Disease severity prediction system

Abstract

To help predict, by machine learning, whether or not aggravation of a disease is present.SOLUTION: A method for generating a learning model for predicting aggravation according to the present invention generates a learning model for predicting whether or not aggravation of a disease a patient contracts is present. The method comprises: an information acquisition step S102 of acquiring analytic object patient information for each of a plurality of items obtained from a plurality of analytic object patients each having a disease contract history; a first learning step S103 of constructing a primary learning model, using the analytic object patient information as an explanatory variable, and presence or absence of the aggravation as an objective variable; a contribution acquisition step S104 of acquiring, in the primary learning model, contribution levels; a selection step S106 of selecting a plurality of specific items from the plurality of items on the basis of the contribution levels; and a second learning step S107 of constructing a secondary learning model serving as an aggravation prediction-purpose learning model using the analytic object patient information corresponding to the plurality of specific items as the explanatory variable, and the presence or absence of the aggravation as the objective variable.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention relates to a method, a program, and a method for generating a learning model for predicting the severity of a disease, a disease severity prediction device, a disease severity prediction program, a disease severity prediction method, and a disease severity prediction system. [Background technology]

[0002] Generally, the severity of a disease is determined by a doctor based on clinical information such as the degree of subjective symptoms, the measurement values ​​of various tests, the number of exacerbations (e.g., attacks), etc. The worsening of the disease leads to a decrease in quality of life (QOL) and an increase in medical expenses due to worsening subjective symptoms, emergency visits due to exacerbations, and hospitalization, which places a social and economic burden on patients.

[0003] Various pieces of information obtained from patients (such as gender and age, patient behavioral history such as smoking, and disease-related test data) have been reported as factors related to the worsening and severity of disease. However, comprehensive analysis of these factors and prediction of the severity of disease for each individual patient depends on individual differences based on the doctor's years of experience in the department and the number of cases they have experienced. Therefore, predicting the severity of disease is difficult.

[0004] A system has been proposed to date that supports the selection of a therapeutic drug based on multiple pieces of information obtained from a patient with bronchial asthma (see, for example, Patent Document 1).

[0005] The system disclosed in Patent Document 1 classifies bronchial asthma phenotypes and supports the selection of therapeutic drugs, including the use of inhaled steroids, by measuring and analyzing multiple pieces of information obtained from bronchial asthma patients (gene expression data, airway and blood eosinophil counts, immunoglobulins, periostin, etc.). However, this system does not predict the worsening of bronchial asthma. Therefore, this system cannot be applied to preventing the worsening of bronchial asthma by implementing early therapeutic intervention based on predictions. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Special Publication No. 2011-523350 Summary of the Invention [Problem to be solved by the invention]

[0007] The present invention aims to use machine learning to assist in predicting whether a disease will worsen. [Means for solving the problem]

[0008] The method for generating a learning model for predicting the worsening of a disease according to the present invention is a method for generating a learning model for predicting the worsening of a disease that a patient is suffering from, and is characterized by comprising: an information acquisition step for acquiring patient information for each of a plurality of items obtained from a plurality of patients to be analyzed who have a history of the disease; a first learning step for constructing a primary learning model for predicting the worsening of a disease using the patient information for the plurality of items as an explanatory variable and the presence or absence of the worsening of a disease as the objective variable; a contribution acquisition step for acquiring the contribution of each of the plurality of items to improving the prediction accuracy of the worsening of a disease in the primary learning model; a selection step for selecting multiple specific items from the multiple items based on the contribution; and a second learning step for constructing a secondary learning model that is a learning model for predicting the worsening of a disease using the patient information for the plurality of specific items as an explanatory variable and the presence or absence of the worsening of a disease as the objective variable. [Effects of the Invention]

[0009] According to the present invention, machine learning can be used to assist in predicting whether a disease will become severe. [Brief explanation of the drawings]

[0010] [Figure 1]1 is a network configuration diagram showing an embodiment of a disease severity prediction system according to the present invention. [Figure 2] FIG. 2 is a functional block diagram of a learning model generation device for predicting the progression of a disease, which is included in the system of FIG. 1. [Figure 3] 3 is a schematic diagram showing an example of information stored in a storage unit included in the device for generating a learning model for predicting progression of illness shown in FIG. 2. FIG. [Figure 4] FIG. 2 is a functional block diagram of a disease aggravation prediction device included in the system of FIG. 1. [Figure 5] 5 is a schematic diagram showing an example of information stored in a storage unit included in the severity prediction device of FIG. 4. FIG. [Figure 6] 1 is a flowchart of a learning model generation process for predicting the severity of a disease, showing an embodiment of a learning model generation method for predicting the severity of a disease according to the present invention. [Figure 7] 1 is a flowchart of a disease aggravation prediction process, showing an embodiment of a disease aggravation prediction method according to the present invention. [Figure 8] 5 is a schematic diagram showing an example of information displayed on a display unit provided in the disease aggravation prediction device of FIG. 4. FIG. [Figure 9] 1 is a graph showing the relationship between explanatory variables and contribution rates in an example of the present invention. [Figure 10] (a) is the ROC curve of the second-order learning model generated using the random forest model, and (b) to (g) are the ROC curves of the second-order learning model generated using the logistic regression model. [Figure 11] This is a table showing the results of validation using the 10-fold cross-validation method for a secondary learning model (random forest model) generated based on combinations of items within the top five rankings. [Figure 12] This is a table showing the results of validation using a 10-fold cross-validation method for a second-order learning model (logistic regression model) generated based on a combination of items within the top five rankings. DETAILED DESCRIPTION OF THE INVENTION

[0011] Below, with reference to the drawings, embodiments of the disease severity prediction system (hereinafter referred to as "the system") according to the present invention, a learning model generation device for predicting disease severity (hereinafter referred to as "the generation device"), a learning model generation program for predicting disease severity (hereinafter referred to as "the generation program"), a learning model generation method for predicting disease severity (hereinafter referred to as "the generation method"), a disease severity prediction device (hereinafter referred to as "the prediction device"), a disease severity prediction program (hereinafter referred to as "the prediction program"), and a disease severity prediction method (hereinafter referred to as "the prediction method") will be described.

[0012] The present invention generates a learning model (hereinafter referred to as a "learning model for predicting aggravation") for predicting whether or not a disease suffering from a patient will become severe, and predicts whether or not the disease suffering from the patient will become severe using the generated learning model for predicting aggravation.

[0013] A "patient" is a subject whose condition is predicted to worsen according to the present invention.

[0014] A "disease" is an abnormality that appears in the mind or body of a patient. In the present invention, a disease is an abnormality whose prediction of whether it will become severe or not is particularly likely to vary depending on the doctor's years of experience and the number of cases he or she has experienced. In the following explanation, the disease is, for example, bronchial asthma, which is one of the chronic respiratory diseases with the highest prevalence and incidence rates worldwide.

[0015] ●Severe illness prediction system● ●Configuration of the severity prediction system FIG. 1 is a network configuration diagram showing an embodiment of this system.

[0016] The present system 1 generates a learning model for predicting the severity of a disease that a patient is suffering from, and predicts the severity of the disease using the model. The present system 1 includes a generation device 10 and a prediction device 20.

[0017] The generating device 10 generates a learning model for predicting the severity of a disease that a patient is suffering from. The specific configuration and operation of the generating device 10 will be described later.

[0018] The prediction device 20 predicts whether or not the disease that the patient is suffering from will become severe. The specific configuration and operation of the prediction device 20 will be described later.

[0019] The external device 30 is, for example, an information storage device such as a server that stores information necessary for the operation of the present generation device 10 and the present prediction device 20 (for example, analysis target patient information, patient information, etc., which will be described later).

[0020] The network N is, for example, a communication network such as the Internet, a mobile communication network, a local area network (LAN), a wide area network (WAN), Wi-Fi (registered trademark), or Bluetooth (registered trademark).

[0021] Configuration of the learning model generation device for predicting progression FIG. 2 is a functional block diagram of the generation device 10.

[0022] The generating device 10 is realized, for example, by a personal computer. The generating program runs on the generating device 10, and the generating program cooperates with the hardware resources of the generating device 10 to realize the generating method.

[0023] Here, by causing a computer (not shown) to execute this generation program, the generation program can cause the computer to function in the same manner as the generation device 10, and cause the computer to execute the generation method.

[0024] The generation device 10 includes a communication unit 11, a storage unit 12, a control unit 13, an operation unit 14, and a display unit 15.

[0025] The communication unit 11 is connected to the prediction device 20 and the external device 30 via a network N. The communication unit 11 is configured by, for example, a communication module and a communication interface.

[0026] The storage unit 12 stores information (e.g., an analysis target patient information database DB1, which will be described later) necessary for the generation device 10 to execute the generation method, which will be described later. The storage unit 12 is configured, for example, by a recording device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive) provided in the generation device 10, and / or a storage medium such as a flash memory.

[0027] The "analysis subject patient information database (hereinafter sometimes simply referred to as "DB1")" is a database in which, for example, information for each of a plurality of items obtained from a plurality of analysis subject patients (hereinafter referred to as "analysis subject information") is stored for each disease. In other words, DB1 stores a plurality of pieces of analysis subject information for each disease. In this embodiment, DB1 is generated, for example, by the user of the generation device 10 based on information held by medical institutions such as hospitals and health checkup centers. DB1 is generated in advance, for example, for each disease and stored in the storage unit 12.

[0028] A "subject to analysis" is a patient from whom information (i.e., training data) that serves as the explanatory variables and objective variables of the machine learning algorithm is obtained. In other words, the subject to analysis is a patient who has a history of a common disease for which the severity of the disease is to be predicted, and whose degree of severity over time (i.e., whether or not the disease has progressed) is known. Here, the disease that the subject to analysis is suffering from (had suffered from) is the same as the disease that the patient is suffering from.

[0029] "Items" are types of information obtained from the patient and the patient. Items include attributes of the patient (e.g., age, sex, etc.), behavioral history of the patient (e.g., smoking history, etc.), and test items that the patient (e.g., respiratory function test, medical interview items, etc.) has undergone.

[0030] FIG. 3 is a schematic diagram showing an example of information (DB1) stored in the storage unit 12. As shown in FIG. "Disease ID" is identification information specific to each disease. "Patient ID to be analyzed" is identification information specific to each patient to be analyzed. "Item ID" is identification information specific to each item. DB1 stores disease IDs, patient IDs to be analyzed, and item IDs in association with each other.

[0031] The figure shows, for example, that the information corresponding to the item ID "Y001" of the analysis target patient ID "X001" who has a history of suffering from the disease ID "A001" is "y01", and the information corresponding to the item ID "Y002" is "y02". For example, by referring to DB1 using the disease ID "A001", the generation device 10 can read (acquire) the analysis target person information stored in the memory unit 12 in association with the disease ID "A001" for each item.

[0032] Return to Figure 2. The control unit 13 controls the overall operation of the generation device 10 and executes the generation method described below. The control unit 13 is configured, for example, with a CPU (Central Processing Unit) provided in the generation device 10, a RAM (Random Access Memory) that functions as a work area for the CPU, and a ROM (Read Only Memory) that stores various information such as the generation program. The control unit 13 includes a generation information acquisition unit 131, a first learning unit 132, a contribution degree acquisition unit 133, a selection unit 134, a second learning unit 135, a model verification unit 136, a model selection unit 137, and an output unit 138.

[0033] The generated information acquisition unit 131 acquires the patient information to be analyzed for each of a plurality of items obtained from a plurality of patients to be analyzed. The specific operation of the generated information acquisition unit 131 will be described later.

[0034] The first learning unit 132 constructs a primary learning model for predicting whether or not a patient will become severe, using analysis target patient information corresponding to multiple items as explanatory variables and whether or not a patient will become severe as a response variable. Specific operations of the first learning unit 132 will be described later.

[0035] The contribution degree acquiring unit 133 acquires the contribution degree to improving the prediction accuracy of whether or not the condition will worsen for each of a plurality of items in the primary learning model. Specific operations of the contribution degree acquiring unit 133 will be described later.

[0036] The "contribution" is the degree to which each explanatory variable (analysis target patient information corresponding to an item) contributes (influences) to improving the prediction accuracy of whether or not a patient will become severe in the primary learning model. In this embodiment, the contribution is permutation importance.

[0037] It should be noted that the contribution in the present invention is not limited to the permutation importance, and may be, for example, a SHAP value.

[0038] The selection unit 134 selects a plurality of specific items from the plurality of items based on the degree of contribution. The specific operation of the selection unit 134 will be described later.

[0039] The "specific items" are the items corresponding to the patient information to be analyzed that are used as explanatory variables when generating the secondary learning model described below.

[0040] The second learning unit 135 constructs a secondary learning model for predicting whether or not a patient will become severe, using analysis target patient information corresponding to a plurality of specific items (hereinafter referred to as "specific target patient information") as explanatory variables and whether or not a patient will become severe as a response variable. The specific operation of the second learning unit 135 will be described later.

[0041] Here, the primary learning model and the secondary learning model are constructed using a common machine learning model. In this embodiment, the machine learning model is, for example, a random forest model.

[0042] Note that the machine learning model in the present invention is not limited to a random forest model, and may be, for example, a logistic regression model, a support vector machine model, or a naive Bayes model.

[0043] The model verification unit 136 verifies the constructed secondary learning model. The specific operation of the model verification unit 136 will be described later.

[0044] Based on the verification results of the constructed one or more secondary learning models, the model selection unit 137 selects the corresponding secondary learning model as a learning model for predicting the severity of the corresponding disease. The specific operation of the model selection unit 137 will be described later.

[0045] The output unit 138 outputs various information obtained by executing the generation method described below (for example, information requiring selection by the user of the generation device 10). The specific operation of the output unit 138 will be described later.

[0046] The operation unit 14 is a device that is operated (for example, to input or select information) by a user (for example, a medical professional such as a doctor or nurse) of the generation device 10. The operation unit 14 is, for example, a keyboard, a mouse, or a touch panel.

[0047] The display unit 15 is a device that displays information output by the control unit 13. The display unit 15 is, for example, a monitor or a display.

[0048] The operation unit and display unit in the generation device may be configured, for example, as a touch panel display.

[0049] Configuration of the severity prediction device FIG. 4 is a functional block diagram of the prediction device 20.

[0050] The present prediction device 20 is realized, for example, by a personal computer. The present prediction program runs in the present prediction device 20, and the present prediction program cooperates with the hardware resources of the present prediction device 20 to realize the present prediction method.

[0051] Here, by causing a computer (not shown) to execute the present prediction program, the present prediction program can cause the computer to function in the same manner as the present prediction device 20, and cause the computer to execute the present prediction method.

[0052] The prediction device 20 includes a communication unit 21 , a storage unit 22 , a control unit 23 , an operation unit 24 , and a display unit 25 .

[0053] The communication unit 21 is connected to the generation device 10 and the external device 30 via a network N. The communication unit 21 is configured by, for example, a communication module and a communication interface.

[0054] The storage unit 22 stores information (e.g., a patient information database DB2, a trained learning model for predicting aggravation, etc.) necessary for the prediction device 20 to execute the prediction method described below. The storage unit 22 is configured, for example, by a recording device such as an HDD or SSD provided in the prediction device 20, and / or a storage medium such as a flash memory.

[0055] The "patient information database DB2 (hereinafter sometimes simply referred to as "DB2")" is a database in which, for example, information on a plurality of items obtained from a patient is stored for each patient. That is, DB2 stores a plurality of pieces of patient information corresponding to one disease for each patient. In this embodiment, DB2 is generated, for example, by a user of the prediction device 20 based on information held by a medical institution. DB2 is generated in advance for each patient, for example, and stored in the storage unit 22.

[0056] FIG. 5 is a schematic diagram showing an example of information (DB2) stored in the storage unit 22. As shown in FIG. "Patient ID" is patient-specific identification information corresponding to each patient. DB2 stores patient IDs, disease IDs, and item IDs in association with each other. For example, by referring to DB2 using patient ID "Z001," the prediction device 20 can read (obtain) the patient information stored in association with patient ID "Z001" for each item.

[0057] Return to Figure 4. The control unit 23 controls the overall operation of the prediction device 20 and executes the prediction method described below. The control unit 23 is configured, for example, with a CPU provided in the prediction device 20, a RAM that functions as a work area for the CPU, and a ROM that stores various information such as the prediction program. The control unit 23 includes a prediction information acquisition unit 231, a prediction unit 232, and an output unit 233.

[0058] The predicted information acquisition unit 231 acquires patient information for each of a plurality of items obtained from the patient. The specific operation of the predicted information acquisition unit 231 will be described later.

[0059] The prediction unit 232 predicts whether or not the condition will worsen based on the patient information, using a trained learning model (a learning model for predicting the worsening condition, which will be described later). Specific operations of the prediction unit 232 will be described later.

[0060] The output unit 233 outputs various information (for example, prediction results) obtained by executing the prediction program described later. Specific operations of the output unit 233 will be described later.

[0061] The operation unit 24 is a device that is operated (for example, to input and select information) by a user (for example, a medical professional such as a doctor or nurse) of the prediction device 20. The operation unit 24 is, for example, a keyboard, a mouse, or a touch panel.

[0062] The display unit 25 is a device that displays information (for example, prediction results) output by the output unit 233. The display unit 25 is, for example, a monitor or a display.

[0063] The operation unit and the display unit in the present prediction device may be configured, for example, as a touch panel display.

[0064] Operation of the severity prediction system Next, the operation of the present system 1, that is, the operation of the present generation device 10 and the present prediction device 20, will be described.

[0065] Operation of the learning model generation device for predicting progression First, a description will be given of the operation of the generation device 10, that is, the generation method executed by the generation device 10. In the following description of the generation method, reference will also be made to FIG.

[0066] The generation device 10 executes a learning model generation process for predicting severity (S1) as the generation method. The "learning model generation process for predicting severity (S1)" is a process in which the generation device 10 generates a learning model for predicting severity of a disease by machine learning based on patient information to be analyzed.

[0067] FIG. 6 is a flowchart of a learning model generation process (S1) for predicting progression, showing an embodiment of the generation method.

[0068] First, the generation information acquisition unit 131 acquires information (disease ID) of the disease for which a learning model is to be generated (S101). Specifically, for example, the generation information acquisition unit 131 acquires the disease ID from the disease name input by the user of the generation device 10 via the operation unit 24. As a result, the generation information acquisition unit 131 acquires information (such as the name) of the disease for which the severity is to be predicted based on the acquired disease ID.

[0069] Next, the generated information acquisition unit 131 acquires the analysis target patient information (S102: information acquisition step). Specifically, the generated information acquisition unit 131 acquires the analysis target patient information stored in DB1 in association with the disease ID, for example, by referring to DB1 using the disease ID.

[0070] Next, the first learning unit 132 constructs a primary learning model for predicting whether or not the condition will worsen, using the patient information to be analyzed corresponding to each item as an explanatory variable and the presence or absence of worsening as a target variable (S103: first learning step). Here, the presence or absence of worsening is preset by the user of the generation device 10 based on, for example, the presence or absence of an item that is an indicator of the worsening of the disease, the magnitude of the value corresponding to the item, the relationship between the value corresponding to the item and a threshold (above the threshold, below the threshold, etc.), etc.

[0071] Next, the contribution degree obtaining unit 133 obtains the contribution degree to the prediction of the presence or absence of aggravation for each item in the learned primary learning model generated by the first learning unit 132 (S104: contribution degree obtaining step). Specifically, the contribution degree obtaining unit 133, for example, calculates the permutation importance for each item, obtains the calculated value as the contribution degree, and stores it in the storage unit 12 in association with the item. Here, the contribution degree (permutation importance) is calculated by, for example, performing a total of 100 iterations.

[0072] Next, the contribution degree obtaining unit 133 compares the contribution degrees for all the items and ranks each of the items based on the magnitude of the contribution degree (S105: ranking step). The items are ranked in descending order of contribution degree.

[0073] Next, the selection unit 134 selects a plurality of specific items from the plurality of items based on the contribution degree (S106: selection step). Specifically, the selection unit 134 selects a plurality of items ranked within the top "n" (n is an integer equal to or greater than 2) in terms of contribution degree as the specific items. Here, the selection unit 134 selects, for example, one combination of a plurality of items from the items ranked within the top "n". In other words, the specific items include a plurality of items ranked within the top "n".

[0074] When the combination of the value of "n" and 1 is preset, the selection unit of the present invention may select the preset combination. For example, the selection unit of the present invention may select specific items by always including the item in the "1" position in the combination. Furthermore, the selection unit of the present invention may select an item selected by the user of the generation device via the operation unit as the specific item.

[0075] Next, the second learning unit 135 constructs a secondary learning model for predicting whether or not the patient will become severe, using the patient information to be analyzed (specific patient information) corresponding to the multiple specific items among the patient information to be analyzed acquired in step S102 as explanatory variables and the presence or absence of the patient becoming severe as a response variable (S107: second learning step). The secondary learning model is stored in the storage unit 12.

[0076] Next, the model validation unit 136 validates the constructed secondary learning model (S108). Specifically, the model validation unit 136 performs ROC analysis of the secondary learning model using, for example, a 10-fold cross-validation method, and validates the secondary learning model by calculating the prediction accuracy of the secondary learning model. The validation results (AUC, accuracy rate) are associated with the secondary learning model and stored in the storage unit 12.

[0077] Next, the control unit 13 predicts whether or not there are any unselected combinations (S109). If there are any unselected combinations ("Y" in S109), the generation method returns to step S106.

[0078] On the other hand, if there are no unselected combinations ("N" in S109), the model selection unit 137 selects the secondary learning model with the best verification result (AUC, accuracy rate) from among the one or more constructed secondary learning models as the learning model for predicting the severity of the corresponding disease (S110). In other words, the generation device 10 constructs one secondary learning model as the learning model for predicting the severity of the corresponding disease. The learning model for predicting the severity of the corresponding disease is stored in the storage unit 12, for example, in association with the corresponding disease ID.

[0079] Next, the control unit 13 transmits the learning model for predicting aggravation to the present prediction device 20 via the communication unit 11 (S111). The transmitted learning model for predicting aggravation is stored in the storage unit 22 of the present prediction device 20.

[0080] The present generation method is not limited to the present embodiment. For example, in the process of selecting specific items, the selection unit of the present invention may select, as the specific items, items selected by a user of the present generation device via an operation unit. In this case, the output unit of the present prediction device outputs, for example, the contribution degree of each item to a display unit.

[0081] Furthermore, for example, the process of verifying a secondary learning model and the process of selecting a secondary learning model may be performed by a user of the prediction device, rather than by the prediction device.

[0082] Furthermore, for example, in the selection process of the secondary learning model, the model selection unit of the present invention may select the secondary learning model selected by the user of the present generation device via the operation unit as the learning model for predicting aggravation. In this case, the output unit of the present prediction device may output the verification result for each secondary learning model to the display unit, for example.

[0083] ● Operation of the severity prediction device Next, a description will be given of the operation of the present prediction device 20, that is, the present prediction method executed by the present prediction device 20. In the following description of the present prediction method, reference will also be made to FIG.

[0084] The present prediction device 20 executes a severity prediction process (S2) as the present prediction method. The "severity prediction process (S2)" is a process in which the present prediction device 20 predicts the severity of a disease based on a learning model for severity prediction and patient information.

[0085] FIG. 7 is a flowchart of the aggravation prediction process (S2) showing an embodiment of the present prediction method.

[0086] First, the prediction information acquisition unit 231 acquires information (disease ID) of the disease for which the progression is predicted (S201). Specifically, for example, the prediction information acquisition unit 231 acquires the disease ID from the disease name input by the user of the prediction device 20 via the operation unit 24. As a result, the prediction information acquisition unit 231 acquires information (such as the name) of the disease for which the progression is predicted based on the acquired disease ID.

[0087] Next, the prediction information acquisition unit 231 acquires patient information of the patient to be predicted (S202: prediction information acquisition step). Specifically, the prediction information acquisition unit 231 refers to DB2 using, for example, the patient ID (and disease ID) and acquires the patient information stored in the storage unit 22 in association with the patient ID.

[0088] Next, the prediction unit 232 reads out the learning model for predicting aggravation stored in the storage unit 22 (S203). Specifically, the prediction unit 232 refers to the storage unit 22 using, for example, a disease ID, and reads out from the storage unit 22 the learning model for predicting aggravation that is associated with the disease ID and stored in the storage unit 22. As described above, the learning model for predicting aggravation is generated by the generation device 10 executing the generation method.

[0089] The prediction unit in the present invention may read from the storage unit, for example, a learning model for predicting aggravation that is selected by the user of the prediction device via the operation unit.

[0090] Next, the prediction unit 232 predicts whether or not the patient's condition will worsen based on the learning model for predicting aggravation and the patient information (S204: prediction step). The prediction result is stored in the storage unit 22, for example, in association with the patient ID. Here, the prediction result is information indicating whether or not the patient will become severely ill.

[0091] Next, the output unit 233 outputs the prediction result to the display unit 25 (S205).

[0092] Next, the display unit 25 displays the prediction result (S206).

[0093] FIG. 8 is a schematic diagram showing an example of information (prediction results) displayed on the display unit 25. As shown in FIG. The figure shows that there is a "yes" possibility that patient "A's" illness "B" will worsen.

[0094] In this way, the present prediction device 20 predicts whether or not the patient's condition will worsen by executing the present prediction method. A user of the present prediction device 20 can determine (estimate) whether or not the patient's condition will worsen by viewing the prediction result displayed on the display unit 25. As a result, for example, a doctor who obtains the prediction result of the present prediction device 20 can formulate a treatment plan for the patient in advance based on the result of the determination of whether or not the patient's condition will worsen.

[0095] The prediction information acquisition unit in the present invention may acquire patient information input by the user of the prediction device via the operation unit.

[0096] ●Example● Next, an embodiment of the present invention will be described below, taking as an example a case where a learning model for predicting progression is generated using actual clinical data. In the following embodiment, the clinical data used was obtained from a three-year prospective observational study (FLOAT: Factors affecting the Long-term Asthma Therapy) conducted on 128 patients with stable bronchial asthma aged 25 years or older. Rapid decliners are defined as those with an average annual decline (decline over time) of -40 mL / year or less in forced exhaled volume in 1 second (FEV1). The objective variable is defined as whether or not a patient is a rapid decliner three years later. On the other hand, the explanatory variables (items) were "age," "gender," "BMI," "smoking history (Packyear)," "asthma symptom score ACT (Asthma Control Test)," "% predicted forced expiratory volume in 1 second (%FEV1)," "% predicted forced vital capacity (%FVC)," "exhaled nitric oxide (FeNO)," "white blood cell count," "blood eosinophil count," "IgE," "number of exacerbations," "presence or absence of rhinitis," "presence or absence of sinusitis," "presence or absence of reflux esophagitis," and "presence or absence of atopy."

[0097] First, a primary learning model was generated based on the explanatory variables and the target variable. Next, the contribution of each explanatory variable to improving the prediction accuracy (permutation importance) was obtained and ranked.

[0098] FIG. 9 is a graph showing the relationship between explanatory variables and contribution rates in an embodiment of the present invention. The figure shows that "number of exacerbations" ranked first, followed by "BMI," "age," "FeNO," and "blood eosinophil count" in second to fifth place.

[0099] Next, the top five contributing items (number of exacerbations, BMI, age, FeNO, and blood eosinophil count) were selected as specific items, and a secondary learning model was generated.

[0100] Figure 10(a) is an ROC curve diagram of a second-order learning model generated using a random forest model, and (b) to (g) are ROC curves of second-order learning models generated using a logistic regression model.

[0101] In the figure, (a) and (b) show an example in which the top five items are selected as specific items, (c) shows a comparative example in which only "number of exacerbations" is selected as a specific item, (d) shows a comparative example in which only "BMI" is selected as a specific item, (e) shows a comparative example in which only "age" is selected as a specific item, (f) shows a comparative example in which "blood eosinophil count" is selected as a specific item, and (g) shows a comparative example in which "FeNO" is selected as a specific item. As shown in Figure 5, the prediction accuracy is better when a combination of all of the top five items is selected as a specific item than when a single item in the top five contributions is selected as a specific item.

[0102] FIG. 11 is a table showing the results of validation using the 10-fold cross validation method for the secondary learning model (random forest model) generated based on the combination of the top five items. FIG. 12 is a table showing the results of validation using the 10-fold cross validation method for a second-order learning model (logistic regression model) generated based on a combination of items within the top five rankings.

[0103] As shown in Figures 11 and 12, the secondary learning model generated based on a combination of the top five items achieved a high AUC and accuracy rate by selecting at least "number of exacerbations" and "age" as specific items.

[0104] Thus, in the severity prediction learning model generated by the present invention for predicting whether or not a bronchial asthma patient will become severe, the random forest model yielded verification results of AUC: 0.80 (95% CI: 0.67-0.93) and accuracy rate: 78% (95% CI: 70-86%), and the logistic regression model yielded verification results of AUC: 0.84 (95% CI: 0.65-1.00) and accuracy rate: 81% (95% CI: 72%-90%). In other words, the present invention generated a severity prediction learning model capable of predicting the severity of bronchial asthma with high accuracy.

[0105] Summary According to the embodiment described above, the generation method executed by the generation device 10 includes an information acquisition step, a first learning step, a contribution degree acquisition step, a selection step, and a second learning step. In the information acquisition step, the generation device 10 acquires analysis target patient information for each of a plurality of items obtained from a plurality of analysis target patients who have a history of disease. In the first learning step, the generation device 10 constructs a primary learning model for predicting whether or not a condition will worsen, using analysis target patient information corresponding to the plurality of items as explanatory variables and whether or not a condition will worsen as a response variable. In the contribution degree acquisition step, the generation device 10 acquires, in the primary learning model, the contribution degree to improving the prediction accuracy of whether or not a condition will worsen. In the selection step, the generation device 10 selects multiple specific items from the multiple items based on the contribution degree. In the second learning step, the generation device 10 constructs a secondary learning model, which is an aggravation prediction learning model used to predict whether or not a condition will worsen, using analysis target patient information corresponding to the plurality of specific items as explanatory variables and whether or not a condition will worsen as a response variable. According to this configuration, the generation method (the generation device 10) can generate a highly accurate learning model for predicting the severity of a disease by selecting specific items based on their contributions and generating a secondary learning model. In other words, the present invention can support prediction of the severity of a disease through machine learning.

[0106] Furthermore, according to the embodiment described above, the contribution degree acquisition step includes a ranking step. In the ranking step, the generation device 10 ranks each of the multiple items based on the magnitude of the acquired contribution degree. In the selection step, the generation device 10 selects multiple specific items based on the ranking. According to this configuration, the generation method (the generation device 10) selects a combination of items with high contribution degrees in the primary learning model as a combination of specific items, and generates a secondary learning model, thereby generating a learning model for predicting severity with higher accuracy. In other words, the present invention can assist in predicting whether a disease will worsen through machine learning.

[0107] Furthermore, according to the embodiment described above, the multiple specific items include multiple items ranked within the top "n" (n is an integer) ranking. With this configuration, the generation method (the generation device 10) selects a combination of items with high contributions in the primary learning model as a combination of specific items, and generates a secondary learning model, thereby generating a more accurate learning model for predicting the worsening of a disease. In other words, the present invention can use machine learning to assist in predicting whether a disease will worsen.

[0108] Furthermore, according to the embodiment described above, the items include the attributes of the patient to be analyzed, the behavioral history of the patient to be analyzed, and the test items undergone by the patient to be analyzed. With this configuration, not only the test items but also the attributes and behavioral history of the patient to be analyzed can be selected as explanatory variables for the learning model. As a result, the range of explanatory variables to be selected for the learning model is increased. Therefore, this generation method (this generation device 10) can generate a learning model for predicting progression with higher accuracy.

[0109] Furthermore, according to the embodiment described above, the prediction method includes a prediction information acquisition step and a prediction step. In the prediction information acquisition step, the prediction device 20 acquires patient information for each of a plurality of items obtained from the patient. In the prediction step, the prediction device 20 predicts whether or not the disease will worsen using a trained learning model for predicting worsening, based on the patient information. With this configuration, the prediction method (the prediction device 20) can assist in predicting whether or not the disease will worsen by providing the user of the prediction device 20 with prediction results using the learning model for predicting worsening generated by machine learning.

[0110] In the above-described embodiment, the generation device 10 is configured as a separate entity (a separate computer) from the prediction device 20. Alternatively, the generation device may function as the prediction device, or the prediction device may function as the generation device. That is, one computer may function as both the generation device and the prediction device (i.e., the system).

[0111] Furthermore, the output destination of the output unit in the present invention is not limited to a display unit. That is, for example, the output unit in the present invention may output information to a printer. Also, for example, the output unit in the present invention may output information to an external device (e.g., a portable information processing terminal) via a communication unit.

[0112] Furthermore, DB1 and DB2 in the present invention are not limited to this embodiment.

[0113] Furthermore, the generated information acquisition section in the present invention may acquire the analysis subject information from an external device instead of DB1.

[0114] Furthermore, in the above-described embodiments, the generation device 10 and the prediction device 20 are each configured by a single computer. Alternatively, the generation device and the prediction device may each be configured by multiple computers. That is, for example, the generation device may be configured by a group of multiple computers that function as the generation device. Specifically, for example, the generation device (group of computers) may be configured by a computer including a storage unit and a computer including a control unit that executes the generation method. Also, for example, the multiple computers may have the functions of the generation information acquisition unit, first learning unit, contribution degree acquisition unit, selection unit, second learning unit, model verification unit, model selection unit, and output unit in a distributed manner. In this case, the multiple computers that make up the group of computers may send and receive information via a network, or may exchange information using a portable storage medium. [Explanation of symbols]

[0115] 1. Severity prediction system 10. Learning model generation device for predicting severity 12 Storage section 131 Generation information acquisition unit 132 First Learning Section 133 Contribution Acquisition Unit 134 Selection Section 135 Second Learning Section 20. Severity prediction device 22 Memory section 231 Prediction Information Acquisition Unit 232 Prediction Department

Claims

1. A method for generating a learning model for predicting the severity of a disease, which is executed by a learning model generation device for predicting the severity of a disease that a patient is suffering from, and which generates a learning model for predicting the severity of a disease, comprising: The device for generating a learning model for predicting progression of an illness, an information acquisition step of acquiring analysis subject patient information for each of a plurality of items obtained from a plurality of analysis subject patients who have a history of suffering from the disease; a first learning step of constructing a primary learning model for predicting whether or not the patient is becoming severe, using the patient information to be analyzed corresponding to the plurality of items as explanatory variables and the presence or absence of the patient's condition becoming severe as a response variable; a contribution degree acquisition step of acquiring a contribution degree to improvement of the prediction accuracy of the presence or absence of the aggravation for each of the plurality of items in the primary learning model; a selection step of selecting a plurality of specific items from the plurality of items based on the degree of contribution; a second learning step of constructing a secondary learning model that serves as the learning model for predicting the severity of the disease, using the patient information to be analyzed corresponding to the plurality of specific items as explanatory variables and the presence or absence of the severity of the disease as a response variable; and the disease is bronchial asthma, The items include the number of exacerbations, BMI, age, exhaled nitric oxide, blood eosinophils, white blood cell count, % predicted forced expiratory volume in one second, % predicted forced vital capacity, presence or absence of sinusitis, and smoking history; The specific items include the number of exacerbations and the age, The presence or absence of the aggravation is defined by whether or not the secular decrease in forced expiratory volume in one second is equal to or greater than a predetermined threshold. A method for generating a learning model for predicting the progression of a disease, characterized by:

2. a ranking step of ranking each of the plurality of items based on the magnitude of the acquired contribution degree; and In the selecting step, the plurality of specific items are selected based on the ranking. The method for generating a learning model for predicting the progression of a disease according to claim 1.

3. The plurality of specific items are: A plurality of the items whose ranking is within the top "n" (n is an integer of 2 or more), Including, The method for generating a learning model for predicting the progression of a disease according to claim 2.

4. The items include attributes of the patient to be analyzed, a behavioral history of the patient to be analyzed, and test items undergone by the patient to be analyzed. A method for generating a learning model for predicting the progression of a disease according to any one of claims 1 to 3.

5. A learning model generation device for predicting the severity of a disease that generates a learning model for predicting the severity of a disease that a patient is suffering from, a generated information acquisition unit that acquires analysis subject patient information for each of a plurality of items obtained from a plurality of analysis subject patients who have a history of suffering from the disease; a first learning unit that constructs a primary learning model for predicting whether or not the patient is becoming severe, using the analysis target patient information corresponding to the plurality of items as explanatory variables and the presence or absence of the patient's condition becoming severe as a response variable; a contribution degree acquisition unit that acquires a contribution degree to improvement of the prediction accuracy of the presence or absence of the aggravation for each of the plurality of items in the primary learning model; a selection unit that selects a plurality of specific items from the plurality of items based on the degree of contribution; a second learning unit that constructs a secondary learning model that serves as the learning model for predicting the worsening of the condition, using the analysis target patient information corresponding to the plurality of specific items as explanatory variables and the presence or absence of the worsening of the condition as a target variable; and the disease is bronchial asthma, The items include the number of exacerbations, BMI, age, exhaled nitric oxide, blood eosinophils, white blood cell count, % predicted forced expiratory volume in one second, % predicted forced vital capacity, presence or absence of sinusitis, and smoking history; The specific items include the number of exacerbations and the age, The presence or absence of the aggravation is defined by whether or not the secular decrease in forced expiratory volume in one second is equal to or greater than a predetermined threshold. A learning model generation device for predicting the progression of a disease, characterized by:

6. causing a computer to function as the learning model generation device for predicting the severity of a disease according to claim 5; A learning model generation program for predicting the progression of a disease, characterized by:

7. A disease aggravation prediction device that predicts whether or not a disease suffering from a patient will become severe, a prediction information acquisition unit that acquires patient information for each of a plurality of items obtained from the patient; A prediction unit that predicts whether or not the condition worsens using a trained learning model based on the patient information; and The learning model is generated by the method for generating a learning model for predicting progression of a disease according to claim 1, The prediction unit predicts the presence or absence of the aggravation based on the patient information corresponding to a plurality of the specific items among the patient information. A disease severity prediction device characterized by:

8. A computer is caused to function as the disease severity prediction device according to claim 7. A disease severity prediction program characterized by:

9. A disease aggravation prediction method executed by a disease aggravation prediction device that predicts whether or not a disease suffering from a patient will aggravate, The severity prediction device includes: a storage unit that stores a trained learning model used to predict whether or not the condition worsens; Equipped with The severity prediction device, a prediction information acquisition step of acquiring patient information for each of a plurality of items obtained from the patient; a prediction step of predicting whether or not the condition worsens using a trained learning model based on the patient information; and The learning model is generated by the method for generating a learning model for predicting progression of a disease according to claim 1, In the prediction step, the presence or absence of the aggravation is predicted based on the patient information corresponding to a plurality of the specific items among the patient information. A method for predicting the severity of a disease, comprising:

10. A disease severity prediction system that predicts whether or not a disease suffering from a patient will become severe, The device for generating a learning model for predicting progression according to claim 5; The severity prediction device according to claim 7; consisting of A disease severity prediction system characterized by:

Images