JP7630256B2 - Server, terminal, method, program, and method for generating trained model for predicting changes in patient condition - Google Patents

Description

The present invention relates to a server, a terminal, a method, a program, and a method for generating a trained model for predicting changes in a patient's condition.

Chronic obstructive pulmonary disease (COPD) is a chronic disease that is common among elderly men with a history of smoking. COPD patients are repeatedly re-admitted to hospital due to repeated exacerbations caused by respiratory infections, etc., which reduces their quality of life (QOL) and progresses the disease, requiring appropriate intervention at key points. The number of COPD patients is on the rise, and it is said that by 2030, ischemic heart disease, stroke, COPD, and respiratory infections will be the four leading causes of death in the world.

When COPD patients are repeatedly readmitted to hospital, medical costs mount. Therefore, it is important to distinguish between patients who have been discharged from the hospital after contracting COPD and who are at high risk of being readmitted to hospital and those who are not, to warn those at high risk, and to prevent readmission.

Up to now, methods for predicting the risk of a patient suffering from COPD have been reported (for example, Patent Document 1). Patent Document 1 discloses an apparatus for assessing the risk of exacerbation and/or hospitalization of a subject who is a COPD patient, the apparatus including an input unit for receiving time-dependent physical activity data of the subject, an activity data conversion unit for acquiring frequency-dependent activity data, a moment determination unit for calculating moments of the frequency-dependent activity data, and a risk assessment unit for assessing the risk of exacerbation and/or hospitalization of the subject based on at least the moments.

The device disclosed in Patent Document 1 predicts an increase in the risk of exacerbation or hospitalization from a decrease in a patient's moment, based on the fact that the decrease in moment corresponds to a decrease in the frequency and intensity of the activity cycle.

Patent No. 6367982

However, because factors other than a decline in the patient's activity cycle also affect the exacerbation or hospitalization of COPD patients, there is a problem in that the method described in Patent Document 1 makes it difficult to accurately predict risk.

The present invention aims to provide a server, terminal, method, program, and method for generating a trained model that can accurately predict changes in the condition of patients suffering from COPD.

The server according to an embodiment of the present disclosure is characterized by having an acquisition unit that acquires data on a first test item related to respiratory function and a second test item related to a comorbid disease of a patient suffering from COPD, and a prediction unit that predicts a change in the patient's condition using the data on the first and second test items.

In a server according to an embodiment of the present disclosure, it is preferable that predictions by the prediction unit are made using a trained model created based on data on the first and second test items of multiple patients suffering from COPD and data on changes in the conditions of the multiple patients.

In a server according to an embodiment of the present disclosure, it is preferable that the trained model is created using a Boosting model.

In a server according to an embodiment of the present disclosure, it is preferable that the Boosting model is an XGBoost model.

In a server according to an embodiment of the present disclosure, it is preferable that the acquisition unit acquires data of the patient before he or she is discharged from the hospital, and the prediction unit predicts changes in the patient's condition after he or she is discharged from the hospital.

In a server according to an embodiment of the present disclosure, a change in a patient's condition preferably includes the patient being re-admitted to hospital.

In a server according to an embodiment of the present disclosure, it is preferable that the acquisition unit further acquires data regarding a third test item related to a frailty factor.

In a server according to an embodiment of the present disclosure, it is preferable that predictions by the prediction unit are made using a trained model created based on data on the first, second, and third test items of multiple patients suffering from chronic obstructive pulmonary disease and data on changes in the conditions of the multiple patients.

In a server according to an embodiment of the present disclosure, the data on the second test item related to a comorbid disease preferably includes at least one of data on the name of the comorbid disease, data on a blood test, data on a vital sign test, and prescription data.

In a server according to an embodiment of the present disclosure, the data on the third test item related to frailty factors preferably includes data on at least one of age, body mass index (BMI), activities of daily living (ADL), nutritional status, and whether or not the person lives alone.

In a server according to an embodiment of the present disclosure, it is preferable to further include a determination unit that determines which of a plurality of predetermined hierarchies the predicted result of the change in the patient's condition predicted by the prediction unit corresponds to.

In a server according to an embodiment of the present disclosure, it is preferable to further include an output unit that outputs the predicted state change.

In a server according to an embodiment of the present disclosure, it is preferable that the prediction unit predicts the patient's physical condition and/or its transition probability.

The terminal according to the embodiment of the present disclosure is characterized by having an input unit for inputting data on a first test item related to respiratory function of a patient suffering from COPD and a second test item related to a comorbid disease, a transmission unit for transmitting the input data to a server, a reception unit for predicting a change in the patient's condition using the data on the first and second test items, respectively, and receiving the predicted change in condition from the server, and an output unit for outputting the received change in condition.

The method for predicting a change in condition according to an embodiment of the present disclosure is characterized by including obtaining data on a first test item related to respiratory function and a second test item related to a comorbid disease for a patient suffering from COPD, and calculating a predicted value of the change in the patient's condition using the data on the first and second test items.

A program according to an embodiment of the present disclosure is characterized in that it causes a computer to acquire data on a first test item related to respiratory function and a second test item related to a comorbid disease of a patient suffering from COPD, and to predict a change in the patient's condition using the data on the first and second test items.

The method for generating a trained model according to an embodiment of the present disclosure is characterized in that it acquires training data including data on a first test item related to respiratory function and a second test item related to a comorbid disease for multiple patients suffering from COPD, and data on changes in the condition of the multiple patients, and generates a trained model in which the data on the first and second test items are input and the data on the changes in condition are output.

The server, terminal, method, program, and method for generating a trained model according to the embodiments of the present disclosure can accurately predict changes in the condition of patients suffering from COPD.

FIG. 1 is a configuration diagram of a system including a server according to an embodiment of the present disclosure. FIG. 2 is a configuration diagram of a terminal according to an embodiment of the present disclosure. 1A and 1B are diagrams for explaining a calculation method using an XGBoost model used in a server according to an embodiment of the present disclosure, in which FIG. 1A shows a model image and FIG. 1B shows a learning image. 11 is a flowchart for explaining an example of a procedure for creating a learning model used in a server according to an embodiment of the present disclosure. FIG. 11 is a configuration diagram of a server according to a first modified example of an embodiment of the present disclosure. FIG. 13 is a diagram illustrating an example of an output result by a server according to a first modified example of an embodiment of the present disclosure. 11 is a flowchart for explaining an operation procedure of a server according to an embodiment of the present disclosure. This table classifies patient conditions into five categories based on vital signs and blood test items. 9 is a table showing the probability of the five states in FIG. 8 undergoing a state transition seven days after the pre-transition state. 9 is a diagram showing changes in a patient's condition suggested by the state sequence obtained by the statistical method, and the state transition probabilities in FIG. 9 and the vital signs and blood test items in the classification in FIG. 8 .

Below, we will explain the server, terminal, method, program, and method for generating a trained model for predicting changes in a patient's condition according to the present invention, with reference to the drawings. However, please note that the technical scope of the present invention is not limited to these embodiments, but extends to the inventions described in the claims and their equivalents.

[System configuration]
1 shows a configuration diagram of a system 1000 including a server 101 according to an embodiment of the present disclosure. The system 1000 includes the server 101 and a plurality of terminals 300. The server 101 and the plurality of terminals 300 are connected by wire or wirelessly via a communication network 200. Although a plurality of terminals 300 are illustrated in FIG. 1, the number of terminals 300 may be one.

The server 101 includes a control unit 10, a communication interface (I/F) 20, and a memory unit 30, which are connected by a bus 40.

The communication I/F 20 (output unit) is connected to multiple terminals 300 via the communication network 200 to exchange data.

The storage unit 30 may be a storage device such as a RAM or ROM, or a storage device such as an optical disk. The storage unit 30 stores a program for controlling the server 101. It is also preferable that the storage unit 30 stores electronic medical record information of the patient. This also includes an operation in which items required for this risk prediction program are extracted from the electronic medical record information of the patient and the anonymized information is stored.

The control unit 10 includes a CPU. The control unit 10 controls the server 101 by executing a program stored in the storage unit 30.

As shown in FIG. 1, the control unit 10 of the server 101 has an acquisition unit 11 and a prediction unit 12. The acquisition unit 11 and the prediction unit 12 are realized as software (programs) by a computer provided in the server 101, which includes a CPU, a ROM, a RAM, and the like.

[Device configuration]
Next, a terminal according to an embodiment of the present disclosure will be described. Fig. 2 shows a configuration diagram of a terminal 300 according to an embodiment of the present disclosure. The terminal 300 includes a control unit 31, an input unit 32, a transmission unit 33, a reception unit 34, an output unit 35, and a storage unit 36, which are connected by a bus 37.

The control unit 31 includes a CPU. The control unit 31 controls the terminal 300 by executing a program stored in the memory unit 36.

The input unit 32 can be an input device such as a touch panel, a mouse, a touchpad, or a keyboard.

A display device such as a liquid crystal display device can be used as the output unit 35. Alternatively, an output device such as a printer can be used as the output unit 35.

[Data used to predict changes in patient condition]
The system 1000 including the server 101 according to the embodiment of the present disclosure predicts a change in the condition of a patient suffering from COPD by using data on a first test item related to respiratory function and a second test item related to a comorbid disease. The system 1000 preferably further predicts a change in the condition of the patient by using data on a third test item related to a frailty factor. Each data will be described in detail below.

The data on the first test item related to respiratory function includes at least one of data on the GOLD (Global Initiative for Chronic Obstructive Lung Disease) classification, other data on spirometry measurements, data on the mMRC (Medical Research Council), and data on the CAT (COPD Assessment Test), and is preferably the GOLD classification and mMRC.

GOLD is a global initiative for chronic obstructive pulmonary disease and provides diagnostic criteria for COPD. The GOLD classification is divided into four categories. A predicted forced expiratory volume in one second (FEV1) of 80 or more is GOLD category 1, 50-79 is GOLD category 2, 30-49 is GOLD category 3, and less than 30 is GOLD category 4.

mMRC is the modified MRC shortness of breath scale, a questionnaire with a modified MRC grade classification from 0 to 4. It is used as a scale to evaluate the severity of dyspnea.

The data on the second test item related to the comorbid disease includes at least one of data on the name of the comorbid disease, data on blood tests, data on vital signs tests, and prescription data.

Data on the names of comorbid diseases is data on the names of diseases other than COPD. COPD patients frequently suffer from respiratory complications and systemic comorbidities. This is thought to be caused by systemic inflammation due to COPD. Therefore, COPD needs to be managed as a systemic disease. Comorbid diseases include, for example, cardiovascular disease and cancer. For data on disease names, it is a good idea to refer to the names of ICD-10 codes (International Classification of Diseases).

The blood test data includes at least one of the following data: BUN (blood urea nitrogen), Alb (albumin), CRP (C-reactive protein), ChE (cholinesterase), pH (hydrogen ion index), NT-proBNP (N-terminal pro-B-type natriuretic peptide), IgM (immunoglobulin M), IgG (immunoglobulin G), LDH (lactate dehydrogenase), PT seconds (prothrombin time), γ-GTP (gamma-glutamyl transpeptidase), total protein, ALP (alkaline phosphatase), cTnI (cardiac troponin I), cTnT (cardiac troponin T), A/G ratio (albumin/globulin ratio), hematocrit, hemoglobin, white blood cell count, PaCO2 (arterial blood carbon dioxide partial pressure), and PaO2 (arterial blood oxygen partial pressure), and is preferably BUN, CRP, ChE, and pH.

BUN indicates the state of kidney and liver function. If the value is above the standard value, kidney disease is suspected, and if it is below the standard value, liver disease is suspected. Values also tend to rise during dehydration. It is also one of the items in the severity classification of pneumonia (A-DROP system).

Alb is a useful indicator for comprehensively assessing the nutritional status of the entire body. If the value decreases, the patient is presumed to be malnourished, and if the value increases, the patient is presumed to be recovered.

CRP appears in the blood when an inflammatory response or tissue destruction occurs in the body, and high levels may indicate infection, cancer, myocardial infarction, etc.

ChE is an enzyme produced in the liver. Low levels indicate decreased liver function, while high levels suggest fatty liver or dyslipidemia caused by overnutrition. ChE, along with Alb, is also used as an indicator of nutritional status.

pH is measured by arterial blood gas analysis to determine whether blood is acidic or alkaline. Abnormal pH is the result of serious diseases such as respiratory failure or renal failure and is therefore used as an indication for treatment.

The data relating to the vital signs test includes at least one of data relating to respiratory rate, pulse rate, body temperature, blood pressure, and SpO2, and is preferably respiratory rate and SpO2.

Respiratory rate is the number of breaths taken per minute, and COPD patients tend to have an increased respiratory rate because they have a lower ventilation volume per breath.

SpO2 stands for percutaneous arterial oxygen saturation, the normal value is 96-99%, and if it is below 90%, respiratory failure is suspected.

Prescription data includes data about medications prescribed to patients. When using the medication data, it can be classified using medication classification codes, etc.

The data on the third test item related to the frailty factor includes data on at least one of age, BMI, ADL, nutritional status, and whether or not the person lives alone, and preferably ADL and nutritional status.

Data on ADL is an indicator used in medical and nursing care settings to measure a patient's degree of independence. Methods for assessing ADL include the Barthel Index (BI).

In addition to BMI, blood test data such as ChE and Alb can be used as data on nutritional status.

[Procedure for predicting changes in a patient's condition]
The system 1000 including the server 101 according to the embodiment of the present disclosure predicts a change in the condition of a patient suffering from COPD by using data on a first test item related to respiratory function and a second test item related to a comorbid disease. The system 1000 preferably further predicts a change in the condition of the patient by using data on a third test item related to a frailty factor. The procedure for executing the prediction will be described below.

The input unit 32 of the terminal 300 inputs data on a first test item related to respiratory function and a second test item related to comorbid diseases for a patient suffering from COPD. It is preferable that the input unit 32 also inputs data on a third test item related to frailty factors.

As shown in FIG. 1, the transmission unit 33 transmits the above data input by the input unit 32 to the server 101 via the communication network 200.

The acquisition unit 11 acquires the transmitted data on the first test item related to respiratory function and the second test item related to a comorbid disease for the patient suffering from COPD. The acquisition unit 11 may also process or convert the data acquired by the acquisition unit 11 so that processing in the prediction unit 12 described below proceeds smoothly. Note that, although the processing has been described as being performed on the server side in one embodiment, it may also be performed on the terminal side (e.g., the control unit 31, the input unit 32, the transmission unit 33, etc.).

It is preferable that the acquisition unit 11 further acquires data regarding a third test item related to frailty factors.

The prediction unit 12 predicts changes in the patient's condition using data on the first and second test items, respectively. That is, the prediction unit 12 predicts changes in the patient's condition using not only data on the first test item related to respiratory function for a patient suffering from COPD, but also data on the second test item related to comorbid diseases. Although COPD is a respiratory disease, it is also known as a systemic inflammatory disease, and therefore changes in the patient's condition may occur not only as an exacerbation of COPD, but throughout the entire body. Thus, the server according to an embodiment of the present disclosure predicts changes in the patient's condition more accurately by not only using data related to respiratory function, but also using data related to comorbid diseases.

It is preferable that the prediction unit 12 further uses data on a third test item related to the frailty factor to predict changes in the patient's condition. The reason for adding the frailty factor is that it is said that decline in lung function and complications are also related to aging factors, and by using data related to the frailty factor, changes in the patient's condition can be predicted more accurately.

The prediction by the prediction unit 12 is made using a trained model created based on data on the first and second test items of multiple patients suffering from COPD and data on changes in the condition of the multiple patients. That is, a trained model is created in advance by machine learning based on data on the first and second test items of multiple patients suffering from COPD and data on changes in the condition of the multiple patients, and is stored in the storage unit 30. The method of creating the model will be described later.

It is preferable that the acquisition unit 11 acquires data of the patient before discharge from the hospital, and the prediction unit 12 predicts changes in the patient's condition after discharge from the hospital. The server 101 according to the embodiment of the present disclosure predicts changes in the patient's condition after discharge from the hospital using data on a first test item related to the patient's respiratory function before discharge from the hospital and a second test item related to a comorbid disease, and preferably further using data on a third test item related to a frailty factor. Here, "before discharge from the hospital" preferably means immediately before discharge from the hospital, for example, "at the time of discharge from the hospital". This is because data on test items related to the patient's respiratory function and comorbid diseases immediately before discharge from the hospital is considered to best represent the condition of the patient being discharged from the hospital.

A change in a patient's condition includes the patient being readmitted to hospital. Here, "readmission" includes not only readmission due to an exacerbation of COPD, but also readmission due to factors other than an exacerbation of COPD. This is because patients with COPD may be readmitted to hospital after being discharged due to factors other than an exacerbation of COPD.

"Readmission" also includes being readmitted within a specified period of time. Here, the specified period may be, for example, 30 days after discharge. However, it is not limited to this example and may be any period of time, such as 60 days after discharge or 90 days after discharge.

The server 101 outputs the predicted state change. For example, it transmits it to the terminal 300 via the communication network 200. As another output destination, the predicted state change may be transmitted to a display device such as a liquid crystal display device, or an output device such as a printer, without passing through the communication network 200.

The receiving unit 34 receives the predicted state change from the server 101.

The output unit 35 outputs the state changes received from the server 101.

As described above, the server 101 for predicting changes in a patient's condition according to an embodiment of the present disclosure uses not only data on a first test item related to respiratory function, but also data on a second test item related to a comorbid disease, or data on a third test item related to a test item related to frailty factors, making it possible to more accurately predict changes in the condition of patients suffering from chronic obstructive pulmonary disease.

[Learning model creation method]
The trained model is preferably created by a Boosting model. Boosting is a method of continuously training a weak learner to create a learning algorithm that is stronger than an algorithm using a single weak learner. Finally, a final model is created by weighting the model with the highest accuracy.

The boosting model may be an XGBoost (eXtreme Gradient Boosting) model, a GBDT (Gradient Boosting Decision Tree), a LightGBM (Light Gradient Boosting Model), etc., with the XGBoost model being preferred. The XGBoost model is an ensemble learning model that combines gradient boosting and random forests.

XGBoost is a method of constructing weak learners one by one in order, and is a model that uses the results of all weak learners constructed up to that point when constructing a new weak learner. Figures 3(a) and 3(b) are diagrams for explaining a calculation method using XGBoost. The vertical axis of Figure 3(a) indicates a tree structure, the vertical axis of Figure 3(b) indicates a predicted value y _i ', and the horizontal axis both indicate round t. In XGBoost, trees such as those shown in Figure 3(a) are added one after another so as to compensate for the difference between the true value and the predicted value (y ₁ ', y ₂ ', y ₃ ', ...) as shown in Figure 3(b).

First, it is assumed that a predicted value y ₁ ' is obtained using the decision tree 41. At this time, if the true value is y ₀ , the error is y ₀ - _{y 1} '. Next, the second decision tree 42 is created so as to cancel this error. That is, the second decision tree 42 is created so that the accuracy is improved when the results of the first decision tree 41 and the second decision tree 42 are added together. At this time, it is assumed that a predicted value y ₂ ' is obtained using the decision tree 42, the error is y ₀ - _{y 2} '. Furthermore, a third decision tree 43 is created so as to cancel this error, and a predicted value y ₃ ' is calculated. As described above, since XGBoost is an algorithm that fills in residuals, high accuracy is obtained.

When performing data analysis to predict state changes on input data used in XGBoost, if too many explanatory variables are used, multicollinearity and noise can become a problem. In addition, there are a wide variety of test items, and it is not efficient to predict the probability of state changes using all of the test items. For this reason, it is preferable to calculate partial correlation coefficients using partial correlation analysis, for example, and extract important correlated explanatory variables.

Furthermore, it is preferable to set thresholds for continuous values of blood tests and vital signs measurements and grade the extracted variables, for example, by decision tree analysis. In this way, the calculations by XGBoost can be made more efficient.

This article describes the results of using machine learning to build a 30-day readmission and death risk prediction model and to search for risk factors using discharge data on the first test item related to respiratory function, the second test item related to comorbid diseases, and the third test item related to frailty factors. As described above, basic information, respiratory function, disease name, prescribed medication, and device prescription information are added to the blood tests, vital signs, and ADL items extracted by decision tree analysis, and machine learning (XGBoost analysis) is performed.

By using various parameters related to the first test item related to respiratory function, the second test item related to comorbid diseases such as disease name, prescribed medications, blood and vital signs, and the third test item related to frailty factors such as ADL and basic information in the analysis, it is estimated that it is possible to comprehensively determine the risk of re-hospitalization or death within 30 days (including hospitalization for reasons other than exacerbation) for patients with COPD as their underlying disease, rather than simply "predicting the patient's risk of COPD exacerbation."

In addition, by using blood and vital signs test items in the analysis, it is possible to link the patient's objective physical condition at the time of discharge with condition transitions and outcomes (readmission).

Next, a method for creating a learning model by a server according to an embodiment of the present disclosure will be described. FIG. 4 shows a flowchart for explaining the procedure for creating a learning model used by a server according to an embodiment of the present disclosure.

First, in step S101, data on a first test item related to respiratory function for a patient suffering from COPD, a second test item related to a comorbid disease, and a third test item related to frailty factors are acquired. As an example, 1,306 pieces of data were used as a dataset. For hold-out validation, 999 pieces of training data and 429 pieces of test data were used. A total of 1,428 pieces of training data and test data were used, and the ratio of labels was adjusted to be label 0:label 1 = 1:1. Here, label 0 is a label that does not have the outcome of readmission or death within 30 days, and label 1 is a label that has the outcome of readmission or death within 30 days.

Next, in step S102, partial correlation coefficients are calculated from the data on the second and third test items by partial correlation analysis, and important explanatory variables are extracted. A partial correlation coefficient is a correlation coefficient obtained by removing the influence of a third variable from two variables when the correlation between the two variables is increased or decreased by the third variable.

Next, in step S103, a threshold is set for the continuous values of the data related to the second test item using decision tree analysis, and grading is performed. For example, the continuous values are divided into high values, reference values, and low values based on the set threshold.

Next, in step S104, the first test item is added to the extracted data on the second and third test items, and machine learning (XGBoost analysis) is performed. The analysis method using XGBoost is as described above.

Next, in step S105, important explanatory variables are extracted and the ROC-AUC is calculated. The ROC-AUC is a statistical method for analyzing how useful a discrimination method is. When the ROC-AUC is 1, the predictive and diagnostic capabilities are the best, and for a random, completely invalid model, the value is 0.5.

First, we used the XGBoost model to calculate the importance of variables. As a result, we found that among the multiple variables with high importance, in addition to variables related to respiratory function, there were also variables related to comorbid diseases and variables related to frailty factors.

Next, the ROC-AUC of the created XGBoost model was calculated. The result was 0.758. Generally, a model is considered to be valid when this value is 0.7 or more, so an effective learning model was obtained by learning using data on the first test item related to respiratory function, the second test item related to comorbid diseases, and the third test item related to frailty factors. In contrast, the ROC-AUC was 0.65 when the learning model was created using only data on the first test item related to respiratory function. From this, it can be said that a more accurate learning model can be created by creating a learning model using data on the first test item related to respiratory function, the second test item related to comorbid diseases, and the third test item related to frailty factors.

In the above explanation, an example was described in which a learning model was created using data on a first test item related to respiratory function, a second test item related to a comorbid disease, and a third test item related to a frailty factor for a patient suffering from COPD. However, a learning model may also be created using data on the first test item related to respiratory function and the second test item related to a comorbid disease.

[Modification 1]
5 shows a configuration diagram of the server 102 according to the first modified example of the embodiment of the present disclosure. The server 102 preferably further includes a determination unit 13 that determines to which of a plurality of predetermined hierarchies the prediction result of the patient's condition change predicted by the prediction unit 12 corresponds. For example, the probability of the patient's condition change can be divided into a plurality of levels, ranging from a low level to a high level. By dividing the prediction result of the patient's condition change into a plurality of hierarchies, the patient and medical personnel can objectively recognize which of the plurality of hierarchies they are currently in.

In addition, if the change in the patient's condition is that the patient will be readmitted to hospital, the risk of the patient being readmitted to hospital may be divided into multiple hierarchies. Specifically, for example, the possibility that the patient will be readmitted to hospital within 30 days after being discharged may be divided into four hierarchies.

As shown in FIG. 5, the server 102 preferably further includes an output unit 50 that outputs the predicted state change. The output unit 50 may be a display device, a printer, or the like. The predicted state change may also be output to the terminal 300 via the communication network 200 as shown in FIG. 1.

Figure 6 shows an example of an output result by the server 102 according to the first variation of the embodiment of the present disclosure. As an example of the output result, Figure 6 shows an example of an output screen 400 to a display device (not shown). On the output screen 400, the output result is displayed in four areas, a first area 401 to a fourth area 404.

The first area 401 shows the COPD readmission risk level. For example, the possibility of readmission within 30 days after discharge is displayed in four stages, from low to high.

In addition, a triangular mark indicates the patient's risk level of readmission out of the four risk levels. Patients with a high risk level of readmission can recognize from the output results that they have a high risk of being readmitted. For example, in the example shown in Figure 6, the patient's risk level of readmission is displayed as being at the second highest risk level out of the four levels.

The second area 402 indicates whether each piece of information included in the patient's basic information is at high or low risk. As the patient's basic information, for example, information on age, smoking history, whether the patient lives alone or with other people, and information on ADL may be displayed.

The third area 403 shows the results of the patient's respiratory function test, including the "mMRC" shortness of breath scale and the "GOLD classification" stage classification indicating the severity of COPD. The level of functional decline of the patient can also be displayed for these test items.

The fourth area 404 shows the SpO2 and respiratory rate test results as vital sign test results, and the BUN and pH test results as blood test results, among the patient's test items. The normal range is hatched.

By showing not only the test results related to respiratory function shown in the third area 403 but also the test results related to vital signs and blood tests shown in the fourth area 404 in the output result 400, it is possible to display the impact that not only the test results related to respiratory function but also the test results related to vital signs and blood tests have on the risk level of re-hospitalization.

As described above, by displaying the output screen 400 as shown in FIG. 6 as the output result from the output unit 50 of the server 102 according to the first modified embodiment of the present disclosure, it is possible to display which of multiple hierarchies the change in the patient's condition is in.

[Method for determining risk level of readmission]
Next, a procedure for predicting a patient's readmission probability and determining a risk level of readmission using a server according to an embodiment of the present disclosure will be described. Fig. 7 shows a flowchart for explaining an operation procedure of the server according to an embodiment of the present disclosure. First, in step S201, data on a first test item related to respiratory function, a second test item related to comorbid diseases, and a third test item related to frailty factors for a patient suffering from COPD are acquired.

Next, in step S202, the probability of a patient being readmitted is predicted using a trained model created based on data on the first to third test items of multiple patients and data on readmission of multiple patients. The trained model may be the learning model described using FIG. 3. Here, readmission occurs within 30 days after discharge.

Next, in step S203, it is determined which of a plurality of predetermined risk levels the predicted result of the patient's readmission probability corresponds to. The plurality of risk levels (hierarchies) can be divided into, for example, four risk levels. Specifically, for example, in the case of the possibility that a patient will be readmitted within 30 days after discharge, the XGBoost analysis algorithm is used to set a predicted probability (probability of readmission within 30 days) of 0-10% as the first level, 10-30% as the second level, 30-60% as the third level, and 60-100% as the fourth level. In this way, by stratifying patients according to their risk, it is possible to easily grasp which hierarchy, from high to low, the patient is in in terms of the probability of readmission within 30 days after discharge.

Next, in step S204, the predicted risk level of readmission is output. The output of the risk level can be displayed as shown in FIG. 6.

In the above explanation, an example has been described in which the probability of a patient being readmitted to hospital within 30 days is calculated using data on a first test item related to respiratory function, a second test item related to a comorbid disease, and a third test item related to a frailty factor for a patient suffering from COPD, but the example is not limited to this. In other words, the probability of a patient being readmitted to hospital within 30 days may be calculated using data on a first test item related to respiratory function and a second test item related to a comorbid disease.

[Modification 2]
The above embodiment and modification 1 predict the change in the patient's condition using a trained model created using machine learning, but other methods for predicting the change in the patient's physical condition can also be used. For example, a change in the patient's condition can be predicted by performing a time-series analysis from the time of discharge using data related to vital signs and blood tests. Examples of time-series analysis methods include a state space model and a hidden Markov model. When the input data is huge, a method such as a convolutional encoder can also be used. When a hidden Markov model is used, a distribution function capable of estimating the patient's condition, a state transition probability at a certain time interval, and various outcomes, for example, a state sequence for readmission within 30 days, can also be obtained.

The results obtained from the above analysis can also be used to estimate changes in condition related to the patient's physical condition, re-admission, death, and exacerbation. For example, a time-series analysis is performed in advance based on data on the first and second test items of multiple patients suffering from COPD and data on changes in the conditions of multiple patients, a model of changes in the patient's condition is created, and stored in the memory unit 30 of the server 101. The prediction unit 12 can use the stored model to predict the patient's physical condition and changes in condition from, for example, the patient's information at the time of discharge, and it is also thought that the model can be used as a basis for deciding whether or not to intervene with the patient, i.e., as a monitoring index.

Furthermore, by utilizing the changes in the patient's physical condition obtained from time series analysis, together with information on the prediction of changes in the patient's condition using a trained model such as that shown in Figure 3, and preferably the state transition probability and state sequence described below, it becomes possible to determine and monitor whether or not to intervene in a way that is more tailored to the patient's condition.

[Time series analysis method]
Next, a method for time-series analysis of the effect of test data of a patient suffering from COPD after discharge on the deterioration of the patient's condition or re-admission will be described. The subject of the time-series analysis is test data related to vital signs and blood test items after the first discharge date. Specifically, the test data narrowed down by the above-mentioned learning model (XGBoost model) and items related to vital signs (respiratory rate, SpO2) were used. The vital signs test values used were data from the hospitalization period and after discharge. The blood test items extracted were BUN, Alb, CRP, and ChE. The extracted items are only an example, so the method can be performed in the same way even if other items are added or subtracted.

A hidden Markov model was used for the time series analysis. The hidden Markov model models the probability distribution of states and outputs at each point in time in time series data, and is a statistical method for predicting hidden states from observable variables.

Data from approximately 1,000 patients was used in the time series analysis. Data for each test item was divided into seven-day intervals from the date of initial discharge, and the average value was used. As hyperparameters for the hidden Markov model, a normal distribution was assumed for the distribution function, and there were five states (S0 to S4). Time series analysis was used to determine the estimated transition probability between the five states, and the state sequence of the time series data. The state sequence was calculated by calculating the proportion of each state (S0 to S4) for various outcomes such as readmission within 30 days and the presence or absence of exacerbation.

Figure 8 shows a heat map of test values related to the patient's vital signs and blood test items. The test values for respiratory rate, SpO2, Alb, ChE, CRP, and BUN were classified into high, slightly high, low, and slightly low values compared to normal values. In Figure 8, "-" indicates that the value is within the normal range. The normal ranges for these values are as follows:

また、図８において、太枠で示した、状態Ｓ１、Ｓ４、及びＳ０は、検査値が異常である検査項目を多く含んでいることから、処置が必要な状態である可能性が示唆され、特にＳ０は正常値からの乖離が大きいので重篤な状態を示唆しているものと推察される。また、Ｓ２は呼吸数が多く、ＳｐＯ２が低値傾向であることから、呼吸機能が健常者と比較して良くない傾向にあると推察される。しかし、抽出されたバイタル、血液検査項目のみから判断は出来ないため、状態遷移確率、状態シーケンスの結果を総合的に分析して各状態の意味合いを考察していく。

In addition, in Fig. 8, states S1, S4, and S0, which are indicated by thick frames, include many test items with abnormal test values, suggesting the possibility that treatment is required, and in particular, S0 is presumed to suggest a serious condition because it is significantly different from the normal value. In addition, S2 has a high respiratory rate and a tendency for SpO2 to be low, so it is presumed that the respiratory function tends to be poorer than that of a healthy person. However, since it is not possible to make a judgment based only on the extracted vital signs and blood test items, the meaning of each state will be considered by comprehensively analyzing the results of the state transition probability and state sequence.

The state transition probability for each of states S0 to S4 is shown in Figure 9. Figure 9 also shows the probability [%] of transitioning from each of pre-transition states S0 to S4 to each of states S0 to S4 after 7 days. As shown in Figure 9, there is a high probability that states S0 to S4 after 7 days will remain as they are.

Focusing on state S3, the probability of transitioning from S3 to another state is low, so state S3 can be said to be a stable state. Conversely, looking at the probability of reaching state S3 after a transition, we can see that the probability of reaching state S3 from a state other than S3 is extremely low.

Furthermore, looking at the probability of reaching state S2 after a transition, the probability of going from S0 to S2 (1.52%) is higher than the probability of going from S2 to S0 (0.342%), and the probability of going from S4 to S2 (2.68%) is higher than the probability of going from S2 to S4 (1.16%), so state S2 is inferred to be a stable state. Since S2 has a poorer SpO2 value than S1 and S4, it is considered to be on the more severe side in terms of the severity of COPD, but it is suggested that it maintains a more stable state than S4.

Figure 10 shows the results of a comprehensive analysis of the five patient conditions, including the results of the condition sequence, in addition to the results of Figures 8 and 9. The horizontal axis indicates the severity of COPD, suggesting that the severity increases in the order of S3 (mild) → S1 → S4 → S2 → S0 (severe). In particular, it shows that from S1 to S0, the condition repeatedly worsens and becomes more severe.

States S1 and S4 were considered to be states requiring treatment and prone to exacerbation, S2 was a state of medium to high severity but stable, and S0 was a state requiring treatment (severe). In Figure 10, states requiring treatment are hatched, and stable states are not hatched. Furthermore, states S3 and S1 indicate states of overnutrition and a tendency towards obesity, while S4, S2, and S0 indicate states of malnutrition. It is known that as the condition worsens, frailty progresses and malnutrition occurs, so it was considered that the present results showed a similar tendency.

If the patient's condition is S0, S1, or S4, it is recognizable that treatment is required, and treatment can be administered immediately. In contrast, in condition S2, the condition is stable, but is more severe than condition S4. This shows that while condition S2 is stable and does not require immediate treatment, it is on the more severe side, and so the patient's condition needs to be continuously managed.

Time series analysis suggests that changes in important test items affect the probability of readmission. Therefore, it is possible to infer a patient's condition from changes in important test items. In addition, because the five states of the hidden Markov model represent the physical condition, it is thought that they can also be used as monitoring indicators to infer a patient's condition.

In variant example 2, the server 101 stores in the memory unit 30 an algorithm for creating a heat map of the test values related to the patient's vital signs and blood test items shown in FIG. 8.

The prediction unit 12 predicts a change in the patient's condition by using data on the first and second test items. Specifically, as in the embodiment and the first modified example, the acquisition unit 11 acquires not only data on the first test item related to respiratory function for a patient suffering from COPD, but also data on the second test item related to comorbid diseases, and compares this with each test value corresponding to each state of the heat map in FIG. 8 to determine whether the current state of the patient is S0 to S4. Then, the change in the patient's condition is predicted based on the state transition probability from that state. Although COPD is a respiratory disease, it is also said to be a systemic inflammatory disease, so that the change in the patient's condition may occur not only as an exacerbation of COPD but also throughout the entire body. Therefore, the server according to the embodiment of the present disclosure predicts the change in the patient's condition more accurately by not only predicting the change in the patient's condition based on data on respiratory function, but also by using data on comorbid diseases.

It is preferable that the prediction unit 12 further uses data on a third test item related to the frailty factor to predict changes in the patient's condition. The reason for adding the frailty factor is that it is said that decline in lung function and complications are also related to aging factors, and by using data related to the frailty factor, changes in the patient's condition can be predicted more accurately.

As described above, the server 101 for predicting changes in a patient's condition according to an embodiment of the present disclosure uses not only data on a first test item related to respiratory function, but also data on a second test item related to a comorbid disease, or data on a third test item related to a test item related to frailty factors, making it possible to more accurately predict changes in the condition of patients suffering from chronic obstructive pulmonary disease.

REFERENCE SIGNS LIST 10 Control unit 11 Acquisition unit 12 Prediction unit 20 Communication I/F
30 Storage unit 40 Bus 101, 102 Server 200 Communication network 300 Terminal

Claims (13)

An acquisition unit that acquires data on a first test item related to respiratory function and a second test item related to a comorbidity of a patient suffering from chronic obstructive pulmonary disease (COPD) at the time of discharge from the patient,
a prediction unit that predicts a change in condition of the patient after discharge from the hospital by using data on the first and second test items, respectively;
having
The acquisition unit further acquires data on a third examination item related to a frailty factor,
The prediction by the prediction unit is made using a trained model created based on data on the first, second, and third test items of a plurality of patients suffering from chronic obstructive pulmonary disease (COPD) and data on changes in the condition of the plurality of patients after discharge from the hospital.
A server comprising: The server according to claim 1 , wherein the trained model is created using a Boosting model. The server of claim 2 , wherein the Boosting model is an XGBoost model. The server of claim 1 , wherein the change in the patient's status after discharge from the hospital includes the patient being re-admitted to the hospital. The server according to claim 1 , wherein the data on the third examination item includes data on at least one of age, body mass index (BMI), activities of daily living (ADL), nutritional status, and whether or not the person lives alone. The server according to claim 1 , wherein the data on the second test item related to the comorbid disease includes at least one of data on the name of the comorbid disease, data on a blood test, data on a vital sign test, and prescription data. The server according to claim 1 , further comprising a determination unit that determines which of a plurality of predetermined hierarchies the predicted result of the change in the patient's condition after discharge from the hospital predicted by the prediction unit corresponds to. The server according to claim 1 , further comprising an output unit for outputting the predicted state change. The server according to claim 1, wherein the prediction unit predicts the patient's physical condition and/or its transition probability. An input unit for inputting data on a patient suffering from chronic obstructive pulmonary disease (COPD) at the time of discharge , the data relating to a first test item related to respiratory function , a second test item related to comorbidities, and a third test item related to frailty factors ;
A transmission unit that transmits the input data to a server;
A receiving unit that predicts a change in condition of a patient after discharge from a hospital by using trained models created based on data on the first , second , and third test items of a plurality of patients suffering from chronic obstructive pulmonary disease (COPD) and data on changes in condition of the plurality of patients after discharge from a hospital, and receives the predicted changes in condition from the server;
an output unit that outputs the received state change;
A terminal comprising: Acquiring discharge data of a patient suffering from chronic obstructive pulmonary disease (COPD) on a first test item related to respiratory function , a second test item related to comorbidities, and a third test item related to frailty factors ;
Calculating a predicted value of a change in condition of a patient after discharge from a hospital by using trained models created based on data on the first , second , and third test items of a plurality of patients suffering from chronic obstructive pulmonary disease (COPD) and data on a change in condition of the plurality of patients after discharge from a hospital;
A method for predicting a state change, comprising: On the computer,
Acquiring discharge data of a patient suffering from chronic obstructive pulmonary disease (COPD) on a first test item related to respiratory function , a second test item related to comorbidities, and a third test item related to frailty factors ;
predicting a change in condition of a patient after discharge from a hospital by using trained models created based on data on the first , second , and third test items of a plurality of patients suffering from chronic obstructive pulmonary disease (COPD) and data on a change in condition of the plurality of patients after discharge from a hospital;
A program characterized by executing the above. Acquire teacher data including data on a first test item related to respiratory function of a plurality of patients suffering from chronic obstructive pulmonary disease (COPD) , a second test item related to a comorbid disease , and a third test item related to a frailty factor , the data including data on the patients at the time of discharge from the hospital and data on changes in condition of the plurality of patients after discharge from the hospital;
Generate a trained model in which data on the patient regarding the first , second , and third test items is input and data on changes in the condition of the patient after discharge is output.
A method for generating a trained model.

Images