JP6986650B1 - Method for predicting the risk of complications in Kawasaki disease or IgA vasculitis, trained model and its generation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for predicting the risk of complications in Kawasaki disease or IgA vasculitis, a trained model and a method for generating the same. SOLUTION: The trained model includes a step of inputting clinical information into a trained model and outputting a predicted value of a sample score, and the trained model contains clinical information and a sample score of a subject determined to have a coronary artery dilated lesion (CAL). The relationship with the calculated value is machine-learned, and clinical information includes three or more types such as gender, age, coronary artery diameter, vasculitis marker, hypercytomicemia marker, and score of immunoglobulin refractory prediction score for intravenous injection. The calculated value of the sample score is directly significant to the judgment result by performing the covariance structure analysis with 4 or more observation variables including 3 or more kinds of clinical data and the judgment result of CAL occurrence. It is a method for predicting the risk of CAL occurrence in Kawasaki disease, which is a factor score of a latent variable having a causal relationship. It is a method for predicting the risk of occurrence of IgA vasculitis with nephritis or nephritis with high proteinuria. [Selection diagram] FIG. 6

Description

The present invention is a predictive method for predicting the risk of complications in Kawasaki disease (hereinafter also referred to as "KD") patients or IgA vasculitis (hereinafter also referred to as "IgAV") patients. It relates to a trained model for prediction and a method for generating the trained model.

Kawasaki disease (KD) is also known as pediatric acute febrile mucocutaneous lymph node syndrome (MCLS). IgA vasculitis (IgAV) is also known as Henoch-Schoenlein purpura (HSP), anaphylactoid purpura, allergic purpura, or vascular purpura. KD and IgAV are types of vasculitis, respectively, and they are common in children because they are common in children and the clear cause is still unknown. Vasculitis is also called vasculitis syndrome or systemic vasculitis, and is a general term for diseases in which inflammation is observed in the blood vessels themselves. In addition, the acute phase is the time when symptoms begin to appear rapidly after becoming ill. Complications are other illnesses caused by one illness. For example, in the acute phase of IgAV, IgA-type immune complexes are deposited on the vascular endothelium, causing small arteritis and glomerulonephritis. In 30% to 50% of cases of IgAV, the complication purpura nephritis (hereinafter also referred to as "PN") develops from about 30 days after the onset of IgAV. PN may be simply referred to as "nephritis" as an IgAV complication (see Non-Patent Document 9), or it may be referred to as Henoch-Schoenline purpura nephritis (HSPN). PN is said to be the most common case of secondary glomerulonephritis in children, and when it develops, it is accompanied by hematuria. In cases of PN with highly proteinuria in hematuria, nephrotic syndrome may be further present and become severe, leading to renal failure.

In the acute phase of KD, vasculitis (middle arterial vasculitis) may be induced in the coronary arteries of the heart from about 7 days after the onset of KD. Furthermore, from about 10 days after the onset of KD, a lesion in which the coronary artery diameter is enlarged (coronary artery lesion: hereinafter also referred to as “CAL”) may occur due to middle arterial vasculitis. CAL with a large degree of expansion of coronary artery diameter is also called coronary artery aneurysm (hereinafter referred to as "CAA") and can cause death due to myocardial infarction. Even if they do not lead to myocardial infarction, most children diagnosed with cardiovascular sequelae due to CAL development will continue to be at risk of some coronary event for the rest of their lives. For this reason, KD is said to be the leading cause of acquired pediatric heart disease. Doctors in charge of acute care for KD or IgAV are required to end vasculitis at an early stage from the viewpoint of suppressing the occurrence of complications.

FIG. 31 shows the algorithm of the conventional KD acute care S20. The doctor in charge admits a patient diagnosed with KD S21 and determines whether the patient has a persistent fever (S22). If it is determined that the fever has already disappeared, the patient is administered with acetylsalicylic acid (hereinafter, also referred to as "ASA") and followed up (S23). If the fever persists, the attending physician determines whether the patient is at low risk for the IVIG refractory case described below (S24). If it is judged to be low risk, as standard established first-line treatment S30, IVIG therapy in which a large amount of intravenous immunoglobulin (hereinafter also referred to as "IVIG") is administered together with moderate dose of ASA administration. At this time, the combination therapy of administering another anti-inflammatory drug is not performed (S31). Alternatively, if determined to be at high risk, the attending physician may administer IVIG and ASA as first-line treatment S30, as well as concomitant therapy with other anti-inflammatory agents to avoid CALs (). S32 to S35). After the first-line treatment S30, the doctor in charge determines whether or not the patient falls under the category of "IVIG reaction example", which is a case in which the patient has a fever of 37.5 ° C or lower and does not relapse within 24 hours or more and 36 hours after the end of IVIG administration. (S36). Patients with IVIG reaction will be further administered with ASA and followed up (S23).

On the other hand, some cases in the acute phase of KD do not resolve or relapse after first-line treatment S30 and are referred to as "IVIG refractory cases". The attending physician will administer the second-line treatment S40 and, if necessary, the third-line treatment and subsequent S50 to patients refractory to IVIG. Nevertheless, some of the IVIG refractory cases lead to the development of CAL. For IVIG refractory cases in the acute phase of KD, the standard treatment method has not yet been established. Therefore, for the doctor in charge, not only IVIG re-administration (S41, S51) but also other anti-inflammatory therapies with more excellent anti-inflammatory effect are available as treatment options that can be taken in S40 after the second-line treatment and S50 after the third-line treatment. There are also (S42 to S44, S52 to S56). The doctor in charge has decided that it is suitable for each patient based on the clinical information that differs for each patient who is refractory to IVIG, and based on the non-patent document 1 and research papers on KD acute care (so-called evidence). We have selected and implemented anti-inflammatory therapy.

Kobayashi score as a risk score (hereinafter also referred to as "IVIG refractory prediction score") for predicting whether or not a patient is IVIG refractory at the time when the doctor in charge determines before the first treatment S30 (S24). (Non-Patent Document 3), Egami Score (Non-Patent Document 4), and Sano Score (Non-Patent Document 5) were proposed (see Non-Patent Documents 1 and 2). Some doctors administer IVIG and ASA in subsequent first-line treatment S30 to patients judged to be at high risk based on the IVIG refractory prediction score score at the time of judgment (S24) prior to first-line treatment S30. Not only that, we recommend the combined use of S34 for intravenous methylprednisolone (hereinafter referred to as “IVMP”) administration (see Non-Patent Document 6).

Japanese Patent No. 6703421

Dai Miura, 14 others, "Guidelines for Acute Treatment of Kawasaki Disease of the Japanese Society of Pediatric Cardiology (2020 Revised Edition)", Pediatric Cardiology and Cardiac Surgery, 2020, Vol. 36, Supplement 1. Japanese Circulation Society, "2020 Revised Guidelines for Diagnosis and Treatment of Kawasaki Disease Cardiovascular Sequelae", [online], March 2020, [Search on July 8, 2020], Internet, <URL: https //www.j-circ.or.jp/old/guideline/pdf/JCS2020_Fukazawa_Kobayashi.pdf> Tohru Kobayashi, 7 others, "Prediction of intravenous immunoglobulin unresponsiveness in patients with Kawasaki disease", Circulation, June 2006, Vol. 113, No. 22, pp. 2606 to 2612 Kimiyasu Egami, 6 others, "Prediction of resistance to intravenous immunoglobulin treatment in patients with Kawasaki disease", The Journal of Pediatrics, August 2006, Vol. 149, No. 2, pp. 237-240 Tetsuya Sano, 7 others, "Prediction of non-responsiveness to standard high-dose gamma-globulin therapy in patients with acute Kawasaki disease before starting initial treatment", European Journal of Pediatrics, February 2007, Vol. 166, No. 2. No., pages 131 to 137 Keiko Okada, 9 others, "Pulse methylprednisolone with gammaglobulin as an initial treatment for acute Kawasaki disease", European Journal of Pediatrics, February 2009, Vol. 168, No. 2, pp. 181-185 Japan Kawasaki Disease Society, Japan Kawasaki Disease Research Center, Health and Labor Sciences Research Group on Intractable Vasculitis, "Kawasaki Disease Diagnosis Guide Revised 6th Edition", [online], June 2019, [Search on June 25, 2nd year of Reiwa], Internet, <URL: http://www.jskd.jp/info/pdf/tebiki201906.pdf> Tohru Kobayashi, 17 others, "A New Z Score Curve of the Coronary Arterial Internal Diameter Using the Lambda-Mu-Sigma Method in a Pediatric Population.", Journal of the American Society of Echocardiography, August 2016, Volume 29. , No. 8, pp. 794-801, e29 Mitsuaki Isobe, 49 others, "Clinical Guidelines for Vasculitis Syndrome (2017 Revised Edition)", [online], June 2018, [Search on July 12, 2018], Internet, <URL: https //www.j-circ.or.jp/cms/wp-content/uploads/2020/02/JCS2017_isobe_h.pdf>

However, conventionally, in view of the sensitivity and specificity of the IVIG refractory prediction score, it is high for the patient in charge at the time of determining (S24) that the patient will develop CAL after the first-line treatment S30 before the first-line treatment S30. It was difficult to predict with accuracy. For example, conventionally, the doctor in charge predicts that he / she is a “low-risk patient” from the score of the IVIG refractory prediction score at the time of judgment (S24) before the first-line treatment S30, but still for this patient just in case. Even if IVIG, ASA and IVMP are administered in combination (S34) in the first-line treatment S30, it may become clear that this patient is refractory to IVIG after the first-line treatment S30 and leads to the occurrence of CAL. There was. For this reason, if the doctor in charge can predict with high accuracy that "even if IVMP is used in combination with S34 in the first treatment S30, CAL will occur" at the time of judgment (S24) before the first treatment S30, for example. The inventor of the present application considered that there may have been room to consider changing the treatment policy so that the first-line treatment S30 is combined with another anti-inflammatory therapy (S32, S33, S35).

On the other hand, other anti-inflammatory therapies (S32 to S35, S42 to S44, S52 to S56) have the advantage of being more excellent in anti-inflammatory effect than IVIG administration (S31, S41, S51), but have side effects? Alternatively, there is a drawback that the labor and cost burden for implementation are large. For this reason, conventionally, the doctor in charge of performing KD acute care S20 generally first performs first-line treatment S30 of S31 without combination treatment, and then for patients who are found to be IVIG refractory. It was customary to try other anti-inflammatory therapies (S42-S44, S52-S56) for the first time after the second-line treatment S40. For example, even if the attending physician proposes to the boss or the guardian of the patient the combined use of IVMP (S34) in the first-line treatment S30, the boss or the guardian may be concerned about the side effects of IVMP, which is a type of steroid drug. In many cases, the proposal was rejected due to concerns that the combined use of IVMP (S34) was not covered by insurance. In the past, there was not enough evidence that combination therapy could suppress the occurrence of CAL with a high probability, so the doctor in charge could not persuade the people concerned with the policy of implementing combination therapy, and in fact, first-line treatment S30 and It is considered that the administration of IVIG and ASA was limited to the second-line treatment S40 (S31, S41), but in the end, the patient did not resolve the fever and developed CAL.

Therefore, conventionally, for the doctor in charge of KD acute care S20, in the process of determining the initial treatment policy, other anti-inflammatory therapies (S3 2 to S51) having a better anti-inflammatory effect than IVIG administration (S31, S41, S51). The inventor of the present application considered that it is often difficult to make a diagnosis by selecting S35, S42 to S44, S52 to S56) and implementing it at an early stage, and as a result, CAL may occur in the patient. In the process of deciding the initial treatment policy in IgAV acute care, there is a similar problem, and it is considered that the patient may develop highly proteinuria.

In view of the above-mentioned problems, an object of the present invention is to assist a doctor in charge of acute care in Kawasaki disease or IgA vasculitis to easily make a diagnosis to make an early decision on a treatment policy to suppress the occurrence of complications. It is an object of the present invention to provide a predictive method capable of predicting the occurrence risk of complications in each case with as high accuracy as possible, a trained model for making this prediction, and a method for generating the same.

In order to solve the above-mentioned problems, the prediction method according to the embodiment of the present invention is a prediction method for predicting the risk of complications in a patient with Kawasaki disease, and the prediction method is a trained model. The complication comprises a step of inputting clinical information of the patient into an information processing apparatus having the above-mentioned, and executing a process of outputting a predicted value in the patient for a sample score regarding the risk of occurrence of the complication. The trained model is an enlarged lesion, and the trained model shows the relationship between the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined and the calculated value of the sample score in each subject. The clinical information of each of the subjects was machine-learned, and the clinical information of each subject was gender, age at the time of acute stage of Kawasaki disease and not receiving anti-inflammatory therapy, coronary artery diameter at the time, and coronary artery diameter at the time. Systemic vasculitis marker test value, hypercytomicemia marker test value at the time point, score of intravenous immunoglobulin (IVIG) refractory prediction score at the time point, received the anti-inflammatory therapy after the time point. Includes clinical data of 3 or more subjects selected from the group consisting of the number of times and data indirectly indicating any of these, and the calculated value of the sample score in each of the subjects is the subject of 3 or more. Observations related to the determination results when covariance structure analysis is performed with four or more observation variables including clinical data and determination results regarding the presence or absence of the complications in each of the subjects. It is data on the factor score of the latent variable which has a direct significant causal relationship with the variable, and the clinical information of the patient is the sex, the age at the time point, the coronary artery diameter at the time point, and the coronary artery diameter at the time point. Systemic vasculitis marker test value, hypercytomicemia marker test value at the time point, score of IVIG refractory prediction score at the time point, number of times to receive the anti-inflammatory therapy after the time point, and any of these It is a prediction method including three or more kinds of patient clinical data selected from the group consisting of data indirectly showing.

According to the method of predicting such a configuration, by using a trained model generated by machine learning the relationship between the clinical data of three or more kinds of subjects and the calculated value of the sample score calculated by the covariance structure analysis. , It becomes possible to obtain a predicted value of a sample score, which is an index of whether or not CAL occurs in a KD patient, with higher accuracy than before. Therefore, for example, even before the first-line treatment of a KD patient in KD acute care, it can be used as a judgment material at the time of diagnosis so that the attending physician can easily make a diagnosis to determine a treatment policy for suppressing the occurrence of CAL at an early stage. It is possible to support this doctor by outputting information on the prediction result (predicted value of the sample score).

The prediction method according to another embodiment of the present invention is a prediction method for predicting the risk of complications in a patient with IgA vasculitis, and the prediction method is a prediction method for an information processing apparatus having a trained model. It comprises the step of inputting clinical information of the patient and performing a process of outputting a predicted value in the patient for a sample score regarding the risk of developing the complication, and the complication is purpura vasculitis or the purpura. It is a case of pathogenic nephritis accompanied by proteinuria, and the learned model is the clinical information of each subject among the subjects whose presence or absence of the complication has already been determined, and the sample score of each subject. The relationship with the calculated value in the above was machine-learned, and the clinical information of each of the above subjects was the sex, the age at the time of acute stage of IgA vasculitis and not receiving anti-inflammatory therapy, and the above-mentioned time point. Systemic vasculitis marker test value at the time point, presence or absence of abdominal pain at the time point, presence or absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, said Includes clinical data of four or more subjects selected from the group consisting of the number of times the anti-inflammatory therapy was received after the time point and data indirectly indicating any of these, and the sample score was given in each of the subjects. The calculated values are provided with five or more observation variables including the clinical data of the four or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects, and the co-dispersion structure analysis is performed. In this case, it is data on the factor score of the latent variable in which a directly significant causal relationship is recognized with respect to the observed variable related to the determination result, and the clinical information of the patient is the sex, the age at the time, and the above. Systemic vasculitis marker test value at the time point, presence or absence of abdominal pain at the time point, presence or absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, A predictive method comprising 4 or more patient clinical data selected from the group consisting of the number of times the anti-inflammatory therapy is scheduled to be received after the time point and data indirectly indicating any of these.

According to the method of predicting such a configuration, by using a trained model generated by machine learning the relationship between the clinical data of four or more kinds of subjects and the calculated value of the sample score calculated by the covariance structure analysis. For IgAV patients, it is possible to obtain a predicted value of a sample score, which is an index of whether or not PN occurs or a case with proteinuria occurs in PN, with higher accuracy than before. Therefore, for example, even before the initial treatment of IgAV patients in IgAV acute care, it is easy for the doctor in charge to make an early diagnosis to determine the treatment policy to suppress the occurrence of PN or the occurrence of cases with proteinuria in PN. In addition, it is possible to support this doctor in charge by outputting information on the prediction result (predicted value of the sample score) that is used as a judgment material at the time of diagnosis.

The anti-inflammatory therapy according to the predictive method according to the present invention includes administration of acetylsalicylic acid and / or a salt thereof, IVIG administration, intravenous methylprednisolone pulse administration, prednisolone administration, infliximab administration, ulinastatin administration, cyclosporin A administration, and plasma. It can be one or more treatment methods selected from the group consisting of exchanges. According to the prediction method of such a configuration, the prediction result (prediction value of the sample score) of the risk of complications in the patient may differ depending on the number of scheduled anti-inflammatory treatments listed here. For this reason, for example, the doctor in charge of acute care refers to the prediction result (predicted value of the sample score) as a judgment material at the time of diagnosis so that the diagnosis for early determination of the treatment policy for suppressing the occurrence of complications can be easily performed. Cheap.

In the prediction method according to the present invention, when the predicted value of the sample score in the patient is equal to or higher than a predetermined cutoff value, a process of outputting the predicted result leading to the occurrence of the complication in the patient is executed. The information processing apparatus can be made to function. According to the prediction method of such a configuration, for the person who implements the prediction method, the prediction result is obtained in a form that makes it easy for the person who implements the prediction method to understand whether or not the patient will develop complications, rather than the numerical data itself of the predicted value of the sample score. Can be done.

Generating method according to an embodiment of the present invention, there is provided a method of generating a learned model for predicting the risk of complications in patients with Kawasaki disease, the generation method, already the occurrence of the complications When the clinical information of each subject in the subjects judged to be present or absent is input to the input layer, the output layer outputs the calculated value for each subject regarding the sample score regarding the risk of developing the complication. The complications include a step of training to generate the trained model, the complications are coronary dilated lesions, and the clinical information of each of the subjects is gender, acute phase of Kawasaki disease and receiving anti-inflammatory therapy. Age at no time, coronary artery diameter at the time point, systemic vasculitis marker test value at the time point, hypercytomicemia marker test value at the time point, score of IVIG refractory prediction score at the time point, Includes clinical data of three or more subjects selected from the group consisting of the number of times the anti-inflammatory therapy was received after the time point and data indirectly indicating any of these, and the sample score was given in each of the subjects. The calculated value of is provided with four or more observation variables including the clinical data of the three or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects, and the co-dispersion structure analysis is performed. This is a method for generating a trained model, which is data on factor scores of latent variables in which a directly significant causal relationship is recognized with respect to the observed variables related to the determination result. According to the method for generating a trained model having such a configuration, all of the subject clinical data given here as examples are information available to, for example, a doctor in charge of KD acute care. In addition, the calculated value of the sample score can be calculated by covariance structure analysis from the clinical data of three or more kinds of subjects and the determination result regarding the presence or absence of CAL. Therefore, it is possible to generate a trained model in which the risk of CAL occurrence can be predicted with higher accuracy than before, for example, in the pediatrics department of a conventional general hospital.

The method for generating a trained model according to another embodiment of the present invention is a method for generating a trained model for predicting the risk of complications in a patient with IgA vasculitis, and the method for generating the trained model has already been described above. When the clinical information of each subject in the subjects who were judged to have complications was input to the input layer, the output layer calculated the sample score for the risk of complications in each subject. The complications include purpura nephritis or a case of purpura nephritis with proteinuria, each of which comprises a step of machine training to output to generate the trained model. Clinical information includes gender, age at the time of acute IgA vasculitis and no anti-inflammatory therapy, systemic vasculitis marker test values at the time point, presence or absence of abdominal pain at the time point, said time point. The presence or absence of immediate allergic disease in, the blood IgA test value at the time point, the blood IgE test value at the time point, the number of times the anti-inflammatory therapy was received after the time point, and any of these indirectly. Includes clinical data of 4 or more subjects selected from the group consisting of the data shown in the above, and the calculated values for the sample score in each of the subjects are the clinical data of the four or more subjects and the clinical data of each of the subjects. When a covariance structure analysis is performed with 5 or more types of observation variables including the judgment result regarding the presence or absence of the complication, a directly significant cause and effect with respect to the observation variable related to the judgment result. It is a method of generating a trained model, which is data on factor scores of latent variables for which relationships are recognized. According to the method for generating a trained model having such a configuration, the subject clinical data given here as an example are all information available to, for example, a doctor in charge of IgAV acute care. In addition, the calculated value of the sample score can be calculated by covariance structure analysis from the clinical data of four or more subjects and the judgment result regarding the presence or absence of PN occurrence or the judgment result regarding the presence or absence of case occurrence with proteinuria in PN. Is. Therefore, it is possible to generate a trained model that can predict the risk of PN occurrence or the risk of case occurrence with proteinuria in PN with higher accuracy than before, for example, in the pediatrics department of a conventional general hospital.

The trained model according to the embodiment of the present invention is a trained model for predicting the risk of complications in a patient with Kawasaki disease, and the trained model already determines the presence or absence of the complications. The weighted values are machine-learned so that the clinical information of each subject in the determined subjects is input to the input layer and the output layer outputs the calculated value for each subject for the sample score regarding the risk of developing the complication. When the clinical information of the patient is input to the input layer, an operation based on the weighted value is performed on the input clinical information of the patient, and the output layer is used to perform the calculation. The purpose is to make the information processing device function so as to output the predicted value of the sample score in the patient, the complication is coronary artery dilated lesion, and the clinical information of each subject is gender and Kawasaki disease. Age at the time of acute phase and not receiving anti-inflammatory therapy, coronary artery diameter at the time point, systemic vasculitis marker test value at the time point, hypercytostatic marker test value at the time point, said Clinical data of three or more subjects selected from the group consisting of the score of the IVIG refractory prediction score at the time point, the number of times the anti-inflammatory therapy was received after the time point, and the data indirectly indicating any of these. Including,
Regarding the sample score, the calculated value in each of the subjects includes four or more observation variables including the clinical data of the three or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. This is data on the factor scores of latent variables for which a directly significant causal relationship is observed with respect to the observed variables related to the determination result when the covariance structure analysis is performed. , Gender, age at the time point, coronary artery diameter at the time point, systemic vasculitis marker test value at the time point, hypercytomicemia marker test value at the time point, IVIG refractory prediction score at the time point. A trained model that includes three or more patient clinical data selected from the group consisting of scores, the number of times the anti-inflammatory therapy is scheduled to be received after the time point, and data indirectly indicating any of these. According to the trained model having such a configuration, for the reason already described, it is possible to generate a trained model in which the risk of CAL occurrence can be predicted with higher accuracy than before, for example, in the pediatrics department of a conventional general hospital.

The trained model according to another embodiment of the present invention is a trained model for predicting the risk of developing complications in a patient with IgA vasculitis, and the trained model already has the occurrence of the complications. The clinical information of each subject in the subjects judged to be present is input to the input layer, and the weighted value is set so that the output layer outputs the calculated value for each subject regarding the sample score regarding the risk of developing the complication. When the clinical information of the patient is input to the input layer, which is machine-learned, the input clinical information of the patient is calculated based on the weighted value, and the output layer is used to perform the calculation. The purpose is to make the information processing device function so as to output the predicted value of the sample score in the patient, and whether the complication is purpura nephritis or the case of purpura nephritis accompanied by proteinuria. The clinical information of each of the above subjects includes gender, age at the time of acute IgA vasculitis and no anti-inflammatory therapy, systemic vasculitis marker test value at the time, and at the time. Presence or absence of abdominal pain, presence or absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, number of times of receiving anti-inflammatory therapy after the time point, and these The calculated values of each of the four or more subjects selected from the group consisting of data indirectly indicating any of them are included, and the calculated values for the sample scores are the four or more subjects clinical data and the above. When a covariance structure analysis is performed with five or more types of observation variables including the judgment result regarding the presence or absence of the complication in each subject, the observation variable related to the judgment result is directly used. It is data on the factor score of the latent variable which has a significant causal relationship, and the clinical information of the patient includes sex, age at the time point, systemic vasculitis marker test value at the time point, and time point. Presence or absence of abdominal pain, presence or absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, number of times to receive the anti-inflammatory therapy after the time point, and these It is a trained model containing 4 or more patient clinical data selected from the group consisting of data indirectly indicating any of the above. According to the trained model having such a configuration, it is possible to generate a trained model that can predict the risk of PN occurrence or the risk of case occurrence with proteinuria in PN with higher accuracy than before, for example, in the pediatric department of a conventional general hospital. Is.

The information processing apparatus having the trained model according to the embodiment of the present invention has a storage unit in which the trained model is stored and, when the clinical information of the patient is input, the input clinical information of the patient is input. It may be provided with a processing unit that is applied to the trained model and executes a process of outputting a predicted value in the patient for the sample score. According to an information processing device having a trained model with such a configuration, for example, in the pediatric department of a conventional general hospital, it is easy for a doctor in charge of acute care to make a diagnosis to make an early decision on a treatment policy to suppress the occurrence of complications. , It is possible to support this doctor in charge by outputting information on the prediction result (predicted value of the sample score) that will be the judgment material at the time of diagnosis.

As described above, the present invention provides a predictive method capable of predicting the risk of complications for each patient with as high accuracy as possible, a trained model for making this prediction, and a method for generating the same. can do.

The flowchart which shows the example of the generation method of the trained model which concerns on one Embodiment of this invention. A path diagram showing an example of a path model of an exploratory factor analysis model that can be created when performing covariance structure analysis in structural equation modeling (SEM). A path diagram showing an example of a path model of a confirmatory factor analysis model that can be created when performing covariance structure analysis with SEM. A path diagram showing an example of a path model of a quadratic factor model including two exploratory factor analysis models that can be created when performing covariance structure analysis with SEM. The schematic diagram which shows an example of the structure in the artificial neural network (ANN) which can be used in the generation method of the trained model which concerns on one Embodiment of this invention. The flowchart which shows the first example of the prediction method which concerns on one Embodiment of this invention. The flowchart which shows the 2nd example of the prediction method which concerns on one Embodiment of this invention. The block diagram which shows the functional structure about the example of the information processing apparatus which has the trained model which concerns on one Embodiment of this invention. The figure explaining the processing of the subjects in the 1st stage study to the 3rd stage study about the prediction of the risk of occurrence of CAL. A model diagram showing a confirmed path model by performing covariance structure analysis with SEM as an observation variable, with clinical data of 6 types of subjects and judgment results regarding the presence or absence of CAL occurrence in the first phase study on CAL occurrence risk prediction. In each of FIGS. 10 to 12, FIG. 20, FIG. 21, FIG. 25, FIG. 26, and FIG. 30, a rectangle has an observation variable, e has an error variable, and an arrow drawn with a thin solid line has a causal relationship. Means a path. In FIGS. 10-12, 21 and 26, the ellipse means a latent variable. The arrow drawn by the thick broken line means the path where the causal relationship exists, but it is different from FIGS. 11 and 12. A model diagram showing a confirmed path model by performing covariance structure analysis with SEM using 6 types of subject clinical data and judgment results regarding the presence or absence of CAL as observation variables in the second phase study on CAL occurrence risk prediction. .. The arrow drawn by the thick solid line means the path where the causal relationship exists, but it is different from FIGS. 10 and 12. A model diagram showing a confirmed path model by performing covariance structure analysis with SEM using 6 types of subject clinical data and judgment results regarding the presence or absence of CAL as observation variables in the 3rd phase study on CAL occurrence risk prediction. .. For 106 subjects in the first phase study, when it was judged that CAL had occurred when the Z score of the maximum coronary artery diameter at the time of judgment was 3.0 SD or more, CAL occurred with the subjects judged to have CAL. A graph comparing the calculated values of the sample scores for the risk of CAL occurrence with the subjects judged to be none. In each of FIGS. 13 to 17, the vertical axis of the graph indicates the magnitude of the sample score calculation value for the risk of CAL occurrence. For 208 subjects in the second phase study, when it was judged that CAL had occurred when the Z score of the maximum coronary artery diameter at the time of judgment was 3.0 SD or more, CAL occurred with the subjects judged to have CAL. A graph comparing the sample score calculation values related to the risk of CAL occurrence with the subjects judged to be none. For 314 subjects in the 3rd phase study, when it was judged that CAL had occurred when the Z score of the maximum coronary artery diameter at the time of judgment was 3.0 SD or more, CAL occurred with the subjects who were judged to have CAL. A graph comparing the sample score calculation values related to the risk of CAL occurrence with the subjects judged to be none. For 314 subjects in the 3rd phase study, when it was judged that CAL had occurred when the Z score of the maximum coronary artery diameter at the time of judgment was 2.5 SD or more, CAL occurred with the subjects who were judged to have CAL. A graph comparing the sample score calculation values related to the risk of CAL occurrence with the subjects judged to be none. For 314 subjects in the 3rd phase study, when it was judged that CAL had occurred when the Z score of the maximum coronary artery diameter at the time of judgment was 2.0 SD or more, CAL occurred with the subjects who were judged to have CAL. A graph comparing the sample score calculation values related to the risk of CAL occurrence with the subjects judged to be none. FIG. 6 is a schematic diagram illustrating the configuration of ANN machine-learned in Example 1-1 for CAL occurrence risk prediction. For 314 subjects in the 3rd phase study, the relationship between the sample score calculation value calculated by the mean covariance structure analysis and the sample score prediction value output from the trained model according to Example 1-1 is shown. graph. Arrows indicate outlier plots. In FIGS. 19, 24, and 29, the vertical axis of the graph shows the size of the calculated sample score, and the horizontal axis of the graph shows the size of the predicted sample score. The model figure which shows the path model which concerns on the comparative example 1-1. Regarding the prediction of the risk of occurrence of cases with high proteinuria in PN, covariance structure analysis was performed by SEM using the clinical data of 6 subjects and the judgment results regarding the presence or absence of cases with high proteinuria in PN as observation variables. , A model diagram showing a confirmed path model. A graph comparing the sample score calculation values of 93 subjects who developed IgAV between the subjects who were determined to have a case with severe proteinuria by PN and the subjects who were determined not to have this case. The vertical axis of the graph shows the magnitude of the sample score calculation value for the risk of developing a case with high proteinuria in PN. The schematic diagram explaining the composition of ANN which was machine-learned in Example 2-1 about the risk prediction of the case occurrence with high proteinuria by PN. Sample score calculation values calculated by mean covariance structure analysis and sample score prediction output from the trained model according to Example 2-1 for 93 subjects related to the prediction of the risk of occurrence of cases with high proteinuria in PN. A graph showing the relevance to the value. The model figure which shows the path model which concerns on the comparative example 2-1. A model diagram showing a confirmed path model by performing covariance structure analysis with SEM using 5 types of subject clinical data and judgment results regarding the presence or absence of PN occurrence as observation variables for PN occurrence risk prediction. A graph comparing the sample score calculation values between the subjects determined to have PN and the subjects determined not to have PN for 93 subjects who developed IgAV. The vertical axis of the graph shows the size of the sample score calculation value for the risk of PN occurrence. The schematic diagram explaining the structure of the ANN which was machine-learned in Example 3-1 about the PN occurrence risk prediction. For 93 subjects related to PN occurrence risk prediction, the relationship between the sample score calculation value calculated by the mean covariance structure analysis and the sample score prediction value output from the trained model according to Example 3-1 is shown. graph. The model figure which shows the path model which concerns on the comparative example 3-1. A flow diagram illustrating an algorithm for KD acute care. See Non-Patent Document 1.

[Learned model generation method for CAL occurrence risk prediction in KD]
The inventor of the present application considered as follows in the process of completing the present invention. Conventionally, it has been difficult to accurately predict the risk of KD acute phase patients leading to CAL development at the time of determination (S24) before the first-line treatment S30 in FIG. 31. It was considered that the cause was that unknown factors related to the occurrence of CAL were hidden in each patient, that is, different predispositions (genetic predisposition and environmental factors) were latent in each patient. Based on this idea, the inventor of the present application focused on covariance structure analysis. Conventionally, covariance structure analysis has been utilized in fields such as sociology, psychology, or marketing to analyze complex problems containing several unknown factors (see, for example, Patent Document 1). As far as the inventor of the present application can know, there is no precedent used in the field of clinical medicine to explain the causal relationship between latent variables and constitution. In pediatric clinical medicine, complex associations may exist between multiple clinical indicators. Therefore, for example, based on clinical information such as test values of KD acute phase patients that can be obtained by a pediatrician in the field of daily medical care, CAL occurrence after the first-line treatment S30 is detected by the covariance structure analysis before the first-line treatment S30. I thought that it might be possible to solve the complicated problem of predicting.

Therefore, initially, the inventor of the present application used the clinical information obtained at the time of performing the KD acute phase medical treatment S20 for the subjects who had performed the KD acute phase medical treatment S20 in the past and evaluated the presence or absence of CAL. A distributed structure analysis was performed. The covariance structure analysis has an advantage that variables (clinical data) that may have a causal relationship with the occurrence of CAL can be arbitrarily extracted from clinical information and this causal relationship can be verified. However, in the statistical model created by covariance structure analysis, it is not possible to apply the clinical data of new KD patients to this statistical model to calculate the predicted value of whether or not new KD patients will develop CAL. There were drawbacks. Here, the clinical information of the subjects whose presence or absence of CAL has already been evaluated by the inventor of the present application on a trial basis and the sample score calculated by covariance structure analysis using this clinical information (sample score: hereinafter "SS"". The calculated values (also referred to as) were machine-learned by an artificial neural network (hereinafter also referred to as “ANN”) as training data. That is, the inventor of the present application tried a combination of covariance structure analysis and ANN analysis. As a result, the generated trained model was able to unexpectedly obtain a predicted value of SS having a high correlation with the occurrence of CAL. Therefore, the inventor of the present application has found that it is possible to provide a predictive method capable of predicting the risk of complications for each patient with as high accuracy as possible, a trained model for making this prediction, and a method for generating the same. , Has led to the creation of the present invention. Hereinafter, embodiments of the present invention will be described with reference to the drawings.

The trained model generation method S1a according to the embodiment of the present invention shown in FIG. 1 includes a clinical information acquisition step S2a of the subjects, a pre-learning normalization step S3a, a covariance structure analysis step S4a, and a learning step. Includes S6a and sorting step S7a.

In the clinical information acquisition step S2a of the subjects, for example, for the subjects whose presence or absence of CAL was determined after receiving at least the first-line treatment S30 in the KD acute care S20 shown in FIG. 31 in order to create learning data. , The clinical information of each subject included in these subjects and the determination result regarding the presence or absence of CAL in each subject are acquired. The first-line treatment S30 is also referred to as the first line. The clinical information and judgment results acquired in step S2a shown in FIG. 1 are the results of diagnosis and treatment of each subject by a doctor (including a person who has been instructed by a doctor; the same applies hereinafter) as described below. , Information that has already been generated. The steps of diagnosis and treatment performed by the doctor in the process until this information is generated are not included in the present invention or step S2a.

From the viewpoint of selecting data suitable for machine learning of ANN, each subject in the generation method S1a may be a patient diagnosed as a typical example KD or an incomplete type KD (KD type example), and is preferably a typical example KD. The patient was diagnosed by a doctor. According to the KD diagnostic criteria described in Non-Patent Document 7, a examinee who presents with 5 or more of the 6 main symptoms of KD described below is diagnosed as a typical KD and has 4 types of symptoms. Those who present with the above are diagnosed with typical KD or incomplete KD. A examinee who presents with three types of symptoms is diagnosed with incomplete KD when other diseases are denied and KD is most likely based on "symptoms or findings". The six major symptoms of KD are (1) fever, (2) bilateral ocular conjunctival swelling, (3) lip or oral findings: flushing of the lips, strawberry tongue, or diffuse redness of the oropharyngeal mucosa, (4). Rash (including redness of BCG inoculation scars), (5) Changes in extremities: rigid edema of limbs or erythema of palm or fingertips in the acute phase, and (6) non-purulent cervical region in the acute phase Lymph node swelling. The "symptoms or findings" here are, for example, an increase in transaminase level in the early stage of the disease, an increase in the urinary leukocyte count in an infant, an increase in brain natriuretic hormone (hereinafter also referred to as "BNP") or NT-proBNP, and cardiac ultrasound. One or more symptoms or findings selected from the group consisting of mitral regurgitation or pericardial effusion, pericardial swelling, and hypoalbuminemia or hyponatriemia on examination.

The number of subjects is, for example, 30 or more, preferably 50 or more, more preferably 100 or more, and the larger the number is, the more the number of subjects is, from the viewpoint of enriching the learning data and improving the prediction accuracy. More preferred. From the same point of view, the subjects included IVIG refractory subjects who did not develop CAL after receiving second-line or third-line treatment, and IVIG-impaired who received second-line or third-line treatment but did not develop CAL. It is preferable to include as many subjects as possible not only for the response subjects but also for the subjects with IVIG reaction who did not develop CAL due to the primary treatment mainly by IVIG therapy.

From the viewpoint that the clinical information of each subject in the generation method S1a is clinical data that tends to have a direct or indirect causal relationship with the occurrence of CAL, gender, KD acute phase, and anti-inflammatory therapy. Age at the time of not receiving (hereinafter also referred to as "pre-first-line treatment"), coronary artery diameter at the time before the first-line treatment, systemic vasculitis marker test value at the time before the first-line treatment, high cytokines at the time before the first-line treatment Select from a group consisting of blood marker test values, IVIG refractory prediction score score before first-line treatment, number of times anti-inflammatory therapy was received before first-line treatment, and data that indirectly indicate any of these. 3 or more types of subject clinical data, for example, 4 or more types may be used, preferably 5 or more types, and more preferably 6 or more types of subject clinical data are included. The "indirectly shown data" here is data that can be roughly estimated by a person skilled in the art, or data that can be estimated by performing some conversion. For example, the age of each subject (for example, 1.5 years) can be estimated to be the age of the moon (for example, 18 months) by multiplying the numerical value by 12, and therefore corresponds to the data indirectly indicating the age of the moon. In this way, the data that correlates with the target data can be said to be a kind of data that indirectly indicates the target data.

From the viewpoint of improving the prediction accuracy, the time point before the first-line treatment described above is, for example, substantially within 7 days from the day of KD onset, with the day of onset of KD (the day of onset of KD) as the 0th day, and the first-line treatment is performed. It is a time when I have not received it. As used herein, the term "substantially" means that an exception is allowed as long as it does not interfere with the object or action of the present invention in terms of content or essence. For example, if a child is taken to a guardian for the first time a few days after the onset of KD, the doctor cannot determine the exact date of onset of KD unless the guardian remembers the date of onset of symptoms. However, for example, a time when a doctor can determine that "probably within 7 days from the onset of KD" from the symptoms or findings of a child corresponds to "substantially within 7 days from the onset of KD". The main symptoms of KD did not appear sufficiently at the first visit, and KD was not diagnosed. The date of onset of the symptoms that caused the diagnosis.

The sex given as an example of the subject clinical data is data indicating a biological sex that can be determined from, for example, the physical structure at birth or the sex chromosome. For example, the test value of the blood sex hormone concentration in each subject corresponds to the data indirectly indicating the sex because a person skilled in the art can roughly estimate the sex of each subject based on this test value. ..

The coronary artery diameter given as an example of the subject clinical data is, for example, the average value, the median value, the maximum value, or any of these obtained by measuring the coronary artery diameter (coronary artery inner diameter) at multiple points where diameter expansion is suspected. Although the data may be indirectly shown, the maximum value obtained by measuring the diameter of the coronary artery is preferable from the viewpoint of improving the prediction accuracy. As data indirectly indicating the coronary artery diameter, for example, a measured value of the coronary artery radius can be mentioned. Conventionally, in the field of daily medical care in pediatrics, the doctor in charge (pediatrician) has measured the coronary artery diameter by utilizing the zoom function of the inspection device by the tomographic echocardiography method when diagnosing KD. The coronary artery diameter data is preferably measured by tomographic echocardiography from the viewpoint of convenience for the doctor in charge at the site, or from the viewpoint of more accurate measurement, coronary angiography by cardiac catheterization, contrast CT examination, etc. Alternatively, it is also preferable that it is a measured value in an inspection using MRI.

From the viewpoint of the site where CAL is likely to occur, the sites of the coronary artery whose diameter is measured are the left main coronary trunk artery (hereinafter referred to as "LMT") and the left anterior descending coronary artery (hereinafter referred to as "LMT"). It is preferably one or more sites selected from the group consisting of a proximal region (referred to as "LAD"), a left coronary artery circumflex artery, and a proximal region of the right coronary artery (hereinafter referred to as "RCA"). The data of the coronary artery diameter is preferably the maximum value data among all the sites of the coronary arteries listed here or one or more of the sites of the coronary arteries, from the viewpoint of improving the prediction accuracy later. Alternatively, from the viewpoint of making it possible to predict the risk of CAL for each site of the coronary artery, data on the maximum value of the coronary artery diameter for each site mentioned here may be acquired for each subject.

For example, the average value of the coronary artery diameter in the hypothetical population (estimated average value of the coronary artery diameter in Japanese children described in the research paper) becomes 0 by the following formula 1 for the measured value data of the coronary artery diameter, and the hypothetical population. It may be converted into a Z score so that the standard deviation (SD) of the coronary artery diameter in is 1.0. When the measured value data of the coronary artery diameter is acquired for each site of the coronary artery, it may be converted into a Z score for each site.

As a child grows older and grows in size and increases the body surface area calculated from height and weight, the diameter of the coronary artery in a healthy state also increases. Therefore, for example, even if the same coronary artery diameter is measured in a 1-year-old child and a 5-year-old child, it may be determined that the 1-year-old child has CAL and that the 5-year-old child does not have CAL. For subjects of different ages and body surface areas, the measured value of the coronary artery diameter is calculated in the following formula 2 from the viewpoint that a corrected Z score can be obtained in consideration of the factor that the coronary artery diameter expands with the growth of the physique. It is preferable to convert to a Z score by the LMS method shown. In the LMS method, the distribution of the measured values of the coronary artery diameter for each body surface area is normally distributed one by one, and is distributed by the skewness (λ, L), the median (M), and the coefficient of variation (S) (SD / M). It is a statistical method to change.

From the viewpoint of the number of samples and research method, the inner diameter of the coronary artery created as a research result of the Z score project (for example, Non-Patent Document 8) or its 2nd stage from the viewpoint of being widely used in the daily clinical practice of pediatrics with high reliability. It is more preferable to convert the coronary artery diameter in each subject into a Z score based on the calculation formula of the LMS method considering the standard curve of. At the time of filing the application, for example, a website (https://kwsd.info/) where the pediatric coronary artery inner diameter Z Score calculation application can be used, or a website where the Coronary Z Score Calculator can be downloaded (for example, http: //). Go to raise.umin.jp/zsp2/data/zsp_calculator_version4_full.xlsx or http://raise.umin.jp/zsp2/data/zsp_calculator_version4_light.xlsx) and access each subject's gender, height, weight, and coronary artery parts. By inputting the data of the measured values of, the Z score of each part of the coronary artery can be calculated. From the same point of view, if version 5 or later of the Coronary Z Score Calculator is published after the application of this application, or if the formula for calculating the coronary artery diameter created as a research result of developing the Z score project 2nd stage is published, etc. , It is preferable to convert the coronary artery diameter of each subject into a Z score by the latest calculation method recommended by the Kawasaki Disease Society of Japan.

The systemic vasculitis marker test value given as an example of the subject clinical data is a test value that is an index of systemic vasculitis and associated tissue (for example, vascular endothelium) destruction. As systemic vasculitis markers, for example, erythrocyte sedimentation rate, serum complement titer, or blood or serum, pentraxin superfamily concentration, immune complex concentration, procalcitonin concentration, fibrin degradation product concentration, and these. One of them is a marker test value that can be substituted. Examples of the pentraxin superfamily include pentraxin-3, amyloid P, C-reactive protein (hereinafter, also referred to as “CRP”) and the like. From the viewpoint of easy and widespread use in pediatric clinical practice and easy clinical application, the systemic vasculitis marker test value is preferably the blood cell sedimentation rate or the CRP concentration in blood or serum. , The serum CRP concentration is more preferable.

The hypercytokineemia marker test value given as an example of the subject clinical data is a test value that is an index of so-called cytokine storm or cytokine release syndrome. Examples of the hypercytokineemia marker test value include a test value for the concentration of anti-inflammatory cytokines such as TNF-α, interferon, IL-1β, MCP-1, or IL-6 in blood or serum. From the viewpoint of improving the prediction accuracy, the IL-6 concentration in blood or serum or the β2 microglobulin concentration in urine is preferable as the test value for the hypercytokine storm marker. Since the hypercytosanemia marker test value has been easily and widely used in general tests in the field of pediatric clinical practice in general hospitals (secondary medical institutions that can be hospitalized and treated), the test cost is low and the test value is low. The ratio of β2 microglobulin to creatinine in urine from the viewpoint of clinical application because it is easy to obtain test results in general hospitals, for example, within 3 days from the time a patient is diagnosed with KD and hospitalized to the first-line treatment. Hereinafter, it is more preferably referred to as “urinary β2MG / Cr”).

As an example of the subject clinical data, the score of the IVIG refractory prediction score includes, for example, the score data calculated by the risk score for predicting the IVIG refractory case such as the Iwasa score or the Harada score (non-patentable). See Document 2). The scores of the IVIG refractory prediction score are the Kobayashi score (Non-Patent Documents 1 and 3 and Table 1 below) and the Egami score (Non-Patent Documents 1 and 4) from the viewpoint of being utilized in the field of daily medical care in pediatrics. The score data is preferably calculated from the following Table 2) or the Sano score (Non-Patent Documents 1 and 5 and the following Table 3), and more preferably the score data calculated from the Sano score. .. As data that indirectly indicates the score of the IVIG refractory prediction score, there is a determination result of whether or not the risk corresponding to IVIG refractory is high risk based on this score. For example, a subject with a Sano score of 2 or more before the first-line treatment is determined to be at high risk of being immune to IVIG after the first-line treatment (high risk), or a subject with less than 2 points is the primary. It is determined that the risk of refractory to IVIG after treatment is low (low risk). Regarding Tables 1 to 3 below, the "disease day" is a unit indicating how many days from the onset date of KD, with the day on which KD develops (KD onset date) as the 0th day. "Sensitivity" is the ratio of those having a disease to be detected by the test that are correctly determined to be positive by the test. "Specificity" is the ratio of those who do not have the disease to be detected by the test and those who are correctly judged to be negative by the test.

The anti-inflammatory therapy herein is a therapeutic method for which an anti-inflammatory effect has been reported in guidelines for medical treatment of diseases (for example, Non-Patent Document 1) or research papers (so-called evidence). As an example of subject clinical data, the number of times the patient received anti-inflammatory therapy after the time before the first-line treatment is the data on the number of times the patient received the anti-inflammatory therapy from the first-line treatment to the time when the presence or absence of CAL was determined. Is. As data indirectly indicating the number of times of receiving anti-inflammatory therapy, for example, data on whether or not anti-inflammatory therapy is performed can be mentioned. As anti-inflammatory therapy, for example, as illustrated in FIG. 31, administration of acetylsalicylic acid (ASA) and / or a salt thereof, IVIG administration, IVMP administration, prednisolone (hereinafter also referred to as “PSL”) administration, infliximab ( 1 selected from the group consisting of Infliximab (hereinafter also referred to as "IFX") administration, ulinastatin (hereinafter also referred to as "UTI") administration, cyclosporine A (hereinafter also referred to as "CsA") administration, and plasma exchange 1 There are more than one type of treatment method (see Non-Patent Document 1). ASA is also known as aspirin®. The salt of ASA may be any pharmacologically acceptable salt, preferably a sodium salt or a potassium salt. Anti-inflammatory therapy here is selected from the group consisting of IVMP, CsA, and plasmapheresis from the perspective of making it predictable whether CAL will occur after first-line treatment in patients predicted to be refractory to IVIG. It is more preferable that one or more therapeutic methods are used, and it is further preferable that the one or more therapeutic methods are used in combination with IVIG administration and / or ASA administration.

In each of ASA and / or its salt administration, IVIG administration, IVMP administration, IFX administration, UTI administration, and CsA administration, the number of times the subject received anti-inflammatory therapy because the subject is usually administered intravascularly at the time of administration. Easy to count. In plasma exchange, the subject is usually given a predetermined amount of blood product intravascularly at the time of implementation, so that it is easy to count the number of times of receiving anti-inflammatory therapy. On the other hand, in anti-inflammatory therapy in which an anti-inflammatory drug is orally administered to a patient, such as PSL, when the anti-inflammatory drug is orally administered to the patient for a certain period of time, the patient is counted as having received one anti-inflammatory therapy. Then, once the medication is discontinued, the patient is counted as having received a second anti-inflammatory therapy when the anti-inflammatory therapy is resumed and the anti-inflammatory drug is orally administered again for a certain period of time. In addition, from the viewpoint of making it possible to verify how the risk of CAL development changes depending on which anti-inflammatory therapy the doctor in charge selects in KD acute care, each subject receives multiple types of anti-inflammatory therapy. When receiving, in the data of the number of times the subject received the anti-inflammatory therapy, a plurality of items may be provided for each type of the anti-inflammatory therapy received by the subject. For example, subject clinical data may be obtained for a subject, such as once receiving IVIG, once receiving ASA, twice receiving IVMP, and so on.

The clinical information of each subject included "gender, age before first-line treatment, coronary artery diameter before first-line treatment, systemic vasculitis marker test values before first-line treatment, and high before first-line treatment. Three or more subjects selected from the group consisting of cytokineemia marker test values, IVIG refractory prediction score scores, number of times anti-inflammatory therapy was received before and after first-line treatment, and data indirectly indicating these. In addition to "data", clinical data that can be obtained from each subject during the acute phase of KD and that has a direct or indirect causal relationship with the occurrence of CAL in the acute phase of KD. , Since it can contribute to the improvement of prediction accuracy, it may be acquired in the form of being included in clinical information. Examples of such clinical data include heart failure marker test values at the time before the first-line treatment, data indirectly indicating the test values, and the like. From heart failure marker test values, for example, from cardiac troponin, atrial natriuretic peptide (ANP), N-terminal (NT) -proANP, proANP, BNP, proBNP, NT-proBNP, troponin, and blood urea nitrogen (BUN). Examples include test values for one or more biomarkers selected from the group. The systemic vasculitis markers (eg, CRP) and hypercytokinemia markers already mentioned herein are not included in the heart failure markers. The heart failure markers listed here are the test values of special tests that are not familiar in the field of pediatric clinical practice, and they require outsourcing at general hospitals (secondary medical institutions), which is expensive and the subject is KD. From the point of view of facilitating the clinical application of the production method S1a, since it is often too late to obtain the test results from the subcontractor within, for example, within 3 days from the diagnosis of hospitalization to receiving the first-line treatment, each subject It is preferable that the clinical information does not include heart failure marker test values or data that indirectly indicate this.

The judgment result regarding the presence or absence of CAL was determined by a doctor inspecting whether or not CAL was formed in each subject within a predetermined period after the first treatment (for example, after the first treatment and within 90 days from the onset of KD). The resulting data. From the viewpoint that CAL is often formed within 30 days from the onset of KD in cases leading to the onset of CAL, "within a predetermined period after the first-line treatment" here is after the first-line treatment and within 30 days from the onset of KD. Is preferable. The determination result is preferably the determination result at the site corresponding to the above-mentioned "coronary artery diameter at the time before the first-line treatment". For example, in each subject, when the diameter was measured at multiple coronary arteries before the first-line treatment and only one maximum value data was obtained, the judgment result of the presence or absence of CAL was also found among several sites. It is good that the result is determined by acquiring only one maximum diameter value in. Alternatively, in each subject, when the maximum value of the diameter is obtained for each site of the coronary artery before the first-line treatment, the maximum value of the diameter is also obtained for each site in the judgment result of the presence or absence of CAL. The result is good.

The judgment result regarding the presence or absence of CAL is, for example, if the body surface area of each subject is about the same, and if the maximum value of the coronary artery diameter at the time of judgment is equal to or more than a predetermined threshold value (for example, 2 mm), CAL is positive (with CAL). The data may be determined to be CAL negative (no CAL occurrence) when the maximum value is less than a predetermined threshold value. From the viewpoint of determining the presence or absence of CAL in consideration of the factor that the coronary artery diameter expands with the growth of the physique, the determination result is based on the above-mentioned LMS method or the calculation formula which is the research result such as the Z score project. This is the result of converting the coronary artery diameter at the time of judgment by the subject into a Z score, and determining that the Z score is CAL positive when the value is equal to or more than the predetermined value, or CAL negative when the value is less than the predetermined value. Is preferable. From the viewpoint of predicting the risk of developing even a CAL that does not lead to a coronary aneurysm (CAA), for example, if the Z score is 1.5 SD or more, it may be determined to be CAL positive, and the Z score is 2.0 SD as in the United States. If it is the above, it is also preferable that it is determined to be CAL positive. From the viewpoint of further improving the prediction accuracy, it is desirable to determine CAL positive when the Z score is preferably 2.5 SD or more, and more preferably when the Z score is 3.0 SD or more. When predicting the risk of CAA occurrence, the criterion is preferably 5.0 SD or more. When predicting the risk of developing a giant aneurysm in CAA, the criterion is preferably 10.0 SD or more.

In the pre-learning normalization step S3a, in order to efficiently perform covariance structure analysis and machine learning later, three or more types of subject clinical data included in the clinical information obtained in the previous step S2a and the presence or absence of CAL have occurred. The judgment results related to the above are normalized so that they can be easily calculated by the information processing device. Normalization is the process of transforming data to make it easier to use based on certain rules. Since the post-normalized data can generally reproduce the pre-normalized data, it corresponds to the data indirectly indicating the pre-normalized data. For example, it is preferable to normalize non-numeric data such as gender by, for example, converting males to 1 and females to 0. For example, it is preferable that the numerical data such as the age of the moon is normalized so as to be included in the range of 0 or more and 1.0 or less by the following mathematical formula 3.

When machine learning is performed later, from the viewpoint that it is easy to calculate if the input variable does not include an approximate value of 0 or 1.0, three or more types of subject clinical data included in the clinical information of each subject are obtained as 0. It is more preferable to normalize it so that it is included in the range of 10 or more and 0.90 or less. For example, non-numerical data such as gender may be normalized by converting males to 0.90 and females to 0.10. From the same viewpoint, it is more preferable that the numerical data such as the age of the moon included in the clinical information is normalized so as to be included in the range of 0.10 or more and 0.90 or less by the following mathematical formula 4.

The determination result regarding the presence or absence of CAL in each subject is used in the later covariance structure analysis step, but is not particularly used in the subsequent machine learning. Therefore, for example, it is preferable to normalize the determination result so that either one of CAL positive and CAL negative is converted to 0, and the remaining one is converted to 1.0. When the determination result of the presence or absence of CAL generation is obtained for each site of the coronary artery in each subject, it is preferable to normalize to 0 or 1.0 for each site depending on whether the site is CAL positive or CAL negative.

If it is possible to suddenly obtain the normalized data of three or more types of subject clinical data included in the clinical information of each subject and the determination result of the presence or absence of CAL in each subject in the previous step S2a. , The normalization step S3a before learning can be omitted. For example, if the data of the determination result regarding the presence / absence of CAL is already 0 or 1.0 at the time of acquisition in the previous step S2a, it is not particularly necessary to further normalize this data.

Covariance structure analysis is one of the statistical analysis methods that can examine the relationship between multiple variables (multiple types of data). In the covariance structure analysis step S4a, a covariance structure analysis is performed using an information processing device in order to calculate an SS calculated value having a high correlation with the occurrence of CAL for each subject. For this purpose, four or more kinds of observation variables are provided for each subject. The observed variable is the data actually observed (measured) or the normalized data thereof. Here, the breakdown of the four or more types of observation variables is a combination of the clinical data of three or more types of subjects included in the clinical information of each subject and the "judgment result regarding the presence or absence of CAL occurrence" in each subject. The data used as each observed variable in the covariance structure analysis may be numerical data, and may not necessarily be numerical data in the range of 0 or more and 1.0 or less. In addition, in order to perform covariance structure analysis, latent variables that are assumed to have a direct causal relationship with the observation variables related to the "judgment result regarding the presence or absence of CAL occurrence" in each subject are provided. Latent variables are hypothetical data that have not actually been observed. After providing four or more types of observed variables and latent variables, the information processing device is made to execute arithmetic processing so as to perform covariance structure analysis, and the observed variables related to the "judgment result regarding the presence or absence of CAL occurrence" are Then, the factor score data is calculated for the latent variables for which a direct statistically significant causal relationship is found. Significance here is, for example, p <0.05, preferably p <0.001. The numerical data of the factor score calculated here, or the data obtained by normalizing the numerical data itself within a reproducible range, is used as the SS calculated value regarding the risk of CAL occurrence in each subject.

The information processing apparatus used for the covariance structure analysis is not particularly limited as long as the arithmetic processing of the covariance structure analysis is possible. For example, in a commercially available personal computer, a covariance structure using statistical analysis software such as AMOS (Analysis of Moment Structure), SAS (Statistical Analysis System), LISREL (Linear Structure Relations), or EQS (Structural Equation Modeling Software). Can perform analysis. From the viewpoint of improving the prediction accuracy later, it is preferable to perform the average covariance structure analysis in the covariance structure analysis step S4a. The mean covariance structure analysis introduces an intercept, which is the difference between the mean of one variable and the mean of another variable, into the relationship between the variables and includes this intercept in the path coefficients and variances to be estimated. It is an analysis. From the viewpoint that statistical analysis can be performed intuitively using a path model, it is preferable to perform covariance structure analysis by structural equation modeling (hereinafter also referred to as “SEM”), and average by SEM. It is more preferable to perform covariance structure analysis. For example, if AMOS is used, the average covariance structure analysis can be performed by SEM.

As an SEM path model, for example, an exploratory factor analysis model in which one latent variable has a direct causal relationship with a plurality of observed variables as illustrated in FIG. 2, or each observation as illustrated in FIG. A confirmatory factor analysis model in which two or more latent variables have a direct causal relationship to a variable and a correlation between the latent variables, or an exploratory factor analysis model as illustrated in FIG. 4 is included. Examples include a secondary factor model consisting of two, or a model modified so that a causal relationship recognized as significant between at least some of the observed variables is defined in these models. In the path model, the unidirectional arrow represents a causal relationship, and it is assumed that the variable at the source of the arrow affects the variable at the tip of the arrow. Although omitted in FIGS. 2 to 4, each unidirectional arrow in the path model is given a path coefficient, and the path coefficient given to each observed variable from the latent variable to the unidirectional arrow means any of them. It is assumed that there is something (statistically significant). That is, in the path model, the causal relationship from the latent variable to each observed variable is significant. Also, in the path model, the double-headed arrow represents the correlation. Although omitted in FIGS. 2 to 4, each of the bidirectional arrows in the path model is given a correlation coefficient, so that the magnitude of each correlation can be discriminated. From the viewpoint of easily solving the discrimination problem (whether or not the relational expression that defines the relationship between the observed variable and the latent variable determined by the path model has a mathematical solution), the path model is an exploratory factor shown in FIG. It is preferable to use an analytical model or a model modified so that the existence of a causal relationship between some observed variables is assumed in this model.

In the covariance structure analysis step S4a shown in FIG. 1, in order to evaluate the validity of the calculation result of the covariance structure analysis and whether the path model fits the data, for example, it is referred to as chi-square (hereinafter referred to as “χ ² ”). .) Statistics, Root Mean square Residual (hereinafter referred to as "RMR"), Goodness of Fit Index (hereinafter referred to as "GFI"), Adjusted Goodness of Fit Index: AGFI), Akaike's Information Criterion (hereinafter referred to as "AIC"), or Root Mean Square Error of Approximation (hereinafter referred to as "RMSEA") may be used. In general, for example, ^{the hypothesis that "the path model fits the data" is not rejected from the χ 2} statistic, and the GFI or AGFI values are greater than or equal to a predetermined threshold (eg 0.9), or The more conditions such as the RMSEA value being less than a predetermined threshold value (for example, 0.080) are satisfied, the more reliable the calculation result and path model of the covariance structure analysis are. In order to evaluate the reliability of the path coefficient included in the path model, for example, a t-test or a Wald test may be performed. For example, when covariance structure analysis is performed by SEM using AMOS, the reliability of the calculation results and path model of covariance structure analysis can be examined by the indexes and methods exemplified here.

In the covariance structure analysis step S4a, when the RMSEA value in the created path model is, for example, less than 0.080, preferably less than 0.050, it is used for covariance structure analysis as a reliable statistical model. Is good. If the RMSEA value is 0.080 or higher, the path model cannot be trusted as it is. Therefore, reconsider the combination of clinical data of three or more subjects used as observation variables so that the RMSEA value is less than 0.080. It is better to modify the path model. From the same viewpoint, it is preferable that the combination of three or more types of subject clinical data when the RMSEA value becomes 0.080 or more is not used as a combination of three or more types of input variables in later machine learning. For example, when the inventor of the present application tried, for each subject, the total amount of bilbilin before the first treatment, the number of leukocytes before the first treatment, and the presence or absence of KD relapse after the first treatment were observed. Even if the mean covariance structure analysis was performed by SEM using it as a variable, no significant causal relationship was found with the latent variable, so it is preferable not to use the data as an observation variable, and it is preferable to use it as an input variable in later machine learning. It is preferable not to use it. Step S2a, S3a, and S4 a combination of can function as a step S5a to produce a machine learning data.

In the learning step S6a, three or more types of subject clinical data included in the clinical information of each subject are used as input variables (independent variables), and the SS calculated value obtained in the previous covariance structure analysis step S4a is used for each subject. As an output variable (dependent variable), let ANN machine learn the relationship between the input variable and the output variable. ANN is an information processing system based on a computational technique that imitates neurological processing in the human brain, and is used for modeling a system in which input variables and output variables exist. Examples of the ANN that can be utilized in the learning step S6a include a multi-layer perceptron in which perceptrons are connected in layers. As an example of the multilayer perceptron in FIG. 5, a three-layer type ANN1 having an input layer 2, an intermediate layer 5 of one layer, and an output layer 8 can be mentioned. Each of these layers (2, 5, 8) is provided with some artificial neurons (hereinafter referred to as "AN"). AN is also called a node. Each AN (3a to 3d) provided in the input layer 2 is connected to each AN (6a to 6c) provided in the intermediate layer 5 via the network 4. Each AN (6a to 6c) provided in the intermediate layer 5 is connected to the AN9 provided in the output layer 8 via the network 7.

When the ANN1 is machine-learned, for example, one of the subject clinical data to be used as an input variable is input to each AN (3a to 3d) provided in the input layer 2. Further, for example, the SS calculated value used as an output variable is input to AN9 provided in the output layer 8. Then, when the information processing apparatus is made to perform an operation, the input variables input to each AN (3a to 3d) provided in the input layer 2 are output toward the intermediate layer 5. Each network (4, 7) has a weighted value _Wij . In (from 6a 6c) AN provided in the intermediate layer 5 and provided at the output layer 8 AN9, as exemplified by the following equation 5, the product-sum of the weighting values W _ij and the input value S _i from the previous layer The calculation and the variable transformation using the sigmoid function are performed, and the calculation formula is output in the output layer 8. Then, a value calculated by this formula, correct the error between the (output variable which is input to the output layer 8 in advance) is calculated, so the error is zero, and the threshold h _i The weight _Wij is modified. In the machine-learned ANN, the relationship that exists between the input variable (independent variable) and the output variable (dependent variable) has been found. Therefore, in the learning step S6a, the trained model can be generated.

P_i ＡＮが発火する確率
W_ij 前層のＡＮと次層のＡＮ間の重み付け値（シナプス結合計数）
S_i 前層のＡＮからの入力値
h_i しきい値
T シグモイド関数の傾き

Probability of firing P _{i AN}
W _ij Weighted value between AN in the previous layer and AN in the next layer (synaptic bond count)
Input value from AN in the front layer of S _i
h _i threshold
Slope of T sigmoid function

Not limited to ANN1 illustrated in FIG. 5, in the learning step S6a shown in FIG. 1, for example, an ANN having an input layer, one or more intermediate layers, and an output layer may be machine-learned. Adopting only a simple perceptron consisting of two layers, an input layer and an intermediate layer, as the machine learning ANN should be avoided because it causes a significant deterioration in prediction accuracy. From the viewpoint of avoiding overfitting, the number of intermediate layers in ANN may be 4 layers or less or 3 layers or less, preferably 2 layers or less. Unless contrary to the object of the present invention, machine learning may be performed in a state where two or more hierarchical ANNs having an input layer, one or more intermediate layers, and an output layer are combined. The operation function used in ANN may be, for example, a radial basis function or a snake function, but from the viewpoint of high reliability, a sigmoid function such as the above-mentioned equation 5 is preferable. The machine learning method may be, for example, a conjugated gradient descent method, a quasi-Newton method, a Levenberg-Marquardt method, etc. It is preferable to use the error propagation method or the error back propagation method, and it is more preferable to use the error propagation method and the repair learning method together from the viewpoint of shortening the learning time. From the viewpoint of easily extracting the rules hidden in the learning data, it is more preferable to use the error propagation method and the growth suppression learning method together.

From the viewpoint of improving the prediction accuracy later, for example, 4 or more types of subject clinical data included in the clinical information of each subject, preferably 5 or more types of subject clinical data, more preferably 6 as input variables in machine learning. It is more preferred to use subject clinical data of more than one species. From the viewpoint of avoiding deterioration of learning efficiency, the subject clinical data used as an input variable is, for example, 20 types or less, preferably 15 types or less, and more preferably 10 types or less. In order to obtain a trained model with even higher prediction accuracy, it was found by having an information processing device repeatedly (for example, 50 times or more) execute arithmetic processing from input variables and output variables to error correction. It is preferable to generate a plurality of trained models in which the relationship between the input variable and the output variable is different.

Sorting step S7a, when to produce a plurality of the learned model in the previous learning step S 6 a, in order to further improve the prediction accuracy, to verify the height of the prediction accuracy for each trained model, relatively Select a trained model with high prediction accuracy. For example, using a commercially available statistical analysis software performs simple cross-validation or K-fold cross-validation for each trained model (e.g. five-fold cross-validation), to calculate the coefficient of determination R ² for each learned model , preferably selected the most R ² value is one of the learned model large. R ² value in selecting the step S7a is an index input variables representing the availability described how the output variables can not be explained as a value close to 0, it means that can be described as the value close to 1.0. From the viewpoint of selecting trained models with a simple structure and small output error, AIC and Schwartz's Bayesian information criterion (hereinafter referred to as "BIC") are verified for each trained model, and the AIC value is verified. Alternatively, it is preferable to exclude the trained model showing a relatively high BIC value from the selection candidates. AIC and BIC can be verified using commercially available statistical software. A trained model with lower AIC and BIC values has higher prediction accuracy and is preferable as a statistical model. If only one trained model is generated in the previous learning step S6a, or if the later prediction accuracy is sufficient to some extent, the selection step S7a may be omitted.

According to the generation method S1a described above, the sex, age, coronary artery diameter, systemic vasculitis marker test value, hypercytokinemia marker test value, and IVIG refractory given as examples of clinical data of three or more subjects. All the scores of the predicted scores are information that can be obtained by a doctor in charge of KD acute care (for example, a pediatrician working in a hospital) by interview or examination before the first treatment. The coronary artery diameter is information that can be obtained, for example, when a doctor in charge diagnoses KD by using a test method or test device such as tomographic echocardiography, which has been widely used in the field of daily medical care in pediatrics. See Patent Document 1 and Non-Patent Document 2). Since each subject has already received the first-line treatment and the presence or absence of CAL has been determined, the number of anti-inflammatory treatments received before and after the first-line treatment and the determination result of the presence or absence of CAL are also KD acute. Information available to the doctor in charge of acute care in the pediatrics department of the hospital. The SS calculated value regarding the risk of CAL occurrence in each subject can be calculated by covariance structure analysis from the clinical data of three or more kinds of subjects and the determination result of the presence or absence of CAL occurrence. Therefore, according to the generation method S1a, it is possible to generate a trained model by using information such as test values that are familiar in the field of daily medical care in conventional pediatrics. By utilizing this trained model as described below, it is possible to predict with as high accuracy as possible whether or not CAL will occur in the KD acute phase patient who is the prediction target before the first-line treatment.

[CAL occurrence risk prediction method in KD]
The prediction method S10a according to the embodiment of the present invention shown in FIG. 6 is a prediction method for predicting the risk of CAL occurrence in a KD acute phase patient by using an information processing apparatus having a learned model. The prediction method S10a includes a clinical information acquisition step S2a of the subjects, a normalization step S3a before learning, a covariance structure analysis step S4a, a learning step S6a, a selection step S7a, and a patient clinical information acquisition step S12a. , The pre-prediction normalization step S13a and the prediction step S14a may be included. Steps S2a to S7a may be performed in the same manner as in the generation method S1a described above.

In the patient clinical information acquisition step S12a, for example, for a KD acute phase patient who wants to predict whether or not CAL will occur after the first-line treatment, clinical information of the patient available before the first-line treatment is obtained. The patient clinical information obtained here includes gender, age before first-line treatment, coronary artery diameter before first-line treatment, systemic vasculitis marker test values before first-line treatment, and high levels before first-line treatment. Indirectly, the cytokineemia marker test value, the score of the IVIG refractory prediction score before the first-line treatment, the number of scheduled anti-inflammatory therapy after the time before the first-line treatment (after the first-line treatment), and any of these. 3 or more patient clinical data selected from the group consisting of the data shown in. From the viewpoint of improving the prediction accuracy, it is desirable to acquire patient clinical data of, for example, 4 or more, preferably 5 or more, and more preferably 6 or more of the clinical information of the patients exemplified here. The step of diagnosis performed by a doctor in the process until this clinical information is generated is not included in the present invention or step S12a.

From the viewpoint of improving the prediction accuracy, the clinical data of 3 or more types of patients obtained in the clinical information acquisition step S12a of the patient is almost the same as the clinical data of 3 or more types of each subject acquired in the previous step S2a. It is preferable that the data is. It can be said that the “scheduled number of anti-inflammatory therapy”, which is an example of patient clinical data, is almost the same as the “number of anti-inflammatory therapy”, which is an example of subject clinical data. Generally, the doctor in charge of acute care for KD diagnoses a patient at least twice from the diagnosis of KD to the first-line treatment. For example, in step S12a, if the patient clinical data of the planned number of times of anti-inflammatory therapy is to be performed is acquired based on the treatment plan, the prescription plan, or the draft thereof already prepared by the doctor in the diagnosis before the final diagnosis before the first treatment. good. The process in which a doctor prepares a treatment plan, a prescription plan, or a draft thereof is not included in the present invention or step S12a. In addition, if the so-called evidence regarding KD acute care recommends a specific number of anti-inflammatory therapies based on the test values before the first-line treatment of KD acute care patients, it is recommended. It may be set so that the "scheduled number of times of anti-inflammatory therapy" is automatically determined according to the standard set without the judgment of a doctor. Other than that, the patient's clinical information acquisition step S12a is the same as that already described in the previous step S2a. However, unlike the previous step S2a, in the patient clinical information acquisition step S12a, it is not necessary to acquire the determination result of the presence or absence of CAL in the patient.

In the pre-prediction normalization step S13a, in order to make an efficient prediction, three or more types of patient clinical data obtained in the previous patient clinical information acquisition step S12a are normalized so that they can be easily calculated by the information processing apparatus. For the same reason, the pre-prediction normalization step S13a normalizes the clinical data of three or more patients in much the same way as the previous pre-learning normalization step S3a normalized the clinical data of three or more subjects. It is preferable to do. Unlike the previous pre-learning normalization step S3a, the pre-prediction normalization step S13a does not particularly require normalization of the determination result regarding the presence or absence of CAL in the patient. If three or more types of patient clinical data suddenly normalized in the previous patient clinical information acquisition step S12a can be acquired, the pre-prediction normalization step S13a can be omitted.

In the prediction step S14a, three or more types of patient clinical data included in the clinical information of the KD acute phase patient are input to the information processing apparatus having the trained model, and the SS regarding the CAL occurrence risk is predicted in the KD acute phase patient. Execute the process to output the value. For example, one type of patient clinical data or its normalized data (input variable) is input for each AN provided in the input layer of the trained model, and an information processing device having the trained model is made to execute an operation, and the output layer is used. In the AN provided in the KD acute phase patient, the SS predicted value (output variable) regarding the risk of CAL occurrence is output.

The SS prediction value in the KD acute phase patient output in the prediction step S14a is numerical data that is an index for predicting whether or not this patient will develop CAL after the first-line treatment. If the SS predicted value is equal to or more than the predetermined cutoff value, the predicted result is CAL positive (to lead to the occurrence of CAL after the first-line treatment), or if it is less than the predetermined cutoff value, the predicted result is CAL negative. (CAL does not occur after first-line treatment). However, the predetermined value of this cutoff value fluctuates according to the reference value when the determination result of the presence or absence of CAL occurrence in each subject is determined. For example, regarding the coronary artery diameter after receiving at least the first-line treatment in each subject, if the Z score value is 3.0 SD or more, it is determined to be CAL positive, or if it is 2.5 SD or more, it is determined to be CAL positive. The predetermined value of the cutoff value fluctuates according to the reference value of whether it is determined or whether it is determined to be CAL positive when it is 2.0 SD or more. The higher the reference value, the higher the predetermined cutoff value. Therefore, further, in the prediction step S14a, when the SS prediction value in the KD acute phase patient is equal to or higher than the predetermined cutoff value, a process of outputting a prediction result indicating that CAL occurs in this patient is executed. As such, it is preferable to operate an information processing device having a trained model. When it is made to function in this way, for example, even if the practitioner of the prediction method S10a is not accustomed to statistical processing, highly accurate prediction results can be easily obtained.

According to the prediction method S10a described above, the clinical data of three or more kinds of subjects included in the clinical information of each subject and the SS calculated value in each subject output in the covariance structure analysis step S4a are obtained. By using the trained model generated by the machine learning used, the SS predicted value, which is an index value of whether or not CAL occurs after the first treatment, is obtained for the KD acute phase patient who is the prediction target before the first treatment. It is possible to obtain high accuracy at that point. Therefore, for example, it is possible to support a doctor in charge of KD acute care so that it is easy to make a diagnosis for early determination of a treatment policy for suppressing the occurrence of CAL. That is, the prediction method S10a, for example, provides information called a prediction result (SS prediction value) as one of the information that can be used as a judgment material when making a diagnosis to determine an initial treatment policy for a doctor in charge of KD acute care. , It can be said that it is a suitable method to obtain before the first-line treatment. If the doctor in charge obtains a prediction result that KD acute phase patients will develop CAL after the first treatment before the first treatment, the doctor in charge will consider changing the treatment plan, prescription plan, or the draft of these plans that have already been prepared. In other words, it will be possible to consider changes in the type of anti-inflammatory therapy that was planned to be given as the first-line treatment and the number of times that it will be given. It is also possible to acquire the patient clinical data of the changed "planned number of times to receive anti-inflammatory therapy" based on the changed plan or the plan thereof, and to carry out the prediction method S10a again. It is also possible to change the plan or its proposal and implement the prediction method S10a again until the prediction result of CAL negative is obtained after the first-line treatment.

From the viewpoint of making it easily feasible, the prediction method S10b including the step S1b for preparing the pre-generated trained model is, as shown in FIG. 7, instead of the steps S2a to S7a in the prediction method S10a. preferable.

[Trained model for predicting CAL occurrence risk in KD]
The trained model according to the embodiment of the present invention is a trained model for predicting whether or not CAL occurs in a KD patient. As described above, this trained model is based on ANN1 whose clinical information is shown in FIG. 5 for subjects who have already been determined to have CAL after receiving at least first-line treatment in KD acute care. The weighted value _Wij shown in the above-mentioned equation 5 is machine-learned so that the input layer 2 is input to the input layer 2 and the output layer 8 outputs the calculated value for each subject for the SS related to the CAL occurrence risk. Further, when the clinical information of the KD acute phase patient is input to the input layer, this trained model performs a calculation based on the _{weighted value Wij on the input clinical information of the KD acute phase patient and outputs it.} The purpose is to make the information processing device function so as to output the SS predicted value in the KD acute phase patient from the layer.

A user who implements the prediction method S10b of FIG. 7 (for example, a doctor in charge of KD acute care, or a medical worker who assists the doctor in charge of KD) uses the above-mentioned trained model 67 as illustrated in FIG. You may use the information processing apparatus 60 which has. The information processing device 60 is an information processing device capable of various information processing and information transmission / reception, and is, for example, a personal computer, a server computer, or the like. The information processing device 60 includes, for example, a control unit 61, a display unit 62, an input unit 63, and a storage unit 65.

The control unit 61 is a processor that performs various arithmetic processing, control processing, and the like related to the information processing apparatus 60 by reading and executing the program 66 stored in the storage unit 65, and is, for example, one or a plurality of CPUs ( Central Processing Unit) etc. The storage unit 65 is a memory for storing various data. The memory includes a RAM (Random Access Memory) and a non-volatile memory. The RAM temporarily stores data necessary for the control unit 61 to execute arithmetic processing. The non-volatile memory includes, for example, a hard disk drive, and stores and holds a program 66 and a trained model 67 for the control unit 61 to execute processing. The non-volatile memory may further store the clinical information database 68 of the subjects, which was used as training data when the trained model 67 was generated. The display unit 62 is a display device such as a liquid crystal display or an organic EL display, and displays an image given by the control unit 61. The input unit 63 is an input interface such as a keyboard or a touch panel, and receives an operation input from a user. When the user inputs the clinical information of the KD acute phase patient to the information processing apparatus 60 via the input unit 63, the control unit 61 applies the input clinical information to the trained model 67 to generate CAL. It functions as a processing unit that executes a process of outputting a predicted value in a KD acute phase patient regarding SS related to the presence or absence. The output SS predicted value may be displayed as it is on the display unit 62 as numerical data. From a user-friendly point of view, the display unit 62 displays a prediction result indicating that CAL occurs after the first-line treatment in a KD acute phase patient when the output numerical data of the SS prediction value is equal to or more than a predetermined cutoff value. It is preferable that the information processing apparatus 60 functions so that a prediction result indicating that CAL does not occur when the cutoff value is less than a predetermined cutoff value is displayed.

The information processing device 60 preferably includes a communication unit 64. The communication unit 64 includes a processing circuit for performing processing related to communication, and transmits / receives information to / from at least one user terminal 71 via a communication network 70 such as the Internet or an intranet in a hospital. The communication unit 64 may include an antenna for this transmission / reception. The user terminal 71 is, for example, a smartphone owned by a pediatrician, a personal computer provided in a doctor's office of a hospital, or the like. When the user inputs clinical information of a KD acute phase patient to the user terminal 71, the input clinical information is transmitted from the user terminal 71 and received by the information processing device 60 via the communication network 70 and the communication unit 64. , The control unit 61 applies the received clinical information to the trained model 67 to execute a process of outputting the predicted value in the KD acute phase patient for the SS regarding the presence or absence of CAL, and the output SS predicted value is the communication unit. It is preferable that the information is transmitted from 64 and received by the user terminal 71 via the communication network 70, and the prediction result of the CAL occurrence risk based on the SS prediction value is displayed on the display of the user terminal 71.

[Learned model generation method for predicting the occurrence of PN with IgAV or proteinuria with PN]
In explaining the trained model generation method S1e according to another embodiment of the present invention with reference to FIG. 1, the common matters with the above-mentioned generation method S1a are appropriately omitted, and the method is different from the generation method S1a. I will mainly explain the matters. Generation method S1e is used to generate a trained model for predicting the risk of developing purpura nephritis (PN) in patients with IgA vasculitis (IgAV) or for predicting the risk of developing cases with proteinuria in PN. The subjects may include clinical information acquisition step S2e, pre-learning normalization step S3e, covariance structure analysis step S4e, learning step S6e, and selection step S7e.

In the clinical information acquisition step S2e of the subjects, in order to create learning data, for each of the subjects who have already been determined to have a case with proteinuria in PN or PN 30 days or more after the onset of IgAV. The clinical information of each subject and the judgment result regarding the presence or absence of the occurrence of a case with proteinuria in PN or PN in each subject are acquired. The process of diagnosis and treatment performed by a doctor in the process until the clinical information and the determination result are generated is not included in the present invention or step S2e. In IgA nephropathy, which is a type of chronic nephritis, the immune complex containing IgA is mainly deposited in the kidney, whereas in purpura nephritis (PN), which is an IgAV complication, the immune complex containing IgA is present. The difference is that it can deposit on systemic tissues, including the kidneys.

From the viewpoint of making it possible to predict the risk of occurrence of various cases of PN, it is preferable that the subjects of the generation method S1e include a large number of subjects having different degrees of aggravation due to PN after the onset of IgAV. Generally, PN is accompanied by hematuria, and the higher the amount of urinary protein, the more likely it is to become severe. If the amount of urinary protein is ^{less than 0.5 g / 1.73 m 2} / day or the protein / creatinine (Cr) ratio in early morning urine is less than 0.5, PN is diagnosed as a case with mild proteinuria. If the amount of urinary protein is 0.5 g / 1.73 m ² / day or more and 1.0 g / 1.73 m ² / day or less, or if the protein / Cr ratio of early morning urine is 0.5 or more and 1.0 or less, PN The patient is diagnosed with moderate proteinuria. If the amount of urinary protein is 1.0 g / 1.73 m ² / day or more, or if the protein / Cr ratio of early morning urine is larger than 1.0, it is diagnosed as a case with high proteinuria by PN (Non-Patent Document). 9). In particular, cases with PN having a urinary protein amount of 3.0 g / 1.73 m ² / day or more are likely to exhibit nephrotic syndrome and become severe. Generation method S1e subjects include PN-negative subjects, PN-positive subjects with no abnormal urinary protein content, PN-positive subjects with mild proteinuria, PN-positive subjects with moderate proteinuria, and PN-positive subjects with severe proteinuria. Subjects with proteinuria, PN-positive subjects with urinary protein content of 2.0 g / 1.73 m ² / day or more, and PN-positive subjects with urinary protein content of 3.0 g / 1.73 m ² / day or more, respectively. It is preferable to include as many people as possible. The preferred number of subjects is the same as that of the generation method S1a described above.

The clinical information of each subject in the generation method S1e is gender, IgAV acute phase, and from the viewpoint of clinical data that tends to have a causal relationship with the occurrence of PN or the occurrence of cases with proteinuria in PN. Age at the time of not receiving anti-inflammatory therapy (hereinafter also referred to as "before the first treatment"), systemic vasculitis marker test value before the first treatment, presence or absence of abdominal pain before the first treatment, before the first treatment Presence or absence of immediate allergic disease at the time point, blood IgA test value before the first treatment, blood IgE test value before the first treatment, number of times of receiving anti-inflammatory therapy after the time before the first treatment, and these It is the clinical data of 4 or more kinds of subjects selected from the group consisting of the data which indirectly shows any of. Gender, age, and indirect data are the same as in step S1a described above.

The test values for systemic vasculitis markers in the production method S1e are almost the same as those described in the generation method S1a, but clinical studies that tend to have a causal relationship with the occurrence of PN or the occurrence of cases with proteinuria in PN. The concentration of D-dimer (hereinafter also referred to as "FDP / D-dimer"), which is a decomposition product of fibrin, from the viewpoint of data and the viewpoint of test values that have been easily and widely used in pediatric clinical practice and are easy to apply clinically. The test value is preferable. Generally, the blood FDP / D-dimer concentration is used for determining thrombosis. The presence or absence of abdominal pain before the first treatment and the presence or absence of immediate allergic disease before the first treatment are obtained, for example, based on the findings of the medical condition when the doctor has already diagnosed IgAV for each subject. Possible information. It is said that 50% of cases are accompanied by acute abdominal pain during the acute phase of IgAV. Immediate allergic diseases include, for example, diseases such as anaphylactic shock, allergic rhinitis, conjunctivitis, bronchial asthma, urticaria, or atopic dermatitis. The blood IgA test value and the blood IgE test value are concentration data in blood or serum, respectively. The anti-inflammatory therapy received by each subject before and after the initial treatment was, for example, from the group consisting of PSL administration, IVMP administration, cyclophosphamide administration, azathioprine administration, mycophenolate mofetil administration, CsA administration, and plasma exchange. One or more selected anti-inflammatory therapies can be mentioned (see Non-Patent Document 9).

Judgment results regarding the presence or absence of cases with proteinuria in PN or PN are determined in each subject within a predetermined period after the initial treatment (for example, after the initial treatment and within 30 to 90 days after the onset of IgAV). It is the data about the result which the doctor has already judged whether or not the case with proteinuria was caused by PN. Usually, if hematuria is accompanied within this predetermined period, it is determined that PN has occurred. Judgment results regarding the occurrence of cases with proteinuria in PN make it possible to predict the risk of occurrence of PN cases that are likely to become severe, so that the doctor in charge of acute phase IgAV medical care can make a diagnosis to determine the treatment policy to avoid the aggravation at an early stage. From the viewpoint of assisting in facilitation, for example, it may be a judgment result as to whether or not PN has led to the occurrence of cases with moderate or higher proteinuria, and preferably whether or not PN has led to the occurrence of cases with high proteinuria. This is the judgment result of whether or not a case with a urinary protein amount of 2.0 g / 1.73 m ² / day or more was reached with PN, and more preferably with PN. It is desirable that it is a judgment result as to whether or not a case of 0 g / 1.73 m ^{2 / day or more has occurred.} In addition, the details of the clinical information acquisition step S2e of the subjects are the same as those of step S2a already described, except for the matters relating to the coronary artery diameter.

In the pre-learning normalization step S3e, four or more types of subject clinical data included in the clinical information of each subject and the determination result regarding the occurrence of PN or the occurrence of a case with proteinuria in PN are processed by an information processing apparatus. Normalize to make it easier to calculate. When normalizing the judgment result, for example, one of "with PN (PN positive)" and "without PN (PN negative)" is converted to 0, and the remaining one is converted to 1.0. May be normalized to. As mentioned above, from the viewpoint of predicting the risk of developing PN cases that tend to become severe, when normalizing the judgment results, for example, "positive cases with moderate or higher proteinuria in PN" and "moderate or higher in PN". "Negative case with proteinuria", preferably "Positive case with high proteinuria in PN" and "Negative case with high proteinuria in PN", and more preferably "Amount of urinary protein in PN 2. "Positive cases of 0 g / 1.73 m ² / day or more" and " ^{Negative cases of 2.0 g / 1.73 m 2} / day or more of PN", and even more preferably "3. "Positive case of 0 g / 1.73 m ² / day or more" and "Negative case of urinary protein amount 3.0 g / 1.73 m ² / day or more with PN", one of which is converted to 0 and the remaining one It is desirable to convert to 1.0. Other details of the pre-learning normalization step S3e are the same as those of step S3a already described.

In the covariance structure analysis step S4e, in order to calculate the SS calculated value having a high correlation with the occurrence of PN or the occurrence of a case with proteinuria in PN for each subject, five or more kinds are calculated for each subject. An observed variable and a latent variable are provided. Here, the breakdown of the five or more observation variables is, for each subject, the clinical data of four or more subjects and the judgment result regarding the occurrence of PN or the occurrence of a case with proteinuria in PN. In addition, one or more latent variables that are assumed to have a direct causal relationship are provided for the observation variables related to "determination results regarding the occurrence of PN or the occurrence of cases with proteinuria in PN". After that, the information processing device is made to execute arithmetic processing so as to perform covariance structure analysis, and the "judgment result regarding the presence or absence of PN occurrence" or the "judgment result regarding the presence or absence of case occurrence with proteinuria in PN" is For latent variables that have a direct and significant causal relationship, factor score data is calculated. The numerical data of this factor score itself, or the data obtained by normalizing the numerical data itself within a reproducible range, is used as the SS calculated value regarding the risk of PN occurrence in each subject.

For the five or more observation variables used in the covariance structure analysis step S4e, the preferred items are the same as in the previous normalization step S3e before learning. When covariance structure analysis is performed with two or more latent variables, the judgment result of PN positive or negative is positive for cases with proteinuria from the viewpoint of making it possible to predict the degree of PN aggravation step by step. Judgment result of negative or negative, judgment result of positive or negative in case with moderate or higher proteinuria in PN, judgment result of positive or negative in case with high proteinuria in PN, urinary protein amount 2.0 g in PN 1.73 m ² / day or more cases positive or negative or determination results, and, PN in urine protein amount 3.0g / 1.73m ² / day or more cases positive or negative of the determination result, from the group consisting of It is preferable to use two or more selected judgment results as latent variables, and in this case, the more types of judgment results used as latent variables, the more types of observation variables are provided from six or more, and a covariance structure. It is preferable to perform an analysis. Other details of the covariance structure analysis step S4e are the same as those of step S4a already described. The combination of steps S2e, S3e, and S4e can function as step S5e to generate machine learning data.

In the learning step S6e, four or more kinds of subject clinical data included in the clinical information of each subject are input to the input layer of ANN as input variables, and the SS obtained in the previous covariance structure analysis step S4e for each subject is obtained. ANN is machine-learned about the relationship between the input variable and the output variable so that the calculated value is output from the output layer as an output variable, and a trained model is generated. Other details of the learning step S6e are the same as those of the already described step S6a. When a plurality of trained models are generated, from the viewpoint of further improving the prediction accuracy, the high prediction accuracy is verified for each trained model, and the trained model having a relatively high prediction accuracy is selected in the selection step S7e. Is good to do. Other details of the sorting step S7e are the same as those of step S7a already described.

According to the generation method S1e described above, sex, age, systemic vasculitis marker test value, presence / absence of abdominal pain, presence / absence of immediate allergic disease, blood IgA test given as examples of clinical data of four or more subjects. Both the value and the blood IgE test value are information that can be obtained by a doctor in charge of IgAV acute care (for example, a pediatrician working at a hospital) by interview or test before the first treatment. Since each subject had already received the initial treatment and was determined to have a case with proteinuria in PN or PN, the number of anti-inflammatory treatments received before and after the initial treatment and PN or PN The determination result regarding the presence or absence of a case with proteinuria is also information available to the doctor in charge of IgAV acute care in the pediatric department of the hospital. The SS calculated value regarding the risk of occurrence of proteinuria in PN or PN in each subject is determined from the clinical data of 4 or more subjects and the judgment result of the presence or absence of the occurrence of proteinuria in PN or PN. It can be calculated by covariance structure analysis. Therefore, according to the generation method S1e, it is possible to generate a trained model by using information such as test values that are familiar in the field of daily medical care in conventional pediatrics. By utilizing this trained model as described below, it is possible to determine whether or not PN or PN will lead to the occurrence of proteinuria in patients in the acute phase of IgAV, who are the predictors, with the highest possible accuracy before the initial treatment. It becomes predictable.

[Method for predicting the risk of occurrence of PN in IgAV or proteinuria in PN]
In the prediction method S10e according to another embodiment of the present invention shown in FIG. 6, the risk of PN occurrence in IgAV acute phase patients or the risk of PN accompanied by proteinuria by using an information processing apparatus having a trained model is used. It is a prediction method for predicting. The prediction method S10e includes a clinical information acquisition step S2e of the subjects, a normalization step S3e before learning, a covariance structure analysis step S4e, a learning step S6e, a selection step S7e, and a patient clinical information acquisition step S12e. , The pre-prediction normalization step S13e and the prediction step S14e may be included. Steps S2e to S7e may be performed in the same manner as in the generation method S1e described above.

In the patient clinical information acquisition step S12e, for example, an IgAV acute phase patient who wants to predict whether PN will occur within a predetermined period after the initial treatment or whether PN will lead to the occurrence of a case with proteinuria will be obtained before the initial treatment. Obtain clinical information on possible patients. The clinical information of the patients obtained here is gender, age before the first treatment, systemic vasculitis marker test value before the first treatment, presence or absence of abdominal pain before the first treatment, and before the first treatment. blood IgA test values, blood IgE test values in the first pre-treatment time, the presence or absence of immediate allergic diseases in the initial pre-treatment time, execution scheduled anti-inflammatory therapy after the initial treatment before point (first treatment later) 4 or more patient clinical data selected from the group consisting of data indicating the number of times and any of these indirectly. The step of diagnosis performed by a doctor in the process until this clinical information is generated is not included in the present invention or step S12e.

From the viewpoint of improving the prediction accuracy, the four or more types of patient clinical data acquired in the patient clinical information acquisition step S12e are substantially the same type of data as compared with the four or more types of subject clinical data acquired in the previous step S2a. Is preferable. Regarding the "scheduled number of times of anti-inflammatory therapy", data may be acquired based on the treatment plan, prescription plan, or the plan already prepared by the doctor in the diagnosis before the initial treatment. If the so-called evidence for IgAV acute care recommends a specific number of anti-inflammatory therapies based on pre-first treatment test values for IgAV acute care patients, the recommended criteria. Therefore, the "scheduled number of times of anti-inflammatory therapy" may be automatically determined without the judgment of a doctor. In addition, the patient's clinical information acquisition step S12e is the same as that described above for step S12a and the previous step S2e. However, unlike the previous step S2e, in the patient clinical information acquisition step S12e, it is not necessary to acquire the determination result regarding the occurrence of PN in the patient or the occurrence of a case with proteinuria in PN.

In the pre-prediction normalization step S13e, in order to make an efficient prediction, four or more types of patient clinical data obtained in the previous patient clinical information acquisition step S12e are normalized so that they can be easily calculated by the information processing apparatus. Other than that, the normalization step S13e before prediction is the same as that already described for step S13a.

In the prediction step S14e, four or more types of patient clinical data included in the clinical information of IgAV acute phase patients are input to an information processing apparatus having a trained model, and a PN development risk or a case occurrence with proteinuria in PN occurs. For SS related to risk, the process of outputting the predicted value in IgAV acute phase patients is executed. For example, one type of patient clinical data or its normalized data (input variable) is input for each AN provided in the input layer of the trained model, and an information processing device having the trained model is made to execute an operation, and the output layer is used. In the AN provided in, the SS prediction value (output variable) regarding the risk of PN development in IgAV acute phase patients or the risk of case occurrence with proteinuria in PN is output.

Proteinuria with PN from the viewpoint of helping doctors in charge of acute IgAV medical care to make early decisions on treatment policies to avoid aggravation by making it possible to predict the risk of developing PN cases that are likely to become severe. The SS predicted value regarding the risk of occurrence of a case with PN may be, for example, the risk of occurrence of a case with moderate or higher proteinuria in PN, preferably the risk of occurrence of a case with high proteinuria in PN, and more preferably. Is related to the risk of developing a case with a urinary protein content of 2.0 g / 1.73 m ² / day or more in PN, and more preferably, a case occurrence with a urinary protein content of 3.0 g / 1.73 m ² / day or more in PN. It should be about risk. Depending on which risk the SS prediction value relates to, the predetermined value of the cutoff value, which is a criterion for determining whether the IgAV acute phase patient is positive (leading to development) or negative (not leading to development), differs. .. Therefore, in the prediction step S14e, when the SS prediction value in the IgAV acute phase patient is equal to or higher than the predetermined cutoff value, the prediction result indicating that the PN case to be predicted in this patient is generated is output. It is preferable to operate the information processing apparatus having the trained model so as to execute the processing. Other than that, the prediction step S14e is the same as that already described for step S14a.

According to the prediction method S10e described above, the clinical data of four or more kinds of subjects included in the clinical information of each subject and the SS calculated value in each subject output in the covariance structure analysis step S4e are obtained. By using the trained model generated by the machine learning used, the SS predicted value, which is an index value of whether or not PN occurs after the initial treatment, is obtained for the IgAV acute phase patient who is the prediction target before the first treatment. It is possible to obtain high accuracy at that point. Therefore, for example, it is possible to support a doctor in charge of IgAV acute care so that it is easy to make a diagnosis for early determination of a treatment policy for suppressing the occurrence of PN or the occurrence of cases with proteinuria in PN. From the viewpoint of making it easily feasible, the prediction method S10 f includes the step S1f for preparing a pre-generated trained model as shown in FIG. 7, instead of the steps S2e to S7e in the prediction method S10e. Is preferable.

[Learned model for predicting the risk of PN development in IgAV or proteinuria in PN]
The trained model according to another embodiment of the present invention is a trained model for predicting the risk of developing PN in IgAV patients or the risk of developing a case with proteinuria in PN. This trained model shows the clinical information of each subject who has already been determined to have PN or a case with proteinuria in PN within a predetermined period after the initial treatment in IgAV acute care. 5 is input to the input layer 2 of ANN1 shown in the output layer 8 so as to output the calculated value in each subject for SS relates PN risk weighting value W _ij indicated by equation 5 described above are machine learning It is a thing. The learned model, if the clinical information of IgAV acute patients to the input layer is inputted, performs a calculation based on the weighting value W _ij relative to the clinical information IgAV acute patient input, from the output layer The purpose is to make the information processing apparatus function so as to output the SS predicted value in the IgAV acute phase patient. For example, a person who implements the prediction method S10e shown in FIG. 6 or the prediction method S10f shown in FIG. 7 will use an information processing apparatus having this trained model. This information processing device may be configured in the same manner as the information processing device 60 already described with reference to FIG.

The present invention can also be carried out in a mode in which various improvements, modifications, or modifications are added based on the knowledge of those skilled in the art without departing from the spirit of the present invention. Further, within the range in which the same action or effect is produced, any of the invention-specific matters may be replaced with another technique.

Examples of the present invention will be described below, but the present invention is not limited to the following examples.

[Prediction of CAL occurrence in KD: Phase 1 and Phase 2 studies]
The inventor of the present application decided to conduct a retrospective study using opt-out at Minoh City Hospital. Retrospective studies are a type of research method that investigates the relationship between disease factors and onset. Looking back from the start of the study to the past, we decided to compare clinical data between subjects who developed CAL and those who did not develop CAL, and to verify the factors that led to the occurrence of CAL. The Minoh City Hospital Institutional Review Board approved the study and also abandoned the need for informed consent. All methods were carried out according to the guidelines relevant to the Declaration of Helsinki.

As shown in FIG. 9, the first study was conducted between March 2002 and 2005 and the period between July 2008 and April 2012, and the second study was conducted between July 2014 and December 2018. In the meantime, we examined whether 375 children who were clinically diagnosed with KD and admitted to Minoh City Hospital were suitable as subjects. 29 patients who met 3 or less of the 6 major symptoms in the KD diagnostic criteria, 16 patients who were transferred to a tertiary care institution, and patients with serious complications such as anaphylaxis and drug-induced hypersensitivity syndrome. 16 people were excluded. The remaining 314 children (185 boys and 129 girls) were the subjects. These subjects were divided into 106 subjects from the first stage study and 208 subjects from the second stage study, according to the time of the first-line treatment. In addition, as a comprehensive study, the subjects of the 1st and 2nd phase studies were combined, and 314 subjects were selected as the subjects of the 3rd phase study, which will be described later.

According to the Sano score shown in Table 3 above, subjects who met 2 or more of the 3 items were classified as high risk of IVIG refractory. The remaining subjects were classified as IVIG refractory low risk according to the Sano score. As shown in FIG. 9, in the first phase study, between March 2002 and 2005, 13 high-risk subjects received IVIG therapy (2 g / day of IVIG and moderate dose of ASA per kg of body weight). Combination therapy with a single dose of IVMP (30 mg / kg) was given, and 34 low-risk subjects received only IVIG therapy. On the other hand, in the first phase study, between July 2008 and April 2012, 17 high-risk subjects received IVIG therapy and two IVMP doses in combination, and 42 low-risk subjects. Was given a combination therapy of IVIG therapy and a single dose of IVMP. In the second phase study, 20 high-risk subjects received IVIG therapy and two IVMP doses in combination, whereas serum CRP concentration was 7 mg / dL or higher before the first-line treatment. The 113 subjects received IVIG therapy in combination with a single IVMP administration, and the 75 subjects with low-risk serum CRP concentration of less than 7 mg / dL before the first-line treatment were treated with IVIG. I was only on therapy.

At the time of KD acute care for the subjects, in order to evaluate the dilation of the coronary arteries for each subject, before the first treatment, immediately after the first treatment, and at the time of discharge (after the first treatment and within 30 days from the onset of KD) ), The diameter of the left anterior descending artery (LAD), and the diameter of the proximal right coronary artery (RCA) are determined by tomographic echocardiography. I was measuring. The coronary artery diameter of each subject was input into the above-mentioned Coronary Z Score Calculator and converted into a Z score calculated by the LMS method. When the maximum value of Z score is 3.0SD or more in LMT, LAD, or RCA, it is defined as CAL occurrence. In addition, clinical trials were conducted at the above three time points, and serum components were measured by the standard method used at Minoh City Hospital. Urinary β2MG / Cr was measured using a latex immunoassay kit manufactured by Fujifilm Wako Pure Chemical Industries, Ltd.

Hereinafter, in the statistical analysis, when p <0.05, it is recognized as statistically significant, and when 0.05 <p <0.10, it is recognized that there is a tendency. Hereinafter, when the Student's t-test or the χ ² test was performed, JMP version 8.0 software (manufactured by SAS) was used. Hereinafter, SPSS version 23.0 (manufactured by IBM-SPSS) was used when performing analysis using a linear mixed model. The profiles of the subjects in the first and second phase studies are shown in Table 4 below. In the subjects of the first phase study, the urinary β2MG / Cr logarithmic value (Log urinary β2MG / Cr) before the first treatment and the IVIG refractory prediction score (Sano score) before the first treatment were high. Met. In the subjects of the second stage study, the age at the time before the first treatment and the maximum value of the coronary artery diameter at the time before the first treatment were high.

In the first phase study, the coronary artery diameter was measured by several doctors, so the measured values fluctuated and there were some clinical data deficiencies such as gender. Therefore, in the first phase of the study, analysis was performed using a linear mixed model. On the other hand, in the second phase study, one doctor measured the coronary artery diameter with almost no deletion of the coronary artery diameter data, so the Student's t-test and the χ ² test were performed. From the viewpoint of statistical significance and the viewpoint of the tendency of characteristics common to CAL, we decided to select variables that identify the risk factors that led to the occurrence of CAL from the first and second phase studies. Preliminarily, we investigated the possibility that test value data such as neutrophil count, blood sodium concentration, and albumin would be useful by logistic regression analysis, but we could not find any significant effect in predicting the occurrence of CAL. The hematocrit value was searched in PubMed Central®, but no usefulness as a predictor of CAL occurrence was found.

As a result of analysis using a linear mixed model, there were no variables related to CAL generation in the first phase study. However, in the linear mixed model, the mean value of β2MG / Cr in Log urine showed a statistically significant positive association with the occurrence of CAL (p = 0.034), and the deviation value of the serum CRP concentration was also the occurrence of CAL. A statistically significant positive association was shown (p = 0.002). In addition, the mean aspartate aminotransferase (AST) concentration also showed a statistically significant positive association with CAL development (p = 0.064). In the second phase study, showed a statistically significant positive association maximum value of Log urine 32MG / C r of the primary pre-treatment time point and CAL generation (p = 0.0175), CAL generating a significantly Relapse of related disease showed a statistically significant positive association with CAL development (p = 0.0175), and the maximum serum CRP concentration before first-line treatment was also statistically significant with CAL development. It showed a positive association (p = 0.0855). Based on the characteristics common to both the 1st and 2nd studies, the maximum value of urinary β2MG / Cr before the first-line treatment and the maximum value of β2MG / Cr at the time before the first-line treatment are candidates for risk factors related to the occurrence of CAL. The blood CRP concentration was selected. In multivariate logistic regression analysis, the maximum value of β2MG / Cr in urine is a factor showing a statistically significant positive association with CAL development (p = 0.063), respectively, and the coronary arteries at the time before first-line treatment. It was suggested that the maximum diameter is a factor showing a statistically significant positive association with CAL development (p <0.0001).

A path model was created using AMOS23.0 (manufactured by IBM-SPSS) in order to perform mean covariance structure analysis by SEM. At the time of preparation, gender and age at the time before first-line treatment were selected as the clinical background of each subject. As factors related to KD, the presence or absence of high risk of IVIG refractory based on the score of the IVIG refractory prediction score (Sano score) at the time before the first-line treatment and the number of times of receiving IVMP in the first-line treatment were selected. As explanatory factors, the maximum value of Log urinary β2MG / Cr at the time before the first-line treatment and the serum CRP concentration at the time before the first-line treatment were selected. The clinical data of the 6 subjects selected here and the coronary artery diameter measured immediately after the first treatment or at the time of discharge (after the first treatment and within 30 days from the onset of KD) in each subject, the Z score value is 3.0 SD or more. A path model is created as an observation variable based on the result of judgment based on the criteria that CAL has occurred in the case of I let you.

Using the Student's t-test, the statistical significance was analyzed between the judgment result regarding the presence or absence of CAL occurrence and the SS calculated value. In the path model of the first study shown in FIG. 10, for a plurality of variables showing a correlation coefficient of more than 0.3 and a significant p-value of p <0.002, a plurality of paths were set between these variables. Selected. In the path model of the second study shown in FIG. 11, the latent variable and the CAL occurrence determination result are standardized based on the correlation coefficient exceeding 0.3 and the significant p-value of p <0.0001. Multiple paths showing the maximum coefficient value were selected.

Hereinafter, when the average covariance structure analysis in retrospective studies by the present inventors has performed less than RMSEA value 0.080 and, where R ² values for adaptation of the path model exceeds 0.95, statistically significant We decided to treat it as a sexual and reliable statistical model. Further, in RMSEA value is less than 0.050, in AIC value is less than 70, CFI (comparative fit index) value is greater than 0.95, and to fit the path model when ^{R 2} value is 1.000 , Furthermore, it was judged to be a statistically significant and reliable statistical model. The path model of the first phase study shown in FIG. 10 and the path model of the second phase study shown in FIG. 11 showed excellent conformity (statistical significance) based on the following values. The RMSEA value was p <0.0001 in the first and second phase studies, respectively. The AIC was 65 in the first study and 61 in the second study. The CFI was 1.000 for each of the first and second phase studies. R ² values for adaptation of the path model using a seven observed variables and one latent variables were respectively 1.000 in the first phase studies and Phase 2 study. As a direct effect of standardization, the standardized path coefficient between the latent variable and the observed variable related to the "judgment result regarding the presence or absence of CAL occurrence" is 0.80 (p <0.001) in the first phase study and 0 in the second phase study. It was .76 (p <0.001). From these results, the latent variables provided in the path model of the first phase study shown in FIG. 10 and the latent variables provided in the path model of the second phase study shown in FIG. 11 were described as "presence or absence of CAL occurrence". A direct and significant causal relationship was found with respect to the observed variables related to "judgment results".

As shown in FIG. 13, the SS calculated values of the subjects who were judged to have CAL because the Z score of the coronary artery diameter was 3.0 SD or more in the first stage study showed that the Z score was less than 3.0 SD. Therefore, it was significantly higher than the SS calculated value of the subjects who were judged not to have CAL (p <0.0001). In the second phase study, as shown in FIG. 14, the SS calculated values of the subjects who were judged to have CAL when the Z score was 3.0 SD or more were judged to have no CAL when the Z score was less than 3.0 SD. Compared with the SS calculated values of the subjects, the values were significantly higher (p <0.0001). Therefore, it was suggested that the determination result regarding the presence or absence of CAL occurrence can be inferred from the value of the SS calculated value. A receiver operating characteristic (hereinafter referred to as “ROC”) analysis was performed to determine a cutoff value for determining whether the value is CAL positive or CAL negative based on the value of the SS calculated value. In the ROC analysis, when "(sensitivity of latent variable) + (specificity of latent variable) -1" is calculated using the SS calculated value of each subject and the maximum value among the subjects is calculated, the maximum value is calculated. The SS calculated value of the subject, which is the source of the maximum value, becomes the cutoff value of the SS calculated value of the subjects. The cutoff value was 2.0 in the first phase study (Fig. 13) and 2.1 in the second phase study (Fig. 14). Since the cutoff value differs between the 1st and 2nd phase studies, it is better to obtain a correlation coefficient of 0.30 or more for the path model rather than obtaining a significant p-value for clinical application. Is considered important.

[Phase 3 study (Example 1-1)]
In order to verify whether the SS calculated value has a potential value that can accurately determine whether or not CAL has occurred, in the 3rd phase study, all the data from the 1st phase study and the 2nd phase study are integrated and averaged. The predictability of CAL occurrence risk was reexamined by covariance structure analysis. In the path model of the third phase study shown in FIG. 12, a significant p-value of p <0.0001 and a correlation coefficient of 0.33 or more were used as a reference. In this path model, RMSEA value 0.032, AIC value 65, CFI value ^{R 2} values for path model adaptation using the 0.98,7 or observed variables and one latent variables in 1.000 there were. As a direct standardization effect, the standardized path coefficient between the latent variable and the observed variable related to the "judgment result regarding the presence or absence of CAL occurrence" is 1.47 (p <0.001), and the standardized path coefficient is 0.741 as the overall standardization effect. It was (p <0.001). From these results, it was found that the latent variables provided in the path model of the 3rd phase study had a direct and significant causal relationship with the observed variables related to the "judgment result regarding the presence or absence of CAL occurrence". In the third phase study, as shown in FIG. 15, it is excellent in discriminating the occurrence of CAL using the SS calculated value (p <0.0001), and the cutoff value of the SS calculated value by ROC analysis is 2.0. there were.

In addition, in the path model of the third stage study shown in FIG. 12, it is determined whether the coronary artery diameter is CAL positive or CAL negative depending on whether the Z score of the coronary artery diameter is 2.5 SD or more as the "judgment result regarding the presence or absence of CAL occurrence". After changing, the mean covariance structure analysis was performed to calculate the SS calculated value. As a result, as shown in FIG. 16, the SS calculated values of the subjects who were determined to have CAL when the Z score was 2.5 SD or more were the subjects who were determined to have no CAL when the Z score was less than 2.5 SD. It was significantly higher than the SS calculated value of (p <0.0001), and the cutoff value of the SS calculated value by ROC analysis was 1.3. Similarly, in the path model shown in FIG. 12, the "judgment result regarding the presence or absence of CAL occurrence" is changed so as to determine whether the Z score is CAL positive or CAL negative depending on whether the Z score is 2.0 SD or more, and the average covariance structure is changed. The SS calculated value was calculated by the analysis. As a result, as shown in FIG. 17, the SS calculated values of the subjects who were determined to have CAL when the Z score was 2.0 SD or more were the subjects who were determined to have no CAL when the Z score was less than 2.0 SD. It was significantly higher than the SS calculated value of (p <0.0001), and the cutoff value of the SS calculated value by ROC analysis was 0.38. In this way, as the definition of CAL generation was reduced from 3.0 SD or more (FIG. 15) to 2.5 SD or more (FIG. 16) or 2.0 SD or more (FIG. 17), the value of the SS calculated value also decreased, and thus CAL. It is considered that the potential of the SS calculated value with respect to the determination of the presence or absence of occurrence has also decreased to some extent. In addition, in order to explain whether the latent factor (latent variable) is related to the maximum value of the coronary artery diameter at the time before the first treatment, another SEM analysis including the maximum value of the coronary artery diameter at the time of admission (before the first treatment). However, no direct relationship between the latent variable and the maximum coronary artery diameter at admission was found (standardized path coefficient 0.095, p = 0.25).

In the ANN analysis, of the 7 observation variables used in the mean covariance structure analysis in the path model of the 3rd phase study shown in FIG. 12, the remaining 6 types are excluded except for the "judgment result regarding the presence or absence of CAL occurrence". The observed variables were selected as 6 types of input variables as shown in FIG. In addition, the maximum value of the coronary artery diameter at the time before the first-line treatment was also selected as one type of input variable. The maximum value of this coronary artery diameter was input into the above-mentioned Coronary Z Score Calculator and converted into a Z score by the LMS method, and used as an input variable. Each of the remaining 6 observation variables was normalized by the above-mentioned equation 4 to obtain 6 input variables. The output variable was the SS calculated value calculated by the average covariance structure analysis by SEM when the Z score was defined as CAL occurrence when the Z score was 3.0 SD or more. In a three-layer ANN with four ANs in the middle layer, seven types of input variables are input to the input layer, and machine learning is performed so that the output variables (SS calculated values) are output in the output layer, and a statistical model is used. (Trained model) was generated. When this machine learning was repeated to generate multiple trained models, the overfit penalties of 0.01, 0.02, and 0.04 selected two to four nodes in the middle layer. A plurality of the learned model, 0.81 or more ^{R 2} values (r = 0.9), and were measured five-fold cross-validation of the ^{R 2} value 0.64 (r = 0.8). The best statistical model selected was the trained model according to Example 1-1. In this trained model, the mathematical formula 6 defined by the parameter (intermediate variable) H1 shown in FIG. 18 is illustrated below.

In the learned model according to Example 1-1, overfitting penalty 0.02, ^{R 2} value by five-fold cross-validation was 0.64. The input variables used in the machine learning were input to the input layer of this trained model, and the "SS predicted value regarding the risk of CAL occurrence" was output for each subject in the third phase study. As shown in FIG. 19, the SS calculated value calculated by the mean covariance structure analysis and the SS predicted value output from the trained model generated by machine learning with the training data including this SS calculated value are used. ^{, The high value of R 2} = 0.89 was shown as the strength of the relevance. When the presence or absence of CAL was determined in each subject based on this SS predicted value, as shown in Table 5 below, the C statistic (c index) was 0.860, the sensitivity was 72.7% (8/11), and the sensitivity was 7/11. The specificity was 99.1% (232/234). In addition, some subjects lacked clinical data of, for example, urinary β2MG / Cr. In addition, the trained model cannot output SS calculated values for subjects lacking even a part of a plurality of input variables used for prediction (for example, subjects lacking clinical data of urinary β2MG / Cr). Therefore, the numerical values of sensitivity and specificity shown in Table 5 below are the results for the subjects who have no defects in the 6 types of clinical data used for the prediction.

From the results of the third phase study (Example 1-1), the inventor of the present application has learned the importance of the SS calculated value that can discriminate the CAL occurrence risk with high accuracy and the CAL occurrence risk can be predicted with high accuracy. We found the usefulness of the model. The prediction accuracy using the trained model of Example 1-1 is considered to be superior to the conventional IVIG refractory prediction score (Tables 1 to 3 described above). However, there is room for improvement because there were some outliers in FIG. If observation variables and input variables that are more suitable for prediction can be found in the future, it will be possible to improve the sensitivity. The CAL occurrence risk prediction method utilizing the trained model generated by the inventor of the present application is, for example, in the KD acute phase medical treatment S20 shown in FIG. From (S32 to S35, S42) other anti-inflammatory therapies (S32 to S35, S42) that are superior in anti-inflammatory effect to the usual IVIG therapy (S31, S41, S51) at the time of (S36) and at the time of judgment (S45) after the second treatment. It is expected that it can contribute to the diagnosis of early determination of the treatment policy for performing S44, S52 to S56).

In order to examine whether the trained model according to Example 1-1 is clinically applicable, it was verified with a new data set. The new dataset includes clinical data from 38 KD acute care patients who received first-line treatment at Minoh City Hospital between 2019 and 2020. However, these 38 patients did not include patients who developed CAL. For this reason, the new data set included two KD acute care patients who had CALs but were excluded due to the exclusion criteria and were not analyzed and were transferred to a tertiary care institution. One KD acute phase patient (boy) was treated at Minoh City Hospital with IVIG therapy and two IVMP doses, but fever continued after the first-line treatment. The inventor of the present application transferred the patient (boy) to a tertiary medical institution, and the patient received plasma exchange therapy at the tertiary medical institution. The remaining one KD acute phase patient (girl) received a combination therapy of IVIG therapy and IVMP administration at Minoh City Hospital, but the five main symptoms of KD persisted even after the first-line treatment. The patient (girl) was transferred to a tertiary care center where she received oral CsA and further IVIG therapy. When a new data set for 40 people was input to the trained model according to Example 1-1 and it was determined from the output SS predicted value whether or not CAL occurred, the sensitivity was 50% (1/2). And the specificity was 100% (38/38). The findings are consistent with the results described above in Table 5, given the small number of KD acute care patients included in the new dataset, 40.

[Comparative Example 1-1]
In the past, studies investigating the relationship between disease factors and onset have attempted multivariate logistic regression analysis, and it was not possible to set latent variables in this analysis. In consideration of this, a path model according to Comparative Example 1-1 excluding latent variables was created as shown in FIG. 20 as compared with the path model of the third stage study shown in FIG. An average covariance structure analysis was attempted using this path model, but as shown in Table 5 above, it was a bad statistical model that did not fit at all, so there was no possibility that the risk of CAL occurrence could be predicted.

[Comparative Example 1-2]
In Comparative Example 1-2, the risk of CAL occurrence was predicted by multivariate logistic regression analysis as before. To do this, based on the clinical information of the subjects in the Phase 3 study, gender, age before first-line treatment, serum CRP concentration before first-line treatment, and Log urinary β2MG / Cr before first-line treatment. Independent clinical data of 7 subjects, including the presence or absence of IVIG refractory high risk based on the Sano score before the first-line treatment, the maximum coronary artery diameter before the first-line treatment, and the number of times IVMP was administered in the first-line treatment. Multivariate logistic regression analysis was tried by extracting as a variable and extracting the judgment result regarding the presence or absence of CAL occurrence as an objective variable. However, as shown in Table 5 above, since R ² = 0.0615 and the C statistic was less than 0.80, the prediction accuracy of the CAL occurrence risk was low as before.

[Examples 1-2 to 1-6]
In Example 1-1 described above, six types of subject clinical data were used as observation variables in addition to the “judgment result regarding the presence or absence of CAL” in the mean covariance structure analysis. In Examples 1-2 to 1-6, it was verified whether or not the risk of CAL occurrence could be predicted even if one of the six observation variables was reduced. As a result, as shown in Table 6 below, it was suggested that the prediction could be made with a certain degree of high accuracy.

[Examples 1-7 to 1-15]
It was verified by the mean covariance structure analysis whether the statistical model fits when 2 or 3 of the 6 observation variables excluding the "judgment result regarding the presence or absence of CAL occurrence" in Example 1-1 are reduced. .. As a result, as shown in Tables 7 and 8 below, they are compatible to some extent, and it is considered that the risk of CAL occurrence can be predicted with some accuracy by further combining with machine learning.

When Examples 1-13 shown in Table 8 were verified, the standardized path coefficient between the latent variable and the observed variable related to the “judgment result regarding the presence or absence of CAL generation” as a direct standardization effect was 1.413 (p <0.001). As a total standardization effect, the standardization path coefficient was 0.768 (p <0.001). That is, even if the types of subject clinical data used as observation variables are reduced from 6 types in Example 1-1 to 3 types in Examples 1-13, observations related to "judgment results regarding the presence or absence of CAL occurrence" are observed with latent variables. A direct and significant causal relationship was found for the variables. Therefore, in the other examples listed in Tables 6 to 8, if the standardized path coefficient is examined, the latent variable is directly related to the observation variable related to the “judgment result regarding the presence or absence of CAL occurrence”. It is considered that a significant causal relationship will be observed. On the other hand, although no data are shown, when 4 or more of the 6 observation variables were reduced, the mean covariance structure analysis did not converge. Therefore, it is considered that three or more types of clinical data are required to predict the risk of CAL occurrence.

[Prediction of occurrence of cases with high proteinuria in PN in IgAV: Example 2-1]
The inventor of the present application conducted a retrospective study in the same manner as the above-mentioned third phase study in order to verify whether the risk of developing a case with high proteinuria in PN can be predicted with high accuracy in patients in the acute phase of IgAV. The subjects were 93 children (male: female = 43:50) who visited Minoh City Hospital between May 1995 and October 2015 and were diagnosed with IgAV and received inpatient treatment. These children were 6.2 ± 2.2 years old (mean ± SD) at admission (before initial treatment), had a history of immediate allergic disease at admission 25%, and had abdominal pain at admission 60. %Met. In addition, 30% of patients received PSL in the initial treatment, that is, 70% of the patients received PSL orally for a certain period of time, and 70% of the patients received PSL orally. It was 30%. When the diagnosis was made after the initial treatment and about 30 days after the onset of IgAV, 26 patients (28%) had PN onset. A urinary chemistry test was performed on these 26 subjects, and when the urinary protein / Cr ratio was 1.0 or higher, it was determined that a case with severe proteinuria was generated by PN, and it was determined that this case occurred. There were 10 (11% of 93 children).

Univariate analysis showed significant differences (p <0.05) between peripheral blood, blood biochemistry, and FDP / D dimer values at admission (before initial treatment) and cases with severe proteinuria in PN. When the path model shown in FIG. 21 was created from the above items and the reports reported in the previous report, the observed variables were gender, age at the time before the first treatment, and blood FDP at the time before the first treatment. D-dimer concentration, presence or absence of immediate allergic disease before initial treatment, presence or absence of abdominal pain before initial treatment, presence or absence of PSL administration in initial treatment, and high proteinuria in PN after initial treatment The judgment result regarding the presence or absence of the occurrence of cases accompanied by was selected. The mean covariance structure analysis by SEM was performed using this path model, and the SS calculated value which is the factor score of the latent variable was calculated. In addition, some of the subjects lacked some clinical data (for example, blood FDP / D-dimer concentration) among the 6 types of observation variables excluding the judgment result.

When the reliability of the path model shown in FIG. 21 was examined, the AIC value was 67, the CFI value was 0.986, the C statistic of the latent variable was 0.958, the RMSEA value was 0.052, and 7 kinds of observed variables. since R ² value on path model fit using a one latent variables were 1.000, suggesting that a reliable statistical model. As a direct standardization effect, the standardization path coefficient between the latent variable and the observation variable related to "judgment result regarding the presence or absence of cases with high proteinuria in PN" is 1.37 (p <0.001), and the standardization total effect is as follows. The standardized path coefficient was 0.583 (p <0.001). From these results, a direct and significant causal relationship was found with the latent variables provided in the path model for the observation variables related to "judgment results regarding the presence or absence of cases with high proteinuria in PN". As shown in FIG. 22, the SS calculated value, which is the factor score for the latent variable for which a direct causal relationship was found in the mean covariance structure analysis, is excellent in discriminating the occurrence of cases with high proteinuria in PN. (P <0.0001), the cutoff value of the SS calculated value by ROC analysis was 1.36.

As shown in FIG. 23, 6 types of observation variables excluding "judgment result regarding the presence or absence of cases with high proteinuria in PN" are used as 6 types of input variables, and SS calculated values are used as output variables, and JMP version 8. Using 0, a three-layer ANN was machine-learned to generate a trained model according to Example 2-1. Six types of input variables used in machine learning were input to the input layer of this trained model, and "SS predicted value regarding the risk of developing a case with high proteinuria in PN" was output for each subject. As shown in FIG. 24, the SS calculated value calculated by the mean covariance structure analysis and the SS predicted value output from the trained model generated by machine learning with the training data including this SS calculated value are used. ^{, The high value of R 2} = 0.92 was shown as the strength of the relevance. Cross-validation ^{R 2} value was 0.47. Based on this SS prediction value, the presence or absence of cases with high proteinuria in PN in each subject was determined. As shown in Table 9 below, the C statistic was 0.90 and the sensitivity was 87.5% (7). / 8), and the specificity was 98% (49/50).

After the above verification, the inventor of the present application inputs clinical data into the trained model according to Example 2-1 for seven children who were diagnosed with IgAV at Minoh City Hospital and received inpatient treatment around 2019. Then, the SS predicted value was output. Based on the output SS prediction values, the risk of developing cases with advanced proteinuria was predicted by PN in 7 children, and the C statistic was 1.00, the sensitivity was 100% (1/1), and the specificity was 80%. (4/5), the C statistic was 1.00. Although the number of cases is small and it is a preliminary study result, it is suggested that the trained model according to Example 2-1 can be used for predicting the occurrence of cases with severe proteinuria in PN in patients in the acute phase of IgAV. rice field. It is expected that this trained model can be used for predictive diagnosis of risk assessment that causes severe proteinuria and renal dysfunction in the near future about 30 days after the initial diagnosis of patients in the acute phase of IgAV.

[Comparative Example 2-1]
Compared with the path model according to Example 2-1 shown in FIG. 21, the path model according to Comparative Example 2-1 excluding the latent variable was created as shown in FIG. 25. An attempt was made to analyze the mean covariance structure using this path model, but since it was a poor statistical model that did not fit as described above in Table 9, there was no possibility that PN could predict the risk of developing cases with high proteinuria. ..

[Comparative Example 2-2]
In Comparative Example 2-2, the conventionally Multivariate logistic regression analysis was attempted to predict the patients risk with high proteinuria in PN. For this purpose, from the clinical information of 93 subjects, gender, age before the first treatment, blood FDP / D-dimer concentration before the first treatment, and immediate allergic disease before the first treatment. Six types of subject clinical data, such as the presence or absence, the presence or absence of abdominal pain before the initial treatment, and the presence or absence of PSL administration in the treatment, were extracted as independent variables, and "Case occurrence with severe proteinuria in PN after treatment". The multivariate logistic regression analysis was tried by extracting "judgment result regarding the presence or absence of" as the objective variable. However, as described above in Table 9, since R ² = 0.492, the accuracy of predicting the risk of developing a case with high proteinuria in PN was low.

[Examples 2-2 to 2-5 and Reference Example 2-6 ]
Statistical by mean covariance structure analysis when one or two of the six observation variables except "judgment result regarding the presence or absence of cases with high proteinuria in PN" in Example 2-1 are reduced. We verified that the model fits. As a result, as shown in Table 10 below, it was suitable to some extent, and it is considered that the risk of occurrence of cases with high proteinuria in PN can be predicted with some accuracy by further combining with machine learning.

When Examples 2-4 shown in Table 10 were verified, the standardized path coefficient between the latent variable and the observed variable related to "judgment result regarding the presence or absence of the occurrence of a case with high proteinuria in PN" as a direct standardization effect was 1.099. It was (p <0.001), and the standardized path coefficient was 0.671 (p <0.001) as a total standardization effect. Similarly, when verified in Example 2-5, the standardized path coefficient between the latent variable as a direct standardization effect and the observation variable related to "judgment result regarding the presence or absence of cases with high proteinuria in PN" is 1.352 (p <. It was 0.001), and the standardized path coefficient was 0.700 (p <0.001) as a total standardization effect. In other words, even if the types of subject clinical data used as observation variables are reduced from 6 types in Example 2-1 to 4 types in Examples 2-4 and 2-5, the latent variable is "with high proteinuria in PN." A direct and significant causal relationship was found with respect to the observation variables related to "judgment results regarding the presence or absence of cases". Therefore, in the other examples listed in Table 10, if the standardized path coefficient is examined, the latent variable is also a direct and significant causal effect with respect to the observation variable related to the "judgment result regarding the presence or absence of CAL occurrence". It is believed that the relationship will be recognized. On the other hand, although no data are shown, when 3 or more of the 6 observation variables were reduced, the mean covariance structure analysis did not converge. Therefore, it is considered that four or more kinds of clinical data are required to predict the risk of CAL occurrence.

[Prediction of PN generation in IgAV: Example 3-1]
In order to verify whether the risk of PN development can be predicted with high accuracy in IgAV acute phase patients, the inventor of the present application extracted some different clinical data from the clinical information of the same subjects as in Example 2-1 described above. Similarly, a retrospective study was conducted. When the path model shown in FIG. 26 was created, the observation variables were gender, age at the time before the first treatment, blood IgA concentration at the time before the first treatment, and blood IgE at the time before the first treatment. The concentration, the presence or absence of PSL administration in the initial treatment, and the judgment result regarding the presence or absence of PN generation after the initial treatment were selected. The mean covariance structure analysis by SEM was performed using this path model, and the SS calculated value which is the factor score of the latent variable was calculated. The presence or absence of PN is determined when a commercially available urinary protein test paper is immersed in urine and + (protein concentration of about 30 mg / dL or more) continues to develop for 3 days or more, or when a commercially available urinary occult blood test paper is used in urine. Soaked + (hemoglobin concentration about 0.06 mg / dL or more or red blood cell concentration about 20 / μL or more) or ++ (hemoglobin concentration about 0.15 mg / dL or more or red blood cell concentration about 50 / μL or more) Or if the urinary occult blood test paper on the market was soaked in urine for 2 days or more, +++ (hemoglobin concentration of about 0.75 mg / dL or more or red blood cell concentration of about 250 cells / μL or more) continued for 2 days or more. In that case, it was determined that PN had occurred. In addition, many subjects lacked clinical data on blood IgA concentration and / or blood IgE concentration among the five observation variables excluding the determination results. There were 32 subjects with no lack of clinical data for the 5 types of observation variables excluding the judgment results.

When the reliability of the path model shown in FIG. 26 was examined, the CFI value was 0.979, the C statistic of the latent variable was 0.808, the RMSEA value was 0.039, 6 kinds of observed variables and one latent variable. since R ² value was 1.000 regarding path model adapted with, it was suggested that a reliable statistical model. As a direct standardization effect, the standardized path coefficient between the latent variable and the observed variable related to the "judgment result regarding the presence or absence of PN generation" is 2.97 (p <0.001), and the standardized path coefficient is 0.426 as the overall standardization effect. It was (p <0.001). From these results, it was found that the latent variables provided in the path model had a direct and significant causal relationship with the observed variables related to the "judgment result regarding the presence or absence of PN generation". As shown in FIG. 27, the SS calculated values for which a direct causal relationship was found in the mean covariance structure analysis are excellent in discriminating the occurrence of PN (p <0.0001), and the SS calculated values by ROC analysis are excellent. The cutoff value was −0.26.

Five types of observation variables excluding "judgment result regarding the presence or absence of PN generation" are used as five types of input variables as shown in FIG. 28, and SS calculated values are used as output variables, and a three-layer type is used using JMP version 8.0. ANN was machine-learned to generate a trained model according to Example 3-1. Five types of input variables used in machine learning were input to the input layer of this trained model, and "SS predicted value regarding PN occurrence risk" was output for each subject. As shown in FIG. 29, the SS calculated value calculated by the mean covariance structure analysis and the SS predicted value output from the trained model generated by machine learning with the training data including this SS calculated value are used. ^{, The high value of R 2} = 0.99 was shown as the strength of the relevance. Cross-validation ^{R 2} value was 0.92. When the presence or absence of PN was determined in each subject based on this SS predicted value, as shown in Table 11 below, the C statistic was 0.725, the sensitivity was 66.7% (6/9), and the specificity was 78. It was 0.3% (18/23). The C statistic is less than 0.80 but greater than 0.70, so it is considered to be within the permissible range.

After the above verification, for 28 IgAV patients different from the subjects, the patient clinical data was input to the trained model according to Example 3-1 and the SS predicted value was output, and the sensitivity was 75% (sensitivity 75% (). It was 3/4) and had a specificity of 50% (12/24). Although the number of cases is small and it is a preliminary study result, it is suggested that the trained model according to Example 3-1 can be used for predicting the occurrence of PN in patients in the acute phase of IgAV. It is expected that this trained model can be used for predictive diagnosis of risk assessment leading to PN development in the near future, about 30 days after the initial diagnosis of patients in the acute phase of IgAV.

[Comparative Example 3-1]
Compared with the path model according to Example 3-1 shown in FIG. 26, the path model according to Comparative Example 3-1 excluding the latent variable was created as shown in FIG. Attempted to mean covariance structure analysis by the path model, because it was bad statistical model which is not at all adapted as shown in Table 11, expected had no predictive of P N onset raw risk.

[Comparative Example 3-2]
In Comparative Example 3-2, we tried to predict the risk of PN generation by multivariate logistic regression analysis as before. For this purpose, from the clinical information of 93 subjects, the sex, the age at the time before the first treatment, the blood IgA concentration at the time before the first treatment, the blood IgE concentration at the time before the first treatment, and the treatment. Multivariate logistic regression analysis was tried by extracting the clinical data of 5 types of subjects, that is, the presence or absence of PSL administration, as independent variables, and the "judgment result regarding the presence or absence of PN generation after treatment" as the objective variable. .. However, as shown in Table 11, ^{the low value of R 2} = 0.122 and the specificity of 4.3% were too low, so that the accuracy of predicting the risk of PN occurrence was low.

Claims (9)

It is a predictive method for predicting the risk of complications in patients with Kawasaki disease.
The prediction method includes a step of inputting clinical information of the patient into an information processing apparatus having a trained model and executing a process of outputting a predicted value in the patient for a sample score regarding the risk of developing the complication. Including,
The complication is a dilated coronary artery lesion
In the trained model, the relationship between the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined and the calculated value of the sample score in each subject is machine-learned. And
The clinical information of each subject is gender, age at the time of acute stage of Kawasaki disease and not receiving anti-inflammatory therapy, coronary artery diameter at the time, systemic vasculitis marker test value at the time, Hypercytokine morbidity marker test value at the time point, score of intravenous immunoglobulin (IVIG) refractory prediction score at the time point, number of times the anti-inflammatory therapy was received after the time point, and any of these. Includes clinical data of 3 or more subjects selected from the group consisting of indirectly shown data
Regarding the sample score, the calculated value in each of the subjects includes four or more observation variables including the clinical data of the three or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. This is data on the factor scores of latent variables for which a directly significant causal relationship is observed with respect to the observed variables related to the above-mentioned determination results when the covariance structure analysis is performed.
The clinical information of the patient includes sex, age at the time point, coronary artery diameter at the time point, systemic vasculitis marker test value at the time point, hypercytokinemia marker test value at the time point, and hypercytokinemia marker test value at the time point. Prediction, including three or more patient clinical data selected from the group consisting of IVIG refractory prediction score scores, the number of times the anti-inflammatory therapy is scheduled to be received after the time point, and data indirectly indicating any of these. Method. It is a predictive method for predicting the risk of complications in patients with IgA vasculitis.
The prediction method includes a step of inputting clinical information of the patient into an information processing apparatus having a trained model and executing a process of outputting a predicted value in the patient for a sample score regarding the risk of developing the complication. Including,
The complication is purpura nephritis or the case of purpura nephritis with proteinuria.
In the trained model, the relationship between the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined and the calculated value of the sample score in each subject is machine-learned. And
The clinical information of each of the above subjects includes gender, age at the time of acute IgA vasculitis and no anti-inflammatory therapy, systemic vasculitis marker test value at the time, and abdominal pain at the time. Presence / absence, presence / absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, number of times of receiving the anti-inflammatory therapy after the time point, and any of these Includes clinical data from 4 or more subjects selected from the group consisting of data that indirectly indicate
Regarding the sample score, the calculated value in each of the subjects includes five or more observation variables including the clinical data of the four or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. This is data on the factor scores of latent variables for which a directly significant causal relationship is observed with respect to the observed variables related to the above-mentioned determination results when the covariance structure analysis is performed.
The clinical information of the patient includes gender, age at the time point, systemic vasculitis marker test value at the time point, presence or absence of abdominal pain at the time point, presence or absence of immediate allergic disease at the time point, and time point. 4 or more selected from the group consisting of blood IgA test values, blood IgE test values at the time point, the number of times the anti-inflammatory therapy is scheduled to be received after the time point, and data indirectly indicating any of these. Prediction method, including patient clinical data. The anti-inflammatory therapy was selected from the group consisting of administration of acetylsalicylic acid and / or a salt thereof, IVIG administration, intravenous methylprednisolone pulse administration, prednisolone administration, infliximab administration, ulinastatin administration, cyclosporine A administration, and plasma exchange. The prediction method according to claim 1 or 2, which is one or more therapeutic methods. When the predicted value of the sample score in the patient is equal to or higher than a predetermined cutoff value, the information processing apparatus is made to function so as to execute a process of outputting the predicted result leading to the occurrence of the complication in the patient. , The prediction method according to any one of claims 1 to 3. A method for generating a trained model for predicting the risk of complications in patients with Kawasaki disease.
In the generation method, when the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined is input to the input layer, the output layer describes the sample score regarding the risk of the occurrence of the complication. Including the step of generating the trained model by machine learning to output the calculated value in the subject.
The complication is a dilated coronary artery lesion
The clinical information of each subject is gender, age at the time of acute stage of Kawasaki disease and not receiving anti-inflammatory therapy, coronary artery diameter at the time, systemic vasculitis marker test value at the time, It consists of data that indirectly indicate the hypercytokinemia marker test value at the time point, the score of the IVIG refractory prediction score at the time point, the number of times the anti-inflammatory therapy was received after the time point, and any of these. Includes clinical data from 3 or more subjects selected from the group
Regarding the sample score, the calculated value in each of the subjects includes four or more observation variables including the clinical data of the three or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. A method for generating a trained model, which is data on factor scores of latent variables for which a directly significant causal relationship is recognized with respect to the observed variables related to the above-mentioned judgment results when the covariance structure analysis is performed. .. A method for generating a trained model for predicting the risk of complications in patients with IgA vasculitis.
In the generation method, when the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined is input to the input layer, the output layer describes the sample score regarding the risk of the occurrence of the complication. Including the step of generating the trained model by machine learning to output the calculated value in the subject.
The complication is purpura nephritis or the case of purpura nephritis with proteinuria.
The clinical information of each of the above subjects includes gender, age at the time of acute IgA vasculitis and no anti-inflammatory therapy, systemic vasculitis marker test value at the time, and abdominal pain at the time. Presence / absence, presence / absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, number of times of receiving the anti-inflammatory therapy after the time point, and any of these Includes clinical data from 4 or more subjects selected from the group consisting of data that indirectly indicate
Regarding the sample score, the calculated value in each of the subjects includes five or more observation variables including the clinical data of the four or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. A method for generating a trained model, which is data on factor scores of latent variables for which a directly significant causal relationship is recognized with respect to the observed variables related to the above-mentioned judgment results when the covariance structure analysis is performed. .. A trained model for predicting the risk of complications in patients with Kawasaki disease
In the trained model, the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined is input to the input layer, and the output layer is the sample score regarding the risk of the complication. When the weighted value is machine-learned so as to output the calculated value in the subject and the clinical information of the patient is input to the input layer, the input clinical information of the patient is obtained. The purpose is to make the information processing device function so as to perform an operation based on the weighted value and output a predicted value of the sample score in the patient from the output layer.
The complication is a dilated coronary artery lesion
The clinical information of each subject is gender, age at the time of acute stage of Kawasaki disease and not receiving anti-inflammatory therapy, coronary artery diameter at the time, systemic vasculitis marker test value at the time, It consists of data that indirectly indicate the hypercytokinemia marker test value at the time point, the score of the IVIG refractory prediction score at the time point, the number of times the anti-inflammatory therapy was received after the time point, and any of these. Includes clinical data from 3 or more subjects selected from the group
Regarding the sample score, the calculated value in each of the subjects includes four or more observation variables including the clinical data of the three or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. This is data on the factor scores of latent variables for which a directly significant causal relationship is observed with respect to the observed variables related to the above-mentioned determination results when the covariance structure analysis is performed.
The clinical information of the patient includes sex, age at the time point, coronary artery diameter at the time point, systemic vasculitis marker test value at the time point, hypercytokinemia marker test value at the time point, and hypercytokinemia marker test value at the time point. Learning, including three or more patient clinical data selected from the group consisting of IVIG refractory prediction score scores, the number of times the anti-inflammatory therapy is scheduled to be received after the time point, and data indirectly indicating any of these. Finished model. A trained model for predicting the risk of complications in patients with IgA vasculitis.
In the trained model, the clinical information of each subject in the subjects whose presence or absence of the complication has already been determined is input to the input layer, and the output layer is the sample score regarding the risk of the complication. When the weighted value is machine-learned so as to output the calculated value in the subject and the clinical information of the patient is input to the input layer, the input clinical information of the patient is described as described above. The purpose is to make the information processing device function so as to perform an operation based on the weighted value and output the predicted value of the patient for the sample score from the output layer.
The complication is purpura nephritis or the case of purpura nephritis with proteinuria.
The clinical information of each of the above subjects includes gender, age at the time of acute IgA vasculitis and no anti-inflammatory therapy, systemic vasculitis marker test value at the time, and abdominal pain at the time. Presence or absence, presence or absence of immediate allergic disease at the time point, blood IgA test value at the time point, blood IgE test value at the time point, number of times of receiving anti-inflammatory therapy after the time point, and any of these Includes clinical data of 4 or more subjects selected from the group consisting of data indirectly indicating
Regarding the sample score, the calculated value in each of the subjects includes five or more observation variables including the clinical data of the four or more subjects and the judgment result regarding the presence or absence of the complication in each of the subjects. This is data on the factor scores of latent variables for which a directly significant causal relationship is observed with respect to the observed variables related to the above-mentioned determination results when the covariance structure analysis is performed.
The clinical information of the patient includes gender, age at the time point, systemic vasculitis marker test value at the time point, presence or absence of abdominal pain at the time point, presence or absence of immediate allergic disease at the time point, and time point. 4 or more selected from the group consisting of blood IgA test values, blood IgE test values at the time point, the number of times the anti-inflammatory therapy is scheduled to be received after the time point, and data indirectly indicating any of these. Trained model containing patient clinical data from. A storage unit in which the trained model is stored and
When the clinical information of the patient is input, the processing unit that applies the input clinical information of the patient to the trained model and outputs the predicted value of the sample score in the patient. ,
The information processing apparatus having the trained model according to claim 7 or 8.

Images