data group 16. Thereby, even when there is little learning data indicating the results of treatment acts within the country, a learning device 26 that has been sufficiently learned and finely tuned for the country can be obtained.

[0038] The sufficiently learned learning device 26 can accurately predict and output the result of the treatment act for the treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment act for the treatment target.

[0039] In this embodiment, the learning process of the learning device 26 is executed in the server 10, but the learning process of the learning device 26 does not necessarily have to be performed in the server 10. A sufficiently learned learning device 26 may be stored in the memory 14 in another device. In this case, the processor 28 does not necessarily have to have the functions of the learning processing unit 30.

[0040] The statistical data acquisition unit 32 acquires statistical data indicating the prediction result of a certain treatment act by inputting an input data group composed of a plurality of input data into the learned learning device 26 (hereinafter may be simply referred to as the learning device 26).

[0041] FIG. 3 and FIG. 4 are diagrams showing an example of the input data group 38 input to the learning device 26. The input data group 38 is configured to include a plurality of input data 40. Each input data 40 is configured to include input attribute information 42 representing the attributes of the treatment target and input treatment information 44 representing the content of the treatment action for the treatment target. Similar to the learning data 18, in the present embodiment, the input attribute information 42 includes a plurality of data items 42a, and the input treatment information 44 also includes a plurality of data items 44a.

[0042] As shown in FIGS. 3 and 4, in the input data group 38 input by the statistical data acquisition unit 32 to the learning device 26, the input attribute information 42 of each input data 40 is different from each other, and the input treatment information 44 of each input data 40 is the same as each other.

[0043] The statistical data acquisition unit 32 can randomly generate the input attribute information 42 of each input data 40. As described above, in the present embodiment, since the input attribute information 42 has a plurality of data items 42a, the statistical data acquisition unit 32 can randomly generate each data item 42a.

[0044] The treatment action indicated by the input treatment information 44 may be set to be the treatment action desired by the user (the generator of the statistical data) of the server 10. Note that, as one of the treatment actions indicated by the input treatment information 44, "not performing treatment" may be set. In that case, for example, as shown in FIG. 4, the value of the input treatment information 44 (more specifically, each data item 44a) of each input data 40 is set to "0.00".

[0045] By inputting such an input data group 38 into the learning device 26, the learning device 26 outputs an output data group 46. FIG. 5 is a diagram showing an example of the output data group 46 of the learning device 26 with respect to the input data group 38. The output data group 46 is configured to include a plurality of output data 48. Each output data 48 is data output by the learning device 26 for each input data 40, and is data indicating the prediction result of the treatment action for the treatment target indicated by each input data 40. Similar to the learning data 18, each output data 48 may include a plurality of data items 48a.

[0046] The output data group 46 is the prediction result of the treatment action when the same treatment action is performed for each of a plurality of treatment targets having different attributes from each other. That is, it can be said that the output data group 46 is statistical data of the prediction result of the treatment action indicated by the input treatment information 44.

[0047] According to the present embodiment, since the statistical data of the prediction result of the treatment action is obtained using the learning device 26, the cost, labor, or time for obtaining the statistical data can be reduced as compared with the prior art. For example, when a doctor or the like wants to obtain statistical data on the results (effects, side effects, etc.) of a new treatment method, conventionally, the treatment method has been repeatedly tested on a large number of treatment targets to obtain the results. Needless to say, this requires a great deal of cost, labor, and time. According to the present embodiment, it is only necessary to prepare data (that is, the learning data group 16) indicating the results of treatment actions sufficient to sufficiently train at least the learning device 26, and then, by simply inputting the input data group 38 into the trained learning device 26, the statistical data of the prediction result of the treatment action can be obtained.

[0048] In addition, it is also possible to easily obtain the statistical data when a certain treatment action is performed and when the treatment action is not performed, and it is also possible to easily compare the prediction results when a certain treatment action is performed and when the treatment action is not performed.

[0049] In addition, if the treatment target indicated by the input attribute information 42 is a human, the statistical data can be clinical trial statistical data. Further, if the treatment act indicated by the input treatment information 44 is a treatment act using a pharmaceutical product or a medical device, the statistical data can be statistical data for a clinical trial. For example, conventionally, obtaining statistical data for approval of pharmaceutical products, treatment methods, etc. has required a great deal of cost, labor, or time. However, if the statistical data according to this embodiment is recognized for approval, it becomes possible to obtain approval for new pharmaceutical products, treatment methods, etc. more simply or earlier. In particular, in the case where a new infectious disease suddenly spreads, it is preferable to obtain approval for pharmaceutical products, treatment methods, etc. for the infectious disease promptly. According to this embodiment, statistical data for obtaining approval for such pharmaceutical products can be obtained promptly. Also, for pharmaceutical products or treatment methods for rare diseases, since there are few cases, it may be difficult to obtain a sufficient number of data regarding the results (such as the presence or absence of effects and side effects) of the pharmaceutical products or treatment methods. Even in such a case, according to this embodiment, a sufficient number of data for obtaining approval for pharmaceutical products for rare diseases can be obtained.

[0050] The statistical data acquisition unit 32 may generate the value of the input attribute information 42 of each input data 40 so as to follow a predetermined distribution. In this embodiment, since the input attribute information 42 has a plurality of data items 42a, the statistical data acquisition unit 32 generates the value of each data item 42a so as to follow a predetermined distribution.

[0051] FIG. 6 is a diagram showing the distribution of each data item 42a of the input attribute information 42. FIG. 6(a) is a graph showing the distribution of the data item 42a named "A", FIG. 6(b) is a graph showing the distribution of the data item 42a named "B", and FIG. 6(c) is a graph showing the distribution of the data item 42a named "C". In the graphs respectively shown in FIGS. 6(a) to 6(c), the horizontal axis indicates the value of the data item 42a, and the vertical axis indicates the frequency of the value. Such a distribution is stored in the memory 14 in advance by, for example, a user of the server 10. Alternatively, such a distribution may be determined by a distribution adjustment unit 33 described later.

[0052] The statistical data acquisition unit 32 generates the input data group 38 by randomly setting the values of the respective data items 42a of the input attribute information 42 so as to follow the distributions shown in FIGS. 6(a) to 6(c). In this way, the user of the server 10 can obtain statistical data indicating the prediction result of the treatment behavior for the population of the desired treatment target.

[0053] Even for a well-trained learning device 26, an output error may occur. The output error means the difference between the output data 48 of the learning device 26 for a certain input data 40 and the true (correct) result of the treatment behavior for the treatment target indicated by the input data 40. Therefore, the statistical data acquisition unit 32 may acquire the output error distribution, which is the distribution of the output errors of the learning device 26, and correct the output data 48 based on the output error distribution.

[0054] First, the statistical data acquisition unit 32 acquires the learning error distribution of the trained learning device 26. The learning error distribution is information indicating the relationship between the magnitude of the error of the learning device 26 and the probability of occurrence of the error. For example, the learning error distribution can be represented by a graph in which the horizontal axis indicates the magnitude of the error and the vertical axis indicates the probability of occurrence of the error.

[0055] The statistical data acquisition unit 32 acquires the learning error distribution using a plurality of verification data (a combination of attribute information representing the attributes of the treatment target, treatment information representing the content of the treatment behavior for the treatment target, and treatment results indicating the results of the treatment behavior) different from the learning data 18. Specifically, the statistical data acquisition unit 32 inputs the attribute information and the treatment information among the verification data into the trained learning device 26, and acquires the difference between the output data of the learning device 26 for the input and the treatment result of the verification data as the error. By performing such processing for a plurality of verification data, the learning error distribution of the learning device 26 is acquired. Also, by preparing a plurality of trained learning devices 26 (for example, the learning processing unit 30 trains a plurality of learning devices 26) and calculating the statistical values (such as the average value and the standard deviation) of the differences between the output data of each of the plurality of learning devices 26 and the treatment results of the verification data, the learning error distribution can be acquired.

[0056] Based on the acquired learning error distribution of the learning device 26, the statistical data acquisition unit 32 corrects the output data 48. For example, the statistical data acquisition unit 32 generates a random number based on the learning error distribution, and corrects the output data 48 by adding the random number to the output data 48. FIG. 7 shows the output data 48 before correction and the corrected output data 48' corrected based on the learning error distribution of the learning device 26.

[0057] Also, based on the learning error distribution of the learning device 26, the output data 48, and the corrected output data 48', the robustness of the learning device 26 can be evaluated. The robustness of the learning device 26 is an index indicating how resistant the output data 48 (prediction result of the treatment behavior) is to the output error of the learning device 26.

[0058] Returning to FIG. 1, the distribution adjustment unit 33 adjusts the above-described distribution (see FIG. 6) that the statistical data acquisition unit 32 refers to in order to generate the values of the input attribute information 42 (specifically, each data item 42a) of each input data 40 input to the learned learning device 26.

[0059] First, on the premise that the statistical data of the output data of the learned learning device 26, that is, the prediction result of the treatment behavior, can change according to the distribution for determining the value of the input attribute information 42. Also, as the statistical data of the prediction result of the treatment behavior, there is ideal statistical data (for example, when the treatment behavior is medication, the average value of the magnitude of the drug effect when administered to each treatment target is large, and the statistical data with a small variation in the magnitude of the drug effect, etc.). The ideal data can be said to be the target statistical data aimed at by the user of the server 10.

[0060] Therefore, the distribution adjustment unit 33 may adjust the distribution so that the statistical data output by the trained learning device 26 approaches the target statistical data based on a plurality of distributions for determining the value of the input attribute information 42, the statistical data that is the output data of the trained learning device 26 when the input attribute information 42 according to each distribution is included in the input data, and the target statistical data that the user aims for. The process of adjusting the distribution for determining the value of the input attribute information 42 may be referred to as the distribution optimization process if it searches for a distribution such that the statistical data output by the trained learning device 26 becomes the target statistical data. However, the optimization process does not necessarily obtain the optimal solution (in this case, the distribution that can obtain the target statistical data), but means a process of searching towards the optimal solution.

[0061] Various methods can be adopted as the optimization process. For example, the distribution adjustment unit 33 optimizes the distribution for determining the value of the input attribute information 42 by a genetic algorithm. The genetic algorithm is the following process. First, a plurality of "individuals" (also called "chromosomes") to be optimized are randomly prepared. An individual is composed of a plurality of parameters called "genes". The plurality of individuals are called the first generation. A parameter called "fitness" for each of the plurality of individuals in the first generation is calculated by a predetermined evaluation function. Then, based on the calculated fitness of each individual, genetic operations are performed on the plurality of individuals in the first generation. Genetic operations include, for example, "selection" for selecting individuals according to fitness, "crossover" for exchanging genes between a plurality of individuals, and "mutation" for changing some genes of an individual. The plurality of individuals subjected to genetic operations become the next generation (here, the second generation). Then, the fitness for each individual in the second generation is calculated. Such a process is repeated until a predetermined stop condition (such as up to a specific number of generations, or the average value of the fitness of a plurality of individuals in a certain generation being equal to or greater than a predetermined value) is satisfied. As a result, individuals with high fitness can be obtained.

[0062] Specifically, the distribution adjustment unit 33 inputs input data including a plurality of input attribute information 42 (which becomes the "individuals" of the first generation) created based on a plurality of initial distributions into the respective trained learning machines 26. The learning machine 26 outputs statistical data of the prediction results of treatment actions for each input data corresponding to each distribution. The distribution adjustment unit 33 performs statistical analysis on the plurality of statistical data (the plurality of statistical data corresponding to each distribution) thus obtained. For example, if the treatment action is drug administration, the average value of the drug efficacy or the variation in the drug efficacy in each statistical data is acquired. The result of the statistical analysis (the average value of the drug efficacy or the variation in the drug efficacy) corresponds to the "fitness" in the genetic algorithm. Here, it is assumed that the target statistical data is statistical data such that the average value of the drug efficacy is as large as possible and the variation in the drug efficacy is as small as possible. That is, in this case, the optimization process is a process of searching for a distribution such that the average value of the drug efficacy is as large as possible and the variation in the drug efficacy is as small as possible in the statistical data.

[0063] Thereafter, based on the results of the statistical analysis of the statistical data corresponding to each distribution, genetic operations are performed on the plurality of distributions of the first generation, and a plurality of distributions of the second generation are obtained. By repeating such a process, the distribution for determining the values of the input attribute information 42 is optimized.

[0064] In actual clinical trials, from the perspective of cost reduction, there is a desire to reduce the amount of clinical trial data (a combination of information indicating the trial subjects and information indicating the treatment effect of a predetermined treatment action on them) as much as possible. Here, according to the trial subjects according to the optimized distribution that can obtain the above-mentioned target statistical data, the effect that the number of trial subjects (that is, the number of clinical trial data) can be reduced is also achieved. This is because the more ideal the statistical data of the results of the treatment action is, the more the number of clinical trial data can be reduced. For example, the larger the average value of the drug efficacy or the smaller the variation in the drug efficacy, the more the number of clinical trial data can be reduced.

[0065] Based on the differences among the multiple output data 48 output by the trained learner 26 for the multiple input data 40, the contribution degree calculation unit 34 calculates the contribution degree indicating how much the data items 42a of the input attribute information 42 or the data items 44a of the input treatment information 44 included in the input data 40 contribute to the output data 48.

[0066] First, the contribution degree calculation unit 34 selects the first input data 40 from the input data group 38 (see FIG. 3). The input attribute information 42 of the first input data 40 is referred to as the first input attribute information 42, and the input treatment information 44 of the first input data 40 is referred to as the first input treatment information 44. The contribution degree calculation unit 34 inputs the first input data 40 into the learner 26 and obtains the first output data 48 as the first prediction result. Next, the contribution degree calculation unit 34 slightly changes one of the multiple data items 42a and 44a included in the first input attribute information 42 and the first input treatment information 44 to obtain the second input data 40 including the second input attribute information 42 and the second input treatment information 44. Next, the contribution degree calculation unit 34 inputs the second input data 40 into the learner 26 and obtains the second output data 48 as the second prediction result. Then, the contribution degree calculation unit 34 calculates the difference between the first output data 48 and the second output data 48.

[0067] While changing the data items 42a and 44a to be changed between the first input data 40 and the second input data 40, the contribution degree calculation unit 34 calculates the difference between the first output data 48 and the second output data 48 when each of the data items 42a and 44a is changed (hereinafter referred to as "the difference for each of the data items 42a and 44a").

[0068] The table on the left side of the arrow in Fig. 8 shows the absolute values of the differences for each data item 42a, 44a. In this embodiment, the contribution calculation unit 34 calculates the absolute value of the difference for each data item 42a, 44a for each data item 48a that the output data 48 has, and Fig. 8 shows this. Also, as shown in Fig. 8, the contribution calculation unit 34 calculates the total value of the absolute values of the differences for each data item 42a, 44a for each data item 48a of the output data 48. For example, in the example of Fig. 8, the total value of the data item 48a named "a" is "21.71", and the total value of the data item 48a named "b" is "1.53".

[0069] The table on the right side of the arrow in Fig. 8 is a table showing the contribution degrees of each data item 42a, 44a to the output data 48 obtained based on the table on the left side. The contribution calculation unit 34 normalizes the absolute values of the differences for each data item 42a, 44a so that the total value of each data item 48a of the output data 48 becomes "100". The normalized values are shown in the table on the right side of Fig. 8, and the normalized numerical values represent the contribution degrees of each data item 42a, 44a to the output data 48. In this embodiment, the contribution degrees of each data item 42a, 44a are shown for each data item 48a that the output data 48 has.

[0070] Note that in this embodiment, the contribution degree is calculated based on the absolute value of the difference for each data item 42a, 44a, but the contribution degree of each data item 42a, 44a may be calculated by dividing according to the positive or negative of the difference between the first output data 48 and the second output data 48. Also, the contribution calculation unit 34 may perform a statistical analysis (such as a t-test or a p-test, etc.) on the contribution degrees of each data item 42a, 44a and verify the statistical certainty.

[0071] Returning to FIG. 1 again, the search unit 36 inputs a plurality of input data 40 with different input treatment information 44 into the trained learning device 26, and based on a plurality of output data 48 of the learning device 26 obtained thereby, that is, prediction results of different treatment actions for a predetermined treatment target, searches for a treatment action suitable for the predetermined treatment target. In this way, the search unit 36 functions as a treatment action search unit.

[0072] If the process of searching for a treatment action searches for an optimal treatment action for a predetermined treatment target, this process may be referred to as an optimization process of the treatment action. However, here too, the optimization process is not necessarily limited to obtaining an optimal solution (in this case, an optimal treatment action), but means a process of searching towards the optimal solution. Also, the optimal solution may be a certain fixed value or may indicate a range with a certain width.

[0073] As a method for optimizing the treatment action using the learning device 26, various methods can be adopted. Here, an optimization process using a genetic algorithm will be described.

[0074] In the genetic algorithm for searching for a treatment action, the "individual" is the input treatment information 44, the "gene" is the data item 44a of the input treatment information 44, and the "fitness" is calculated based on the output data 48 of the learning device 26 for the input data 40 including the input treatment information 44 as an individual.

[0075] Specifically describe the genetic algorithm for exploring treatment behaviors. First, a plurality of input data 40 are prepared in which the input attribute information 42 is the same for each other and the input treatment information 44 is different for each other. The treatment target indicated by the input attribute information 42 here may be the treatment target desired by the user of the server 10. For example, the attributes of the patient to whom the treatment method to be explored will be applied are set. The input treatment information 44 of each input data 40 may be randomly generated by the exploration unit 36. The plurality of input treatment information 44 included in the plurality of input data 40 prepared in this way correspond to the individuals of the genetic algorithm, and each data item 44a included in the input treatment information 44 corresponds to the gene of the genetic algorithm.

[0076] Next, the exploration unit 36 sequentially inputs the plurality of input data 40 into the learner 26 and obtains a plurality of output data 48 corresponding to each input data 40. The exploration unit 36 calculates the difference between the ideal result data obtained by quantifying the result of the ideal treatment behavior in the form of the output data 48 and each output data 48. Then, based on the difference, the exploration unit 36 calculates a difference parameter, which is a parameter that becomes larger as the difference is smaller and smaller as the difference is larger. The difference parameter is a parameter for the input data 40 (especially the input treatment information 44) and corresponds to the fitness of the genetic algorithm.

[0077] Based on the difference parameter as the fitness of the genetic algorithm for each input treatment information 44 as an individual of the genetic algorithm, the exploration unit 36 performs genetic operations on each input treatment information 44. For example, operations such as selection, crossover, or mutation are performed on each data item 44a included in each input treatment information 44. As a result, a plurality of input data 40 including the input treatment information 44 of the next generation (here, the second generation) are generated. Note that the exploration unit 36 does not change the input attribute information 42. That is, the input attribute information 42 is fixed.

[0078] Thereafter, the search unit 36 repeats the above-described process. That is, a plurality of input data 40 including the second-generation input treatment information 44 are input to the learning device 26, and genetic operations are performed on each input treatment information 44 based on the difference parameters of each input treatment information 44 obtained from the output data 48, thereby generating a plurality of input data 40 including the next-generation input treatment information 44.

[0079] The search unit 36 repeats the above-described process until a predetermined end condition is satisfied, and among the plurality of input treatment information 44 obtained at the end time, the treatment action indicated by the input treatment information 44 having the largest difference parameter is specified as the treatment action suitable for the treatment target.

[0080] Note that the input attribute information 42 of the plurality of input data 40 used in the above-described genetic algorithm may be of one type, or a plurality of input data 40 having a plurality of mutually different input attribute information 42 may be used. For example, the plurality of input attribute information 42 may be determined according to a predetermined distribution (see FIG. 6) or a distribution adjusted by the distribution adjustment unit 33. An example of the plurality of input data 40 in that case is shown in FIG. 9. Also in this case, the basic process is the same as the above-described process, and the plurality of input attribute information 42 are not changed during the genetic algorithm. However, the difference parameter as the fitness is calculated based on a plurality of differences between the ideal result data and each output data 48 for each input data 40. For example, the difference parameter is calculated based on the average value of the plurality of differences.

[0081] In addition, the search unit 36 may input a plurality of input data 40 having mutually different input attribute information 42 to the learned learning device 26, and search for a treatment target suitable for a predetermined treatment action based on the plurality of output data 48 of the learning device 26 obtained thereby, that is, the prediction results of predetermined treatment actions for a plurality of mutually different treatment targets. In this way, the search unit 36 also functions as a treatment target search unit.

[0082] If the process of searching for a treatment target is to search for an optimal treatment target for a predetermined treatment action, then this process can be called an optimization process for the treatment target.

[0083] As a method for optimizing the treatment target using the learning device 26, various methods can be adopted. Here too, an optimization process using a genetic algorithm will be described.

[0084] In the genetic algorithm for searching for a treatment target, the "individual" is the input attribute information 42, the "gene" is the data item 42a of the input attribute information 42, and the "fitness" is calculated based on the output data 48 of the learning device 26 for the input data 40 including the input attribute information 42 as an individual.

[0085] A genetic algorithm for searching for a treatment target will be specifically described. First, a plurality of input data 40 are prepared in which the input attribute information 42 is different from each other and the input treatment information 44 is the same as each other. The treatment action indicated by the input treatment information 44 here may be a treatment action desired by the user of the server 10. The input attribute information 42 of each input data 40 may be randomly generated by the search unit 36. The plurality of input attribute information 42 included in the plurality of input data 40 prepared in this way corresponds to the individuals of the genetic algorithm, and each data item 42a included in the input attribute information 42 corresponds to the gene of the genetic algorithm.

[0086] Next, the search unit 36 sequentially inputs the plurality of input data 40 into the learning device 26 and acquires a plurality of output data 48 corresponding to each input data 40. The search unit 36 calculates the difference between the ideal result data obtained by quantifying the result of the ideal treatment action into the format of the output data 48 and each output data 48. Then, the search unit 36 calculates a difference parameter, which is a parameter such that the smaller the difference, the larger it becomes, and the larger the difference, the smaller it becomes, based on the difference. The difference parameter is a parameter for the input data 40 (particularly the input attribute information 42) and corresponds to the fitness of the genetic algorithm.

[0087] The exploration unit 36 performs genetic operations on each input attribute information 42 based on the difference parameter as the fitness of the genetic algorithm for each input attribute information 42 as an individual of the genetic algorithm. For example, operations such as selection, crossover, or mutation are performed on each data item 42a included in each input attribute information 42. As a result, a plurality of input data 40 including the input attribute information 42 of the next generation (here, the second generation) are generated. Note that the exploration unit 36 does not change the input treatment information 44. That is, the input treatment information 44 is fixed.

[0088] Thereafter, the exploration unit 36 repeats the above-described process. That is, a plurality of input data 40 including the input attribute information 42 of the second generation are input to the learning device 26, and genetic operations are performed on each input attribute information 42 based on the difference parameter of each input attribute information 42 obtained from the output data 48, thereby generating a plurality of input data 40 including the input attribute information 42 of the next generation.

[0089] The exploration unit 36 repeats the above-described process until a predetermined end condition is satisfied, and among the plurality of input attribute information 42 obtained at the end point, the treatment target indicated by the input attribute information 42 having the largest difference parameter is specified as the treatment target suitable for the treatment action.

[0090] Note that the input treatment information 44 of the plurality of input data 40 used in the above-described genetic algorithm may be of one type, or a plurality of input data 40 having a plurality of different input treatment information 44 may be used. For example, the plurality of input treatment information 44 may be determined according to a predetermined distribution (see FIG. 6) or a distribution adjusted by the distribution adjustment unit 33. Also in this case, the basic process is the same as the above-described process, and the plurality of input treatment information 44 is not changed during the genetic algorithm. However, the difference parameter as the fitness is calculated based on a plurality of differences between the ideal result data and each output data 48 for each input data 40. For example, the difference parameter is calculated based on the average value of the plurality of differences.

[0091] As described above, the search unit 36 searches for a treatment action suitable for a predetermined treatment target or a treatment target suitable for a predetermined treatment action. The search unit 36 may search for a treatment action suitable for a predetermined treatment target and a treatment target suitable for a predetermined treatment action.

[0092] As described above, the embodiments according to the present disclosure have been described. However, the present disclosure is not limited to the above embodiments, and various modifications are possible without departing from the spirit of the present disclosure.

Explanation of Reference Numerals

[0093] 10 Server, 12 Communication interface, 14 Memory, 16 Learning data group, 18 Learning data, 20 Learning attribute information, 20a, 22a, 24a, 42a, 44a, 48a Data items, 22 Learning treatment information, 24 Learning treatment result, 26 Learning device, 28 Processor, 30 Learning processing unit, 32 Statistical data acquisition unit, 33 Distribution adjustment unit, 34 Contribution degree calculation unit, 36 Search unit, 38 Input data group, 40 Input data, 42 Input attribute information, 44 Input treatment information, 46 Output data group, 48 Output data, 48’ Corrected output data.

Claims

1. Using learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, an input data group including a plurality of input data including input attribute information representing the attributes of a treatment target and input treatment information representing the content of a treatment action for the treatment target is input to a learning device learned to predict and output the result of the treatment action for the treatment target from the attribute information representing the attributes of the treatment target and the treatment information representing the content of the treatment action for the treatment target. In the input data group, the input attribute information of each input data is different from each other so that its value follows a predetermined distribution, and the input treatment information of each input data is the same as each other. By inputting the input data group, a statistical data acquisition unit that acquires statistical data of the prediction results of the treatment actions indicated by the input treatment information. A statistical data acquisition device characterized by comprising the above.

2. The distribution regarding the value of the input attribute information is determined by the user. The statistical data acquisition device according to claim 1, characterized by the above.

3. Based on a plurality of the distributions for determining the values of the input attribute information, the statistical data which is the output data of the learned learning device when the input attribute information according to each distribution is included in the input data, and the target statistical data desired by the user, a distribution adjustment unit that adjusts the distribution so that the statistical data output by the learned learning device approaches the target statistical data. The statistical data acquisition device according to claim 1, further characterized by comprising the above.

4. The statistical data acquisition unit acquires an output error distribution which is a distribution of output errors of the learned learning device, and corrects the result of the treatment action indicated by the input treatment information based on the output error distribution. The statistical data acquisition device according to claim 1, characterized by the above.

5. The learning device is learned using virtual learning data created based on statistical information on past treatment actions. The statistical data acquisition device according to claim 1, characterized by the above.

6. The learning device is learned using virtual learning data created so as to follow the distribution regarding the treatment action indicated by the statistical information on past treatment actions. The statistical data acquisition device according to claim 5, characterized by the above.

7. After being trained using the first training data group, the trainer is retrained using a second training data group different from the first training data group. The statistical data acquisition device according to claim 1, characterized in that.

8. (Deleted)

9. Using learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, from attribute information representing the attributes of the treatment target and treatment information representing the content of the treatment action for the treatment target, a trainer trained to predict and output the results of the treatment action for the treatment target, each of which includes input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment action for the treatment target. A plurality of input data, wherein the input attribute information of each of the input data is different from each other so that its value follows a predetermined distribution, and the input treatment information of each of the input data is different from each other. A treatment action search unit that searches for a treatment action suitable for the predetermined treatment target based on prediction results of a plurality of different treatment actions for the predetermined treatment target obtained by inputting the plurality of input data. A treatment action search device, characterized by comprising.

10. Using learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, from attribute information representing the attributes of the treatment target and treatment information representing the content of the treatment action for the treatment target, a trainer trained to predict and output the results of the treatment action for the treatment target, each of which includes input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment action for the treatment target. A plurality of input data, wherein the input attribute information of each of the input data is different from each other, and the input treatment information of each of the input data is different from each other so that its value follows a predetermined distribution. Based on prediction results of a predetermined treatment action for a plurality of different treatment targets obtained by inputting the plurality of input data, a treatment target search unit that searches for a treatment target suitable for the predetermined treatment action. A treatment target search device, characterized by comprising.

11. A computer, using learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, from attribute information representing the attributes of a treatment target and treatment information representing the content of a treatment action for the treatment target, to predict and output the result of the treatment action for the treatment target to a learner trained to do so, an input data group including a plurality of input data including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment action for the treatment target, wherein the input attribute information of each of the input data is different from each other such that its value follows a predetermined distribution, and the input treatment information of each of the input data is the same as each other, by inputting the input data group, a statistical data acquisition unit that acquires statistical data of the prediction results of the treatment action indicated by the input treatment information, characterized by causing the computer to function as the statistical data acquisition unit.

12. (Deleted)

13. A computer, using learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, from attribute information representing the attributes of a treatment target and treatment information representing the content of a treatment action for the treatment target, to predict and output the result of the treatment action for the treatment target to a learner trained to do so, a plurality of input data each including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment action for the treatment target, wherein the input attribute information of each of the input data is different from each other such that its value follows a predetermined distribution, and the input treatment information of each of the input data is different from each other, by inputting the plurality of input data, a treatment action search unit that searches for a treatment action suitable for the predetermined treatment target based on the prediction results of a plurality of different treatment actions for the predetermined treatment target obtained, characterized by causing the computer to function as the treatment action search unit.

14. A computer, Using learning data including learning attribute information representing the attributes of past treatment targets, learning treatment information representing the content of treatment actions for the past treatment targets, and learning treatment results representing the results of the treatment actions for the past treatment targets, an attribute information representing the attributes of a treatment target and a treatment information representing the content of a treatment action for the treatment target are used to train a learner to predict and output the results of the treatment action for the treatment target. A treatment target search unit that searches for a treatment target suitable for the predetermined treatment action based on a plurality of input data including input attribute information representing the attributes of the treatment target and input treatment information representing the content of the treatment action for the treatment target, wherein the input attribute information of each of the input data is different from each other, and the input treatment information of each of the input data is different from each other so that its value follows a predetermined distribution. The treatment target search unit obtains a plurality of different input data, and based on the prediction results of a predetermined treatment action for a plurality of different treatment targets obtained by inputting the plurality of different input data, searches for a treatment target suitable for the predetermined treatment action. A treatment target search program characterized by functioning as such.

Images