Health risk prediction system, computer implementation method, and computer program

Abstract

Predicting individual health risks. [Solution] The disclosed health risk prediction system is configured to perform a process that includes acquiring input data including an individual's medical claims data, health checkup data, and long-term care claims data, inputting the acquired input data into a health risk prediction model, and obtaining output data relating to the individual's health risk. The health risk prediction model is machine-trained using training data having feature data including medical claims data, health checkup data, and long-term care claims data, and correct labels which are data relating to health risks corresponding to the feature data, and is configured to output the output data relating to the individual's health risk from the input data including the individual's medical claims data, health checkup data, and long-term care claims data.

Description

Technical Field

[0001] The present disclosure relates to a health risk prediction system, a computer-implemented method, and a computer program.

Background Art

[0002] Non-Patent Document 1 discloses predicting the risk of needing care by utilizing AI and the long-term care insurance system.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

[0004] The problem to be solved by the present disclosure is to predict an individual's health risk or provide related technologies.

[0005] The technology of the present disclosure from a certain aspect includes obtaining input data including an individual's medical receipt data, health examination data, and care receipt data, inputting the obtained input data into a health risk prediction model, and obtaining output data regarding the individual's health risk. The health risk prediction model is machine-learned using learning data including feature data including medical receipt data, health examination data, and care receipt data, and a correct label which is data regarding the health risk corresponding to the feature data, and is configured to output the output data regarding the individual's health risk from the input data including the individual's medical receipt data, health examination data, and care receipt data.

[0006] The technology described herein may be implemented as a system, method, or computer program.

[0007] Further details will be described in the embodiments below. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 is a diagram showing the configuration of the health risk prediction system. [Figure 2] Figure 2 is a flowchart showing the steps for building the training dataset and performing machine learning. [Figure 3] Figure 3 is an explanatory diagram of dimensionality reduction. [Figure 4] Figure 4 shows an example of a summary of disease names and medical procedures. [Figure 5] Figure 5 shows an example of training data. [Figure 6] Figure 6 shows a modified version of the health risk prediction system. [Figure 7] Figure 7 is a flowchart showing the procedure for health risk prediction processing. [Figure 8] Figure 8 shows an example of input data. [Figure 9] Figure 9 shows an example of input data. [Figure 10] Figure 10 shows an example of output data. [Figure 11] Figure 11 shows the prediction performance. [Figure 12] Figure 12 shows the percentage increase in the main features. [Figure 13] Figure 13 shows the various prediction performances in the experimental results. [Modes for carrying out the invention]

[0009] <1. Overview of the health risk prediction system, computer implementation method, and computer program>

[0010] (1) The health risk prediction system according to the embodiment may be configured to perform a process that includes acquiring input data including an individual's medical claims data, health checkup data and long-term care claims data, inputting the acquired input data into a health risk prediction model, and obtaining output data relating to the individual's health risk. The health risk prediction model is machine-trained using training data having feature data including medical claims data, health checkup data and long-term care claims data, and ground truth labels which are data relating to health risks corresponding to the feature data, and may be configured to output the output data relating to the individual's health risk from the input data including the individual's medical claims data, health checkup data and long-term care claims data.

[0011] (2) The health risk prediction model may consist of an explainable artificial intelligence that outputs risk factors for the health risk as output data relating to the health risk of the individual.

[0012] (3) The risk factors mentioned above may be risk factors corresponding to data items included in at least one of the medical claims data, health checkup data, and long-term care claims data.

[0013] (4) The health risks may include the risk of needing long-term care. The risk factors may include risk factors for the risk of needing long-term care that correspond to data items included in at least one of the medical claims data and the health checkup data.

[0014] (5) The health risks may include the risk of needing long-term care. The risk factors may include one or more risk factors selected from the group consisting of disease names, medical procedures, medications, and health tests, which are risk factors for the risk of needing long-term care.

[0015] (6) The health risks may include disease risks. The risk factors may include risk factors for the disease risks that correspond to data items included in the long-term care claims data.

[0016] (7) The health risk may include a disease risk. The risk factor may be a risk factor for the disease risk and may include a risk factor corresponding to a data item included in at least one of the medical receipt data and the health examination data.

[0017] (8) The health risk prediction model may be configured to output a care-required risk and a disease risk as the health risks.

[0018] (9) The input data including an individual's medical receipt data, health examination data, and care receipt data may be data including an individual's medical receipt data, health examination data, and care receipt data in the National Health Insurance Database (KDB).

[0019] (10) The processing may further include performing dimensionality reduction on the input data before inputting the acquired input data into the health risk prediction model.

[0020] (11) In the dimensionality reduction, among the acquired input data, a plurality of predetermined first data items to be aggregated are aggregated so that the number of data items decreases, and a plurality of predetermined second data items other than the first data items maintain the number of data items without being aggregated. This may be included.

[0021] (12) The dimensionality reduction may include not performing dimensionality reduction on the health examination data and the care receipt data among the acquired input data, and performing dimensionality reduction on the medical receipt data.

[0022] (13) The dimensionality reduction may include not performing dimensionality reduction on the health examination data and the care receipt data among the acquired input data, and among the medical receipt data, a plurality of predetermined data items to be aggregated are aggregated so that the number of data items decreases, and other data items maintain the number of data items without being aggregated. This may be included.

[0023] (14) The output data relating to the health risks of the individual may include a care risk value indicating the probability that the individual will become in need of care, and risk factors for the care risk.

[0024] (15) The output data relating to the health risks of the individual may include disease risk values ​​indicating the probability that the individual will develop a particular disease, and risk factors for disease risk.

[0025] (16) A method according to the embodiment may be a computer implementation method executed by a computer. A method according to the embodiment may include acquiring input data including an individual's medical claims data, health checkup data and long-term care claims data, inputting the acquired input data into a health risk prediction model to obtain output data relating to the individual's health risk. The health risk prediction model is machine-trained using training data having feature data including medical claims data, health checkup data and long-term care claims data, and ground truth labels which are data relating to health risks corresponding to the feature data, and may be configured to output the output data relating to the individual's health risk from the input data including the individual's medical claims data, health checkup data and long-term care claims data.

[0026] (17) The computer program according to the embodiment may be a computer program that causes a computer to perform processing. The processing may include acquiring input data including an individual's medical claims data, health checkup data and long-term care claims data, inputting the acquired input data into a health risk prediction model to obtain output data relating to the individual's health risk. The health risk prediction model is machine-trained using training data having feature data including medical claims data, health checkup data and long-term care claims data, and ground truth labels which are data relating to health risks corresponding to the feature data, and may be configured to output the output data relating to the individual's health risk from the input data including the individual's medical claims data, health checkup data and long-term care claims data.

[0027] <2. Examples of health risk prediction systems, computer implementation methods, and computer programs>

[0028] The embodiments will be described in more detail below with reference to the drawings.

[0029] <2.1 Examples of applications for health risk prediction systems>

[0030] The health risk prediction system 10 according to this embodiment uses a health risk prediction model 100 to predict an individual's health risk from their personal health data. The system 10 can, for example, be used to determine the health risks of multiple individuals from their individual data.

[0031] Hereafter, the health risk prediction system 10 will be simply referred to as "system 10," and the health risk prediction model 100 will be simply referred to as "prediction model 100."

[0032] System 10 can be used, for example, by local governments, nursing care facilities, hospitals, and other facilities and organizations that can utilize health risk information. By using System 10, it is possible to understand the health risks of multiple individuals and take necessary measures to avoid or delay the occurrence of health risks.

[0033] Health risks are risks related to an individual's health, such as the risk of needing long-term care or the risk of disease.

[0034] The risk of requiring long-term care is the risk that an individual will become in need of long-term care, for example, the risk of being certified as "Level 2 Long-Term Care" under Japan's Long-Term Care Insurance Act. With the aging population, the importance of preventative care, such as delaying the onset of the need for long-term care, is increasing. If we can predict each individual's risk of requiring long-term care using System 10, we can then take individualized preventative measures (preventative interventions) for individuals at high risk based on that prediction. As a result, effective preventative care becomes possible.

[0035] Disease risk is the risk of an individual developing a disease, such as osteoporosis or diabetes. Using System 10, if we can predict each individual's disease risk, we can identify hidden diseases in individuals who haven't undergone health checkups, and take early countermeasures based on these predictions. As a result, healthcare costs can be reduced.

[0036] System 10, as an example, predicts and outputs an individual's health risk, along with its risk factors. Risk factors are those that increase the health risk. For example, if the health risk is the risk of needing long-term care, the risk factors are diseases that increase the risk of needing long-term care. Examples of diseases that increase the risk of needing long-term care include osteoporosis, constipation, back pain, periodontal disease, or insomnia. If the health risk is the risk of disease, the risk factors are administered medications or other diseases.

[0037] By outputting health risks along with their risk factors, preventive or countermeasures that take these risk factors into account can be taken for individuals at high health risk. For example, if there is a disease that is a risk factor for needing long-term care, treating that disease can reduce the risk of needing long-term care.

[0038] Health risks can be output, for example, as a health risk value indicating the probability of that health risk occurring. This health risk value could be, for example, a caregiving risk value indicating the probability that an individual will require care (within a specified period), or a disease risk value indicating the probability that an individual will develop a specific disease (within a specified period). Furthermore, health risks may also be output in other ways that allow users to understand the prediction results, such as indicating whether or not there is a possibility of that health risk occurring.

[0039] <2.2 Example System Configuration>

[0040] Figure 1 shows an example of the configuration of system 10. System 10 shown in Figure 1 may consist of one or more computers.

[0041] The system 10 shown in Figure 1 may, as an example, include a user device 200 used by users of the system 10 (such as the aforementioned local government). The user device 200 is composed of a computer, such as a personal computer. Users can operate the user device 200 to determine the health risks of multiple individuals.

[0042] The computer comprising the user device 200 may include a processor 210 and a storage device 220 connected to the processor 210.

[0043] The processor 210 is a CPU, GPU, or other type of processor. The storage device 220 includes, for example, primary storage and secondary storage. The primary storage device is, for example, RAM. The secondary storage device is, for example, a hard disk drive (HDD) or a solid-state drive (SSD). The storage device 220 may contain a computer program 221 executed by the processor 210. The processor 210 reads and executes the computer program 221 stored in the storage device 220.

[0044] The computer program 221 contains program code that indicates instructions for the processor 210 to execute processes for health risk prediction, etc. When the computer program 221 is executed by the computer (processor), the computer operates as the user device 200 according to the embodiment and executes each step of the processes for health risk prediction, etc.

[0045] The computer program 221 may also be located on a server connected to the user device 200 (which acts as a client) via a network. In this case, the server executes the computer program 221 to perform processing for health risk prediction, etc. The user can access the server using the user device 200 and receive services such as health risk prediction.

[0046] For example, if a computer program 221 running on a server is configured as a web application that utilizes a web browser, a user device 200 can receive services such as health risk prediction if it has a web browser application.

[0047] The user device 200 may include an output device 230 for outputting output data such as predicted health risks. The output device 230 is, for example, a display. The user device 200 may also include a communication unit 240 for communicating with the outside world. The communication unit 240 is used to communicate with external computers 100, 300 via the Internet or other communication networks. The user device 200 may also include an input device (not shown) for receiving user operations, etc.

[0048] As described above, System 10 predicts health risks using the predictive model 100. In one embodiment, for example, the predictive model 100 is artificial intelligence (neural network) for predicting health risks and is provided by one or more computers outside the user device 200. The user device 200 transmits input data to the predictive model 100 and receives output data from the predictive model 100 via a network such as the Internet. The predictive model 100 may also be built as edge AI on the user device 200.

[0049] Predictive model 100 is configured, for example, as explainable artificial intelligence (explainable AI). Explainable AI is artificial intelligence that can explain the basis for its predictions / inferences. Predictive model 100, configured with explainable AI, can output predicted values ​​of health risks from input data, as well as the factors (risk factors) that underlie the predicted health risks.

[0050] The input data (original input data) provided to the prediction model 100 is, for example, obtained from an external system 300. The user device 200, for example, obtains the input data (original input data) from the external device 300 via a network and performs necessary processing, such as preprocessing, on the obtained input data. The user device 200 then transmits the preprocessed input data (preprocessed input data) to the prediction model 100. Note that preprocessing and other processing may be omitted if not necessary.

[0051] External system 300 is, as an example, the National Health Insurance Database (KDB) system. The KDB system is a system in Japan in which the National Health Insurance Federation utilizes various data such as "health checkups and health guidance," "medical care," and "nursing care" to create "statistical information" and "data on individual health" in order to support the creation and implementation of health programs by national health insurance providers and regional associations for medical care for the elderly. External system 300 may be other devices owned by users such as local governments, which may be devices that import data from the KDB system.

[0052] In one embodiment, as an example, "personal health data" from the KDB system is acquired as input data (original input data). The "personal health data" held by the KDB system includes individual historical data. In addition to information such as the individual's gender and age, the "personal health data" includes, as historical data, "medical claims data," "health checkup data," "long-term care claims data," and "long-term care certification survey data."

[0053] "Medical claims data" refers to data from medical claims (medical fee statements) such as medical claims and dispensing claims. Medical claims data includes data on what diagnoses (disease names), tests performed, treatments given, and how medications were prescribed for patients. For example, medical claims data has 5,697 data items for disease names and 6,069 data items for medical procedures, while dispensing claims data has 690 data items for medications (see original data in Figure 3).

[0054] "Health checkup data" refers to data from health checkup results (health examinations). Health checkup data includes 52 items related to behavioral habits, 21 items related to medical history / medication, and 123 items related to biochemical tests (see original data in Figure 3).

[0055] "Care claims data" refers to data from care insurance claims, including data indicating the care services provided. The care claims data has 621 data items related to care services (see original data in Figure 3).

[0056] "Care needs assessment survey data" (care needs assessment data) includes data indicating the level of care needed (including support needed). Examples of care needs include Level 1 care, Level 2 care, Level 1 support, Level 2 support, etc.

[0057] The original input data acquired for health risk prediction is preferably one or more data selected from the group consisting of "medical claims data," "health checkup data," "long-term care claims data," and "long-term care certification survey data," more preferably two or more or three or more data, and even more preferably all four data: "medical claims data," "health checkup data," "long-term care claims data," and "long-term care certification survey data."

[0058] When using two or more data selected from the group consisting of "medical claims data," "health checkup data," "long-term care claims data," and "long-term care certification survey data" to predict health risks, the prediction model 100 can predict health risks by considering a wider range of factors compared to when using only one data selected from this group. Therefore, an improvement in prediction accuracy can be expected.

[0059] The original input data acquired for health risk prediction preferably includes at least "medical claims data," health checkup data, and "long-term care claims data." In this case, it is preferable that the prediction model 100 considers all of the "medical claims data," "health checkup data," and "long-term care claims data," rather than just one of them, to make health risk predictions.

[0060] For example, if a health risk is the risk of needing long-term care, it might be possible to predict the risk of needing long-term care by considering only "long-term care claims data" related to long-term care. However, since diseases and other factors can also be considered risk factors for needing long-term care, considering "medical claims data" that includes information on diseases and other factors can lead to a more accurate prediction of the risk of needing long-term care. Furthermore, since the results of health examinations can also be considered risk factors for needing long-term care, considering the results of health examinations can also lead to a more accurate prediction of the risk of needing long-term care.

[0061] Furthermore, not only "nursing care claim data" containing information about nursing care, but also "medical claim data" and "health checkup data" containing information about diseases, etc., are suitable for a predictive model 100 composed of explainable artificial intelligence. In other words, by inputting "medical claim data," "health checkup data," and "nursing care claim data" into a predictive model 100 composed of explainable artificial intelligence, it becomes possible not only to output data items related to nursing care as risk factors for the risk of requiring nursing care, but also to output data items related to diseases, etc. and data items related to health checkup results as risk factors for the risk of requiring nursing care, which is preferable.

[0062] Similarly, when health risk refers to disease risk, disease risk can be predicted more appropriately by considering not only "medical claims data" related to the disease, but also "health checkup data" including information on health checkup results and "nursing care claims data" including information on nursing care. Furthermore, the predictive model 100, composed of explainable artificial intelligence, is preferable because it can output data items related to diseases, etc., as risk factors for disease risk, as well as data items related to health checkup results and data items related to nursing care as risk factors for disease risk.

[0063] Thus, the risk factors output from the predictive model 100 may be risk factors corresponding to data items included in medical claims data, health checkup data, or long-term care claims data. For example, if the health risk output by the predictive model 100 is the risk of needing long-term care, the risk factors output by the predictive model 100 may be risk factors for the risk of needing long-term care, and may be risk factors corresponding to data items included in at least one of the medical claims data, health checkup data, or long-term care claims data. The risk factors for the risk of needing long-term care may include one or more risk factors selected from the group consisting of disease name, medical treatment, medication, and health examination. Also, if the health risk output by the predictive model 100 is the risk of disease, the risk factors output by the predictive model 100 may be risk factors for disease, and may be risk factors corresponding to data items included in at least one of the long-term care claims data, medical claims data, or health checkup data.

[0064] In the following example, personal information such as an individual's gender and age, and three data items from the "Care Needs Assessment Survey Data" (Care Level 1, Support Level 2, and Support Level 1), as well as the individual's "Medical Claims Data," "Health Checkup Data," and "Care Claims Data," are acquired as input data (original input data). In addition, the individual's gender, age, Care Level 1 to Support Level 1 certification, "Medical Claims Data," "Health Checkup Data," and "Care Claims Data" are also used as training data for machine learning of predictive model 100.

[0065] <2.3 Machine Learning for Predictive Models>

[0066] Figure 2 shows the procedure for machine learning of the prediction model 100. For machine learning, a training dataset is constructed (steps S21-S24), and the prediction model 100 is trained (machine learning) using this training dataset (step S25).

[0067] In step S21, data acquisition takes place. Here, in addition to personal information such as an individual's gender and age, data from the KDB system 300 is acquired, including that individual's "medical claims data," "health checkup data," "long-term care claims data," and "long-term care certification survey data," which are also acquired as original data for machine learning training. The data acquired in step S21 consists of data from multiple individuals.

[0068] In step S22, personal information is anonymized. The data obtained in step S21 contains personal information, so it is anonymized through processes such as k-anonymization. Regarding age, for example, the data can be stratified in 5-year increments, and those 85 years and older can be grouped together as one age group.

[0069] In step S23, data items are aggregated (dimensionality reduction). As shown in Figure 3, the original data (original input data) for "medical claims data," "health checkup data," and "nursing care claims data" has 13,273 data items. Through the aggregation (dimensionality reduction) of data items in step S23, usable data is obtained with the data items aggregated to 2,573 items.

[0070] As shown in Figure 3, in the original data, the number of data items for "disease name" (disease name) (5,697 items) and "medical procedure" (6,069 items) included in "medical claims data" is an order of magnitude greater than the number of data items for other categories such as "drugs," "nursing care services," and "health checkups." Hereafter, the data items for "disease name" and "medical procedure," which have a large number of data items, will be referred to as "first data items," and the data items for other categories such as "drugs," "nursing care services," and "health checkups" will be referred to as "second data items."

[0071] In one embodiment, to balance the number of data items, the first data item, which has a large number of data items, is targeted for aggregation (dimensionality reduction), while the second data item, which has a small number of data items, is not aggregated and its number of data items is maintained. In other words, medical claims data is dimensionally reduced, but health checkup data and long-term care claims data are not dimensionally reduced, and their number of data items is maintained. This is preferable as it balances the number of data items between medical claims data and long-term care claims data. Furthermore, "health checkup data" is also not dimensionally reduced, and its number of data items is maintained.

[0072] Furthermore, among the data items included in medical claims data, certain data items (first medical data items) are aggregated (dimensionality reduction) to reduce the number of data items, while other data items (second medical data items) are not aggregated and the number of data items remains the same.

[0073] For example, dementia encompasses a diverse range of conditions, and as shown in Figure 4(A), it is classified into nine International Classification of Diseases (ICD) codes for medical fee claims. Of these, three disease names—Alzheimer's disease (ICD code: G309), dementia (ICD code: F03), and Alzheimer's disease in early senile dementia (ICD code: G300)—are more widely recognized than other disease names and are frequently observed in medical claims data. Furthermore, the inventors confirmed using explainable artificial intelligence that these three disease names are output as important features for predicting the risk of requiring long-term care in exploratory machine learning. Therefore, treating them as independent variables (features) as unaggregated second medical data items is highly significant.

[0074] The "aggregation code" in Figure 4(A) shows the code (identifier) ​​for each data item after aggregation in step S23. For Alzheimer's disease (ICD code: G309), dementia (ICD code: F03), and Alzheimer's disease in early senile dementia (ICD code: G300), independent aggregation codes (01021x_1, 01021x_2, 01021x_1) are assigned without aggregation.

[0075] On the other hand, the other six dementia-related disease names (ICD codes: F010, F011, F012, F019, G301, F308) have relatively low general awareness and are observed infrequently. In such cases, these six minor disease names are grouped together as "other dementias" according to the DPC classification (diagnostic group classification) and treated as the first medical data item to be aggregated, resulting in a single data item (aggregation code: 0121x_oth). This reduces the number of data items (number of features) and the complexity of the feature configuration. As a result, it is preferable to reduce the computational load during learning and prediction. Furthermore, instead of outputting the six minor disease names as risk factors from the prediction model 100, the grouped "other dementias" are output as risk factors, making it easier for users to understand the prediction results from the prediction model 100.

[0076] Similarly, medical procedures are also aggregated. For example, in the case of bone mineral density measurement, although it is a test performed for the same purpose, as shown in Figure 4(B), it is classified in detail by test method (medical fee claim code classification: D217-1, 2, 3, 4) when medical fee claims are submitted. Therefore, these four data items are aggregated as the first medical data item subject to aggregation, according to the classification of the medical fee claim code, into a single data item (aggregation code: D217 bone mineral density test). This reduces the number of data items (number of features) and the complexity of the feature configuration. As a result, it is preferable as it reduces the computational load during learning and prediction. Also, as a risk factor output from prediction model 100, instead of outputting the test method by method, the aggregated "bone mineral density test" is output. As a result, users can obtain appropriate information necessary for interpreting the prediction of the risk of needing long-term care, namely "whether or not the test (necessary for the treatment of the relevant disease) was performed and its purpose," which is preferable.

[0077] As described above, among the disease name / medical procedure data items included in the medical claims data, a predetermined number of primary medical data items are aggregated (dimensionality reduced), resulting in usable data (see Figure 3) with 814 disease name / medical procedure data items and 252 medical procedure data items, respectively.

[0078] In step S24, the usage data obtained in step S23 is used as feature data for the training data, and correct labels are assigned to each person's feature data.

[0079] As shown in Figure 5, the feature data here, as an example, indicates the applicability of each data item (feature), with 1 indicating applicability and 0 indicating non-applicability.

[0080] The correct label is selected appropriately according to the health risk to be predicted. For example, if the health risk to be predicted is the risk of needing long-term care, the correct label used is whether the individual is eligible for long-term care (within a specified period). If the health risk to be predicted is a specific disease, the correct label used is whether or not the disease has developed (within a specified period). In Figure 5, as an example, whether or not the individual has been certified as requiring Level 2 long-term care (within two years) is assigned as the correct label. If certified, it is indicated by 1, and if not certified, it is indicated by 0. The presence or absence of Level 2 long-term care certification, which serves as the correct label, is obtained from the "Long-Term Care Certification Survey Data". Note that the presence or absence of Level 1 long-term care certification, Level 2 support certification, and Level 1 support certification in the "Long-Term Care Certification Survey Data" may be included as data items in the feature data rather than being the correct label. By including the presence or absence of Level 1 long-term care to Level 1 support certification in the feature data, the prediction of Level 2 long-term care certification can be made more appropriate.

[0081] In step S25, the predictive model 100 is trained using the training data (training dataset) obtained as described above. This constructs a predictive model 100 that can predict health risks according to the type of correct label in the training data. Furthermore, by training an explainable artificial intelligence as described above, a predictive model 100 is constructed that outputs health risks along with their risk factors.

[0082] The predictive model 100 constructed as described above is an example of a predictive model trained using machine learning with training data that includes feature data, such as medical claims data, health checkup data, and long-term care claims data, and ground truth labels, which are data relating to health risks corresponding to the feature data. This predictive model 100 can output the output data relating to an individual's health risks from the input data, which includes the individual's medical claims data, health checkup data, and long-term care claims data.

[0083] The health risks predicted by the prediction model 100 do not have to be just one; as shown in Figure 6, there may be multiple. In other words, as shown in Figure 6, the prediction model 100 may be a collection of multiple prediction models 100A, 100B, 100C (sub-prediction models), each predicting a different health risk.

[0084] For example, as shown in Figure 6, the prediction model 100 may comprise a long-term care risk prediction model 100A and one or more disease risk prediction models 100B, 100C. If the input data to the prediction model 100 includes all of the medical claims data, health checkup data, and long-term care claims data, the inputs for the long-term care risk prediction model 100A and the disease risk prediction models 100B, 100C can be the same. That is, the long-term care risk prediction model 100A can output the long-term care risk (and risk factors if necessary) from input data that includes both medical claims data, health checkup data, and long-term care claims data, and the disease risk prediction models 100B, 100C can also output disease risk (and risk factors if necessary) from input data that includes both medical claims data and long-term care claims data.

[0085] Furthermore, as shown in Figure 6, the prediction model 100 may include multiple disease risk prediction models 100B and 100C (the long-term care risk prediction model 100A may be omitted). The input data for the multiple disease risk prediction models 100B and 100C may be the same, and from this common input data, the disease risk of different diseases can be predicted. In Figure 6, the first disease risk prediction model 100B predicts the risk of osteoporosis, and the second disease risk prediction model 100C predicts the risk of diabetes.

[0086] <2.4 Health Risk Prediction Processing>

[0087] Figure 7 shows an example of the steps involved in the health risk prediction process performed by the user device 200, etc. In the health risk prediction process, input data is acquired, and the acquired input data is input into the prediction model 100 to obtain output data related to health risks (steps S71 to S75 in Figure 7).

[0088] In the health risk prediction process shown in Figure 7, preprocessing is performed on the input data acquired in step S71 (step S72), and the preprocessed input data is input to the prediction model 100 (step S73). The prediction model 100 outputs output data related to health risks (step S74). The output data is presented to the user by the output device 230 of the user device 200 (step S75). This allows the user to understand their personal health risks, etc.

[0089] The input data acquired in step S71 is, for example, similar to the data acquired in step S21 in Figure 2, and includes personal information such as gender and age of multiple individuals from the data held by the KDB system 300, as well as each individual's "medical claim data," "health checkup data," "nursing care claim data," and "nursing care certification survey data." Regarding personal information, age can be stratified in 5-year increments, for example, and those 85 years and older can be aggregated into a single age group. The "nursing care certification survey data" may include data items indicating whether or not the individual is certified as requiring care level 1, support level 2, or support level 1.

[0090] The preprocessing in step S72 is the same process as the aggregation (dimensionality reduction) of data items shown in step S23 in Figure 2 and Figure 3, reducing the number of data items in the original input data from 13,273 to 2,573. In the dimensionality reduction in step S72, in order to balance the number of data items, the first data item with the most data items is targeted for aggregation (dimensionality reduction), while the second data item with fewer data items is not aggregated and its number of data items is maintained. In other words, medical claims data is dimensionally reduced, but long-term care claims data is not, and its number of data items is maintained. Similarly, "health checkup data" is not dimensionally reduced, and its number of data items is maintained.

[0091] Furthermore, among the data items included in medical claims data, certain first medical data items are aggregated (dimensionality reduced) to reduce the number of data items, while other second medical data items are not aggregated and the number of data items remains the same.

[0092] Figures 8 and 9 show examples of some dimensionality-reduced input data for a given individual. In Figures 8 and 9, data items that are applicable (1) are shown, while data items that are not applicable (0) are omitted. Figure 10 shows an example of output data when the input data, including the data items from Figures 8 and 9, is input into a predictive model 100 that predicts the risk of being certified as requiring Level 2 long-term care within two years. The output data in Figure 10 shows that the risk of requiring Level 2 long-term care (risk value; probability value) is 0.9016371, and that the most influential risk factor is "Dementia_1 Alzheimer's disease".

[0093] <2.5 Experimental Examples>

[0094] The following describes the experimental results regarding health risk prediction. In the experiment, artificial intelligence was trained on numerous patterns ranging from independent living to requiring Level 2 care. Based on the input of personal data, a predictive model 100 was constructed that predicts the risk (probability) and risk factors for reaching Level 2 care, and its performance was verified.

[0095] The experiment used 10 years of data from 2015 to 2025 for approximately 380,000 Kobe City long-term care insurance beneficiaries aged 65 and over, stored in the Kobe City Healthcare Data Linkage System. This data includes medical claims data (National Health Insurance + Late-Stage Elderly), health checkup data (Specific + Late-Stage Elderly), long-term care claims data, and long-term care certification survey data.

[0096] In the experiment, XGBoost and B3 were used as the artificial intelligence platform. B3 is an explainable artificial intelligence manufactured by Hitachi, Ltd.

[0097] For the construction of Predictive Model 100 (machine learning), the training data shown in Figure 5 was obtained using the procedure shown in Figure 2, with the most recent 21 months of medical claims data, health checkup data, and long-term care claims data. The correct labels (whether or not the person was certified as requiring Level 2 long-term care) were created based on the long-term care certification survey data.

[0098] Furthermore, data from 284,660 individuals (21 months' worth, from April 2015 to December 2017), excluding citizens who were certified as requiring Level 2 or higher care as of December 2017 and those presumed to have died during the experimental period, was divided into two sets of data (training and validation) in a ratio of approximately 3:1 (213,098 people: 71,562 people). Of the 213,098 training data points, 201,185 individuals were not certified as requiring Level 2 or higher care, while 11,913 individuals were, indicating an imbalance in the number of data points. To avoid this imbalance in the number of data points in machine learning, the data actually used for training was from 11,961 individuals out of the 201,185 who were not certified as requiring Level 2 or higher care.

[0099] The above training data was reduced in dimensionality as shown in Figure 3. Using the dimensionality-reduced 2,753 training data items, XGBoost machine learning was performed to construct a predictive model 100 that predicts the risk (probability) and risk factors for reaching Level 2 care needs.

[0100] To validate Predictive Model 100, data from 71,562 citizens not used in machine learning was employed. Of these 71,562 citizens, 4,002 received certification for Level 2 or higher long-term care within two years from January 2018, while 67,560 did not. Validation was performed by comparing the prediction results of Predictive Model 100 regarding long-term care needs with the actual long-term care certification status, and calculating the accuracy rate.

[0101] Figure 11 shows the validation results of the prediction performance of prediction model 100. As shown in Figure 11, the prediction accuracy (AUC) of prediction model 100 was 0.878, indicating excellent prediction performance. The sensitivity was 0.796 and the specificity was 0.811.

[0102] Furthermore, to examine the validity of prediction model 100, we analyzed the main features (data items included in the data) in prediction model 100 and their marginal effects (increase rates). Figure 12 shows the results, indicating the features (data items) that showed a high risk increase rate in prediction model 100. Since many of the features shown in Figure 12 coincide with disease names that are major factors in the need for long-term care, the validity of prediction model 100 was confirmed.

[0103] Furthermore, the same machine learning process used in XGBoost was also applied to B3. While the risk prediction accuracy of B3 was the same as that of XGBoost, B3 is an explainable artificial intelligence, so the prediction model 100 built with B3 was able to individually present not only the risk value for Level 2 care needs but also risk factors (disease name, prescribed medications, medical procedures, etc.).

[0104] Figure 13 shows the results of verifying prediction accuracy by changing the conditions.

[0105] Figure 13(A) shows the prediction results of XGBoost machine learning with dimensionality reduced to 2,753 items as described above, as well as the prediction accuracy of XGBoost and B3 machine learning with 13,273 items without dimensionality reduction. As shown in Figure 13(A), even when dimensionality is reduced to 2,753 items, prediction accuracy equal to or better than that of the case without dimensionality reduction is obtained, indicating that there is no decrease in prediction accuracy due to dimensionality reduction, and in fact that the prediction accuracy is slightly improved. Furthermore, dimensionality reduction makes it possible to reduce the processing load during both machine learning and prediction without causing a significant decrease in prediction accuracy. For example, the time required for machine learning was 28.16 seconds for 13,273 items without dimensionality reduction, compared to 15.12 seconds for 2,753 items with dimensionality reduction, showing a reduction in processing load. Similarly, the time required for risk prediction was 6.49 seconds for 13,273 items without dimensionality reduction, compared to 3.48 seconds for 2,753 items with dimensionality reduction, showing a reduction in processing load. These times are based on calculations performed with a CPU: Core i9-14900K processor, GPU: NVIDIA® RTX 5000 Ada generation, and memory: 32GB (16GB x 2) DDR5-4800.

[0106] Figure 13(B) shows the predictive accuracy when XGBoost and B3 are machine-trained using training data reduced to 2,753 dimensionality items as described above, as well as the predictive accuracy when XGBoost and B3 are machine-trained using training data reduced to 1,507 dimensionality items.

[0107] As shown in Figure 3, the 2,573 items represent the sum of disease names (814 items), medical procedures (252 items), medications (690 items), care services (621 items), and health checkups (196 items). On the other hand, the 1,507 items represent the sum of disease names (293 items), medical procedures (178 items), medications (690 items), care services (621 items), and health checkups (196 items). In other words, in the 1,507 items, there are 293 disease names instead of 814, and 178 medical procedures instead of 252.

[0108] For the 293 disease names and 178 medical procedure items, the data consists "only" of the data items that were identified as important features for predicting the risk of needing long-term care in B3's exploratory machine learning. For data items that are not important features, the dimensionality is reduced by deleting them rather than aggregating them, such as by adding 814 disease names and 252 medical procedure items.

[0109] Although the 2,573 data items include data items that are not important features regarding disease names and medical procedures, the prediction accuracy is comparable to that of the 1,507-item dataset. The inclusion of non-important data items in the 2,573-item dataset is advantageous when outputting risk factors. That is, even non-important data items can, in some cases, become risk factors. Therefore, including non-important data items makes it possible to output such risk factors, which is preferable.

[0110] Figure 13(C) shows the predictive accuracy when XGBoost is machine-learned using training data reduced to 2,753 items as described above (training data including medical-related data and nursing care-related data), as well as the predictive accuracy when XGBoost and B3 are machine-learned using training data consisting of nursing care-related data excluding medical-related data. The training data consisting of nursing care-related data excluding medical-related data used only nursing care claims data, data on independence level (level of support / care required), and age.

[0111] As shown in Figure 13(C), using training data (2,753 items) that includes medical-related data and nursing care-related data is preferable because it improves prediction accuracy compared to using training data consisting only of nursing care-related data that does not include medical-related data.

[0112] Figure 13(D) shows the prediction accuracy when XGBoost is trained using machine learning with the training data reduced to 2,753 items as described above (corresponding to "Using disease names and medical procedures selected by B3" in Figure 13(D)), and the prediction accuracy when XGBoost is trained using training data that does not use disease names and medical procedures selected by B3. The training data that does not use disease names and medical procedures selected by B3 contains, in addition to the disease names and medical procedures selected by B3, 239 disease names and 178 medical procedures randomly selected from the disease names and medical procedures not selected by B3, without data aggregation or deletion.

[0113] Figure 13(D) shows that when B3 did not select any disease names (239 items) or medical procedures (178 items), the sensitivity (the ability to classify high-risk individuals) decreased significantly. On the other hand, when XGBoost was machine-trained using training data with reduced dimensionality to 2,753 items (corresponding to "using disease names and medical procedures selected by B3" in Figure 13(D)), the prediction accuracy, including sensitivity, was good, indicating that dimensionality reduction was performed more appropriately.

[0114] The present invention is not limited to the above embodiments, and various modifications are possible. [Explanation of Symbols]

[0115] 10: Health Risk Prediction System 100: Health Risk Prediction Model 100A: Long-term care risk prediction model 100B: First disease risk prediction model 100C: Second disease risk prediction model 200: User device 210: Processor 220: Storage device 221: Computer Program 230: Output device 240: Communications Department 273: Number of data items 300: External system (KDB system)

Claims

1. We acquire input data including individual medical claims data, health checkup data, and long-term care claims data. The acquired input data is input into a health risk prediction model to obtain output data regarding the individual's health risk. It is configured to perform a process that includes the following: The health risk prediction model is machine-trained using training data comprising feature data including medical claims data, health checkup data, and long-term care claims data, and ground truth labels which are data relating to health risks corresponding to the feature data. It is configured to output output data relating to an individual's health risks from the input data, which includes the individual's medical claims data, health checkup data, and long-term care claims data. Health risk prediction system.

2. The health risk prediction model is comprised of explainable artificial intelligence that outputs risk factors for the health risk as output data related to the individual's health risk. A health risk prediction system according to claim 1.

3. The aforementioned risk factors are risk factors corresponding to data items included in at least one of the following: medical claims data, health checkup data, and long-term care claims data. The health risk prediction system according to claim 2.

4. The aforementioned health risks include the risk of needing long-term care. The aforementioned risk factors are risk factors for the risk of requiring long-term care, and include risk factors corresponding to data items included in at least one of the medical claims data and the health checkup data. The health risk prediction system according to claim 2.

5. The aforementioned health risks include the risk of needing long-term care. The aforementioned risk factors are risk factors for the risk of requiring long-term care, and include one or more risk factors selected from the group consisting of disease name, medical treatment, medication, and health examination. The health risk prediction system according to claim 2.

6. The aforementioned health risks include disease risks. The aforementioned risk factors are risk factors for the disease risk, and include risk factors corresponding to data items included in the care claim data. The health risk prediction system according to claim 2.

7. The aforementioned health risks include disease risks. The aforementioned risk factors are risk factors for the disease risk, and include risk factors corresponding to data items included in at least one of the medical claims data and the health checkup data. The health risk prediction system according to claim 2.

8. The aforementioned health risk prediction model is configured to output the risk of requiring long-term care and the risk of disease as health risks. A health risk prediction system according to claim 1.

9. The aforementioned input data, which includes individual medical claims data, health checkup data, and long-term care claims data, is data from the National Health Insurance Database (KDB) that includes individual medical claims data, health checkup data, and long-term care claims data. A health risk prediction system according to claim 1.

10. The process further includes performing dimensionality reduction on the input data before inputting the acquired input data into the health risk prediction model. A health risk prediction system according to claim 1.

11. The dimensionality reduction includes aggregating a predetermined number of first data items from the acquired input data so as to reduce the number of data items, while maintaining the number of data items for a predetermined number of second data items other than the first data items without aggregation. A health risk prediction system according to claim 10.

12. The aforementioned dimensionality reduction includes, among the acquired input data, not reducing the dimensionality of the health checkup data and the long-term care claim data, but reducing the dimensionality of the medical claim data. A health risk prediction system according to claim 10.

13. The dimensionality reduction described above includes, among the acquired input data, not reducing the dimensionality of the health checkup data and the long-term care claim data, and among the medical claim data, aggregating a predetermined number of data items that are subject to aggregation in such a way that the number of data items is reduced, while maintaining the number of data items for other data items without aggregation. A health risk prediction system according to claim 10.

14. The output data relating to the individual's health risks includes a caregiving risk value indicating the probability that the individual will require care, and risk factors for the caregiving risk. A health risk prediction system according to claim 1.

15. The output data relating to the individual's health risks includes a disease risk value indicating the probability that the individual will develop a specific disease, and risk factors for disease risk. A health risk prediction system according to claim 1.

16. A computer implementation method that is performed by a computer, We acquire input data including individual medical claims data, health checkup data, and long-term care claims data. The acquired input data is input into a health risk prediction model to obtain output data regarding the individual's health risk. This includes, The health risk prediction model is machine-trained using training data comprising feature data including medical claims data, health checkup data, and long-term care claims data, and ground truth labels which are data relating to health risks corresponding to the feature data. It is configured to output output data relating to an individual's health risks from the input data, which includes the individual's medical claims data, health checkup data, and long-term care claims data. Computer implementation method.

17. A computer program that causes a computer to perform a process, The aforementioned process is, We acquire input data including individual medical claims data, health checkup data, and long-term care claims data. The acquired input data is input into a health risk prediction model to obtain output data regarding the individual's health risk. This includes, The health risk prediction model is machine-trained using training data comprising feature data including medical claims data, health checkup data, and long-term care claims data, and ground truth labels which are data relating to health risks corresponding to the feature data. It is configured to output output data relating to an individual's health risks from the input data, which includes the individual's medical claims data, health checkup data, and long-term care claims data. Computer program.

Images