Medical support device, method of operating the medical support device, operating program of the medical support device

Abstract

This medical assistance device comprises a processor and a memory connected or built into the processor. The processor acquires a trial period and target input data, which is input data relating to a disease of a candidate for a pharmaceutical trial, inputs the target input data and the trial period to a machine learning model trained by using training data that includes stored input data relating to the disease at two or more points in time, and time intervals of the input data, causes prediction results relating to the disease of the candidate in the trial period to be output from the machine learning model, and in accordance with the prediction results, outputs selection reference information for determining whether the candidate will be made a subject for the trial.

Description

Technical Field

[0001] The technology of the present disclosure relates to a medical support device, a method for operating a medical support device, and an operating program for a medical support device.

Background Art

[0002] With the advent of a full-fledged aging society, efforts are being made to develop drugs (hereinafter abbreviated as anti-dementia drugs) that prevent the onset of diseases such as dementia represented by Alzheimer's disease or slow down the progression of dementia. The efficacy of anti-dementia drugs is evaluated through clinical trials over a certain period, for example, one year and six months (eighteen months). As subjects for this clinical trial, those with relatively rapid progression of dementia are preferred in order to correctly evaluate the efficacy of anti-dementia drugs. This is because if a person has a slow progression of dementia, it is not clear whether the progression is suppressed by the efficacy of the anti-dementia drug or due to reasons specific to that person.

[0003] Then, before conducting a clinical trial of an anti-dementia drug, it is necessary to predict those with relatively rapid progression of dementia and select them as subjects for the clinical trial. As a method for predicting those with relatively rapid progression of dementia, there is a method using a machine learning model. For example, in "M. Nguyen, T. He and L. An et al.: Predicting Alzheimer’s disease progression using deep recurrent neural networks, NeuroImage, Nov. 2020" (hereinafter referred to as Document 1), a technique for predicting the progression of dementia using a recurrent neural network (RNN) as a machine learning model is disclosed. In Document 1, examination data related to dementia at three or more time points (for example, examination data three months ago, two months ago, and one month ago) is given to the RNN as a set of teacher data for learning.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Even in the most popular database for dementia-related test data, ADNI (Alzheimer's Disease Neuroimaging Initiative), the number of contributors is less than 3,000. In other words, the method in Reference 1 has an overwhelmingly insufficient amount of training data. Consequently, the method in Reference 1 could lead to overfitting, significantly reducing the accuracy of predicting the progression of dementia and potentially selecting individuals unsuitable for clinical trials of anti-dementia drugs.

[0005] One embodiment of the technology described herein provides a medical support device, a method for operating the medical support device, and an operating program for the medical support device that can select individuals suitable for clinical trials of pharmaceuticals with high accuracy. [Means for solving the problem]

[0006] The medical support device of this disclosure comprises a processor and memory connected to or built into the processor. The processor acquires target input data, which is input data relating to the disease of a candidate for a drug clinical trial, and the trial period. The processor inputs the target input data and the trial period into a machine learning model that has been trained using training data, which includes accumulated input data relating to the disease at two or more points in time and the time interval of the input data. The machine learning model outputs prediction results regarding the disease of the candidate during the trial period, and the device outputs selection reference information for determining whether or not to include the candidate in the clinical trial based on the prediction results.

[0007] The time interval is preferably set according to the duration of the clinical trial.

[0008] The input data preferably includes at least one of the following: test data showing the results of tests related to the disease, and diagnostic data showing the results of a diagnosis related to the disease.

[0009] Preferably, the machine learning model outputs a score that quantitatively represents the degree of disease progression as a prediction result.

[0010] Preferably, the machine learning model also outputs a class that qualitatively represents the degree of disease progression as part of the prediction results.

[0011] Preferably, in addition to training data, the system has clinical trial suitability data that meets predetermined adoption criteria depending on the drug, and by inputting the input data and time intervals of the clinical trial suitability data into a machine learning model, the machine learning model outputs setting prediction results, and the processor outputs selection reference information according to selection criteria set at least based on the setting prediction result distribution, which is the distribution of the number of data points in the setting prediction results.

[0012] The selection criteria are preferably set based on the exclusion group prediction result distribution, which is the distribution of the number of data points for determining the group of individuals to be excluded from the clinical trial, extracted from the correct data included in the clinical trial suitability data.

[0013] The selection criteria are preferably set based on the selection group prediction result distribution, which is the distribution of the number of data points for determining the group of individuals to be selected as subjects for the clinical trial, extracted from the correct data included in the clinical trial suitability data.

[0014] In the prediction result distribution for setting, multiple provisional selection conditions are set, and for each of the multiple provisional selection conditions, the number of errors in the prediction results for setting with respect to the ground truth data is counted, and it is preferable that the provisional selection condition with the smallest number of errors is set as the selection condition.

[0015] The selection criteria are preferably set based not only on the prediction result distribution for setting, but also on the ground truth data distribution, which is the distribution of the number of ground truth data points included in the clinical trial-compatible data.

[0016] The selection criteria are preferably set by applying the provisional selection criteria set in the ground truth data distribution to the prediction result distribution used for setting.

[0017] The selection criteria are preferably set at the boundary of the region defined in the prediction result distribution for setting, which includes individuals whose disease is progressing rapidly.

[0018] The disease is preferably dementia.

[0019] The method of operating the medical support device of this disclosure includes: acquiring target input data, which is input data relating to the disease of a candidate for a drug clinical trial, and the clinical trial period; inputting the target input data and the clinical trial period into a machine learning model trained using training data that includes accumulated input data relating to the disease at two or more points in time and the time interval of the input data; having the machine learning model output prediction results regarding the disease of the candidate during the clinical trial period; and outputting selection reference information for determining whether or not to include the candidate as a subject of the clinical trial, according to the prediction results.

[0020] The operating program for the medical support device of this disclosure causes a computer to perform the following processes: acquire target input data, which is input data relating to the disease of a candidate for a drug clinical trial, and the clinical trial period; input the target input data and the clinical trial period into a machine learning model that has been trained using training data that includes accumulated input data relating to the disease at two or more points in time and the time interval of the input data; have the machine learning model output prediction results regarding the disease of the candidate during the clinical trial period; and output selection reference information for determining whether or not to include the candidate in the clinical trial based on the prediction results. [Effects of the Invention]

[0021] The technology disclosed herein provides a medical support device, a method for operating the medical support device, and an operating program for the medical support device that can select individuals suitable for clinical trials of pharmaceuticals with high precision. [Brief explanation of the drawing]

[0022] [Figure 1] This diagram shows the clinical trial participant selection support server and user terminals. [Figure 2]It is a diagram showing target input data. [Figure 3] It is a diagram showing the clinical trial period. [Figure 4] It is a diagram showing selected reference information. [Figure 5] It is a block diagram showing a computer constituting a clinical trial subject selection support server. [Figure 6] It is a block diagram showing the processing unit of the CPU of the clinical trial subject selection support server. [Figure 7] It is a block diagram showing the detailed configuration of the dementia progression prediction model. [Figure 8] It is a diagram showing an overview of the processing in the learning phase of the dementia progression prediction model. [Figure 9] It is a diagram for explaining the establishment of the teacher data of the dementia progression prediction model. [Figure 10] It is a diagram for explaining another example of the establishment of the teacher data of the dementia progression prediction model. [Figure 11] It is a diagram showing an overview of the processing in the operation phase of the dementia progression prediction model. [Figure 12] It is a diagram showing selection conditions. [Figure 13] It is a diagram showing selected reference information when the selection conditions are met. [Figure 14] It is a diagram showing selected reference information when the selection conditions are not met. [Figure 15] It is a diagram showing a clinical trial subject selection support screen. [Figure 16] It is a diagram showing a clinical trial subject selection support screen on which a message showing selected reference information is displayed. [Figure 17] It is a diagram showing a clinical trial subject selection support screen on which messages showing selected reference information of two candidate subjects are displayed. [Figure 18] It is a flowchart showing the processing procedure of the clinical trial subject selection support server. [Figure 19] It is a diagram showing another example of selection conditions. [Figure 20] It is a diagram showing another example of score prediction results. [Figure 21] This figure shows yet another example of a score prediction result. [Figure 22] This figure shows another example of the progress prediction results. [Figure 23] This diagram shows the process of generating training data and clinical trial-compatible data from all the data. [Figure 24] This figure shows how a dementia progression prediction model, which has been trained using training data, is input with target input data for setting clinical trial suitability data and a setting clinical trial period, and how the dementia progression prediction model outputs a setting score prediction result. [Figure 25] This graph shows the distribution of correct scores for setting and the distribution of predicted scores for setting. [Figure 26] This figure shows method 1 for setting selection criteria based on the distribution of correct scores for setting and the distribution of predicted scores for setting. [Figure 27] This figure shows the case where the selection criteria set by Method 1 are met. [Figure 28] This figure shows the case where the selection criteria set by Method 1 were not met. [Figure 29] This figure shows the process of generating the score prediction result distribution for setting the exclusion group. [Figure 30] This figure shows method 2 for setting selection criteria based on the score prediction result distribution for setting exclusion groups. [Figure 31] This figure shows the process of generating the score prediction result distribution for setting the selection group. [Figure 32] This figure shows method 3 for setting selection criteria based on the score prediction result distribution for setting the selection group. [Figure 33] This figure shows method 4, in which multiple provisional selection conditions are set in the distribution of predicted scores for setting, the error rate of the predicted score for setting relative to the correct score for setting is calculated for each of the multiple provisional selection conditions, and the provisional selection condition with the smallest error rate is set as the selection condition. [Figure 34] This figure shows method 5 for setting selection criteria at the boundary of a region defined as including individuals with rapidly progressing dementia in the distribution of predicted score results for setting. [Modes for carrying out the invention]

[0023] [First Embodiment] As an example, as shown in Figure 1, the clinical trial subject selection support server 10 is connected to the user terminal 11 via a network 12. The clinical trial subject selection support server 10 is an example of a "medical support device" related to the technology disclosed herein. The user terminal 11 is installed, for example, in a pharmaceutical development facility and is operated by drug discovery staff involved in the development of drugs that prevent the onset or slow the progression of dementia, particularly Alzheimer's disease, i.e., anti-dementia drugs. Examples of dementia include Alzheimer's disease, Lewy body dementia, and vascular dementia. Anti-dementia drugs may also be used for Alzheimer's diseases other than Alzheimer's disease. Specifically, these include the preclinical stage of Alzheimer's disease (PAD) and mild cognitive impairment due to Alzheimer's disease (MCI). Hereafter, Alzheimer's disease will be abbreviated as AD as appropriate. The disease is preferably a brain disorder such as the example dementia. The user terminal 11 has a display 13 and input devices 14 such as a keyboard and mouse. The network 12 is a WAN (Wide Area Network), such as the Internet or a public communication network. Although only one user terminal 11 is connected to the clinical trial subject selection support server 10 in Figure 1, in reality, multiple user terminals 11 from multiple pharmaceutical development facilities are connected to the clinical trial subject selection support server 10.

[0024] The user terminal 11 sends a distribution request 15 to the clinical trial subject selection support server 10. The distribution request 15 includes the target input data 16 and the clinical trial period 17. The distribution request 15 is a request to the clinical trial subject selection support server 10 to distribute selection reference information 18, which is referenced by drug discovery staff when selecting subjects for clinical trials of an anti-dementia drug under development.

[0025] The target input data 16 is input data related to dementia of the target candidate, who is a candidate for inclusion in the clinical trial, and data related to the diagnostic criteria for dementia is preferred.

[0026] Diagnostic criteria for dementia include those described in the "Dementia Disease Treatment Guidelines 2017" supervised by the Japanese Society of Neurology, the "International Statistical Classification of Diseases and Related Health Problems (ICD)-11" (11th edition), the "Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5)" by the American Psychiatric Association, and the "National Institute on Aging-Alzheimer's Association workgroup (NIA-AA) criteria." These diagnostic criteria can be referenced, and their contents are incorporated into this specification.

[0027] Data related to the diagnostic criteria for dementia include the data related to the diagnostic criteria mentioned above. The target input data 16 includes data related to the diagnostic criteria for dementia. Specifically, data related to the diagnostic criteria for dementia include cognitive function test data, morphological imaging test data, brain function imaging test data, blood and cerebrospinal fluid test data, and genetic test data. The target input data 16 preferably includes at least morphological imaging test data, and more preferably includes at least morphological imaging test data and cognitive function test data.

[0028] Cognitive function test data include the Clinical Dementia Rating-Sum of Boxes (CDR-SOB) score, the Mini-Mental State Examination (MMSE) score, and the Alzheimer's Disease Assessment Scale-cognitive subscale (ADAS-Cog) score. Morphological imaging test data include cross-sectional images of the brain obtained by magnetic resonance imaging (MRI) (see Figure 2)28 (see Figure 2), and cross-sectional images of the brain obtained by computed tomography (CT).

[0029] Brain function imaging data includes cross-sectional images of the brain obtained by positron emission tomography (PET) (hereinafter referred to as PET images) and cross-sectional images of the brain obtained by single-photon emission computed tomography (SPECT) (hereinafter referred to as SPECT images). Blood and cerebrospinal fluid (CSF) test data includes the amount of p-tau (phosphorylated tau protein) 181 in the cerebrospinal fluid (hereinafter abbreviated as CSF). Genetic test data includes the results of ApoE gene genotype testing.

[0030] The target input data 16 is entered by drug discovery staff operating the input device 14. Target candidates are, for example, those who have responded to a recruitment call for clinical trials at a pharmaceutical development facility. The clinical trial period 17 is, literally, the period during which clinical trials of an anti-dementia drug are conducted, and is pre-set according to the anti-dementia drug under development. Although not shown in the diagram, the distribution request 15 also includes a terminal ID (Identification Data) to uniquely identify the user terminal 11 that sent the distribution request 15.

[0031] When the distribution request 15 is received, the clinical trial subject selection support server 10 inputs the subject input data 16 and the clinical trial period 17 into the dementia progression prediction model 41 (see Figure 6), and causes the dementia progression prediction model 41 to output prediction results regarding the dementia of the subject candidates. The clinical trial subject selection support server 10 generates selection reference information 18 according to the prediction results and distributes the generated selection reference information 18 to the user terminal 11 that sent the distribution request 15. When the selection reference information 18 is received, the user terminal 11 displays the selection reference information 18 on the display 13 and makes the selection reference information 18 available for viewing by the drug discovery staff.

[0032] As an example, as shown in Figure 2, the target input data 16 includes candidate data 20, test data 21, and diagnostic data 22. Candidate data 20 is data that indicates the attributes of the target candidate, and includes the target candidate's age 23 and gender 24. Note that the target input data 16 is data obtained on the same day as the transmission date of the distribution request 15. The target input data 16 may also be data obtained on the transmission date of the distribution request 15, or data obtained from three days to one week prior to the transmission date. Alternatively, the target input data 16 may be data obtained on the clinical trial start date, or from three days to one week prior to the clinical trial start date.

[0033] The examination data 21 is data showing the results of dementia-related examinations of the target candidate, and includes cognitive ability test scores 25, which are cognitive function test data; cerebrospinal fluid (CSF) test results 26, which are blood and cerebrospinal fluid test data; genetic test results 27, which are genetic test data; and MRI images 28, which are morphological imaging test data. The cognitive ability test score 25 is, for example, the Clinical Dementia Rating-Sum of Boxes (CDR-SOB) score. The CSF test results 26 are, for example, the amount of p-tau (phosphorylated tau protein) 181 in the CSF.

[0034] Genetic test result 27 is, for example, the result of testing the genotype of the ApoE gene. The genotype of the ApoE gene is a combination of two of the three types of ApoE genes (ε2 and ε3, ε3 and ε4, etc.). The risk of developing Alzheimer's disease is estimated to be approximately 3 to 12 times higher for individuals with a genotype that has one or two ε4 genes (ε2 and ε4, ε4 and ε4, etc.) compared to individuals with a genotype that does not have any ε4 genes (ε2 and ε3, ε3 and ε3, etc.).

[0035] The diagnostic data 22 represents the results of the dementia diagnosis made by the physician at this time, based on the examination data 21, etc. The diagnostic data 22 can be one of the following: normal (NC; Normal Control) / pre-symptomatic stage (PAD) / mild cognitive impairment (MCI) / Alzheimer's disease (ADM; Alzheimer Dementia). Thus, there are multiple types of target input data 16, and the dementia progression prediction model 41 is a so-called multimodal machine learning model.

[0036] As an example, as shown in Figure 3, the clinical trial period 17 in this embodiment is one year and six months (eighteen months). The clinical trial period 17 varies depending on the anti-dementia drug, but is approximately one to two years.

[0037] As an example, as shown in Figure 4, the selection reference information 18 is either that the candidate is suitable or unsuitable for the clinical trial.

[0038] As an example, as shown in Figure 5, the computer constituting the clinical trial subject selection support server 10 includes storage 30, memory 31, CPU (Central Processing Unit) 32, communication unit 33, display 34, and input device 35. These are interconnected via a bus line 36. Note that the CPU 32 is an example of a "processor" related to the technology of this disclosure.

[0039] Storage 30 is a hard disk drive built into the computer constituting the clinical trial subject selection support server 10, or connected via cable or network. Alternatively, storage 30 is a disk array consisting of multiple hard disk drives installed in series. Storage 30 stores control programs such as the operating system, various application programs, and various data associated with these programs. A solid-state drive may be used instead of a hard disk drive.

[0040] Memory 31 is work memory for the CPU 32 to execute processing. The CPU 32 loads programs stored in storage 30 into memory 31 and executes processing according to the programs. In this way, the CPU 32 comprehensively controls all parts of the computer. Note that memory 31 may be built into the CPU 32.

[0041] The communication unit 33 controls the transmission of various information to external devices such as the user terminal 11. The display 34 displays various screens. These screens include a GUI (Graphical User Interface). It is equipped with an operation function via an interface. The clinical trial subject selection support server 10 is configured. The computer receives operation instructions from input devices 35 through various screens. The input devices 35 include keyboards, mice, touch panels, and microphones for voice input.

[0042] As an example, as shown in Figure 6, the storage 30 of the clinical trial subject selection support server 10 stores an operating program 40. The operating program 40 is an application program that causes the computer to function as the clinical trial subject selection support server 10. In other words, the operating program 40 is an example of an "operating program for a medical support device" related to the technology of this disclosure. The storage 30 also stores a dementia progression prediction model 41 and selection conditions 42. The dementia progression prediction model 41 is an example of a "machine learning model" related to the technology of this disclosure.

[0043] When the operating program 40 is started, the CPU 32 of the computer constituting the clinical trial subject selection support server 10 works in cooperation with the memory 31 and the like to function as a reception unit 45, a read / write (hereinafter abbreviated as RW (Read Write)) control unit 46, a prediction unit 47, a determination unit 48, and a distribution control unit 49.

[0044] The reception unit 45 receives a distribution request 15 from the user terminal 11. As mentioned above, the distribution request 15 includes the target input data 16 and the clinical trial period 17, so by receiving the distribution request 15, the reception unit 45 obtains the target input data 16 and the clinical trial period 17. The reception unit 45 outputs the target input data 16 and the clinical trial period 17 to the prediction unit 47. The reception unit 45 also outputs the cognitive ability test score 25 from the target input data 16 to the judgment unit 48. Furthermore, the reception unit 45 outputs the terminal ID of the user terminal 11 (not shown) to the distribution control unit 49.

[0045] The RW control unit 46 controls the storage of various data to the storage 30 and the reading of various data from the storage 30. For example, the RW control unit 46 reads the dementia progression prediction model 41 from the storage 30 and outputs the dementia progression prediction model 41 to the prediction unit 47. The RW control unit 46 also reads the selection conditions 42 from the storage 30 and outputs the selection conditions 42 to the determination unit 48.

[0046] The prediction unit 47 inputs the target input data 16 and the clinical trial period 17 into the dementia progression prediction model 41 and outputs a score prediction result 50 from the dementia progression prediction model 41. The prediction unit 47 outputs the score prediction result 50 to the judgment unit 48. The score prediction result 50 is an example of the "prediction result" and "score that quantitatively represents the degree of dementia progression" related to the technology of this disclosure.

[0047] The determination unit 48 determines whether a candidate is suitable to participate in the clinical trial, in accordance with the selection criteria 42, based on the cognitive ability test score 25 from the reception unit 45 and the score prediction result 50 from the prediction unit 47. Based on the determination result, the determination unit 48 generates selection reference information 18 and outputs the generated selection reference information 18 to the distribution control unit 49.

[0048] The distribution control unit 49 controls the distribution of selection reference information 18 to the user terminal 11 that sent the distribution request 15. At this time, the distribution control unit 49 identifies the user terminal 11 that sent the distribution request 15 based on the terminal ID from the reception unit 45.

[0049] As an example, as shown in Figure 7, the dementia progression prediction model 41 includes a feature extraction layer 55, a self-attention (SA) mechanism layer 56, a global average pooling (GAP) layer 57, fully connected (FC) layers 58, 59, and 60, and bilinear (BL) layers. i The system includes a layer 61 (abbreviated as near) and a softmax function (hereinafter abbreviated as SMF (SoftMax Function)) layer 62.

[0050] The feature extraction layer 55 is, for example, DenseNet (Densely Connected This is a Convolutional Network. The MRI image 28 is input to the feature extraction layer 55. The feature extraction layer 55 performs convolution and other operations on the MRI image 28 to convert the MRI image 28 into a feature map 63. The feature extraction layer 55 outputs the feature map 63 to the SA mechanism layer 56.

[0051] The SA mechanism layer 56 performs a convolution on the feature map 63, changing the coefficients of the convolution filter according to the features to be processed in the feature map 63. Hereafter, this convolution process performed by the SA mechanism layer 56 will be referred to as the SA convolution process. The SA mechanism layer 56 outputs the feature map 63 after the SA convolution process to the GAP layer 57.

[0052] The GAP layer 57 applies a global average pooling process to the feature map 63 after the SA convolution process. The global average pooling process calculates the average value of the features for each channel of the feature map 63. For example, if the feature map 63 has 512 channels, the global average pooling process will calculate the average value of 512 features. The GAP layer 57 outputs the calculated average value of the features to the BL layer 61.

[0053] The FC layer 58 is input with candidate data 20, examination data 21A excluding MRI images 28, diagnostic data 22, and the clinical trial period 17. The gender 24 of the candidate data 20 is input numerically, with 1 for male and 0 for female. Similarly, the genetic test results 27 of the examination data 21 are input numerically, with 1 for the combination of ε2 and ε3, and 2 for the combination of ε3 and ε3. The diagnostic data 22 is input numerically in the same way. The FC layer 58 has an input layer with units equal to the number of each data set, and an output layer with units equal to the number of data handled by the BL layer 61. Each unit in the input layer and each unit in the output layer are fully connected to each other, and each has a weight set. Each unit in the input layer is input with candidate data 20, examination data 21A excluding MRI images 28, diagnostic data 22, and the clinical trial period 17. The sum of the products of each of these data and the weights set between each unit becomes the output value of each unit in the output layer. The FC layer 58 outputs the output value of the output layer to the BL layer 61.

[0054] The BL layer 61 applies bilinear processing to the average value of the features from the GAP layer 57 and the output value from the FC layer 58. The BL layer 61 outputs the bilinearly processed values ​​to the FC layers 59 and 60. For details on the BL layer 61 and bilinear processing, please refer to the following literature. <Goto, T.etc, Multi-modal deep learning for predicting progression of Alzheimer’s disease using bi-linear shake fusion, Proc. SPIE 11314, Medical Imaging (2020)>

[0055] The FC layer 59 converts the bilinearly processed values ​​into variables to be handled by the SMF in the SMF layer 62. Similar to the FC layer 58, the FC layer 59 has an input layer with units equal to the number of bilinearly processed values ​​and an output layer with units equal to the number of variables to be handled by the SMF. Each unit in the input layer and each unit in the output layer are fully connected to each other, and each has a weight assigned to it. The bilinearly processed values ​​are input to each unit in the input layer. The sum of the products of the bilinearly processed values ​​and the weights assigned between each unit becomes the output value of each unit in the output layer. These output values ​​are the variables to be handled by the SMF. The FC layer 59 outputs the variables to be handled by the SMF to the SMF layer 62. The SMF layer 62 outputs the progression prediction result 64 by applying the variables to the SMF. The progression prediction result 64, like the diagnostic data 22, indicates whether the target candidate is normal / pre-symptomatic / mild cognitive impairment / Alzheimer's disease. The progression prediction result 64 is an example of a "prediction result" and a "class that qualitatively represents the degree of progression of dementia" related to the technology of this disclosure.

[0056] The FC layer 60 converts the bilinearly processed values ​​into a score prediction result 50. Similar to FC layers 58 and 59, the FC layer 60 has an input layer with units equal to the number of bilinearly processed values, and an output layer for the score prediction result 50. Each unit in the input layer and the output layer are fully connected, and each has a weight assigned to it. Each unit in the input layer is input with the bilinearly processed value. The sum of the products of the bilinearly processed value and the weights assigned between each unit becomes the output value of the output layer. This output value is the score prediction result 50. The score prediction result 50 is a prediction of the cognitive ability test score of the target candidate at the end of trial period 17, in this case the CDR-SOB score itself. The CDR-SOB score takes values ​​from 0 to 18, where 0 is normal and 18 indicates maximum cognitive impairment. Thus, the dementia progression prediction model 41 is a so-called multi-task machine learning model that outputs a progression prediction result 64 and a score prediction result 50.

[0057] As an example, as shown in Figure 8, the dementia progression prediction model 41 is trained in the learning phase by being given training data (also called learning data) 70. The training data 70 is a set of target input data 16L for learning, a learning trial period 17L, correct diagnostic results 64CA for learning, and correct scores 50CA for learning. The target input data 16L for learning is, for example, the target input data 16 of a sample subject (including patients, the same applies hereinafter) stored in a database such as ADNI, at the start of the learning trial period 17L. The learning trial period 17L is an interval set according to the trial period 17. In this example, the learning trial period 17L is one to two years. This one to two year period is ± six months of the one year and six months of the trial period 17.

[0058] The learning-based correct diagnostic result 64CA is the actual dementia diagnosis made by a physician to the sample subject at the end of the learning-based clinical trial period 17L. The learning-based correct score 50CA is the score of the cognitive ability test actually performed by the sample subject at the end of the learning-based clinical trial period 17L. The learning-based subject input data 16L is an example of "accumulated input data related to dementia at two or more points in time" related to the technology disclosed herein. The learning-based clinical trial period 17L is an example of "time interval of input data" related to the technology disclosed herein.

[0059] In the learning phase, the dementia progression prediction model 41 receives 16L of target input data for learning and 17L of the clinical trial period for learning. Based on the 16L of target input data for learning and the 17L of the clinical trial period for learning, the dementia progression prediction model 41 outputs 64L of learning progression prediction results and 50L of learning score prediction results.

[0060] Based on the learning progression prediction results 64L and the learning correct diagnosis results 64CA, a loss calculation is performed on the dementia progression prediction model 41 using a cross-entropy function. Hereafter, the result of this loss calculation will be denoted as loss L1. In addition, based on the learning score prediction results 50L and the learning correct score 50CA, a loss calculation is performed on the dementia progression prediction model 41 using a regression loss function such as mean squared error. Hereafter, the result of this loss calculation will be denoted as loss L2.

[0061] The various coefficients of the dementia progression prediction model 41 are updated according to the losses L1 and L2, and the dementia progression prediction model 41 is updated according to the update settings. The update settings are based on the total loss L shown in equation (1) below. Note that α is the weight. L = L1 × α + L2 × (1 - α) ... (1) In other words, the total loss L is the weighted sum of loss L1 and loss L2. α is, for example, 0.5.

[0062] In the learning phase, the above series of processes—inputting the target input data 16L and the learning clinical trial period 17L into the dementia progression prediction model 41, outputting the learning progression prediction result 64L and the learning score prediction result 50L from the dementia progression prediction model 41, performing loss calculation, setting update parameters, and updating the dementia progression prediction model 41—is repeated, with the training data 70 being exchanged at least twice. The repetition of the above series of processes ends when the prediction accuracy of the learning progression prediction result 64L and the learning score prediction result 50L for the learning correct diagnosis result 64CA and the learning correct score 50CA reaches a predetermined set level. The dementia progression prediction model 41, whose prediction accuracy has reached the set level, is stored in the storage 30 and used by the prediction unit 47. Alternatively, learning may be terminated after the above series of processes has been repeated a set number of times, regardless of the prediction accuracy of the learning progression prediction result 64L and the learning score prediction result 50L for the learning correct diagnosis result 64CA and the learning correct score 50CA.

[0063] While 0.5 is used as an example for α, it is not limited to this value. Furthermore, α is not limited to a fixed value; for example, α may be changed between the initial learning phase and other periods. For instance, α could be set to 1 at the beginning of the learning phase, and then gradually decreased as learning progresses, eventually becoming a fixed value, such as 0.5.

[0064] Figures 9 and 10 are diagrams illustrating the composition of the training data 70. Figure 9 shows the case of sample subject A. Figure 10 shows the case of sample subject B.

[0065] In Figure 9, sample subject A has test data 21 and diagnostic data 22 at four time points T0A, T1A, T2A, and T3A. Specifically, this includes test data 21_T0A (labeled as test data at T0A in the figure) and diagnostic data 22_T0A (labeled as diagnostic data at T0A in the figure) at time T0A, test data 21_T1A (labeled as test data at T1A in the figure) and diagnostic data 22_T1A (labeled as diagnostic data at T1A in the figure) at time T1A, test data 21_T2A (labeled as test data at T2A in the figure) and diagnostic data 22_T2A (labeled as diagnostic data at T2A in the figure) at time T2A, and test data 21_T3A (labeled as test data at T3A in the figure) and diagnostic data 22_T3A (labeled as diagnostic data at T3A in the figure) at time T3A.

[0066] Table 75 shows the time intervals at each point in time. Specifically, the time interval T1A-T0A between time T0A and time T1A for No. 1, the time interval T2A-T1A between time T1A and time T2A for No. 4, and the time interval T3A-T2A between time T2A and time T3A for No. 6 are six months. The time interval T2A-T0A between time T0A and time T2A for No. 2, and the time interval T3A-T1A between time T1A and time T3A for No. 5 are one year. The time interval T3A-T0A between time T0A and time T3A for No. 3 is two years. Of these No. 1 to No. 6, only No. 2 and No. 5, with a time interval of one year, and No. 3, with a time interval of two years, satisfy the condition of one to two years for the learning trial period of 17L.

[0067] Therefore, from sample subject A, it is possible to generate a total of three training data sets 70, No. 2, No. 3, and No. 5, as shown in Table 76. For example, training data 70 No. 2 is data related to time points T0A and T2A. The training target input data 16L is the test data 21_T0A and diagnostic data 22_T0A from time point T0A. The training trial period 17L is one year, the time interval T2A-T0A between time points T0A and T2A. The training correct diagnostic result 64CA is the diagnostic data 22_T2A from time point T2A. The training correct score 50CA is the cognitive ability test score 25 from the test data 21_T2A from time point T2A. In this case, time point T0A corresponds to the start of the training trial period 17L, and time point T2A corresponds to the end of the training trial period 17L.

[0068] For example, training data 70 for No. 5 is data related to time points T1A and T3A. The target input data 16L for learning is the test data 21_T1A and diagnostic data 22_T1A from time point T1A. The learning trial period 17L is one year, the time interval T3A-T1A between time points T1A and T3A. The learning correct diagnostic result 64CA is the diagnostic data 22_T3A from time point T3A. The learning correct score 50CA is the cognitive ability test score 25 from the test data 21_T3A from time point T3A. In this case, time point T1A corresponds to the start of the learning trial period 17L, and time point T3A corresponds to the end of the learning trial period 17L. Note that the numbers No. 1 to No. 6 correspond to the numbers 1 to 6 of the arcs connecting each time point on the time axis. The same applies to Figure 10.

[0069] In Figure 10, sample subject B has test data 21 and diagnostic data 22 at two time points, T0B and T1B. Specifically, these are the test data 21_T0B (labeled as test data at T0B in the figure) and diagnostic data 22_T0B (labeled as diagnostic data at T0B in the figure) from time T0B, and the test data 21_T1B (labeled as test data at T1B in the figure) and diagnostic data 22_T1B (labeled as diagnostic data at T1B in the figure) from time T1B.

[0070] Table 80 shows the time interval between time point T0B and time point T1B. That is, the time interval T1B-T0B between time point T0B and time point T1B is one year and three months. This one year and three months satisfies the condition of one to two years for the learning clinical trial period 17L. Therefore, as shown in Table 81, it is possible to generate one training data 70, No. 1, from sample subject B. That is, training data 70 No. 1 is data related to time points T0B and T1B. The training subject input data 16L is the examination data 21_T0B and diagnostic data 22_T0B at time point T0B. The learning clinical trial period 17L is one year and three months, the time interval T1B-T0B between time points T0B and time point T1B. The training correct diagnostic result 64CA is the diagnostic data 22_T1B at time point T1B. The learning correct score 50CA is the cognitive ability test score 25 from the test data 21_T1B at time T1B. In this case, time T0B corresponds to the start of the learning trial period 17L, and time T1B corresponds to the end of the learning trial period 17L. Thus, the training data 70 includes the test data 21 and diagnostic data 22 from two or more time points of the same sample subject, as well as the interval between the two time points. Although not shown in the diagram, for example, in the case of a sample subject with test data 21 and diagnostic data 22 from six time points, it is possible to generate training data 70 from fifteen data points that satisfy the condition of the learning trial period 17L, with 6C2 = 6 × 5 ÷ 2 = 15. Furthermore, for example, in the case of a sample subject with eight time points of examination data 21 and diagnostic data 22, it is possible to generate training data 70 from 28 data points (8C2 = 8 × 7 ÷ 2 = 28) that satisfy the condition of a learning clinical trial period of 17L.

[0071] Furthermore, the training data 70 is not limited to including input data and time intervals related to dementia for two or more points in time from the same sample subject. Input data and time intervals related to dementia for two or more points in time from multiple sample subjects with the same and / or similar dementia symptoms may be combined to generate input data and time intervals related to dementia for two or more points in time, and this may be used as the training data 70. Examples of sample subjects with the same and / or similar dementia symptoms include sample subjects whose test data 21 and / or diagnostic data 22 are the same and / or similar. Alternatively, input data and time intervals related to dementia for two or more points in time may be combined from multiple sample subjects with the same and / or similar attributes to generate input data and time intervals related to dementia for two or more points in time, and this may be used as the training data 70. Examples of sample subjects with the same and / or similar attributes include sample subjects with the same and / or similar age 23 and / or gender 24. Input data and time intervals related to dementia from multiple sample subjects who have the same and / or similar dementia symptoms and the same and / or similar attributes may be combined to generate input data and time intervals related to dementia at two or more time points, and this may be used as training data 70.

[0072] As an example, as shown in Figure 11, the prediction unit 47 inputs the target input data 16 and the clinical trial period 17 into the dementia progression prediction model 41, and outputs a score prediction result 50 from the dementia progression prediction model 41. The dementia progression prediction model 41 also outputs a progression prediction result 64, but the prediction unit 47 discards the progression prediction result 64 and outputs only the score prediction result 50 to the judgment unit 48. In Figure 11, an example is shown where the score prediction result 50 is 4.5.

[0073] As an example, as shown in Figure 12, selection criterion 42 states that the difference between the cognitive ability test score 25 and the predicted score 50 of the input data 16 is 2 or more. If the difference between the cognitive ability test score 25 and the predicted score 50 of the input data 16 is 2 or more, the progression of Alzheimer's disease is relatively rapid. Therefore, candidates whose difference between the cognitive ability test score 25 and the predicted score 50 of the input data 16 is 2 or more are suitable for the clinical trial. Conversely, candidates whose difference between the cognitive ability test score 25 and the predicted score 50 of the input data 16 is less than 2 are not suitable for the clinical trial because it is unclear whether the progression is being suppressed by the efficacy of the anti-dementia drug or whether the progression is being slowed due to reasons specific to that individual.

[0074] Therefore, as shown in Figure 13 as an example, if the difference between the cognitive ability test score 25 of the target input data 16 and the predicted score result 50 is 2 or more, and the selection condition 42 is met, the judgment unit 48 generates selection reference information 18 indicating that the candidate is suitable to be a subject of the clinical trial. In Figure 13, an example is shown where the cognitive ability test score 25 of the target input data 16 is 0.5, the predicted score result 50 is 4.5, and the difference is 4, which is 2 or more.

[0075] On the other hand, as shown in Figure 14 as an example, if the difference between the cognitive ability test score 25 of the target input data 16 and the predicted score result 50 is less than 2 and the selection condition 42 is not met, the judgment unit 48 generates selection reference information 18 indicating that the candidate is not suitable to be a subject of the clinical trial. Figure 14 illustrates the case where the cognitive ability test score 25 of the target input data 16 is 1, the predicted score result 50 is 1.5, and the difference is 0.5, which is less than 2.

[0076] Figure 15 shows an example of a clinical trial participant selection support screen 85 displayed on the display 13 of the user terminal 11. The clinical trial participant selection support screen 85 includes a pull-down menu 86 for selecting the age 23 of the candidate, a pull-down menu 87 for selecting the gender 24, and an input box 88 for the cognitive ability test score 25. CSFAn input box 89 for the test result 26 and a pull-down menu 90 for selecting the genetic test result 27 are provided.

[0077] The clinical trial subject selection support screen 85 is equipped with a file selection button 91 for selecting an MRI image file 28. When an MRI image file 28 is selected, a file icon 92 is displayed next to the file selection button 91. The file icon 92 is not displayed if no file is selected. The clinical trial subject selection support screen 85 is also equipped with a pull-down menu 93 for selecting diagnostic results (diagnostic data 22).

[0078] The clinical trial subject selection support screen 85 is provided with an "Add Candidate" button 94. When the "Add Candidate" button 94 is selected, a set of pull-down menus 86, 87, 90, and 93, input boxes 88 and 89, and a file selection button 91 are added to the clinical trial subject selection support screen 85 (see Figure 17). The "Add Candidate" button 94 can be selected multiple times. This makes it possible to input subject input data 16 for two or more candidates on a single clinical trial subject selection support screen 85.

[0079] A decision button 95 is located at the bottom of the clinical trial subject selection support screen 85. When the decision button 95 is selected, a distribution request 15 including the target input data 16 and the clinical trial period 17 is sent from the user terminal 11 to the clinical trial subject selection support server 10. The target input data 16 consists of the contents selected in the pull-down menus 86, 87, 90, and 93, the contents entered in the input boxes 88 and 89, and the MRI image 28 selected with the file selection button 91.

[0080] When the clinical trial subject selection support server 10 receives selection reference information 18, the clinical trial subject selection support screen 85 transitions as shown in Figure 16 as an example. Specifically, a message 100 indicating the selection reference information 18 is displayed. Figure 16 illustrates the case where the selection reference information 18 indicates that the candidate is suitable to be a subject of the clinical trial. The clinical trial subject selection support screen 85 disappears when the close button 101 is selected.

[0081] Furthermore, when the "Add Candidate" button 94 is selected and a candidate is added, the clinical trial participant selection support screen 85 will look like Figure 17 as an example. Figure 17 shows an example where a message 100 indicating the selection reference information 18 for two candidate participants is displayed.

[0082] Next, the operation of the above configuration will be explained with reference to the flowchart in Figure 18. First, when the operating program 40 is started in the clinical trial subject selection support server 10, the CPU 32 of the clinical trial subject selection support server 10 functions as a reception unit 45, an RW control unit 46, a prediction unit 47, a determination unit 48, and a distribution control unit 49, as shown in Figure 6.

[0083] First, the reception unit 45 receives the distribution request 15 from the user terminal 11, thereby acquiring the target input data 16 and the clinical trial period 17 (step ST100). The target input data 16 and the clinical trial period 17 are output from the reception unit 45 to the prediction unit 47.

[0084] As shown in Figure 11, in the prediction unit 47, the target input data 16 and the clinical trial period 17 are input to the dementia progression prediction model 41, and the dementia progression prediction model 41 outputs a score prediction result 50 (step ST110). The score prediction result 50 is output from the prediction unit 47 to the judgment unit 48.

[0085] As shown in Figures 13 and 14, the determination unit 48 calculates the difference between the cognitive ability test score 25 of the target input data 16 and the score prediction result 50. Then, it is determined whether the selection condition 42 is met if the difference is 2 or more, or whether the selection condition 42 is not met if the difference is less than 2 (step ST120). If the selection condition 42 is met, as shown in Figure 13, the determination unit 48 generates selection reference information 18 indicating that the candidate is suitable to be a subject of the clinical trial (step ST130). On the other hand, if the selection condition 42 is not met, as shown in Figure 14, the determination unit 48 generates selection reference information 18 indicating that the candidate is not suitable to be a subject of the clinical trial (step ST130).

[0086] The selection reference information 18 is output from the determination unit 48 to the distribution control unit 49. The selection reference information 18 is distributed to the user terminal 11 that sent the distribution request 15 under the control of the distribution control unit 49 (step ST140).

[0087] As described above, the CPU 32 of the clinical trial subject selection support server 10 includes a reception unit 45, a prediction unit 47, and a determination unit 48. The reception unit 45 receives a distribution request 15 and obtains target input data 16, which is input data related to dementia of the candidate for the clinical trial of the anti-dementia drug, and the clinical trial period 17 of the anti-dementia drug. The prediction unit 47 inputs the target input data 16 and the clinical trial period 17 of the anti-dementia drug into the dementia progression prediction model 41 and causes the dementia progression prediction model 41 to output a score prediction result 50, which is a prediction result regarding the dementia of the candidate at the end of the clinical trial period 17. The determination unit 48 outputs selection reference information 18 for deciding whether or not to select the candidate as a subject for the clinical trial according to the score prediction result 50.

[0088] As shown in Figures 8 to 10, the dementia progression prediction model 41 is trained using training data 70 that includes accumulated training input data 16L related to dementia at two or more time points and training clinical trial periods 17L. Because it includes training clinical trial periods 17L as a time interval for the input data, it can improve the prediction accuracy of the score prediction results 50 compared to the method in Reference 1, which trains the RNN by providing a set of training data consisting of test data at three or more time points. Because it is possible to prepare more abundant training data 70 than in the method in Reference 1, overfitting can be prevented. Therefore, it is possible to suppress the decrease in the accuracy of dementia progression prediction, and consequently, to improve the accuracy of dementia progression prediction. As a result, it becomes possible to select individuals who are suitable as subjects for clinical trials of anti-dementia drugs with high accuracy.

[0089] The learning trial period 17L is an interval set according to the trial period 17. Therefore, the dementia progression prediction model 41 can be made into a machine learning model specialized for predictions at time intervals aligned with the trial period 17, thereby improving the accuracy of selecting suitable subjects for the clinical trial.

[0090] The input data includes test data 21 showing the results of tests related to dementia, and diagnostic data 22 showing the results of diagnoses related to dementia. Therefore, it can contribute to improving the prediction accuracy of the score prediction result 50. Note that the input data only needs to include at least one of the test data 21 and the diagnostic data 22.

[0091] The dementia progression prediction model 41 outputs a score prediction result 50, which is a prediction result of a score that quantitatively represents the degree of dementia progression. Therefore, it is possible to follow the conventional method of selecting participants for clinical trials using cognitive ability test scores 25.

[0092] The dementia progression prediction model 41 also outputs a progression prediction result 64, which is a qualitative prediction result of a class representing the degree of dementia progression. Predicting the score prediction result 50 in conjunction with the progression prediction result 64, which has a relatively simple classification of only a few cases (in this example, four cases: normal / pre-symptomatic / mild cognitive impairment / Alzheimer's disease), rather than just predicting the cognitive ability test score 25, which is a continuous quantity with a reasonable range, can improve the prediction accuracy of the score prediction result 50.

[0093] The clinical trial period 17 does not have to be included in the distribution request 15. Since the clinical trial period 17 is known in advance, it may be stored in the storage 30. In this case, the RW control unit 46 retrieves the clinical trial period 17 from the storage 30 to obtain it. The RW control unit 46 outputs the retrieved clinical trial period 17 to the prediction unit 47.

[0094] For example, multiple dementia progression prediction models 41 may be prepared for different clinical trial periods 17, such as a dementia progression prediction model 41 for a one-year clinical trial period 17 and a dementia progression prediction model 41 for a two-year clinical trial period 17.

[0095] The selection reference information is not limited to selection reference information 18, which states whether the example candidate is suitable or unsuitable for the clinical trial. The score prediction result 50 and / or the progression prediction result 64 themselves may be distributed to the user terminal 11 as selection reference information. In this case, the drug discovery staff will determine whether the candidate is suitable for the clinical trial by referring to the score prediction result 50 and / or the progression prediction result 64. In this case, the selection criteria 42 are not required.

[0096] As an example, the selection criteria 105 shown in Figure 19 may be used. Selection criteria 105 is that the progression prediction result 64 is worse than the diagnostic data 22 of the target input data 16, and the difference between the cognitive ability test score 25 of the target input data 16 and the score prediction result 50 is 2 or more. The progression prediction result 64 being worse than the diagnostic data 22 of the target input data 16 means that the diagnostic data 22 of the target input data 16 is normal and the progression prediction result 64 is pre-symptomatic, mild cognitive impairment, or Alzheimer's disease, or that the diagnostic data 22 of the target input data 16 is mild cognitive impairment and the progression prediction result 64 is Alzheimer's disease. It also means that the diagnostic data 22 of the target input data 16 is pre-symptomatic and the progression prediction result 64 is mild cognitive impairment or Alzheimer's disease. Thus, the selection criteria may involve the progression prediction result 64 in place of, or in addition to, the score prediction result 50.

[0097] The score prediction result is not limited to the score prediction result 50 which shows the cognitive ability test score 25 of the first embodiment described above. For example, the score prediction result 110 shown in Figure 20 and the score prediction result 115 shown in Figure 21 may also be used.

[0098] The score prediction result 110 shown in Figure 20 represents the change in the cognitive ability test score 25. By adding this change to the cognitive ability test score 25 of the target input data 16 entered into the dementia progression prediction model 41, or by subtracting it from the cognitive ability test score 25, the cognitive ability test score 25 at the end of the clinical trial period 17 can be calculated. In Figure 20, 2 is shown as an example of the change. Therefore, by adding 2 to the cognitive ability test score 25 of the target input data 16 entered into the dementia progression prediction model 41, the cognitive ability test score 25 at the end of the clinical trial period 17 can be calculated.

[0099] The score prediction result 115 shown in Figure 21 represents the annual rate of change of the cognitive ability test score 25. The annual rate of change is the percentage that indicates how much the cognitive ability test score 25 changes in one year. By multiplying this amount of change by the clinical trial period 17, and adding or subtracting the result of the multiplication from the cognitive ability test score 25 of the target input data 16 entered into the dementia progression prediction model 41, the cognitive ability test score 25 at the end of the clinical trial period 17 can be calculated. In Figure 21, 0.8 / year is given as an example of the annual rate of change. Therefore, by multiplying 0.8 by the clinical trial period 17 and adding the result of the multiplication to the cognitive ability test score 25 of the target input data 16 entered into the dementia progression prediction model 41, the cognitive ability test score 25 at the end of the clinical trial period 17 can be calculated. If the clinical trial period 17 is, for example, one year and six months, the multiplication result will be 0.8 × 1.5 = 1.2. Furthermore, if the clinical trial period is, for example, two years, the multiplication result would be 0.8 × 2 = 1.6.

[0100] The progression prediction results are not limited to the example progression prediction result 64, which states one of the following: normal / pre-symptomatic stage / mild cognitive impairment / Alzheimer's disease. For example, as shown in the progression prediction result 120 in Figure 22, the probability of each of the following may be given: normal / pre-symptomatic stage / mild cognitive impairment / Alzheimer's disease.

[0101] Furthermore, the progression prediction results may not be limited to Alzheimer's disease, but more generally, they may state whether the candidate is normal / pre-symptomatic / mild cognitive impairment / dementia. Subjective cognitive impairment (SCI) and / or subjective cognitive decline (SCD) may also be included as prediction targets. The progression prediction results may also state whether the candidate will develop Alzheimer's disease after two years or not. Alternatively, for example, it may state whether the candidate's progression to dementia after three years is fast or slow. It may also state whether the candidate will progress from normal or pre-symptomatic to MCI, or whether the candidate will progress from normal, pre-symptomatic, or MCI to Alzheimer's disease.

[0102] [Second Embodiment] As an example, as shown in Figure 23, in the second embodiment, clinical trial-compatible data 131 is generated from all data 130 in addition to training data 70. While the training data 70 has no constraints, the clinical trial-compatible data 131 has the constraint that it must meet the inclusion criteria. The inclusion criteria are predetermined according to the anti-dementia drug, and include, for example, individuals aged 65 or older with a Mini-Mental State Examination (MMSE) score of 25 or less. For this reason, the clinical trial-compatible data 131 has fewer data points than the training data 70.

[0103] The clinical trial compliance data 131 is a set consisting of the target input data for setting 16S, the clinical trial period for setting 17S, the correct diagnostic result for setting 64SCA, and the correct score for setting 132SCA. The target input data for setting 16S corresponds to the target input data for learning 16L of the training data 70, and the clinical trial period for setting 17S corresponds to the clinical trial period for learning 17L of the training data 70. The clinical trial period for setting 17S, like the clinical trial period for learning 17L, is an interval set according to the clinical trial period 17. For example, if the clinical trial period 17 is one year and six months, the clinical trial period for setting 17S is one to two years, which is one year and six months ± six months. The target input data for setting 16S is an example of "input data for clinical trial compliance data" relating to the technology of this disclosure. The clinical trial period for setting 17S is an example of "time interval for clinical trial compliance data" relating to the technology of this disclosure.

[0104] The setting-based correct diagnostic result 64SCA corresponds to the learning-based correct diagnostic result 64CA of the training data 70, and the setting-based correct score 132SCA corresponds to the learning-based correct score 132CA of the training data 70. The learning-based correct score 132CA and the setting-based correct score 132SCA are the annual rate of change (hereinafter simply referred to as change) of the cognitive ability test score 25 shown in Figure 21. The setting-based correct score 132SCA is an example of "correct data included in clinical trial-compliant data" related to the technology of this disclosure.

[0105] As an example, as shown in Figure 24, the dementia progression prediction model 41, which has been trained on the training data 70, is input with the target input data 16S and the clinical trial period 17S for setting the clinical trial suitability data 131. As a result, the dementia progression prediction model 41 outputs a setting score prediction result 132S. The setting score prediction result 132S is the change in the cognitive ability test score 25, just like the setting correct score 132SCA. The setting score prediction result 132S is an example of a "setting prediction result" related to the technology of this disclosure.

[0106] As mentioned above, since the training data 70 is generated without any constraints, there is a large amount of data that does not meet the selection criteria. Therefore, there is some error in the setting score prediction result 132S obtained by inputting the setting target input data 16S and setting clinical trial period 17S of the clinical trial suitability data 131 into the dementia progression prediction model 41 trained on such training data 70. This error may also occur in the score prediction result 132 (see Figure 27) output by inputting the target input data 16 and clinical trial period 17 of the target candidate into the dementia progression prediction model 41. Therefore, if the selection criteria are determined without correcting this error, suitable candidates for the clinical trial may be excluded from selection, or conversely, unsuitable candidates may be selected. Therefore, the following describes a method for correcting the above error.

[0107] <Method 1> As an example, as shown in Figure 25, Table 135 summarizes the number of clinical trial compliance data 131 for each of the 0.1 increments of the correct setting score 132SCA. Similarly, Table 136 summarizes the number of clinical trial compliance data 131 for each of the 0.1 increments of the predicted setting score 132S. From Table 135, a correct setting score distribution 137 can be generated, which is the distribution of the number of data for the correct setting score 132SCA. Furthermore, a predicted setting score result distribution 138 can be generated from Table 136, which is the distribution of the number of data for the predicted setting score result 132S. The correct setting score distribution 137 is an example of a "correct data distribution" related to the technology of this disclosure. Furthermore, the predicted setting score result distribution 138 is an example of a "predicted setting result distribution" related to the technology of this disclosure. As can be seen from the setting correct score distribution 137 and the setting score prediction result distribution 138, there is an error between the setting correct score 132SCA and the setting score prediction result 132S. For the sake of explanation, the error is exaggerated in this depiction.

[0108] As an example, as shown in Figure 26, Method 1 first sets a provisional selection condition 140T for the setting correct score distribution 137. Then, the provisional selection condition 140T is applied to the setting score prediction result distribution 138 to obtain the selection condition 140. Specifically, the change shown by line 142, which divides the setting score prediction result distribution 138 at the same rate as line 141 drawn on the change in cognitive ability test score 25 included in the provisional selection condition 140T divides the setting correct score distribution 137, is set as the selection condition 140.

[0109] Figure 26 illustrates a hypothetical selection criterion, 140T, where the change in cognitive ability test score 25 is greater than 0. Here, individuals with a change greater than 0 are those whose dementia has progressed after 17 days of the clinical trial. Conversely, individuals with a change of 0 or less are those whose dementia has not progressed after 17 days of the clinical trial.

[0110] Figure 26 illustrates the case where line 141, drawn at a change of 0, is a line that divides the setting correct score distribution 137 in a 4:6 ratio. In this case, the change of 2.5, shown by line 142 which divides the setting score prediction result distribution 138 in a 4:6 ratio, following line 141, is set as the selection condition 140. That is, the selection condition 140 is that the change in the cognitive ability test score 25 is greater than 2.5.

[0111] As an example, as shown in Figure 27, if the score prediction result 132 obtained by inputting the target candidate's target input data 16 and the clinical trial period 17 into the dementia progression prediction model 41 satisfies the selection criteria 140, the judgment unit 48 generates selection reference information 18 indicating that the target candidate is suitable to be a subject of the clinical trial. Figure 27 illustrates the case where the score prediction result 132 was 3.2.

[0112] On the other hand, as shown in Figure 28 as an example, if the score prediction result 132 obtained by inputting the target candidate's target input data 16 and the clinical trial period 17 into the dementia progression prediction model 41 does not satisfy the selection criteria 140, the judgment unit 48 generates selection reference information 18 indicating that the target candidate is not suitable to be a subject of the clinical trial. Figure 28 illustrates the case where the score prediction result 132 is 1.6.

[0113] Thus, in Method 1, the selection criteria 140 are set based on the setting ground truth score distribution 137, which is the distribution of the number of data points for the setting ground truth score 132SCA included in the clinical trial suitability data 131, and the setting score prediction result distribution 138, which is the distribution of the number of data points for the setting score prediction result 132S. More specifically, the selection criteria 140 are set by applying the provisional selection criteria 140T set in the setting ground truth score distribution 137 to the setting score prediction result distribution 138. This makes it possible to correct errors that occur in the score prediction result 132. The probability of missing suitable candidates from the selection process or, conversely, selecting unsuitable candidates can be significantly reduced.

[0114] <Method 2> As an example, as shown in Figure 29, in Method 2, first, as shown in step ST200, an exclusion group is extracted, which is a group of people who should be excluded from the clinical trial (hereinafter referred to as exclusion recommendations), based on the setting ground truth score 132SCA. Figure 29 illustrates the case in which people with a setting ground truth score 132SCA of 0 or less are extracted as exclusion recommendations. Next, as shown in step ST210, the dementia progression prediction model 41, which has been trained on the training data 70, is input with the setting target input data 16S and the setting clinical trial period 17S of the clinical trial suitability data 131 for exclusion recommendations, and the dementia progression prediction model 41 outputs the setting score prediction result 132S.

[0115] Table 145 summarizes the number of clinical trial suitability data 131 for each of the setting score prediction results 132S output in step ST210. From this Table 145, an exclusion group setting score prediction result distribution 146 can be generated, which is the distribution of the number of data for the exclusion group setting score prediction results 132S. The exclusion group setting score prediction result distribution 146 is an example of an "exclusion group prediction result distribution" related to the technology of this disclosure.

[0116] As an example, as shown in Figure 30, in Method 2, the user establishes a policy 150 that determines how many exclusion candidates are acceptable to be selected as subjects of the clinical trial. Then, the amount of change shown by the line 151, which corresponds to the policy 150 and is drawn on the exclusion group setting score prediction result distribution 146, is set as the selection condition 152.

[0117] Figure 30 illustrates a case where policy 150 is established to keep the probability of selecting individuals recommended for exclusion below 20%. In this case, line 151 divides the prediction result distribution 146 for setting the exclusion group score in an 8:2 ratio. The change of 2.3 shown by this line 151 is set as the selection condition 152. In other words, the selection condition 152 is that the change in the cognitive ability test score 25 is greater than 2.3.

[0118] Thus, in Method 2, the selection criteria 152 are set based on the exclusion group setting score prediction result distribution 146, which is the distribution of the number of data points for the exclusion group setting score prediction result 132S, which is a group extracted based on the setting correct score 132SCA included in the clinical trial suitability data 131 and which should be excluded from the clinical trial. Therefore, the probability of selecting individuals who are not suitable for the clinical trial, i.e., those recommended for exclusion, can be suppressed to a certain extent. Since a looser selection criteria 152 can be set compared to Method 1, which significantly reduces the probability of selecting individuals recommended for exclusion, the number of individuals who can be included in the clinical trial can also be increased compared to Method 1.

[0119] <Method 3> As an example, as shown in Figure 31, in Method 3, first, as shown in step ST250, a selection group is extracted, which is a group of people who should be selected as subjects for the clinical trial (hereinafter referred to as selection recommendations), based on the setting ground truth score 132SCA. Figure 31 illustrates the case in which individuals with a setting ground truth score 132SCA greater than 0 are extracted as selection recommendations. Next, as shown in step ST260, the dementia progression prediction model 41, which has been trained on the training data 70, is input with the setting target input data 16S and the setting clinical trial period 17S of the clinical trial suitability data 131 for selection recommendations, and the dementia progression prediction model 41 outputs the setting score prediction result 132S.

[0120] Table 155 summarizes the number of clinical trial suitability data 131 for each of the setting score prediction results 132S output in step ST260. From this Table 155, a selection group setting score prediction result distribution 156 can be generated, which is the distribution of the number of data for the setting score prediction results 132S of the selection group. The selection group setting score prediction result distribution 156 is an example of a "selection group prediction result distribution" related to the technology of this disclosure.

[0121] As an example, as shown in Figure 32, in Method 3, the user sets a policy 160 regarding how many recommended candidates should be included in the clinical trial. Then, the amount of change shown by the line 161, which is drawn on the score prediction result distribution 156 for setting the selection group according to the policy 160, is set as the selection condition 162.

[0122] Figure 32 illustrates a case where policy 160 is established to secure more than 80% of the candidates to be recommended for selection. In this case, line 161 divides the score prediction result distribution 156 for setting the selection group in a 2:8 ratio. The change of 3.1 shown by this line 161 is set as the selection condition 162. In other words, the selection condition 162 is that the change in the cognitive ability test score 25 is greater than 3.1.

[0123] Thus, in Method 3, the selection criteria 162 are set based on the selection group setting score prediction result distribution 156, which is the distribution of the number of data points in the selection group setting score prediction result 132S, which is a group extracted based on the setting correct score 132SCA included in the clinical trial suitability data 131 and which should be selected as subjects for the clinical trial. As a result, a certain number of suitable subjects for the clinical trial, i.e., recommended subjects, can be secured as subjects for the clinical trial.

[0124] <Method 4> As an example, as shown in Figure 33, in Method 4, multiple provisional selection conditions are set by changing the amount of change in the cognitive ability test score 25 in increments of 0.1 in the setting score prediction result distribution 138, as shown by multiple lines 165. Then, as shown in Table 166, the number of errors in the setting score prediction result 132S relative to the setting correct answer score 132SCA is counted for each of the multiple provisional selection conditions. The number of errors is the sum of the number of data points where the setting correct answer score 132SCA is 0 or less, but the setting score prediction result 132S is greater than the provisional selection condition, and the number of data points where the setting correct answer score 132SCA is greater than 0, but the setting score prediction result 132S is less than or equal to the provisional selection condition. The former case, where the setting correct answer score 132SCA is 0 or less, but the setting score prediction result 132S is greater than the provisional selection condition, is when a person who is actually a person to be excluded is mistakenly selected as a person to be selected. On the other hand, the latter case, where the setting score 132SCA is greater than 0, but the setting score prediction result 132S is less than or equal to the provisional selection criteria, is a case where a person is actually a recommended candidate but is incorrectly listed as a recommended candidate for exclusion.

[0125] Then, the provisional selection criterion with the minimum number of counted errors is set as selection criterion 167. Figure 33 illustrates the case where the minimum number of errors is 5 when the change in the cognitive ability test score 25 for the provisional selection criterion is 2.7. In this case, selection criterion 167 would be that the change in the cognitive ability test score 25 is greater than 2.7.

[0126] Thus, in Method 4, multiple provisional selection criteria are set in the setting score prediction result distribution 138. For each of the multiple provisional selection criteria, the number of errors in the setting score prediction result 132S relative to the setting correct score 132SCA is counted, and the provisional selection criterion with the smallest number of errors is set as the selection criterion 167. Therefore, the probability of excluding suitable candidates from the clinical trial or, conversely, selecting unsuitable candidates can be further significantly reduced.

[0127] Furthermore, the methods described in either document A or document B below may be used to search for selection criterion 167. These methods in documents A or B are commonly used to find the optimal solution (selection criterion 167 in this case) from among multiple candidates (in this case, multiple provisional selection criteria). Reference A: J Kittler, J Illingworth, J Foglein, Threshold selection based on a simple image statistic, Computer Vision, Graphics, and Image Processing, Vo l 30, Issue 2, May 1985, pp. 125-147 Literature B: Nobuyuki Otsu (1979). “A threshold selection method from gray-level histograms”. IEEE Trans. Sys. Man. Cyber. 9 (1): pp. 62-66.

[0128] <Method 5> As an example, as shown in Figure 34, Method 5 sets the selection condition 171 at the boundary of the region 170 defined as containing individuals whose dementia is progressing rapidly in the setting score prediction result distribution 138. Figure 34 illustrates Method 1 shown in Figure 26.

[0129] Region 170, or more specifically, the boundary line 172 of region 170, is defined by the user. The user should define region 170 (line 172) based on the pharmacology of the anti-dementia drug or the results of clinical trials such as animal experiments conducted prior to the clinical trial using the dementia progression prediction model 41. Line 172 is, for example, a line drawn at a distance of +2σ (σ is the standard deviation) or +3σ from the mean of the setting score prediction result distribution 138. The change amount of 4.4 shown by this line 172 and the change amount of 2.5 shown by line 142 are set as selection conditions 171. In other words, selection condition 171 means that the change amount of the cognitive ability test score 25 is greater than 2.5 and less than 4.4.

[0130] Individuals with rapidly progressing dementia are likely to have advanced brain nerve damage that renders anti-dementia drugs ineffective, making it impossible to properly verify the efficacy of these drugs, and therefore they are unsuitable as subjects for clinical trials. Therefore, in Method 5, selection criteria 171 are set at the boundary of region 170, which is defined in the setting score prediction result distribution 138 as including individuals with rapidly progressing dementia. This reduces the probability of selecting individuals with rapidly progressing dementia as subjects for clinical trials. While Figure 34 illustrates Method 1, Method 5 may also be applied to Methods 2-4.

[0131] Although not shown in the diagram, in methods 2 to 5, the determination unit 48 also outputs selection reference information 18 according to each of the selection conditions 152, 162, 167, and 171.

[0132] The process of inputting the target input data 16S and the setting clinical trial period 17S for the clinical trial suitability data 131 into the dementia progression prediction model 41, which has been trained on the training data 70 as shown in Figure 24, and outputting the setting score prediction result 132S from the dementia progression prediction model 41, may be performed on the clinical trial subject selection support server 10 or on a device other than the clinical trial subject selection support server 10. Furthermore, the setting of the selection condition 140 by Method 1 shown in Figures 25 and 26, the setting of the selection condition 152 by Method 2 shown in Figures 29 and 30, and the setting of the selection condition 162 by Method 3 shown in Figures 31 and 32 may also be performed on the clinical trial subject selection support server 10 or on a device other than the clinical trial subject selection support server 10. In addition, the setting of the selection condition 167 by Method 4 shown in Figure 33, and the setting of the selection condition 171 by Method 5 shown in Figure 34 may also be performed on the clinical trial subject selection support server 10 or on a device other than the clinical trial subject selection support server 10.

[0133] Clinical trial-compatible data 131 may be prepared in the following way: the entire data 130 is divided, for example, into 80% training data 70 and 20% test data. Then, data that meet the inclusion criteria are extracted from the test data to be clinical trial-compatible data 131.

[0134] The score prediction results are not limited to the example change amounts. They may also be the probabilities for normal / pre-symptomatic / mild cognitive impairment / Alzheimer's disease as shown in Figure 22. Alternatively, they may be a weighted sum of the change amount and the probabilities for normal / pre-symptomatic / mild cognitive impairment / Alzheimer's disease.

[0135] Instead of distributing the selection reference information 18 from the clinical trial subject selection support server 10 to the user terminal 11, screen data of the clinical trial subject selection support screen 85 shown in Figure 16, etc., may be distributed from the clinical trial subject selection support server 10 to the user terminal 11.

[0136] The method of making the selection reference information 18 available for viewing by drug discovery staff is not limited to the clinical trial subject selection support screen 85. The selection reference information 18 may be provided to drug discovery staff as a printed copy, or an email with the selection reference information 18 attached may be sent to the drug discovery staff's mobile device.

[0137] The training of the dementia progression prediction model 41 shown in Figure 8 may be performed on the clinical trial subject selection support server 10, or on a device other than the clinical trial subject selection support server 10. Furthermore, the training of the dementia progression prediction model 41 may be continued even after deployment.

[0138] The clinical trial subject selection support server 10 may be installed at each pharmaceutical development facility, or it may be installed in a data center independent of the pharmaceutical development facilities. Furthermore, the user terminal 11 may perform some or all of the functions of the processing units 45-49 of the clinical trial subject selection support server 10.

[0139] A cognitive ability test score of 25 can also be expressed using Rivermead Behavioural Memory Test (RBMT) scores, Activities of Daily Living (ADL) scores, etc. Alternatively, a cognitive ability test score of 25 can also be expressed using ADAS-Cog scores, MMSE scores, etc.

[0140] The CSF test result 26 is not limited to the amount of p-tau181 as exemplified. It could also be the amount of t-tau (total tau protein) or the amount of Aβ42 (amyloid-beta protein).

[0141] The MRI image 28 may be an image of a portion of the brain, such as an image of the hippocampus. Alternatively, instead of or in addition to the MRI image 28, a PET image or SPECT image may be used as the examination data 21.

[0142] As described in International Publication No. 2022 / 071158, for example, images of anatomical regions of the brain, such as the hippocampus, may be extracted from medical images such as MRI images 28. The extracted images of anatomical regions may be input into a feature derivation model such as a convolutional neural network to perform convolution operations and output features. These features may then be input as target input data 16 into a dementia progression prediction model 41, thereby outputting a score prediction result 50 from the dementia progression prediction model 41. The features accurately represent the shape and texture characteristics of the anatomical regions, such as the degree of hippocampal atrophy. Therefore, the prediction accuracy of the score prediction result 50 can be further improved. The images of anatomical regions to be extracted are preferably not limited to images of the hippocampus, but also include images of multiple other anatomical regions such as the parahippocampal gyrus, frontal lobe, anterior temporal lobe (anterior part of the temporal lobe), occipital lobe, thalamus, hypothalamus, and amygdala. The images of anatomical regions to be extracted preferably include at least images of the hippocampus, and more preferably include at least images of the hippocampus and images of the anterior temporal lobe. In this case, a feature derivation model is prepared for each image of multiple anatomical regions. This method of extracting images of brain anatomical regions from medical images, inputting the extracted images of anatomical regions into the feature derivation model to output features, and then inputting these features as target input data 16 into the dementia progression prediction model 41 is particularly effective for predicting progression from MCI.

[0143] Predictions regarding dementia include predictions of cognitive function, such as how much a subject's cognitive function will decline in two years, and predictions of the risk of developing dementia, such as the subject's risk of developing dementia.

[0144] While dementia is used as an example of a disease, it is not limited to this. The disease could be, for example, cerebral infarction. In this case, the target input data 16 includes Stroke Assessment Scale (NIHSS (National Institutes of Health Stroke Scale)) scores and Japan Stroke Assessment Scale (JSS (Japan Stroke Scale)) scores, CT images and MRI images, etc. Furthermore, the machine learning model is not limited to one that accepts multiple types of target input data 16 related to the disease, such as the dementia progression prediction model 41. Thus, medical support may also include support for selecting clinical trial subjects for diseases other than dementia. The diseases could include neurodegenerative diseases such as cerebral infarction, or Parkinson's disease, as well as neurovascular diseases including cerebrovascular diseases.

[0145] However, dementia has become a social problem with the arrival of the aging society in recent years. For this reason, the dementia progression prediction server 10, which uses a dementia progression prediction model 41 into which target input data 16 related to dementia is input, can be said to be in a form that matches the current social problem.

[0146] In each of the above embodiments, the hardware structure of the Processing Unit that performs various processes, such as the reception unit 45, the RW control unit 46, the prediction unit 47, the determination unit 48, and the distribution control unit 49, can be the following types of processors. As mentioned above, the types of processors include a CPU 32, which is a general-purpose processor that executes software (operation program 40) and functions as various processing units, as well as programmable logic devices (PLDs), such as FPGAs (Field Programmable Gate Arrays), which are processors whose circuit configuration can be changed after manufacturing, and dedicated electrical circuits, such as ASICs (Application Specific Integrated Circuits), which are processors with circuit configurations specifically designed to perform specific processes.

[0147] A single processing unit may consist of one of these various processors, or it may consist of a combination of two or more processors of the same or different types (for example, a combination of multiple FPGAs, and / or a combination of a CPU and an FPGA). Alternatively, multiple processing units may be composed of a single processor.

[0148] Examples of configuring multiple processing units with a single processor include, firstly, a configuration where one or more CPUs and software combine to form a single processor, which then functions as multiple processing units, as exemplified by client and server computers. Secondly, a configuration using a processor that realizes the functions of the entire system, including multiple processing units, on a single IC (Integrated Circuit) chip, as exemplified by System-on-a-Chip (SoC). Thus, various processing units are configured, in terms of hardware structure, using one or more of the above-mentioned processors.

[0149] Furthermore, the hardware structure of these various processors can more specifically utilize electrical circuits, which are combinations of circuit elements such as semiconductor devices.

[0150] The technology of this disclosure can be appropriately combined with the various embodiments and / or variations described above. Furthermore, it is understood that various configurations can be adopted without departing from the spirit of the invention, and the invention is not limited to the embodiments described above. Moreover, the technology of this disclosure extends not only to programs but also to storage media for storing programs non-temporarily.

[0151] The descriptions and illustrations presented above are detailed explanations of the technical aspects of this disclosure and are merely examples of the technical aspects. For example, the above descriptions of the structure, function, operation, and effect are examples of the structure, function, operation, and effect of the technical aspects of this disclosure. Therefore, it goes without saying that you may delete unnecessary parts, add new elements, or replace elements in the descriptions and illustrations presented above, as long as you do not deviate from the essence of the technical aspects of this disclosure. Furthermore, in order to avoid confusion and facilitate understanding of the technical aspects of this disclosure, explanations of common technical knowledge and the like that do not require special explanation to enable the implementation of the technical aspects of this disclosure have been omitted from the descriptions and illustrations presented above.

[0152] In this specification, "A and / or B" is synonymous with "at least one of A and B." That is, "A and / or B" means that it may be A alone, or B alone, or a combination of A and B. Furthermore, in this specification, the same concept as "A and / or B" applies when expressing three or more things linked by "and / or."

[0153] All documents, patent applications, and technical standards described herein are incorporated by reference to the same extent as if each individual document, patent application, and technical standard were specifically and individually noted to be incorporated by reference.

Claims

1. Processor and The processor comprises, The aforementioned processor, The target input data, which is input data related to the disease of the candidate for a drug clinical trial, and the clinical trial period are obtained. A machine learning model trained using training data that includes input data related to diseases at two or more time points in time for sample subjects, the time interval of the input data, and ground truth data indicating the state of the disease of the sample subject at the time point corresponding to the time interval, The aforementioned target input data and the clinical trial period are input, and the machine learning model outputs the predicted results regarding the disease of the target candidate during the clinical trial period. Based on the prediction results, the system outputs selection reference information for determining whether or not to include the candidate in the clinical trial. Medical support device.

2. The medical support device according to claim 1, wherein the time interval is an interval set according to the clinical trial period.

3. The medical support device according to claim 1, wherein the input data includes at least one of the following: test data showing the results of tests related to a disease, and diagnostic data showing the results of a diagnosis related to a disease.

4. The medical support device according to claim 1, wherein the machine learning model outputs a score that quantitatively represents the degree of disease progression as a prediction result.

5. The medical support device according to claim 4, wherein the machine learning model also outputs a class that qualitatively represents the degree of disease progression as a prediction result.

6. In addition to the aforementioned training data, the system has clinical trial suitability data that meets predetermined recruitment conditions according to the aforementioned pharmaceutical product. By inputting the aforementioned clinical trial suitability data and time intervals into the machine learning model, the machine learning model outputs a prediction result for setting purposes. The aforementioned processor, The medical support device according to claim 1, which outputs selection reference information in accordance with selection conditions set based at least on a setting prediction result distribution which is a distribution of the number of data points of the setting prediction results.

7. The medical support device according to claim 6, wherein the selection criteria are set based on the exclusion group prediction result distribution, which is the distribution of the number of data points of the setting prediction result for a group of persons to be excluded from the subject of the clinical trial, extracted based on the correct answer data included in the clinical trial suitability data.

8. The medical support device according to claim 6, wherein the selection criteria are set based on the selection group prediction result distribution, which is the distribution of the number of data points of the setting prediction result for a group of persons to be selected as subjects of the clinical trial, extracted based on the correct answer data included in the clinical trial suitability data.

9. The medical support device according to claim 6, wherein a plurality of provisional selection conditions are set in the setting prediction result distribution, the number of errors in the setting prediction result relative to the correct data is counted for each of the plurality of provisional selection conditions, and the provisional selection condition with the smallest number of errors is set as the selection condition.

10. The medical support device according to claim 6, wherein the selection criteria are set based not only on the prediction result distribution for setting, but also on the ground truth data distribution, which is the distribution of the number of ground truth data included in the clinical trial suitability data.

11. The medical support device according to claim 10, wherein the selection conditions are set by applying the provisional selection conditions set in the correct data distribution to the prediction result distribution for setting.

12. The medical support device according to claim 6, wherein the selection criteria are set at the boundary of a region defined in the setting prediction result distribution as including individuals whose disease is progressing rapidly.

13. The medical support device according to claim 1, wherein the disease is dementia.

14. A method for operating a medical support device performed by a computer, To obtain target input data, which is input data related to the disease of candidate subjects for a drug clinical trial, and the clinical trial period. A machine learning model trained using training data that includes input data related to diseases at two or more time points in time for sample subjects, the time interval of the input data, and ground truth data indicating the state of the disease of the sample subject at the time point corresponding to the time interval, The machine learning model outputs a prediction result regarding the disease of the candidate during the clinical trial period, and the target input data and the clinical trial period are input and the machine learning model is output, Based on the prediction results, output selection reference information for determining whether or not to include the candidate in the clinical trial. A method for operating a medical support device, including the device itself.

15. To obtain target input data, which is input data related to the disease of candidate subjects for a drug clinical trial, and the clinical trial period. A machine learning model trained using training data that includes input data related to diseases at two or more time points in time for sample subjects, the time interval of the input data, and ground truth data indicating the state of the disease of the sample subject at the time point corresponding to the time interval, The machine learning model outputs a prediction result regarding the disease of the candidate during the clinical trial period, and the target input data and the clinical trial period are input and the machine learning model is output, Based on the prediction results, output selection reference information for determining whether or not to include the candidate in the clinical trial. An operating program for a medical support device that causes a computer to perform a process including [specific details].

Images