Disease risk determination system using fundus image, machine learning model generation device, disease risk determination device, and disease risk determination method

Abstract

Provided is a disease risk determination system capable of determining disease risk from a fundus image. The disease risk determination system according to the present invention comprises a machine learning model generation device that generates a trained model for determining disease risk, and a disease risk determination device that determines disease risk using the generated trained model. The machine learning model generation device comprises: a learning data acquisition unit; a first trained model generation unit that generates a first trained model; and a second trained model generation unit that generates a second trained model. The disease risk determination device comprises: a determination data acquisition unit; a health checkup data prediction unit that inputs a determination fundus image and determination health checkup data to the first trained model, and outputs predicted health checkup data; a disease risk prediction unit that inputs the predicted health checkup data and the determination fundus image to the second trained model, and outputs predicted disease risk data; and a prediction result presentation unit.

Description

Disease risk assessment system using fundus images, machine learning model generation device, disease risk assessment device, and disease risk assessment method

[0001] The present invention relates to a disease risk assessment system, and more particularly to a disease risk assessment system, a machine learning model generation device, a disease risk assessment device, and a disease risk assessment method that assess disease risk from fundus images using a trained model based on a neural network.

[0002] Conventionally, systems have been proposed that use fundus images to analyze the health status of a subject, specifically, by analyzing fundus images to output the presence or likely progression of a specific medical condition.

[0003] Such an analysis system uses fundus images to assess a subject's risk for a health event or to assess their overall health. The system uses a fundus image processing machine learning model to generate data that explains the basis for the predictions made by the system, i.e., specific predictions. The machine learning model processes multiple fundus image data and the corresponding health conditions of the subjects to generate a learning model for the subjects. Any fundus image data of the subjects is then input into the learning model, and health analysis data is generated from the model output.

[0004] However, conventional systems require the use of data such as blood tests in addition to fundus images to obtain health analysis data related to specific diseases.

[0005] As an example of a system for analyzing the health condition of a subject using fundus images, Patent Document 1 discloses a "system capable of generating health analysis data for a patient, and optionally other patient data, from input including one or more fundus images of the patient," which "in order to generate health analysis data for a given patient, the system uses a fundus image processing machine learning model to process the one or more fundus images and, optionally, other patient data to generate model output for the patient, and then generates health analysis data from the model output."

[0006] Furthermore, as a system for analyzing the health condition of a subject, particularly symptoms associated with infectious diseases and signs of worsening, using eye image data such as fundus images, for example, Patent Document 2 discloses a medical system that "includes a data acquisition unit that acquires at least two pieces of data from a patient, including blood oxygen data, auscultatory sound data, eye image data, and eye blood flow data, and a data processing unit that processes the at least two pieces of data acquired by the data acquisition unit in order to detect changes in the state of the circulatory system associated with infectious diseases."

[0007] However, neither Patent Document 1 nor Patent Document 2 mentions achieving accuracy equivalent to that achieved by an invasive test such as a blood test without performing an invasive test such as a blood test on the body.

[0008] JP 2019-528113 A JP 2021-176056 A

[0009] Therefore, the present invention solves the above-mentioned problems and provides a disease risk determination system that can determine disease risk from data obtained through non-invasive testing alone, without performing a blood test. That is, the present invention provides a disease risk determination system that can make predictions equivalent to those made by performing a blood test, without performing a blood test, by predicting health check data obtained from a blood test from a fundus image and using the predicted data as virtual blood test data. This is expected to reduce the physical and mental burden on subjects caused by the test and also contribute to promoting opportunities to learn about disease risk.

[0010] In addition, according to the present invention, a disease risk determination system is provided that can predict unobtained health check data from a subject's fundus image using a trained model, and can determine disease risk from the subject's fundus image and health check data predicted by the trained model using a different trained model.

[0011] Furthermore, one aspect of the present invention provides a disease risk determination system that can present disease risk data to a user terminal and provide the user terminal with predicted health checkup data corresponding to predicted diseases based on the predicted disease risk data.

[0012] In order to solve the above problems, the present invention provides a disease risk determination system, comprising: a machine learning model generation device that generates a trained model for determining the risk of disease; and a disease risk determination device that determines the risk of disease using the trained model generated by the machine learning model generation device, wherein the machine learning model generation device includes a training data acquisition unit that acquires training fundus images, training health checkup data, and training disease risk data from a training database; a first trained model generation unit that uses the training health checkup data as correct data and inputs the training fundus images as observed data into a neural network to generate a first trained model; and a second trained model generation unit that uses the training disease risk data as correct data and inputs the training fundus images as observed data into a neural network to generate a first trained model. and a second trained model generation unit that inputs the fundus image for determination and medical examination data for determination as data into a neural network and generates a second trained model, and the disease risk determination device comprises a data acquisition unit for acquiring a fundus image for determination and medical examination data for determination from a database for determination, a medical examination data prediction unit that inputs the fundus image for determination and the medical examination data for determination into the first trained model and outputs predicted medical examination data, a disease risk prediction unit that inputs the predicted medical examination data output from the medical examination data prediction unit and the fundus image for determination into the second trained model and outputs predicted disease risk data, and a prediction result presentation unit that outputs at least the predicted disease risk data to a user terminal.

[0013] In a disease risk determination system according to one aspect of the present invention, the prediction result presentation unit further outputs predicted health checkup data corresponding to the predicted disease based on the predicted disease risk data to a user terminal.

[0014] The present invention also provides a machine learning model generation device, comprising: a training data acquisition unit that acquires training fundus images, training health check data, and training disease risk data from a training database; a first trained model generation unit that uses the training health check data as correct data, inputs the training fundus images as observed data into a neural network, and generates a first trained model; and a second trained model generation unit that uses the training disease risk data as correct data, inputs the training fundus images as observed data into the neural network, and generates a second trained model.

[0015] The present invention also provides a disease risk assessment device comprising: a data acquisition unit for assessment that acquires a fundus image for assessment and health check data for assessment from a database for assessment; a health check data prediction unit that inputs the fundus image for assessment and the health check data for assessment into a first trained model and outputs predicted health check data; a disease risk prediction unit that inputs the predicted health check data output from the health check data prediction unit and the fundus image for assessment into a second trained model and outputs predicted disease risk data; and a prediction result presentation unit that outputs at least the predicted disease risk data to a user terminal, wherein the first trained model is generated by using the health check data for training as correct data and inputting the fundus image for training as observation data into a neural network, and the second trained model is generated by using the disease risk data for training as correct data and inputting the fundus image for training as observation data into the neural network.

[0016] In the disease risk assessment device according to one aspect of the present invention, the prediction result presentation unit further outputs predicted health check data corresponding to the predicted disease based on the predicted disease risk data to the user terminal.

[0017] The present invention also provides a disease risk assessment method, comprising the steps of: a data acquisition unit for assessment acquiring a fundus image for assessment and medical examination data for assessment from a database for assessment; a medical examination data prediction unit inputting the fundus image for assessment and the medical examination data for assessment into a first trained model and outputting predicted medical examination data; a disease risk prediction unit inputting the predicted medical examination data output from the medical examination data prediction unit and the fundus image for assessment into a second trained model and outputting predicted disease risk data; and a prediction result presentation unit outputting the predicted disease risk data and predicted medical examination data corresponding to the predicted disease based on the predicted disease risk data to a user terminal, wherein the first trained model is generated by using the medical examination data for training as correct data and inputting the fundus image for training as observation data into a neural network, and the second trained model is generated by using the disease risk data for training as correct data and inputting the fundus image for training as observation data into the neural network.

[0018] A program is provided that causes a computer to execute each step of the above method.

[0019] In the present invention, "disease risk" refers to the risk of contracting a disease, and includes the risk of already having the disease and the risk of contracting the disease in the future.

[0020] In the present invention, unless otherwise specified below, "determining" a "disease risk" means obtaining data (e.g., a score, a numerical value, an index, a rank, etc.) indicating the disease risk derived by the system, device, or method of the present invention, and does not include the medical act of actually diagnosing the disease by a doctor or the like.

[0021] In the present invention, "health checkup data" refers to basic health checkup data of a subject measured in a health checkup or the like. The health checkup data includes, for example, sex, age at checkup, height, weight, abdominal circumference, BMI, systolic blood pressure, diastolic blood pressure, etc. The health checkup data includes health checkup items and numerical data for those items.

[0022] In the present invention, "disease risk data" refers to data indicating the risk of a disease, and refers to data in which the risk is scored according to the type of disease. The scored data is typically expressed as a numerical value. For example, if the disease is "depression," the numerical value of the PHQ-9 (Patient Health Questionnaire-9), which is used to diagnose depression, may be used.

[0023] In the present invention, a "fundus image" refers to a photograph of the fundus of a subject's eye. The fundus of the eye refers to the inner surface of the eye opposite the lens. The fundus of the eye also includes the retina and optic nerve head.

[0024] According to the present invention, disease risk can be determined solely from data obtained through non-invasive testing, without the need for a blood test. In other words, by predicting health checkup data obtained through a blood test from a fundus image and using the data as virtual blood test data, predictions equivalent to those made through a blood test can be made without the need for a blood test. Therefore, it is expected that this will reduce the physical and mental burden on subjects due to the test and also contribute to promoting opportunities to learn about disease risk.

[0025] According to the present invention, it is possible to predict unobtained health check data from a subject's fundus image using a trained model, and it is possible to determine disease risk from the subject's fundus image and the health check data predicted by the trained model using a different trained model.

[0026] Furthermore, according to the present invention, disease risk data can be presented to a user terminal, and predictive health checkup data corresponding to predicted diseases based on the predicted disease risk data can be provided to the user terminal. Other objects, features, and advantages of the present invention will become apparent from the following description of the embodiments of the present invention taken in conjunction with the accompanying drawings.

[0027] Fig. 1 is a schematic diagram showing the entire disease risk determination system according to the present invention. Fig. 2 is a diagram showing the process of generating a trained model in a machine learning model generation device according to the present invention. Fig. 3 is a diagram showing the prediction process in a disease risk determination device according to the present invention. Fig. 4 is a flowchart showing the flow of the disease risk determination process according to the present invention.

[0028] 1 is a schematic diagram showing the entire disease risk assessment system according to the present invention. The disease risk assessment system 1 includes a machine learning model generation device 20 that generates a trained model for assessing disease risk, and a disease risk assessment device 30 that assesses disease risk using the trained model generated by the machine learning model generation device 20.

[0029] The machine learning model generation device 20 includes a training data acquisition unit 201, a first trained model generation unit 202, and a second trained model generation unit 203. The training data acquisition unit 201 acquires training fundus images, training health check data, and training disease risk data from the training database 41. The first trained model generation unit 202 uses the training health check data as correct answer data and inputs the training fundus images as observed data into a neural network to generate a first trained model. The second trained model generation unit 203 uses the training disease risk data as correct answer data and inputs the training fundus images as observed data into a neural network to generate a second trained model. The first trained model is a trained model for predicting health check data from fundus images, etc. The second trained model is a trained model for predicting disease risk from fundus images and predicted health check data.

[0030] The disease risk assessment device 30 comprises a data acquisition unit 301 for assessment, a medical examination data prediction unit 302, a disease risk prediction unit 303, and a prediction result presentation unit 304. The data acquisition unit 301 for assessment acquires a fundus image for assessment and medical examination data for assessment from the assessment database 42. The medical examination data prediction unit 302 inputs the fundus image for assessment and the medical examination data for assessment into a first trained model M1 and outputs predicted medical examination data. The disease risk prediction unit 303 inputs the predicted medical examination data output from the medical examination data prediction unit and the fundus image for assessment into a second trained model and outputs predicted disease risk data. The prediction result presentation unit 304 outputs at least the predicted disease risk data to the user terminal 50.

[0031] The training database 41 stores training data used to generate a trained model in the machine learning model generation device 20. The training database 41 stores training fundus images, training health check data, and training disease risk data as training data. The training fundus images are captured by an external examination device such as a fundus image capturing device. The training fundus images, training health check data, and training disease risk data are collected in advance from medical institutions, health check institutions, etc. The training database 41 may be constructed on a server or a cloud.

[0032] The assessment database 42 stores assessment data used for assessing disease risk in the disease risk assessment device 30. The assessment database stores assessment fundus images and assessment health check data as assessment data. The assessment fundus images are captured by an external examination device such as a fundus imaging device. The assessment fundus images are fundus images of a subject whose disease risk is to be assessed, and are captured using a fundus imaging device or the like at a medical institution, health checkup institution, or the like. The assessment health check data are health check data of the subject obtained when conducting an examination or health check at a medical institution, health checkup institution, or the like. The assessment health check data are data acquired by an external examination device such as a height gauge, weight scale, body fat gauge, body composition gauge, blood pressure monitor, or the like. The assessment health check data typically include, for example, gender, age at examination, height, weight, abdominal circumference, BMI, systolic blood pressure, and diastolic blood pressure. The assessment health check data are preferably data that can be obtained without conducting a blood test. Another advantage of the present invention is that disease risk can be assessed solely from data obtained through non-invasive testing. The assessment health checkup data may include interview data, which may be data obtained by scoring answers to interview questions, and the scored data may be expressed as a numerical value. The training database 41 may be constructed on a server or cloud.

[0033] The examination device is an external examination device such as a fundus image taking device for taking an image of the fundus of the subject, an external examination device such as a height measuring device, a weight scale, a body fat scale, a body composition scale, or a blood pressure monitor for measuring the subject's height, weight, body fat percentage, blood pressure, etc., or an external instrument such as a tape measure for measuring the subject's abdominal circumference. The examination device may be an existing medical device used for taking fundus images in medical treatments, health checkups, etc. at medical institutions. The training fundus images taken by the examination device and the training health checkup data acquired by the examination device are stored in a training database 41. The determination fundus images taken by the examination device and the determination health checkup data acquired by the examination device are stored in a determination database 42.

[0034] The user terminal 50 may be a personal computer or a mobile terminal such as a smartphone or tablet. The user terminal 50 displays the judgment result output from the prediction result presentation unit 304 of the disease risk judgment device 30. At least the predicted disease risk data predicted by the disease risk prediction unit 303 is output to the user terminal 50 as the disease risk judgment result. The user terminal 50 may further output predicted health check data corresponding to the predicted disease based on the predicted disease risk data.

[0035] 2 is a diagram showing a trained model generation process in the machine learning model generation device according to the present invention. The machine learning model generation device 20 includes a first trained model generation unit 202 and a second trained model generation unit 203. The first trained model generation unit 202 is for generating a first trained model M1, and the second trained model generation unit 203 is for generating a second trained model M2. The first trained model M1 is a trained model for predicting health checkup data from fundus images and the like. The second trained model M2 is a trained model for predicting disease risk from fundus images and predicted health checkup data.

[0036] The first trained model generation unit 202 uses, as training data, the training fundus image and training medical examination data acquired by the training data acquisition unit 201 from the training database 41. The first trained model generation unit 202 uses the training medical examination data as correct answer data, inputs the training fundus image as observation data into the neural network, and generates a first trained model M1.

[0037] The training fundus images are used when generating the first trained model M1 in the first trained model generation unit 202. The training fundus images are captured using an external examination device such as a fundus image capturing device, and are collected in advance at medical institutions, health checkup institutions, etc. and stored in the training database 41.

[0038] The training health checkup data is health checkup data used when generating the first trained model M1 in the first trained model generation unit 202. The training health checkup data refers to basic health checkup data of a subject measured in a health checkup or the like. The health checkup data includes, for example, gender, age at the time of examination, height, weight, abdominal circumference, BMI, systolic blood pressure, and diastolic blood pressure. The health checkup data includes health checkup items and numerical data for those items.

[0039] The neural network used in the first trained model generation unit 202 may be configured to include a first neural network and a second neural network different from the first neural network. In this case, the output from the first neural network may be used as the input to the second neural network.

[0040] The second trained model generation unit 203 uses, as training data, the training fundus image and the training disease risk data acquired by the training data acquisition unit 201 from the training database 41. The second trained model generation unit 203 regards the training disease risk data as correct answer data, inputs the training fundus image as observation data into the neural network, and generates a second trained model.

[0041] The training fundus images are also used when generating the second trained model M2 in the second trained model generation unit 203. The training fundus images, like those used in the first trained model generation unit 202, are captured using an external examination device such as a fundus image capturing device, and are collected in advance at medical institutions, health checkup institutions, etc. and stored in the training database 41.

[0042] The training disease risk data is disease risk data used in the second trained model generation unit 203 when generating the second trained model M2. The training disease risk data may be represented by different indices or scored data depending on the type of disease. The scored data is typically expressed as a numerical value. For example, if the disease is "depression," the training disease risk data may be a numerical value from the Patient Health Questionnaire-9 (PHQ-9), which is used to diagnose depression. The numerical value of the PHQ-9 is a number between 0 and 27.

[0043] The neural network used in the second trained model generation unit 203 may be configured to include a first neural network and a second neural network different from the first neural network. In this case, the output from the first neural network may be used as the input to the second neural network.

[0044] The first neural network used in the first trained model generation unit 202 and the second trained model generation unit 203 may be a neural network that receives as input the output of a layer prior to a certain layer. The first neural network may be, for example, a so-called DenseNet.

[0045] The second neural network used in the first trained model generation unit 202 and the second trained model generation unit 203 may be a neural network having one intermediate layer between the input layer and the output layer. The second neural network may be, for example, a so-called shallow net.

[0046] 3 is a diagram showing the prediction process in the disease risk determination device according to the present invention. The disease risk determination device 30 comprises a health checkup data prediction unit 302 and a disease risk prediction unit 303. The health checkup data prediction unit 302 is for predicting health checkup data using a first trained model M1, and the disease risk prediction unit 303 is for predicting predicted disease risk data using a second trained model M2.

[0047] The medical checkup data prediction unit 302 uses as input data the fundus image for determination and the medical checkup data for determination that the data for determination acquisition unit 301 has acquired from the database for determination 42. The medical checkup data prediction unit 302 inputs the fundus image for determination and the medical checkup data for determination into the first trained model M1, and outputs predicted medical checkup data.

[0048] The disease risk prediction unit 303 uses as input data the fundus image for determination acquired from the determination database 42 by the data for determination acquisition unit 301 and the predicted health check data predicted by the health check data prediction unit 302. The disease risk prediction unit 303 inputs the predicted health check data and the fundus image for determination output from the health check data prediction unit 302 into the second trained model M2, and outputs predicted disease risk data.

[0049] The predicted disease risk data is disease risk data predicted by the disease risk prediction unit 303 using the second trained model M2. The predicted disease risk data may be represented by different indices or scored data depending on the type of disease. The scored data is typically expressed as a numerical value. For example, if the disease is "depression," the predicted disease risk data may be a numerical value of the PHQ-9 (Patient Health Questionnaire-9), which is used to diagnose depression. The numerical value of the PHQ-9 is a value between 0 and 27. The existence of a correlation between fundus images and depression has been recognized in previous research in the medical field and the like.

[0050] In another example, the predicted disease risk data may represent the probability of each disease risk for multiple diseases, such as "depression: 20%, diabetes: 5%, glaucoma: 3%, cardiovascular disease: 1%."

[0051] The disease risk assessment system 1 of the present invention can predict the risk of any disease for which a correlation is found between a fundus image and the disease, and the disease risk assessment system 1 of the present invention can predict the risk of that disease. That is, for diseases for which some symptoms, signs, or changes appear in fundus images when a person has the disease or is at risk of contracting the disease in the future, the disease risk assessment system 1 of the present invention can predict the risk. For example, in addition to depression, diabetes, glaucoma, heart disease, vascular disease, and the like have also been found to be correlated with fundus images in previous research in the medical field, and the risk of these diseases can be predicted using the disease risk assessment system 1 of the present invention.

[0052] The prediction result presentation unit 304 outputs at least the predicted disease risk data predicted by the disease risk prediction unit 303 to the user terminal 50. The prediction result presentation unit 304 may further output to the user terminal 50 predicted health check data corresponding to the predicted disease based on the predicted disease risk data.

[0053] In an example in which the predicted disease risk data is represented by different indices or scored data depending on the type of disease, the scored predicted disease risk data may be output to the user terminal 50. For example, if the disease is "depression," the PHQ-9 numerical value (e.g., a numerical value from 0 to 27) may be displayed on the user terminal 50. Furthermore, predicted health check data corresponding to "depression" may be output to the user terminal 50.

[0054] In an example where the predicted disease risk data indicates the probability of each disease risk for a plurality of diseases, the probability of the disease risk for each disease may be output to the user terminal 50. For example, the probability of having a disease risk for each disease may be shown as a percentage in the form of a table or graph, such as "depression: 20%, diabetes: 5%, glaucoma: 3%, cardiovascular disease: 1%", and output to the user terminal 50. Furthermore, predicted health check data corresponding to each disease may be output to the user terminal 50. For example, for data such as "diabetes: 30%," the predicted health check data "HbA1c: 6.3%" may be output together with the data.

[0055] In this way, by displaying the predicted health checkup data as well, when doctors or subjects themselves use the predicted results to actually judge disease risk, they can know not only the type of disease they are at risk of contracting, but also the related health checkup items and their predicted values, which has the advantage of making judgments more efficient.

[0056] Next, a disease risk determination method according to the present invention will be described. Fig. 4 is a flowchart showing the flow of the disease risk determination process according to the present invention. The disease risk determination process in the disease risk determination method according to the present invention can be broadly divided into a step of acquiring determination data (S401), a step of determining the disease risk using a second trained model (S402), and a step of presenting the disease risk prediction result (S403).

[0057] The disease risk determination method of the present invention is executed by a disease risk determination device 30 including a determination data acquisition unit 301, a health check data prediction unit 302, a disease risk prediction unit 303, and a prediction result presentation unit 304. The disease risk determination method of the present invention includes the steps of: the determination data acquisition unit 301 acquiring a fundus image for determination and health check data for determination from the determination database 42; the health check data prediction unit 302 inputting the fundus image for determination and the health check data for determination into a first trained model M1 and outputting predicted health check data; the disease risk prediction unit 303 inputting the predicted health check data output from the health check data prediction unit and the fundus image for determination into a second trained model M2 and outputting predicted disease risk data; and the prediction result presentation unit 304 outputting the predicted disease risk data and predicted health check data corresponding to the predicted disease based on the predicted disease risk data to a user terminal 50.

[0058] In the disease risk assessment method of the present invention, the first trained model M1 is generated by using training health checkup data as correct data and inputting training fundus images as observed data into a neural network. In the disease risk assessment method of the present invention, the second trained model is generated by using training disease risk data as correct data and inputting training fundus images as observed data into a neural network. The first trained model M1 and the second trained model M2 are generated by the machine learning model generation device 20 of the present invention.

[0059] Furthermore, a program for causing a computer to execute each step of the disease risk assessment method of the present invention may be provided. The program of the present invention may be stored in a computer-readable recording medium. Furthermore, the program of the present invention may be stored in a memory within the disease risk assessment system, or may be stored externally or on a cloud linked to the disease risk assessment system.

[0060] According to the present invention, disease risk can be determined solely from data obtained through non-invasive testing, without the need for a blood test. In other words, by predicting health checkup data obtained through a blood test from a fundus image and using the data as virtual blood test data, predictions equivalent to those obtained through a blood test can be made without the need for a blood test. This is expected to reduce the physical and mental burden on subjects due to the test and promote opportunities to learn about disease risk. That is, even if a subject is unaware of their disease risk, there is an advantage in that they can learn about their disease risk through a low-burden test from the pre-disease stage, before they become ill. The fact that a low-burden test is sufficient increases opportunities for testing, which is expected to contribute to preventive medicine that avoids disease risks at the pre-disease stage.

[0061] Furthermore, according to the present invention, it is possible to predict unobtained medical checkup data from a fundus image of a subject using a trained model, and it is possible to determine disease risk from the fundus image of the subject and the medical checkup data predicted by the trained model using a different trained model. Therefore, there is an advantage in that it is possible to achieve accuracy equivalent to that achieved when medical checkup data is used, without obtaining the medical checkup data.

[0062] Furthermore, according to the present invention, disease risk data can be presented to a user terminal, and predictive health checkup data corresponding to a predicted disease based on the predicted disease risk data can be provided to the user terminal. Therefore, if the user is a physician or the like, when actually diagnosing or advising a subject based on the disease risk prediction results obtained by the disease risk assessment system 1 according to the present invention, the physician can simultaneously obtain not only the disease risk data but also the predictive health checkup data for that disease, thereby improving the efficiency of the diagnosis and advice. Furthermore, if the user is the subject himself / herself, even without medical knowledge, he / she can learn about diseases at high risk and the health checkup data items that require attention for those diseases, which can be useful for managing his / her own health and motivate him / her to undergo further examinations or a medical examination by a doctor, etc. While the above description is based on examples, the present invention is not limited thereto, and it will be apparent to those skilled in the art that various modifications and variations can be made within the scope of the principles of the present invention and the appended claims.

[0063] 1 Disease risk determination system 20 Machine learning model generation device 201 Learning data acquisition unit 202 First trained model generation unit 203 Second trained model generation unit 30 Disease risk determination device 301 Determination data acquisition unit 302 Health check data prediction unit 303 Disease risk prediction unit 304 Prediction result presentation unit 41 Learning database 42 Determination database 50 User terminal M1 First trained model M2 Second trained model

Claims

1. A disease risk determination system, comprising a machine learning model generation device that generates a learned model for determining the risk of a disease, and a disease risk determination device that determines the risk of a disease using the learned model generated by the machine learning model generation device, wherein the machine learning model generation device includes a learning data acquisition unit that acquires learning fundus images, learning health examination data, and learning disease risk data from a learning database, a first learned model generation unit that inputs the learning health examination data as correct answer data and the learning fundus images as observation data into a neural network to generate a first learned model, and a second learned model generation unit that inputs the learning disease risk data as correct answer data and the learning fundus images as observation data into a neural network to generate a second learned model, and the disease risk determination device includes a determination data acquisition unit that acquires determination fundus images and determination health examination data from a determination database, a health examination data prediction unit that inputs the determination fundus images and the determination health examination data into the first learned model to output predicted health examination data, a disease risk prediction unit that inputs the predicted health examination data output from the health examination data prediction unit and the determination fundus images into the second learned model to output predicted disease risk data, and a prediction result presentation unit that outputs at least the predicted disease risk data to a user terminal. The disease risk determination system is characterized by the above.

2. The disease risk determination system according to claim 1, wherein the prediction result presentation unit further outputs the predicted health examination data corresponding to the predicted disease based on the predicted disease risk data to the user terminal.

3. A machine learning model generation device, comprising: A learning data acquisition unit that acquires learning fundus images, learning health examination data, and learning disease risk data from a learning database; A first trained model generation unit that uses the learning health examination data as correct data, inputs the learning fundus images as observation data into a neural network, and generates a first trained model; A second trained model generation unit that uses the learning disease risk data as correct data, inputs the learning fundus images as observation data into a neural network, and generates a second trained model. A machine learning model generation device characterized by comprising the above components.

4. A disease risk determination device, comprising: A determination data acquisition unit that acquires determination fundus images and determination health examination data from a determination database; A health examination data prediction unit that inputs the determination fundus images and the determination health examination data into a first trained model and outputs predicted health examination data; A disease risk prediction unit that inputs the predicted health examination data output from the health examination data prediction unit and the determination fundus images into a second trained model and outputs predicted disease risk data; A prediction result presentation unit that outputs at least the predicted disease risk data to a user terminal. The first trained model is generated by inputting learning health examination data as correct data and learning fundus images as observation data into a neural network. The second trained model is generated by inputting learning disease risk data as correct data and the learning fundus images as observation data into a neural network. A disease risk determination device characterized by the above features.

5. The disease risk determination device according to claim 4, wherein the prediction result presentation unit further outputs the predicted health examination data corresponding to the predicted disease based on the predicted disease risk data to a user terminal.

6. A method for determining disease risk, comprising: a step in which a determination data acquisition unit acquires a determination fundus image and determination health examination data from a determination database; a step in which a health examination data prediction unit inputs the determination fundus image and the determination health examination data into a first learned model and outputs predicted health examination data; a step in which a disease risk prediction unit inputs the predicted health examination data output from the health examination data prediction unit and the determination fundus image into a second learned model and outputs predicted disease risk data; and a step in which a prediction result presentation unit outputs the predicted disease risk data and the predicted health examination data corresponding to the predicted disease based on the predicted disease risk data to a user terminal. The first learned model is generated by inputting learning health examination data as correct answer data and learning fundus images as observation data into a neural network. The second learned model is generated by inputting learning disease risk data as correct answer data and the learning fundus images as observation data into a neural network. A disease risk determination method characterized by the above.

7. A program for causing a computer to execute each step of the method according to claim 6.

Images