JP2018067266A - Program for forecasting onset risk or recurrence risk of disease - Google Patents

Abstract

PROBLEM TO BE SOLVED: To forecast an onset risk or a recurrence risk of a disease within a predetermined period by a simple inspection.SOLUTION: A program forecasts an onset risk or a recurrence risk of the disease within a predetermined period by inputting first data based on one or both of a test sample collected from an analyte and biological information acquired from the analyte including the following step executed by a computer having a calculation part, and includes: a step of generating and acquiring a plurality of second data based on index parameters of two or more types by a calculation part; a step of structuring a plurality of first learning models and a plurality of second learning models by the calculation part; and a step of selecting the plurality of first selection learning models and the plurality of second selection learning models for each of the plurality of first learning models and each of the plurality of second learning models by the calculation part.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a program for predicting the onset risk or recurrence risk of a disease within a predetermined period, a prediction apparatus for executing the program, and a learning model construction method for the program.

  Prediction of disease onset or recurrence risk is important because the onset or recurrence of the disease can result in serious consequences.

  For example, Patent Document 1 below discloses a method for determining the risk of recurrence of major adverse cardiac events within one year after the onset of at least one symptom of acute coronary syndrome.

U.S. Pat. No. 8,658,384

In Patent Document 1, in determining the risk of recurrence of major adverse cardiac events, so-called myocardial markers, cardiac troponin I (cTnI), pro-B-type natriuretic peptide (proBNP) or a cleavage product thereof, high sensitivity C Reactive protein (hsCRP), myeloperoxidase (MPO), placental growth factor (PIGF), estimated glomerular filtration rate (eGFR), homocysteine (HCY), choline, ischemia modified albumin (IMA), soluble CD40 ligand (sCD40L ) And lipoprotein-related phospholipase A ₂ (LpPLA2), the amount of at least three kinds of biomarkers in the test sample is used as an indicator.

  However, since the determination method according to Patent Document 1 uses the amount of the myocardial marker in the test sample as an index, for example, a special and expensive cost is required separately from, for example, a test generally performed in a hospitalization test or a health checkup. It was necessary to carry out a proper inspection, and as a result, the burden for the implementation became large.

  The present invention has been made in view of the above problems. The present inventors do not separately carry out such a high-load test, but by using the program of the present invention only by a simple and inexpensive test that is generally performed in hospitalization tests and medical examinations. The present inventors have found that the problem can be solved and have completed the present invention.

That is, the present invention provides the following [1] to [15].
[1] By using the test data collected from the subject or the first data obtained from the subject, including the following steps executed by a computer having a calculation unit, A program for predicting the risk of disease onset or recurrence within a predetermined period,
Two or more kinds selected from the group consisting of a plurality of indicator parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred within the predetermined period. Generating and acquiring a plurality of second data based on the index parameter;
The computing unit constructing a plurality of first learning models for predicting the onset or recurrence of the disease or unknown, and a plurality of second learning models for predicting the onset or no recurrence of the disease or unknown;
The arithmetic unit includes a step of selecting a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models.
[2] The computing unit processes the first data by the first selection learning model and the second selection learning model, and a first determination result for each of the first selection learning model and the second selection learning model. Step to get the
The arithmetic unit obtaining a second determination result based on a plurality of first determination results of the plurality of first selection learning models and a plurality of first determination results of the plurality of second selection learning models;
The arithmetic unit integrating the second determination result of the first selective learning model and the second determination result of the second selective learning model to obtain a third determination result;
The calculation unit further includes a step of predicting the risk of developing the disease or the risk of recurrence within a predetermined period based on the third determination result, or the risk of developing a disease within the predetermined period according to [1] or A program for predicting the risk of recurrence.
[3] The step of obtaining the second determination result sets the cut-off value of the vote rate of the first selection learning model and the cut-off value of the vote rate of the second selection learning model to different values. The program for predicting the onset risk or recurrence risk within the predetermined period according to [2], which is a performed step.
[4] The calculation unit further includes a step of calculating a reliability of the predicted onset risk or recurrence risk based on the second determination result, within the predetermined period according to [2] or [3] A program for predicting the risk of disease onset or recurrence.
[5] The step of constructing the plurality of first learning models and the plurality of second learning models was used for generating the plurality of partial data and the partial data generated by selecting the plurality of index parameters. Second data including complete data is generated and prepared, the second data is divided into a plurality of data groups, each of the data groups is used, and each of the plurality of data groups is performed with a plurality of different parameter conditions. The program for predicting the onset risk or recurrence risk within a predetermined period according to any one of [1] to [4], which is a step constructed by machine learning.
[6] The step of obtaining the second determination result votes for the plurality of first determination results of the plurality of first selection learning models and the plurality of first determination results of the plurality of second selection learning models, The disease within a predetermined period according to any one of [2] to [5], which is a step of acquiring a second determination result based on a vote rate for each of the first selection learning model and the second selection learning model. A program for predicting the risk of onset or recurrence.
[7] The step of selecting the first selection learning model and the second selection learning model is performed by evaluating the first learning model and the second learning model based on Pareto ranks using sensitivity and positive predictive value as indexes. The program for predicting the onset risk or the recurrence risk of the disease within the predetermined period according to any one of [1] to [6], which is a selection step.
[8] Onset of disease within a predetermined period according to any one of [1] to [7], wherein the disease is a major adverse cardiac event, acute kidney injury, cerebral aneurysm rupture or diabetic nephropathy A program for predicting risk or risk of recurrence.
[9] Two or more index parameters measured using a test sample collected from the subject or obtained from the subject are acquired, and first data is generated and acquired based on the index parameters A first data generation and acquisition unit;
The first data acquired by the first data generation and acquisition unit is processed by a first selection learning model and a second selection learning model, and a first determination result is obtained for each of the first selection learning model and the second selection learning model. A first determination result generation acquisition unit that generates and acquires;
The second determination result based on the plurality of first determination results of the plurality of first selection learning models and the plurality of first determination results of the plurality of second selection learning models acquired by the first determination result generation acquisition unit. A second determination result generation acquisition unit that generates and acquires
A second determination result of the first selection learning model acquired by the second determination result generation acquisition unit and a second determination result of the second selection learning model are integrated to generate and acquire a third determination result. 3 determination result generation acquisition unit;
Based on the third determination result acquired by the third determination result generation acquisition unit, the prediction unit predicts the risk of disease onset or recurrence within a predetermined period, Prediction device for predicting recurrence risk.
[10] Two or more of the indicators selected based on a selection frequency from a group consisting of a plurality of indicator parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred A second data generation and acquisition unit that generates and acquires second data based on the parameters;
Based on the second data acquired from the second data generation / acquisition unit, a plurality of first learning models for predicting the onset or recurrence of the disease or the unknown, and predicting the onset or no recurrence of the disease or the unknown A learning model building unit for building a plurality of second learning models;
A learning model selection unit that selects a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models constructed by the learning model construction unit. Furthermore, the prediction apparatus for estimating the onset risk or recurrence risk of the said disease within the predetermined period as described in [9].
[11] The second data generation / acquisition unit selects a plurality of index parameters and generates partial data, thereby including a plurality of partial data and complete data used for generating the partial data. The prediction device for predicting the onset risk or recurrence risk of the disease within the predetermined period according to [9] or [10], which is a functional unit that generates data.
[12] A method for predicting the risk of disease onset or recurrence within a predetermined period of time by using a test sample collected from a subject or inputting first data obtained from the subject. A method for constructing a learning model to be used,
Based on two or more index parameters selected based on the selection frequency from a group consisting of a plurality of index parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred Generating and acquiring a plurality of second data;
Constructing a plurality of first learning models for predicting the onset or recurrence of the disease or unknown, and a plurality of second learning models for predicting the onset or recurrence of the disease or no unknown;
Selecting a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models.
[13] The step of constructing the plurality of first learning models and the plurality of second learning models performs selection on the plurality of index parameters and generates partial data, thereby generating a plurality of partial data and the partial data. Generating and preparing the second data including the complete data used in the process, dividing the second data into a plurality of data groups, using each of the data groups, and a plurality of parameters different for each of the plurality of data groups A learning model construction method used in the method for predicting disease onset risk or recurrence risk within a predetermined period according to [12], which is a step constructed by machine learning performed under conditions.
[14] In the step of selecting the first selection learning model and the second selection learning model, the first learning model and the second learning model are evaluated based on a Pareto rank using sensitivity and positive predictive value as indexes. A learning model construction method used in the prediction method of disease onset risk or recurrence risk within a predetermined period according to [13], which is a step of selecting.
[15] The disease is a major adverse cardiac event, acute kidney injury, cerebral aneurysm rupture or diabetic nephropathy, and is used in the method for predicting the risk of disease onset or recurrence within a predetermined period according to [14]. How to build a learning model.

  According to the program according to the present invention, it is possible to determine the disease onset risk or the recurrence risk with higher accuracy in simpler steps. As a result, it is possible to provide more accurate consulting on the treatment policy, the frequency of hospital visits, and the like. In addition, it is possible to carry out only simpler and cheaper examinations without carrying out high-load examinations such as measurement of biomarkers that require special and expensive costs, so the burden for implementation can be further reduced. .

FIG. 1 is a flowchart showing steps for predicting the onset risk or recurrence risk of a disease. FIG. 2 is a flowchart for explaining step (S1). FIG. 3 is a flowchart for explaining the step (S0). FIG. 4 is a table showing an example of index parameters according to the second data. FIG. 5 is a schematic diagram illustrating a configuration of complete data for generating the second data. FIG. 6 is a schematic diagram of the second data. FIG. 7 is a flowchart for explaining step (S2). FIG. 8 is a table for explaining the evaluation result of the cutoff value. FIG. 9 is a table for explaining the evaluation result of the cutoff value. FIG. 10 is a schematic block diagram for explaining the configuration of the computer. FIG. 11 is a schematic block diagram for explaining the configuration of the calculation unit.

  Embodiments of the present invention will be described below with reference to the drawings. In addition, each drawing has shown only the shape of the component, the magnitude | size, and arrangement | positioning to such an extent that an invention can be understood. The present invention is not limited to the following description, and each component can be appropriately changed without departing from the gist of the present invention. In the drawings used for the following description, the same components are denoted by the same reference numerals, and overlapping descriptions may be omitted. Moreover, the component concerning embodiment of this invention is not necessarily manufactured or used by arrangement | positioning shown by drawing.

[Explanation of terms]
In this specification, “prediction risk” means within a predetermined period (for example, 3 months, 6 months, 1 year, 1 year, 6 months, 2 years, 3 years or longer) at the first examination or after the first examination. This means that the presence or absence (or unknown) of the disease or the likelihood (or unknown) of the disease is predicted.

  As used herein, “prediction of recurrence risk” refers to further major adverse effects within a specified period (eg, 3 months, 6 months, 1 year, 1 year, 6 months, 2 years, 3 years or longer) after the onset of the disease. It means predicting the presence or absence (or unknown) of a cardiac event or the possibility (or unknown) of a possibility.

  In the present specification, the “subject” means a living body that is a prediction target of the onset risk or recurrence risk of a disease, and specifically includes, for example, a patient.

  In the present specification, the “test sample” means an arbitrary sample capable of obtaining an index parameter described later. Examples of such test samples include liquid samples (eg, blood (whole blood) or blood-derived samples (eg, serum, plasma), urine, saliva, ascites, tissue extract, cell extract), non-liquid samples. (For example, a tissue sample, a cell sample) is mentioned, but a liquid sample is preferable, blood or a blood-derived sample is more preferable, and blood is more preferable. The test sample may be processed in advance before measurement. Examples of such treatment include centrifugation, extraction, concentration, fractionation, cell fixation, tissue fixation, tissue freezing, and tissue thinning.

  In this specification, the “index parameter” refers to the results of various tests performed on the above-described test sample, for example, the results of so-called blood tests (such as biochemical tests, blood glucose tests, general blood tests, and coagulation tests). Content, quantity, characteristics, etc.) and other parameters based on biological information.

  As used herein, “disease” refers to major adverse cardiac events, acute kidney injury (acute kidney injury after surgery), cerebral aneurysm rupture or diabetic nephropathy (cases that have become severe enough to require dialysis) Means.

  As used herein, “major adverse cardiac event” refers to acute myocardial infarction, angina pectoris with coronary revascularization, heart failure requiring hospitalization, atrial fibrillation, stroke (transient ischemic attack (TIA) ), Or death due to cardiovascular reasons.

  Here, an example of the index parameter will be described. Examples of “index parameters” when the disease is, for example, a major adverse event include C-reactive protein (CRP), D-dimer, HDL-cholesterol (HDL-C), LDL-cholesterol (LDL-C), Prothrombin time (international standard ratio (INR)) (PT-INR), γ-glutamyltransferase (γ-GTP), aspartate aminotransferase (AST (GOT)), amylase (AMY), alanine aminotransferase (ALT (ALT) GPT)), alkaline phosphatase (ALP), albumin (ALB), antithrombin (AT), glycohemoglobin (HbA1c), chlor (Cl), triglyceride (TG), fibrinogen (Fbg), fibrin / fibrinogen degradation product (FDP) Activated partial thromboplastin time (APTT), serum creatinine (CRE), blood urea nitrogen (BUN), blood sugar (Glu), total cholesterol (CHO), total bibilin (T · Bil), monocyte count (Mono), Direct bilirubin (D / Bil), lactate dehydrogenase (LD (LDH)), urinary protein (qualitative) (UP), uric acid (UA), urine sugar (qualitative) (US), pH, potassium (K), calcium (Ca), sodium (Na), red blood cell count (RBC), hematocrit value (Ht), hemoglobin (Hb), lymphocyte count (Lym), platelet count (PL), basophil count (Baso), eosinophil Number (Eos), and neutrophil count (Neut). Information on these measured values in the test sample can be obtained by a conventional method.

Examples of the “index parameter” when the disease is a major adverse event include, for example, the history of recurrence of major adverse cardiac events within a predetermined period, gender, reason for hospitalization, current symptoms at hospitalization, history of diabetes, hypertension Further examples include binarized or digitized parameters for biometric information (attributes, qualitative data) such as history of illness, history of dyslipidemia, smoking habit, age, height, weight, heart rate.
Among these, the “reason for hospitalization” can be determined based on, for example, a doctor's diagnosis and treatment when the patient (subject) is hospitalized.

  Specifically, for example, when the subject is a patient who has been diagnosed with acute myocardial infarction (AMI) and is hospitalized, the index parameter may be set to 1, and an angina pectoris in which the subject has undergone coronary revascularization. If the patient is hospitalized for the reason, the index parameter may be 2, and if the subject is a patient who has been diagnosed with heart failure (HF) and is hospitalized, the index parameter may be 3. If the patient is hospitalized because of atrial fibrillation (required myocardial ablation (ablation)), the index parameter may be 4, and the patient is diagnosed with cerebral infarction (CI) and hospitalized. In some cases, the index parameter may be 5, and when the subject is a patient diagnosed with a transient ischemic attack (TIA) and hospitalized, the index parameter may be 6.

  As for “oncology at admission”, for example, the pathological findings based on the electrocardiogram waveform indicate that the index parameter is 0 in the case of “no atrial fibrillation” (sinus rhythm: normal), and there is atrial fibrillation. In this case, the index parameter is 1.

  In the present specification, the “myocardial marker” means an index related to the heart (myocardium) in particular, among the results of examinations performed on test samples. Examples of “myocardial markers” include creatine kinase MB (CKMB), human cardiac fatty acid binding protein, cardiac troponin I, cardiac troponin T, brain natriuretic peptide (BNP), pro-brain natriuretic peptide or its cleavage product, myel Examples include peroxidase, placental growth factor, estimated glomerular filtration rate, homocysteine, ischemia-modified albumin, soluble CD40 ligand, lipoprotein-related phospholipase A2, choline, and highly sensitive C-reactive protein.

[Program for predicting the risk of disease onset or recurrence within a specified period]
Hereinafter, each step included in the program of the present embodiment will be specifically described. In this embodiment, unless otherwise specified, “steps” are executed by a computer (details will be described later).

  Here, the “program” is a data processing method described in an arbitrary language or description method, and may be in any format such as source code or binary code. The “program” is not necessarily limited to a single configuration, and may be distributed as a plurality of modules and libraries, and cooperates with a separate program represented by an OS (Operating System). Sometimes achieve this function.

  The present invention also relates to a storage medium in which a program for predicting the risk of disease onset or recurrence within a predetermined period is recorded.

  The program is normally recorded on a recording medium according to a storage unit described later, and is read by a computer (prediction device) that performs a prediction method as necessary. As a specific configuration for reading the program recorded on the recording medium by the prediction device, a reading procedure, an installation procedure after reading, and the like, a well-known configuration and procedure can be used.

  The “recording medium” includes any “portable physical medium”, any “fixed physical medium”, and “communication medium”. The “portable physical medium” is a flexible disk, a magneto-optical disk, a ROM, an EPROM, an EEPROM, a CD-ROM, an MO, a DVD, or the like. The “fixed physical medium” is a ROM, a RAM, a hard disk drive or the like built in various computer systems. The “communication medium” holds the program for a short period of time like a communication line or a carrier wave in the case of transmitting the program via a network such as a LAN, WAN, or the Internet.

  A program according to an embodiment of the present invention is a first program obtained by measurement using a test sample collected from a subject, including the following steps executed by a computer including a calculation unit: A program for predicting the risk of disease onset or recurrence within a predetermined period by inputting data, and the calculation unit has developed or recurred within the predetermined period, or has not developed or recurred Generating and acquiring a plurality of second data based on two or more types of index parameters selected from a group consisting of a plurality of index parameters for a plurality of subjects based on a selection frequency; A plurality of first learning models for predicting the onset or recurrence of the disease, and a plurality of second learning models for predicting the onset or recurrence of the disease or the unknown A step of constructing a computing unit, for each of the plurality of first learning model and the plurality of second learning model, and a step of selecting a plurality of first selection learning model and the plurality of second selection learning model.

Hereinafter, each step included in the program according to the present embodiment will be specifically described. In this embodiment, unless otherwise specified, “steps” are executed by a computer (details will be described later).
In the following description, “major adverse event” will be described as an example of a disease.

  With reference to FIG. 1, the program for predicting the onset risk or recurrence risk of the disease of this embodiment will be described. FIG. 1 is a flowchart showing steps for predicting the onset risk or recurrence risk of a disease.

  (1) As shown in FIG. 1, first, as a premise, a step (S1) of acquiring first data based on an index parameter is executed.

  Hereinafter, step (S1) will be specifically described with reference to FIG. FIG. 2 is a flowchart for explaining step (S1).

  Prior to step (S1), it is preferable to construct a learning model to be described later.

  As shown in FIG. 2, in step (S1), first, a step (S1-1) of collecting a test sample is performed. The method of selecting the test sample and collecting the test sample in step (S1-1) is not particularly limited on condition that the index parameter can be acquired.

  For example, when the test sample is a sample particularly related to blood, the test sample can be obtained by a normal blood collection method.

  In the present embodiment, an examination result and / or biological information based on a test sample collected from a subject is used as the index parameter.

  For example, regarding major adverse cardiac events, it is preferable to use an index parameter that is not derived from the myocardial marker because it can improve convenience and reduce the load.

  Next, a step (S1-2) of obtaining first data based on an index parameter measured using the obtained test sample or obtained from the subject is performed.

  This step (S1-2) can be performed by analyzing (measuring) the obtained test sample using any conventionally known suitable inspection means (measurement means) and inspection method (measurement method).

  For example, examples of index parameters not derived from a myocardial marker when the disease is a major adverse event include C-reactive protein amount, D-dimer amount, HDL-cholesterol amount, LDL-cholesterol amount, prothrombin time (international standard ratio) (INR)), γ-glutamyl transpeptidase level, aspartate aminotransferase level, amylase level, alanine aminotransferase level, alkaline phosphatase level, albumin level, antithrombin level, glycohemoglobin level, crawl level, triglyceride level, fibrinogen level , Fibrin / fibrinogen degradation product level, activated partial thromboplastin time, serum creatinine level, blood urea nitrogen level, blood glucose level, total cholesterol level, total bilbilin level, monocyte count, direct bilirubin level, lactic acid Hydrogen enzyme (LDH) level, uric acid level, pH, potassium level, calcium level, sodium level, red blood cell count, hematocrit level, hemoglobin level, lymphocyte count, platelet count, basophil count, eosinophil count, and neutrophil The number of balls is mentioned.

  The format of the first data is not particularly limited on condition that it can be applied to step (S2) described later.

  Further, in the step (S1-2) of acquiring the first data, examples of the index parameter not derived from the myocardial marker used in addition to the already described “group of index parameters not derived from the myocardial marker” , Gender, history of diabetes, history of hypertension, history of dyslipidemia, presence or absence of smoking habits, age, height, weight, heart rate, urinary protein (qualitative), and urine sugar (qualitative) Biometric information obtained.

  In addition, as an example of the index parameter not derived from the myocardial marker, which is used in addition to the group of index parameters not derived from the myocardial marker already described, it was obtained from a subject such as the reason for hospitalization and on-hospital disease. Biological information can be mentioned.

  Furthermore, when the disease is a major adverse event, the first data (and second data described later) is further added to the already described myocardial marker in addition to the group of index parameters not derived from the already described myocardial marker. It is good also as data based on the index parameter selected from the group which further contains the derived index parameter.

  Thus, if the index parameter derived from the myocardial marker is further used, for example, the prediction accuracy of the recurrence risk in a relatively short period such as “within 3 months” after the onset of the major adverse cardiac event can be further improved.

  The first data (and second data described later) is data based on an index parameter selected from the group further including an index parameter derived from an electrocardiogram in addition to the index parameter group not derived from the myocardial marker already described. It can be.

  Examples of index parameters derived from such an electrocardiogram include P wave height, R wave interval (RR interval), PQ time, R wave height, QRS width, ST portion change amount (S wave The sum of the height and the height of the T wave), the height of the T wave, and the power spectrum obtained by Fourier transforming the electrocardiogram.

  If the index parameters derived from the electrocardiogram are further used in this way, the number of index parameters can be further increased, so that the effect of improving the prediction accuracy of the recurrence risk of the main adverse cardiac event can be obtained. .

  The first data can be generated based on two or more types of index parameters selected from the group of index parameters already described. Such an index parameter is preferably selected based on a selection frequency described later.

  In the present embodiment, the number of index parameters that can be used can be any suitable number in consideration of the time required for prediction. The number of index parameters that can be used can be determined, for example, by appropriately selecting an index parameter having a higher selection frequency in consideration of prediction accuracy.

  In the present embodiment, as the first data (and second data described later), two or more types of index parameters are used among the index parameters already described. However, the number of indicator parameters that can be used is not particularly limited. Examples of the index parameter include not only two types but also 5 types or more, 10 types or more, 15 types or more, 20 types or more, 25 types or more, 30 types or more, 35 types or more, 40 types or more, 45 types or more or already All described can be used.

  When two or more types of index parameters are used, an appropriate combination of the two or more types of index parameters can be selected by processing such as correlation rule analysis or clustering.

  (2) Next, the first data is processed with a learning model constructed based on the second data to predict the risk of disease onset or recurrence (S2).

[How to build a learning model]
Here, first, with reference to FIG. 3, step (S0) concerning the construction method of the learning model used for step (S2) will be described. FIG. 3 is a flowchart for explaining the step (S0).

  Note that the step (S0) of building the learning model can be performed prior to the step (S1) already described.

  As shown in FIG. 3, first, a step of preparing second data including a plurality of bit strings (S0-1) is performed.

  Specifically, when the disease is a major adverse event, two types selected based on the selection frequency from the group added to the “group consisting of a plurality of index parameters not derived from the myocardial marker” already described. Second data that is a bit string based on the above index parameters is generated and prepared.

  The second data is teacher data for constructing a learning model. Here, the second data when the disease is a major adverse event and the generation thereof will be described.

  First, a case data group collected in advance is prepared. Here, the case data group is a data group configured by collecting a plurality of case data relating to a patient as a subject.

  The case data group is classified as a plurality of case data whose occurrence or recurrence of disease is known within a predetermined period, that is, the disease has occurred or recurred within the predetermined period. “Onset or recurrence” case data and “onset or no recurrence” case data classified as “onset or no recurrence” that did not develop or recurrence of disease within a predetermined period.

  Thus, the case data group consists of a plurality of “onset or recurrence” case data group consisting of a plurality of “onset or recurrence” case data and a “onset or recurrence” case data consisting of a plurality of “onset or recurrence” case data. It consists of a group.

  Next, each of the case data groups, that is, the “onset or recurrence” case data group and the “onset or no recurrence” case data group of the disease, (i) a case data group for constructing a learning model and (ii) for evaluation It is divided into the case data group. Among these, the second data is generated using the divided “case data for constructing a learning model”.

  Here, if the number of “onset or recurrence” case data included in the “case data group for learning model construction” and the number of “onset or recurrence” case data are not equal and biased, the learning model There is a risk of learning bias in building

  Therefore, if there is a bias in this way, an equalization process is performed to equalize the number of “onset or recurrence” case data and the number of “onset or recurrence” case data to the same extent. Prepare an “equalized case data group”. Specifically, for example, when the number of “onset or recurrence” case data is smaller than the number of “onset or recurrence” case data, the same number of “onset” as the number of “onset or recurrence” case data It is preferable to perform equalization processing by extracting case data “or no recurrence” and aligning both numbers.

  Next, the equalized case data group is further divided into a plurality of groups. The number of case data included in each of the divided groups may be approximately the same, and is preferably the same. The total number of groups after division is not particularly limited, but is preferably about 4, for example.

  Next, “partial data” is generated using a part of the divided plurality of equalized case data groups. For example, when the data is divided into four equalized case data groups, “partial data” may be generated using three groups (75%) of them. In this case, the remaining one group (25%) is set as “evaluation data”.

  And a learning model is constructed | assembled using the 2nd data which is the data group containing the complete data (case data) which is the original data used for the production | generation of this some partial data and partial data (for details, see FIG. (It will be described later.)

Here, with reference to FIG. 4, an example of an index parameter related to the second data when the disease is a major adverse heart event will be described.
FIG. 4 is a table showing an example of index parameters according to the second data.
In FIG. 4, in addition to the index parameter name and the index parameter, the definition and unit of the index parameter, and the type of the index parameter are shown. The index parameters are assigned serial numbers (1 to 24) as ID numbers.

  In this case, the second data is “an index on whether the major adverse event recurred within the prescribed period after the onset of the major adverse event or did not recur (history of major adverse events within the prescribed period). The selection of “parameter” is essential. In the first data, such an index parameter cannot be selected because it cannot exist in the first place.

  The group of index parameters according to the second data in this case includes that “index parameters regarding whether or not a major adverse event occurred or not recurred within a predetermined period after the occurrence of a major adverse event” is included. Except for this, it can be the same as the group of index parameters related to the first data already described.

  In the selection of the index parameter in the second data, it is preferable to select the index parameter selected from the group consisting of the same index parameters as the first data based on the selection frequency (details will be described later).

  The second data will be described with reference to FIGS. FIG. 5 is a schematic diagram for explaining the configuration of complete data for generating partial data. FIG. 6 is a schematic diagram of second data that is partial data.

As shown in FIG. 5 and FIG. 6, the second data including complete data and partial data generated from the complete data is a bit string (bit string) for selecting an index parameter and selecting case data used during machine learning. It is data expressed as
In the example shown in FIG. 5, the complete data for generating the partial data uses 10 kinds of index parameters and 10 cases of data.

  Here, the “case data” is data (parameters) for specifying a case (patient) corresponding to the selected index parameter. Note that a plurality of cases for a single patient may exist as separate case data.

  As shown in FIG. 5 and FIG. 6, the bit string for the second data including the complete data and the partial data includes the first partial BSP1 in which the selection or non-selection of the index parameter is described and the selection or non-selection of the case data. It consists of a second part BSP2 in which the selection is described. In this example, the second part BSP2 is described continuously after the first part BSP1. In this example, ten index parameters (ID: 1 to 10) are described in the first part BSP1, and case data (ID: PT1 to PT10) of ten cases used are described in the second part BSP2. Yes.

  In FIG. 5 and FIG. 6, the uppermost number sequence indicates the ID number, and the numerical values on the lower level indicate selection or non-selection of the index parameter and selection or non-selection of the case data.

  The bit string shown in FIG. 5 is complete data showing an example using all 10 kinds of index parameters and case data of 10 cases. Therefore, the numerical values constituting the bit string are all composed of “1”. ing. Here, if considered in association with the index parameter shown in FIG. 4, this bit string specifically means that the index parameter related to (ID: 1 to 10) is used in contrast.

  As shown in FIG. 4, when the obtained test result and / or biological information is acquired as quantitative data (numerical data) such as the content, age, and height of a predetermined component, for example, it is used as an index parameter as it is. Can be used.

  In addition, such quantitative data (numerical data) should be converted into an ordinal scale, interval scale, proportional scale, etc., taking into account the age (elderly, middle-aged, adolescent, boy, young), for example, and used as an index parameter. Can do.

  For attribute data relating to biological information such as a history of diabetes, a history of hypertension, a history of dyslipidemia, and qualitative data such as positive (level) and negative, for example, “Yes = 1, No = 0” and 2 It can be used as an index parameter by converting it into a value.

  Next, generation of second data including complete data and partial data based on the complete data will be described. Specifically, multiple index parameters (data group columns) and case data (data group rows) included in complete data are selected, and multiple partial data different from the original complete data are generated. Thus, the second data which is a data group including the complete data and a plurality of partial data based on the complete data is prepared.

  As described above, the second data includes complete data including all index parameters and all case data in addition to a plurality of partial data, provided that the prediction performance of the corresponding learning model is sufficient. It may be.

Here, the generation of partial data and partial data will be described with reference to FIG.
Here, partial data (bit string) based on complete data (bit string) that can constitute the second data already described with reference to FIG. 5 and generation steps thereof will be described.

  As shown in FIG. 6, in this example, 6 types (ID = 1, 3, 4, 6, 7, and 9) are selected from 10 types of index parameters, and 4 types (ID = 2, 5, 8, and 8) are selected. 10) is not selected, and 4 cases (ID = PT1, PT4, PT7 and PT10) are selected from the case data of 10 cases, and 6 cases (ID = PT2, PT3, PT5, PT6, PT8 and PT9) are selected. ) Is not selected.

  Specifically, for example, the partial data is obtained by generating a plurality of partial data by randomly selecting the selection results so as not to overlap without particularly considering the combination of the selected index parameter and the learning model. be able to. Then, a process of including the obtained partial data in the second data is performed.

  A learning model having a larger variance is obtained by performing the step of obtaining second data that is a data set incorporating a plurality of partial data obtained by re-sampling the index parameter (feature value) in this way. Can be obtained.

  As shown in FIG. 6, in the present embodiment, the second data including partial data is managed and stored as a data set including a plurality of bit strings. FIG. 6 shows bit strings (BS1, BS2, and BS3) of partial data having three patterns included in the data set that is the second data. Details of the processing relating to a data set (second data) including a plurality of bit strings will be described later.

  Next, a step (S0-2) of constructing a learning model by machine learning is performed using the obtained second data.

  The learning model constructed in the present embodiment includes a plurality of types of learning models, that is, a first learning model and a second learning model.

  In the present embodiment, the first learning model is a learning model that predicts “the occurrence or recurrence of a disease within a predetermined period” or “unknown”. The second learning model is a learning model that predicts “no disease onset or recurrence within a predetermined period” or “unknown”.

Hereinafter, step (S0-2) will be specifically described.
Here, complete data that can be included in the second data and each of the plurality of partial data are used as teacher data, and a plurality of learning models, that is, a plurality of first learning models and a plurality of second learning models are constructed by machine learning. .

  In the present embodiment, the first learning model and the second learning model are preferably support vector machines (SVM). Also, as means other than the support vector machine, other means such as a neural network can be used.

  In this embodiment, an example of a support vector machine that can be used for machine learning to construct a learning model is obtained from a website (https://cran.r-project.org/web/packages/e1071). A support vector machine based on the "e1071 package of R language (https://www.R-project.org)" is possible.

  Here, the learning model is constructed by machine learning performed by dividing the second data into a plurality of data groups, using each of the data groups as teacher data, and performing a plurality of parameter conditions different for each of the plurality of data groups. It is preferable.

  In this embodiment, a support vector machine using an RBF kernel (Gauss kernel) as a kernel function can be used for machine learning.

  Specifically, the machine learning is performed under a plurality of different parameter conditions. The “different parameter conditions” can be set by adjusting an optimum adjustment coefficient (hyper parameter) by, for example, random sampling.

  For example, when using a support vector machine employing a Gaussian kernel as described above, a plurality of different learning models can be constructed from the same partial data by adjusting the γ parameter and the C parameter, which are adjustment coefficients. .

  In this case, for example, five parameters, γ = 0.01, γ = 0.02, γ = 0.03, γ = 0.04, and γ = 0.05, are used as γ parameters for adjusting the complexity of the identification boundary line. , And the allowable parameter C of the soft margin is replaced by a discriminant function (described later) used in performance evaluation during machine learning, so that the learning model can be constructed with C = 100. it can.

  For example, by setting the adjustment coefficient as described above, a learning model having a larger variance can be obtained.

  The support vector machine is a learning model that is originally constructed so that the variance is small. However, in the present embodiment, a support vector machine having a large dispersion is constructed, and the processing described later is further performed.

  According to the present embodiment, as a result, prediction accuracy can be further improved by constructing a support vector machine having a large variance. In the following description, processing using a support vector machine as a learning model will be described unless otherwise specified.

  Next, a step (S0-3) of processing the evaluation data with the learning model and evaluating the prediction result based on the Pareto rank is performed.

  In this step (S0-3), the performance of the first learning model and the second learning model is evaluated based on the Pareto rank using the sensitivity and the positive predictive value as indices. Hereinafter, this evaluation step will be specifically described.

  1) First, prediction results are obtained for all learning models using the constructed learning models (first learning model and second learning model) and the evaluation data already described.

2) When dealing with an identification problem that assumes a predictable area and an impossible area, it is important to reduce the degree of misidentification as much as possible and to increase the number of objects that can be correctly predicted as much as possible. Therefore, it is obtained using the objective function O ₁ (function for measuring a prediction error in a predictable region) and the objective function O ₂ (function for counting the number of data that can be correctly predicted in the data space). Evaluate the predicted results.

Here, the objective function O ₁ and the objective function O ₂ will be described. Here, a case will be described in which a support vector machine is used as an identification function to evaluate a first learning model whose prediction target is “the occurrence or recurrence of a disease”.

  As a premise, in the evaluation data (x, y), x is a set of values measured for a certain case, y is a correct answer label, “onset or recurrence of disease = + 1” and “onset of disease or One of the values is assumed as the code data of “no recurrence = −1”.

  If the discriminant function of the first learning model constructed by learning using the predetermined second data is f, the discriminant function f is determined based on the discriminating target of this learning model, “onset or recurrence of disease”. Based on the predicted value f (x) calculated when such x is input, if f (x)> 0, it is predicted that “the disease will occur or recur”, and if f (x) <0. “Unknown”. Conversely, in the second learning model whose identification target is “no disease onset or recurrence”, if the prediction value of the discrimination function is f (x) <0, “no disease onset or recurrence” is predicted, If f (x)> 0, it is determined as “unknown”.

  The first learning model for predicting “with or without disease onset or recurrence” according to the present embodiment always predicts “with the onset or recurrence of disease” for data existing in the predictable region, and out of the predictable region. The data existing in is considered “unknown”.

  Therefore, when there is data that should be determined as “no disease onset or recurrence” in the predictable region of the first learning model, the prediction always fails.

The evaluation based on the objective function O ₁ and the objective function O ₂ in the predictable region of “the onset or recurrence of the disease” in the first learning model will be described.

Note that the evaluation by the objective function O ₁ can be generalized for the learning model using the two-group discriminant function even when a learning model other than the support vector machine is used. Specifically, for example, when a linear discriminant function, a quadratic discriminant function, or a logistic discriminant function is used, evaluation is performed using the distance from the mispredicted data to the identification line as in the case of the support vector machine. be able to.

In the objective function O ₁ , the minimization of the sum total of the distance (SVM Confidence Margin) from the mispredicted data to the identification line is considered.

  The SVM Confidence Margin is defined as the product yf (x) of the predicted value f (x) and the correct answer label y using the predicted value f (x) calculated by the identification function f of the evaluation data (x, y). The

  Here, when the evaluation data is correctly predicted as “the onset or recurrence of the disease”, the SVM Confidence Margin has the predicted value of f (x)> 0 and the correct label “the onset or recurrence of the disease = + 1”. Since it is a product of, it takes a positive value. On the other hand, the correct label y is “no disease onset or recurrence = −1” even though it is predicted that there is a misprediction in the predictable region, that is, f (x)> 0 and “onset or recurrence of disease” is predicted. The product of the predicted value and the correct answer label takes a negative value.

  Similarly, in the second learning model, the SVM Confidence Margin calculates the predicted value of f (x) <0 and the correct label “onset of disease” when the evaluation data is correctly predicted as “no disease onset or recurrence”. Or, since it is a product of “no recurrence = −1”, it takes a positive value, and takes a negative value if mispredicted.

The minimization of the objective function O ₁ by SVM Confidence Margin is expressed by the following equation (1).

であり、ａｂｓ［ｘ］はｘの絶対値を表す。すなわち、式（１）は、ＳＶＭ Ｃｏｎｆｉｄｅｎｃｅ Ｍａｒｇｉｎにおいて、予測可能領域における誤識別の度合いのみを集計するための機能を有する。 In formula (1), for m (y, f (x))

And abs [x] represents the absolute value of x. That is, the expression (1) has a function for counting only the degree of misidentification in the predictable region in the SVM Confidence Margin.

When evaluating the first learning model with “onset or recurrence of disease” as a prediction target, the “onset or recurrence of disease” data is a negative example (f (x _i ) Only the predicted distances of the evaluation data predicted as <0) are tabulated.

Next, regarding the objective function O ₂ , maximization of the number of data for evaluation of “the onset or recurrence of disease” that is correctly predicted while allowing some misprediction to some extent is considered.

When the correctness of the prediction of the discrimination function f is expressed using the correct answer label y and the predicted value f (x), it is expressed by the following formula (2).

Here, maximization of the objective function O ₂ when prediction is performed for k pieces of evaluation data is expressed by the following equation (3).

  In Expression (3), the second term on the right side represents regularization based on the total number of “no disease onset or recurrence” in the predictable region.

  In equation (3), α (1> α> 0), which is a variable for adjusting the tolerance of misprediction, is preferably set to α = 0.3.

  Thus, the prediction result of the learning model is evaluated by the Pareto rank.

  Next, a step (S0-4) of selecting a learning model having a high evaluation and a bit string for which a learning model having a high evaluation can be constructed is performed.

Specifically, a learning model (first learning model and second learning model) in which the evaluation values (O ₁ , 1 / O ₂ ) by the objective functions O ₁ and O ₂ already described are smaller and such a learning model are constructed. The bit string (second data) that can be selected is selected.

  Here, step (S0-4) will be specifically described.

  First, a plurality of evaluation data is prepared which has the same index parameter configuration as that of the first data described above and whose index parameter values do not match the first data.

  The evaluation data used for selecting the learning model belongs to, for example, an equalized case data group that has not been used to generate second data (partial data) among a plurality of divided equalized case data groups Case data can be used.

  Then, the evaluation data is processed by a plurality of first learning models and a plurality of second learning models, respectively, to predict the risk of disease onset or recurrence within a predetermined period.

Next, with respect to the obtained prediction results, the first learning model and the second learning model are evaluated by the Pareto rank using the sensitivity and positive predictive value as indices using the objective functions O ₁ and O ₂ already described.

  Based on the obtained evaluation results, both the sensitivity and the positive predictive value are high, that is, the learning model (the first learning model and the second learning model) having a high evaluation and such a learning model can be constructed. Two data (bit strings) are selected and stored.

  The number of learning models to be selected and the number of corresponding bit strings can be arbitrarily set in consideration of the time required, the implementation scale, and the like. In the case of this embodiment which has already been described, it is preferable that the number is about 40.

  The step (S0-4) of selecting a highly evaluated learning model (first learning model and second learning model) and a corresponding bit string (second data) has a sensitivity of 1 or less and 0.95 or more. 0.7 or more, preferably 0.6 or more, and the false positive rate is 0 or more, preferably 0.4 or less, 0.3 or less, or 0.05 or less.

  The analysis result of analyzing the selection frequency of the index parameter concerning the learning model selected by the step of selecting the first learning model and the second learning model having a high evaluation is used to construct the first data and the second data. Can be used to select the index parameter.

  Specifically, when the first learning model and the second learning model are constructed, if an index parameter having a high selection frequency is selected in advance when the first data and the second data are generated, it is necessary to implement the prediction method. Time can be shortened and prediction accuracy can be further improved.

  In addition, a plurality of indices that can further improve the prediction accuracy by analyzing a combination of the adopted index parameters using the bit strings that have been able to build the first learning model and the second learning model that are highly evaluated. A combination of parameters can be found.

  If a combination of a plurality of index parameters found in this way is selected in advance when generating the first data and the second data, the time required to implement the prediction method can be shortened, and the prediction accuracy can be improved. It can be improved further.

Next, a step (S0-5) of determining whether the learning model satisfies a predetermined requirement is performed.
Specifically, a step (S0-5) for determining whether or not the learning models (the first learning model and the second learning model) selected by performing the step (S0-4) satisfy a predetermined requirement. Is done.

  Specifically, the step (S0-5) is a step in which the first learning model and the second learning model have predetermined requirements, for example, the performance of the bit string, that is, the prediction accuracy of the selected first learning model and second learning model. Is a predetermined prediction accuracy, for example, whether the improvement rate of the prediction accuracy is less than 0.1%, or the number of generations required to update the first learning model and the second learning model (bit string) is an arbitrarily set number of generations A determination is made as to whether an upper limit (eg, 100 generations) is satisfied.

  First, in the step (S0-5), when the selected first learning model and second learning model do not satisfy the predetermined requirements due to the execution of the step (S0-4) (step (S0-5)). In the case of “No” in FIG.

  When the selected first learning model and second learning model (learning model) do not satisfy the predetermined requirements (“No” in step (S0-5)), for example, as a result of the determination in the example, the first When the prediction accuracy of the learning model and the second learning model is less than 0.1% and / or when the number of generations of the first learning model and the second learning model (bit string) does not reach 100 generations), A step (S0-6) of generating a new bit string by performing evolutionary processing on the selected bit string corresponding to the learning model using a genetic algorithm is performed.

  Note that the bit string generation (optimization) step can be performed not only by a genetic algorithm but also by, for example, a search for a combination of all patterns, a random search, or the like.

  Here, the step of optimizing a bit string using a genetic algorithm will be described.

  In the step of optimizing the bit string using the genetic algorithm, the selection of the bit string and the second data including the plurality of selected bit strings (data set) that can construct the learning model having better prediction performance ) Is updated and saved.

(1) First, as already described, based on the evaluation of the learning model by the objective functions O ₁ and O _2, the bit string that can construct the learning model having a higher evaluation, that is, the Pareto rank is higher. Ranking so that

  Specifically, for the bit strings that have already been ranked, for example, rearrange the higher-order bit strings to the upper level, and consider the number of bit strings included in the data set. Edit by excluding from the dataset. Then, the data set (second data) including the updated plurality of bit strings is stored in a state where the bit strings included in the data set can be read.

  Taking the bit string shown in FIG. 6 as an example, bit string BS1 is the most significant bit string, bit string BS2 is the second most significant bit string, and bit string BS3 is the third most significant bit string. .

  (2) Next, the genetic algorithm is used to perform evolutionary processing such as selection, crossover, mutation introduction, and bit string evaluation on the ranked bit strings.

  In the present embodiment, the processing using such a genetic algorithm can be performed using, for example, NSGA-II (Elitist Non-Dominated Sorting Genetic Algorithm). Here, NSGA-II is a non-dominant sort genetic algorithm.

  The processing by the genetic algorithm is performed, for example, as a condition in which the number of models per generation is 500, the archive size is 125, the mutation rate per bit string is 10%, one-point crossover is performed, and update is performed up to 80 generations. be able to.

  Then, the bit string newly generated by the processing by the genetic algorithm is incorporated into the original second data, updated to the latest second data (data set), and the bit string included in the data set can be read. A process of saving as a state is performed (S0-7).

  Next, using the updated second data, the process returns to the step (S0-1) for preparing the second data including the plurality of bit strings already described, and the steps up to step (S0-5) are performed again. If the step is repeated and the determination result in step (S0-5) is “No”, step (S0-1) to step (S0-1) are repeated until the determination result in step (S0-5) is “Yes”. Steps S0-7) are repeated.

  In this way, a bit string capable of constructing a better learning model is selected, and a data set related to better second data can be held.

  In the process of optimizing a bit string by such a genetic algorithm, selection of variables (index data) is performed simultaneously. Specifically, for example, by comparing the variable used in the second data (bit string) with better results (prediction accuracy) and the variable not used, for example, by comparing the selection frequency of the variables. Evaluate the importance of and select variables that are judged to be highly important.

  By such variable selection, it is possible to effectively extract variables considered to contribute to prediction or narrow down combinations of variables.

  If the determination result in step (S0-5) satisfies the predetermined requirement already described and is “Yes” (the determination accuracy in the above example, the prediction accuracy of the first learning model and the second learning model) Is 0.1% or more and / or the number of generations of the first learning model and the second learning model (bit string) reaches 100 generations), the bit string (second data) The update is completed, and a learning model based on the final bit string is selected as a selection learning model (a first selection learning model and a second selection learning model).

Finally, the first selection learning model and the second selection learning model are stored (S0-8). More specifically, the first selection learning model and the second selection learning model that are finally selected are stored in a readable state. Here, the second data (bit string) that has been updated is stored in a state where it can be read out.
By executing step (S0-8), step (S0) is completed.

  Next, step (S2) will be described with reference to FIG. FIG. 7 is a flowchart for explaining step (S2).

  First, the first data is processed by the first selection learning model and the second selection learning model, and the step of obtaining the first determination result for each of the first selection learning model and the second selection learning model (S2-1) is performed. Is called.

  By this step (S2-1), a plurality of first determination results of each of the plurality of first selection learning models and a plurality of first determination results of the plurality of second selection learning models can be acquired.

  Next, voting is performed for each of the plurality of first determination results of the plurality of first selection learning models and the plurality of first determination results of the plurality of second selection learning models, and the first selection learning model and the second selection learning model. A step of acquiring a second determination result based on the vote rate is performed every time. Hereinafter, this step will be specifically described.

  First, voting is performed for each of the plurality of first determination results of the plurality of first selection learning models and the plurality of first determination results of the plurality of second selection learning models (S2-2).

  Specifically, based on the obtained first determination result, the plurality of first selection learning models vote for either “the onset or recurrence of the disease” or “unknown”. The plurality of second selection learning models vote for “no disease onset or recurrence” or “unknown”. Each voting result is totaled for each of the first selection learning model and the second selection learning model.

  Next, a vote rate is calculated for each of the first selection learning model and the second selection learning model. Next, the vote rate is compared with the cut-off value, and it is determined whether the vote rate is equal to or exceeds the cut-off value (cut-off value ≦ voting rate) (S2-3). Details of the cutoff value will be described later.

  As a result, when the vote rate is equal to or exceeds the set cut-off value (in the case of “Yes” in step (S2-3)), for the first selection learning model, A second determination result for determining relapse as “present” is acquired, and a second determination result for determining occurrence or relapse of a disease within a predetermined period as “none” is acquired for the second selective learning model ( S2-4).

  When the vote rate is smaller than the set cut-off value (“No” in step (S2-3)), the disease within a predetermined period for both the first selection learning model and the second selection learning model The 2nd determination result which makes the onset risk or recurrence risk of "unknown" is acquired (S2-5).

  Here, an example of voting results based on a plurality of first determination results for a plurality of first selection learning models will be described with reference to Table 1.

  Table 1 is a table showing an example of the vote rate and the second determination result according to the first selection learning model. Here, an example in which the second determination result is obtained using four first selection learning models for each of the three first data regarding the harmful event will be described. In this example, in order to obtain the second determination result, the vote rate when determining that there is a recurrence of the major adverse event within one year as “Yes” is 75% (3/4), that is, the cut-off value is This is an example of 0.75.

  As shown in Table 1, in this example, for the first data (1), since any of the first learning models (1) to (4) determined that the recurrence of the major adverse cardiac event was “Yes”, “Yes” "" Is 100% (4/4) and, as a result, exceeds the set cut-off value (cut-off value≤voting rate), the second determination result is "Yes".

  For the first data (2), the first learning models (1) to (3) are determined to be “present”, and the first learning model (4) is determined to be “unknown”. 75% (3/4), and as a result, the second determination result is “present”.

  For the first data (3), the first learning models (1) to (3) are determined to be “unknown”, and the first learning model (4) is determined to be “present”. Since it is 25% (1/4) and smaller than the set cut-off value (cut-off value> voting rate), the second determination result is “unknown” as a result.

  The example shown in Table 1 is an example related to the first learning model that predicts “Yes” or “Unknown”, and the second determination result related to the second learning model that predicts “None” or “Unknown”. In Table 1, the first learning models (1) to (4) in Table 1 are replaced with the second learning models (1) to (4), and the second determination result is included and “Yes” is inverted to “No”. This can be explained similarly.

  In the present embodiment, voting is performed on the first determination result, and a step of acquiring the second determination result based on the vote rate is performed. In the present embodiment, the prediction accuracy tends to improve as compared to the prediction performance when a single learning model is used as the number of learning models used increases. Further, when the number of learning models reaches a predetermined number, the number of prediction errors tends to increase as the number of learning models increases thereafter.

  This suggests that there is an optimal value for the number of learning models used. Therefore, to improve prediction accuracy (to reduce the number of prediction errors), for example, by observing fluctuations in the number of learning errors and the number of occurrences of prediction errors as the number of learning models increases, By reducing the number of learning models used rather than the number of learning models that have fluctuated in the direction of increasing the number of occurrences, it is possible to improve prediction accuracy (reduce the number of prediction errors). In the present embodiment, in addition to the increase in the number of occurrences of prediction errors, considering that the number of votes for the prediction of “onset or recurrence” and the prediction of “onset or no recurrence” are aligned, at most about 40 models Each of the first selection learning model and the second selection learning model is preferably used.

  In the present embodiment, the cut-off value can be appropriately adjusted in order to perform prediction more suitable for the purpose. Here, the cut-off value of the vote rate of the first selection learning model and the cut-off value of the vote rate of the second selection learning model can be set to different values.

  In the case of exclusion diagnosis that covers major adverse cardiac events and acute kidney injury as described above, we will exclude as much as possible false negatives, that is, cases that have developed even though we predicted no onset It is necessary. Therefore, the cutoff value of the first learning model is made smaller than the cutoff value of the second learning model, that is, the cutoff value of the second learning model is made larger than the cutoff value of the first learning model. By setting to, the false negative rate can be further reduced. In this way, a prediction method more suitable for “exclusion diagnosis” can be achieved.

  Further, by setting the cutoff value of the first learning model to be larger than the cutoff value of the second learning model, the false positive rate can be further reduced. As a result, a prediction method more suitable for “definite diagnosis” can be achieved as in the case where diabetic nephropathy is targeted for prediction.

  The cut-off value is appropriately adjusted in view of the target disease, the seriousness when the prediction result in such a disease is incorrect, and the like.

  In the present embodiment, both the prediction of the first learning model (onset or recurrence) and the prediction of the second learning model (onset or no recurrence) are determined by voting, so that the cut-off value is too high. No matter what first data is applied, all are predicted to be “unknown”, and if the cut-off value is too low, all may be predicted to be “onset or recurrence”. In view of the above, in the present embodiment, the cutoff value is preferably set in the range of 0.3 to 0.7.

  Specifically, in the present embodiment aimed at exclusion diagnosis of major adverse cardiac events, the cutoff value of the first learning model is set to a range of 0.3 to 0.6, and the cutoff value of the second learning model is set. Is preferably in the range of 0.4 to 0.7 from the viewpoint of giving a gradient to the acceptance rate of the voting results of the first learning model and the second learning model.

  The evaluation of the cutoff value will be described with reference to FIGS. 8 and 9 are tables for explaining the evaluation result of the cutoff value.

  Here, out of a total of 1231 cases, the number of positive cases (the number of patients who recurred major adverse events within one year) was 100, and the number of negative cases (the major adverse events recurred within one year). An example of using a model in which the number of patients who did not exist is 1131 is shown.

  When the optimum value of the cutoff value of the second learning model was searched for when the cutoff value of the first learning model was set to 0.35, the optimum cutoff value of the second learning model was 0.45. .

  As is apparent from FIGS. 8 and 9, when the cutoff value of the first learning model is set to 0.35 and the cutoff value of the second learning model is set to 0.45 (in FIG. 9, P35_N45). The sensitivity was 0.77, and the specificity was 0.45.

  In FIG. 8, the calculated value in which the case where the prediction result is “unknown” is not counted is shown in the upper part, and the calculated value according to the conventional method is shown in the lower part.

  In FIG. 8, “YY” indicates the number of cases where “the first learning model is determined to be“ recurring ”(Y) and the second learning model is determined to be“ no recurrence ”(Y). , “UU” represents the number of cases where the first learning model is determined as “unknown” (U) and the second learning model is determined as “unknown” (U).

  Next, a step (S2-6) of acquiring a third determination result by integrating the second determination result of the first selection learning model and the second determination result of the second selection learning model is performed.

  By this step (S2-6), the second selection result of the classification problem relating to “the onset or recurrence of disease within the predetermined period or unknown” of the first selection learning model (first learning model) and the second selection The second determination result relating to the classification problem of “no disease onset or recurrence within a predetermined period or unknown” in the learning model (second learning model) is integrated. As a result, a third determination result obtained by integrating the second determination result of the first selective learning model and the second determination result of the second selective learning model is acquired.

  Specifically, for example, (1) the second determination result related to the first selection learning model is “onset or recurrence of disease”, and the second determination result related to the second selection learning model is “unknown”. In this case, the third determination result is “onset or recurrence of disease”. (2) When the second determination result concerning the first selection learning model is “unknown” and the second determination result concerning the second selection learning model is “no disease onset or recurrence”, The third determination result is “no disease onset or recurrence”.

  When the second determination result related to the first selective learning model is “with the onset or recurrence of the disease” and the second determination result according to the second selective learning model is “the onset or no recurrence of the disease” When the second determination result concerning the first selection learning model is “unknown” and the second determination result concerning the second selection learning model is also “unknown”, the third determination result is “unknown”. Is done.

Next, based on the third determination result, a step (S2-7) of predicting a disease onset risk or a recurrence risk is performed.
Specifically, based on the third determination result obtained in step (S2-6) already described, in the case of (1) above, “the risk of disease onset or recurrence is high within a predetermined period” In the case of (2) above, it is determined that “the risk of disease onset or recurrence is low within a predetermined period”, and in cases other than (1) and (2) above, within the predetermined period The determination of the risk of disease onset or recurrence in is suspended.

  As already described, this step (S2-7) is usually performed by a computer. However, the onset risk or recurrence risk prediction based on the third determination result, the determination of the frequency of outpatient visit based on the onset risk or recurrence risk prediction, etc., follow a rule definition file or algorithm based on the knowledge of doctors, consultants, etc. Can also be processed.

  The step (S2-8) of calculating the reliability of the predicted onset risk or recurrence risk can be performed based on the already described second determination result.

  Specifically, first, when acquiring the second determination result, when a plurality of learning models determine, for example, “onset or recurrence”, “+1” point, and “onset or recurrence” Is assigned “−1” points, and “0” points are assigned when it is determined as “unknown”, and the sum of scores for all models for the same subject is calculated. The ratio of the sum of the scores of “” or the sum of the scores of “no onset or recurrence” is calculated as the reliability (%).

  The calculated reliability (%) is a value from -100 (%) (small risk of disease onset or recurrence) to 100 (%) (high risk of disease onset or recurrence). I can take it.

  The reliability (%) is calculated based on the past cases by associating the analysis by machine learning with the onset frequency or recurrence frequency in the data used for learning as described above. It is suitable as an index of

[Prediction device]
The program according to the present embodiment is executed by a computer including a calculation unit. Hereinafter, with reference to FIG. 10 and FIG. 11, the configuration of the computer 10 that is a prediction device will be described.

  FIG. 10 is a schematic block diagram for explaining the configuration of the computer. FIG. 11 is a schematic block diagram for explaining the configuration of the calculation unit.

  As shown in FIG. 10, the computer 10 includes an arithmetic unit 12 that processes instructions such as generating data based on the acquired parameters and storing the data.

  The calculation unit 12 is a functional unit corresponding to, for example, a microprocessor (CPU), a graphic processor (GPU), or the like.

  The computer 10 further includes a storage unit 14 that can store input data, generated data, and the like temporarily or for a predetermined period, and stores the input data in a readable state.

  The storage unit 14 corresponds to, for example, a memory (RAM) device, a hard disk drive, an SSD, and the like, and is a functional unit configured to cooperate with the calculation unit 12.

  Examples of data that can be stored and stored in the storage unit 14 in a readable state include first data, second data (complete data, partial data, updated bit string data set), evaluation data, Examples include a support vector machine, a first learning model, a second learning model, a first selection learning model, and a second selection learning model.

  The computer 10 further includes external functional units and functional units such as an input / output unit 16 that is an interface such as serial connection and parallel connection for exchanging data with the device.

  In addition, the computer 10 outputs data generated on an input device 22 such as a keyboard and a mouse, a display device capable of visually displaying data, a paper medium, and the like, which functions when connected to the input / output unit 16. The computer 10 includes a so-called computer hardware resource such as a printer capable of reading and writing, a large-capacity external storage device 32 capable of reading and writing, or a dedicated hardware resource corresponding to each of these components. It can also be set as the structure connected so that it may cooperate with a function part.

  Specifically, for example, an external storage device that stores the database for the first data and the database for the second data already described, or hardware resources that store the first data as a database and the second data as the database The hardware resources to be stored are configured as separate external storage devices, installed in remote locations that are physically separated from the computer installation locations described above, and can cooperate with each other by electric communication lines. You may comprise so that it may connect to.

  Here, “connected by telecommunication line” means connected by connecting data, control signals, etc. via a wired or wireless information line using a medium such as electricity or light. This means that the devices are configured to work together.

  Even if the prediction device for executing the program of the present embodiment is constituted by a single computer, a plurality of computers (including servers, operation terminals, etc.) and other peripheral devices are integrally connected by an electric communication line. It may be configured as a system.

  In addition, among the processes described in the above-described embodiments, all or part of the processes described as being automatically performed can be manually performed, or all of the processes described as being performed manually are performed. Alternatively, a part can be automatically performed by a known method. In addition, the processing procedures, control procedures, specific names, various data, information including parameters, screen examples, and database configurations shown in the above description and drawings may be arbitrarily changed unless otherwise specified. it can.

  The prediction device of this embodiment is a prediction device for predicting the risk of disease onset or the risk of recurrence within a predetermined period.

  As illustrated in FIG. 11, the prediction device 100 of the present embodiment includes a plurality of functional units that function in cooperation with each other in the arithmetic unit 12 that has already been described. The prediction device 100 of the present embodiment includes a first data generation acquisition unit 12a, a second data generation acquisition unit 12b, a learning model construction unit 12d, a learning model selection unit 12e, a first determination result generation acquisition unit 12f, and a second determination result. A generation acquisition unit 12g, a third determination result generation acquisition unit 12h, and a prediction unit 12i are provided.

  The prediction apparatus 100 according to an embodiment of the present invention acquires two or more index parameters measured using a test sample collected from a subject or obtained from the subject, and based on the index parameters The first data generation / acquisition unit 12a that generates and acquires the first data and the first data acquired by the first data generation / acquisition unit 12a are processed by the first selection learning model and the second selection learning model, and the first data A first determination result generation acquisition unit 12f that generates and acquires a first determination result for each of the one selection learning model and the second selection learning model, and a plurality of first selection learning models acquired by the first determination result generation acquisition unit 12f. A second determination result generation and acquisition unit 12g that generates and acquires a second determination result based on the plurality of first determination results and the plurality of first determination results of the plurality of second selection learning models; Generation A third determination result generation acquisition unit 12h that generates and acquires a third determination result by integrating the second determination result of the first selection learning model acquired by 12g and the second determination result of the second selection learning model; And a prediction unit 12i that predicts a disease onset risk or a recurrence risk based on the third determination result acquired by the third determination result generation acquisition unit 12h.

  The prediction apparatus 100 according to another embodiment selects based on the selection frequency from a group consisting of a plurality of index parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred. Based on the second data generated and acquired from the second data generation and acquisition unit 12b, the second data generation and acquisition unit 12b that generates and acquires the second data based on the two or more types of the index parameters, A learning model construction unit 12d and a learning model construction unit 12d constructed a plurality of first learning models for predicting the presence or absence of recurrence and a plurality of second learning models for predicting the onset or no recurrence of disease or unknown. Learning model selection for selecting a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models It may further include a 12e.

  In the prediction device 100 of yet another embodiment, the second data generation / acquisition unit 12b selects a plurality of index parameters and generates partial data, thereby generating a plurality of partial data and partial data. It is preferable that it is a function part which produces | generates the 2nd data which may contain the used complete data (case data).

  In the prediction apparatus 100 according to yet another embodiment, the learning model construction unit 12d divides the second data acquired from the second data generation acquisition unit 12b into a plurality of data groups, and each of the data groups is teacher data. And a functional unit that constructs a plurality of first learning models and a plurality of second learning models by machine learning performed under a plurality of parameter conditions that differ for each of the plurality of data groups.

  In the prediction apparatus 100 according to another embodiment, the second determination result generation acquisition unit 12g includes a plurality of first determination results of the plurality of first selection learning models and a plurality of first of the plurality of second selection learning models. It is preferable that the function unit performs voting on the determination result, and generates and acquires the second determination result based on the vote rate for each of the first selection learning model and the second selection learning model.

DESCRIPTION OF SYMBOLS 10 Computer 12 Calculation part 12a 1st data generation acquisition part 12b 2nd data generation acquisition part 12d Learning model construction part 12e Learning model selection part 12f 1st determination result generation acquisition part 12g 2nd determination result generation acquisition part 12h 3rd determination result Generation acquisition unit 12i Prediction unit 14 Storage unit 16 Input / output unit 22 Input device 32 External storage device 100 Prediction device

Claims (15)

By using the first data obtained from a test sample measured from the subject, including the following steps executed by a computer having a calculation unit, or within a predetermined period A program for predicting the risk of disease onset or recurrence in
Two or more kinds selected from the group consisting of a plurality of indicator parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred within the predetermined period. Generating and acquiring a plurality of second data based on the index parameter;
The computing unit constructing a plurality of first learning models for predicting the onset or recurrence of the disease or unknown, and a plurality of second learning models for predicting the onset or no recurrence of the disease or unknown;
The arithmetic unit includes a step of selecting a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models. The arithmetic unit processes the first data by the first selection learning model and the second selection learning model, and acquires a first determination result for each of the first selection learning model and the second selection learning model. Steps,
The arithmetic unit obtaining a second determination result based on a plurality of first determination results of the plurality of first selection learning models and a plurality of first determination results of the plurality of second selection learning models;
The arithmetic unit integrating the second determination result of the first selective learning model and the second determination result of the second selective learning model to obtain a third determination result;
The said calculating part further includes the step of predicting the onset risk or recurrence risk of the said disease within a predetermined period based on the said 3rd determination result, The onset risk of the disease within the predetermined period of Claim 1 or A program for predicting the risk of recurrence.   The step of obtaining the second determination result is performed by setting the cut-off value of the vote ratio of the first selection learning model and the cut-off value of the vote ratio of the second selection learning model to different values. The program for predicting the onset risk or recurrence risk of the disease within the predetermined period according to claim 2.   The risk of developing a disease within a predetermined period according to claim 2 or 3, wherein the calculation unit further includes a step of calculating a reliability of the predicted onset risk or recurrence risk based on the second determination result. A program for predicting the risk of recurrence.   The step of constructing the plurality of first learning models and the plurality of second learning models selects the plurality of index parameters, thereby generating the plurality of partial data generated and the complete data used for generating the partial data. The second data is generated and prepared, the second data is divided into a plurality of data groups, each of the data groups is used, and machine learning is performed under a plurality of parameter conditions that are different for each of the plurality of data groups. The program for predicting the onset risk or the recurrence risk of the disease in the predetermined period of any one of Claims 1-4 which is a step to construct | assemble.   The step of obtaining the second determination result performs voting on the plurality of first determination results of the plurality of first selection learning models and the plurality of first determination results of the plurality of second selection learning models, The risk of occurrence or recurrence of a disease within a predetermined period according to any one of claims 2 to 5, which is a step of obtaining a second determination result based on a vote rate for each of the selective learning model and the second selective learning model. A program for predicting risk.   The step of selecting the first selection learning model and the second selection learning model is a step of evaluating and selecting the first learning model and the second learning model based on a Pareto rank with indices of sensitivity and positive predictive value. The program for predicting the onset risk or recurrence risk of the disease within the predetermined period according to any one of claims 1 to 6.   The risk of disease onset or recurrence within a predetermined period according to any one of claims 1 to 7, wherein the disease is a major adverse cardiac event, acute kidney injury, cerebral aneurysm rupture or diabetic nephropathy. A program to predict. First data that is measured using a test sample collected from a subject, or two or more index parameters obtained from the subject are acquired, and first data is generated and acquired based on the index parameters A generation acquisition unit;
The first data acquired by the first data generation and acquisition unit is processed by a first selection learning model and a second selection learning model, and a first determination result is obtained for each of the first selection learning model and the second selection learning model. A first determination result generation acquisition unit that generates and acquires;
The second determination result based on the plurality of first determination results of the plurality of first selection learning models and the plurality of first determination results of the plurality of second selection learning models acquired by the first determination result generation acquisition unit. A second determination result generation acquisition unit that generates and acquires
A second determination result of the first selection learning model acquired by the second determination result generation acquisition unit and a second determination result of the second selection learning model are integrated to generate and acquire a third determination result. 3 determination result generation acquisition unit;
Based on the third determination result acquired by the third determination result generation acquisition unit, the prediction unit predicts the risk of disease onset or recurrence within a predetermined period, Prediction device for predicting recurrence risk. Based on two or more index parameters selected based on the selection frequency from a group consisting of a plurality of index parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred A second data generation and acquisition unit for generating and acquiring second data;
Based on the second data acquired from the second data generation / acquisition unit, a plurality of first learning models for predicting the onset or recurrence of the disease or the unknown, and predicting the onset or no recurrence of the disease or the unknown A learning model building unit for building a plurality of second learning models;
A learning model selection unit that selects a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models constructed by the learning model construction unit. Furthermore, the prediction apparatus for predicting the onset risk or recurrence risk of the said disease within the predetermined period of Claim 9 further included.   The second data generation / acquisition unit selects a plurality of index parameters and generates partial data, thereby generating second data including a plurality of partial data and complete data used to generate the partial data. The prediction device for predicting the onset risk or the recurrence risk of the disease within a predetermined period according to claim 9 or 10, which is a functional unit. A learning model that is measured using a test sample collected from a subject, or that is used in a method for predicting the risk of disease onset or recurrence within a predetermined period by inputting first data obtained from the subject. The construction method of
Based on two or more index parameters selected based on the selection frequency from a group consisting of a plurality of index parameters for a plurality of subjects who have developed or relapsed a disease within a predetermined period or have not developed or recurred Generating and acquiring a plurality of second data;
Constructing a plurality of first learning models for predicting the onset or recurrence of the disease or unknown, and a plurality of second learning models for predicting the onset or recurrence of the disease or no unknown;
Selecting a plurality of first selection learning models and a plurality of second selection learning models for each of the plurality of first learning models and the plurality of second learning models.   The step of constructing a plurality of first learning models and a plurality of second learning models is used to generate a plurality of partial data and the partial data by selecting a plurality of index parameters and generating partial data. Generating and preparing second data including complete data, dividing the second data into a plurality of data groups, using each of the data groups, and performing a plurality of parameter conditions for each of the plurality of data groups. The learning model construction method used for the prediction method of the onset risk or recurrence risk of the disease within the predetermined period according to claim 12, which is a step of construction by machine learning.   The step of selecting the first selection learning model and the second selection learning model is a step of evaluating and selecting the first learning model and the second learning model based on a Pareto rank with indices of sensitivity and positive predictive value. A method for constructing a learning model for use in the method for predicting a risk of developing a disease or a risk of recurrence within a predetermined period according to claim 13.   The learning model used for the prediction method of the risk of disease onset or recurrence within a predetermined period according to claim 14, wherein the disease is a major adverse cardiac event, acute kidney injury, cerebral aneurysm rupture or diabetic nephropathy. Construction method.

Images