WO2024089928A1

WO2024089928A1 - Method for assisting advance prediction of target dewatering amount in dialysis, assisting device, and assisting program

Info

Publication number: WO2024089928A1
Application number: PCT/JP2023/020349
Authority: WO
Inventors: 敏男宮田; 翔加藤
Original assignee: 株式会社レナサイエンス
Priority date: 2022-10-28
Filing date: 2023-05-31
Publication date: 2024-05-02

Abstract

The present invention addresses the problem of predicting, in advance, a dewatering amount in dialysis. The above problem is solved by this method which is for assisting advance prediction, with respect to a prediction subject, of a target dewatering amount in dialysis, on the day dialysis is performed. The method enables assistance by using artificial intelligence provided with an algorithm constituted from a first neural network structure and a second neural network structure different from the first neural network structure.

Description

Method, device, and program for supporting prediction of target water removal volume by dialysis

This specification discloses a method, device, and program that uses artificial intelligence to assist in the advance prediction of the target amount of water removal through dialysis.

Intradialytic hypotension is an adverse event that occurs with the removal of fluid during dialysis, and is accompanied by symptoms such as leg cramps, feeling unwell, and loss of consciousness, and also has a negative impact on life prognosis. A drop in blood pressure occurs with a frequency of 5-10%.

Non-Patent Document 1 describes an evaluation of the performance of a Dual-Channel Combiner Network (DCCN) using a dialysis event dataset.
Patent Document 1 describes a personalized DCCN.

International Publication No. 2022-216618A1

If blood pressure drops during dialysis, it can lead to a deterioration in the patient's condition, and so treatment is required from a doctor or nurse. However, while a dialysis room typically houses multiple patients undergoing dialysis, the number of staff members stationed there is limited. For this reason, if a patient's blood pressure suddenly drops during dialysis, many of the staff members stationed in the room will be busy dealing with the situation. Furthermore, if another patient experiences a drop in blood pressure during dialysis, treatment may be delayed.

Therefore, if predictions could be made before dialysis began, the number of staff could be increased in advance, enabling safe and secure dialysis treatment. In addition, if the amount of water to be removed without causing a drop in blood pressure could be known in advance, safer dialysis treatment could be performed.
An object of the present invention is to predict in advance the amount of water removed by dialysis.

An embodiment of the present invention relates to a method for supporting advance prediction of a target water removal amount by dialysis on a dialysis execution date for a prediction subject, the method being supported by artificial intelligence.
The artificial intelligence includes an algorithm composed of a first neural network structure and a second neural network structure different from the first neural network structure. The algorithm is initially trained using the following training data sets (A) and (B) obtained from a plurality of dialysis patients, and the initially trained initial algorithm is further trained using the following (1) and (2) obtained from the prediction subject, thereby constructing the algorithm for each prediction subject.
The artificial intelligence is configured to output the target water removal volume on the dialysis execution day based on the following analysis datasets (i) and (ii):
(A) a dialysis data set obtained from each of the plurality of dialysis patients;
(B) a profile data set including a profile of each of the plurality of dialysis patients and a profile of a dialysis machine used by each of the plurality of dialysis patients;
(1) a data set of multiple past dialysis sessions acquired from the subject;
(2) a profile dataset including a profile of the prediction subject and a profile of a dialysis device used by the prediction subject;
(i) at least one past dialysis data set immediately preceding the date of dialysis administration obtained from the subject;
(ii) a profile dataset including a profile of the predicted subject and a profile of a dialysis machine to be used by the predicted subject;

Preferably, the dialysis data set in (A), (1) and (i) includes the blood flow rate per unit time, the pressure of the dialysis fluid per unit time, the temperature of the dialysis fluid on the day of dialysis, the amount of water removed per set unit time, the target total amount of water removed determined by the doctor, the actual amount of water removed per unit time, the dry weight on the day of dialysis, the previous value of the post-dialysis weight, the weight before the start of dialysis on the day of dialysis, the increase in weight from the previous value of the post-dialysis weight to the start of dialysis on the day of dialysis, the increase from the dry weight on the day of dialysis, the post-dialysis weight on the day of dialysis, the systolic blood pressure and the diastolic blood pressure before the start of dialysis on the day of dialysis, and whether or not a decrease in blood pressure has occurred during dialysis in the past.

Preferably, the profile of each of the multiple dialysis patients in (B) includes the start date of dialysis, the time period during which dialysis is performed, the gender and age group of each dialysis patient, and the profile of the dialysis device includes the type of dialysis method.

Preferably, the profile of the predicted subject in (2) includes the dialysis start date, dialysis time period, gender, and age group of the predicted subject, and the profile of the dialysis device includes the type of dialysis method.

Preferably, the profile of the predicted subject in (ii) includes the predicted subject's dialysis start date, dialysis time period, gender, and age group, and the profile of the dialysis device includes the type of dialysis method.

Preferably, the first neural network has a feedforward neural network structure. The second neural network has a recurrent neural network structure.

More preferably, the feedforward neural network is a multi-layer perceptron, and the recurrent neural network is a Long Short Term Memory (LSTM).

An embodiment of the present invention relates to a support device that supports a prediction of a target water removal amount by dialysis on a dialysis implementation date for a prediction subject. The support device includes a control unit. The control unit outputs the target water removal amount on the dialysis implementation date based on the following analysis datasets (i) and (ii):
(i) at least one past dialysis data set immediately preceding the date of dialysis administration obtained from the subject;
(ii) a profile dataset including a profile of the predicted subject and a profile of a dialysis machine to be used by the predicted subject;
The artificial intelligence comprises an algorithm that is comprised of a first neural network structure and a second neural network structure that is different from the first neural network structure.
The algorithm is initially trained using the following training datasets (A) and (B) obtained from a plurality of dialysis patients, and the initially trained initial algorithm is further trained using the following (1) and (2) obtained from the prediction subject, thereby constructing the algorithm for each prediction subject:
(A) a dialysis data set obtained from each of the plurality of dialysis patients;
(B) a profile data set including a profile of each of the plurality of dialysis patients and a profile of a dialysis machine used by each of the patients;
(1) a data set of multiple past dialysis sessions acquired from the subject;
(2) A profile dataset including a profile of the predicted subject and a profile of a dialysis device to be used by the predicted subject.

An embodiment of the present invention, when executed on a computer, causes the computer to:
The present invention relates to a support program for supporting a prediction of a target water removal amount by dialysis on a dialysis implementation date for a prediction subject, the support program executing a step of outputting the target water removal amount on the dialysis implementation date based on the following analysis datasets (i) and (ii):
(i) at least one past dialysis data set immediately preceding the date of dialysis administration obtained from the subject;
(ii) a profile dataset including a profile of the predicted subject and a profile of a dialysis machine to be used by the predicted subject;
The artificial intelligence comprises an algorithm comprised of a first neural network structure and a second neural network structure different from the first neural network structure.
The algorithm is initially trained using the following training datasets (A) and (B) obtained from a plurality of dialysis patients, and the initially trained initial algorithm is further trained using the following (1) and (2) obtained from the prediction subject, thereby constructing the algorithm for each prediction subject:
(A) a dialysis dataset obtained from each of the plurality of dialysis patients;
(B) a profile data set including a profile of each of the plurality of dialysis patients and a profile of a dialysis machine used by each of the patients;
(1) a data set of multiple past dialysis sessions acquired from the subject;
(2) A profile dataset including a profile of the predicted subject and a profile of a dialysis device to be used by the predicted subject.

For each subject, the target amount of water removed by dialysis on the day of dialysis can be predicted in advance.

1 shows the hardware configuration of the training device 10. 13 shows the process flow of the training program 1042. 2 shows the hardware configuration of the support device 20. 13 shows the process flow of the support program 2042. The breakdown of patients from whom the dataset used in the study was obtained is shown below. This section describes the exclusion rules for the dataset. The breakdown of the dataset is shown below. The consideration steps and prediction accuracy are shown. The prediction accuracy of a conventional method for predicting the amount of water removed and a method for predicting the amount of water removed using a support model are shown. The MAE and MAPE of each group in Verification Example 1 are shown. The results of k-fold cross-validation performed by dividing each of groups 1 to 4 in validation example 1 into five groups are shown below. (A) shows the prediction accuracy of the assistance algorithm constructed using a specimen test data set of 16 items, a dialysis data set, and a profile data set. (B) shows the prediction accuracy of the assistance algorithm constructed using a specimen test data set of 6 items, a dialysis data set, and a profile data set. (C) shows the prediction accuracy of the assistance algorithm constructed using the dialysis data set and a profile data set.

1. Support model for supporting advance prediction of target water removal amount by dialysis One embodiment of the present invention relates to a support model (hereinafter also simply referred to as "support model") that supports advance prediction of a target water removal amount by dialysis on a dialysis implementation day for a prediction subject.
This section describes a training method for training an artificial intelligence to function as an assistance model.

1-1. Overview The artificial intelligence includes an algorithm that is composed of a first neural network structure and a second neural network structure. The first neural network structure and the second neural network structure are different.

For example, since a static parameter data set is input to the first neural network, the first neural network is a neural network used to analyze static parameters. For example, a forward propagation type neural network can be mentioned. Preferably, it is a forward propagation type neural network. Among them, a DNN (Deep Neural Network) is more preferable. More preferably, the first neural network is a fully connected neural network.

For example, the second neural network is a neural network used to analyze dynamic parameters, since a temporal parameter data set is input to the second neural network. For example, a recurrent neural network can be mentioned. A recurrent neural network is preferable. Among them, a Long Short Term Memory (LSTM) is more preferable.

An algorithm with such a different structure is known as a Dual-Channel Combiner Network (DCCN), as described in Non-Patent Document 1 and Patent Document 1. Non-Patent Document 1 and Patent Document 1 are incorporated herein by reference.

The activation function in DCCN can be ReLU (Rectified Linear Unit) or the identity function.

The static parameter dataset may include a profile dataset and a specimen test dataset, preferably a profile dataset. The ad-hoc parameter dataset may include a dialysis dataset. Preferably, the static parameter dataset does not include a specimen test dataset.

The specimen test dataset includes, for example, data from biochemical tests and blood tests. These data can be obtained by specimen tests at ordinary hospitals or testing centers. The specimen test dataset may preferably include serum albumin concentration, red blood cell count, hematocrit value, mean corpuscular hemoglobin (MCH), and platelet count. The specimen test dataset preferably includes serum albumin concentration, serum BUN concentration, serum creatinine concentration, serum calcium concentration, corrected calcium concentration, serum inorganic phosphorus concentration, serum iron concentration, total iron binding capacity (TIBC), serum sodium concentration, serum potassium concentration, serum chloride concentration, white blood cell count, red blood cell count, hemoglobin concentration, hematocrit value, mean corpuscular hemoglobin (MCH), platelet count, and serum CRP concentration. Specimen tests are generally performed about twice a month.

The dialysis dataset includes blood flow rate per unit time, dialysis fluid pressure per unit time, dialysis fluid temperature on the day of dialysis, set amount of water removed per unit time, target total amount of water removed determined by the doctor, actual amount of water removed per unit time, dry weight on the day of dialysis, previous post-dialysis weight, weight before the start of dialysis on the day of dialysis, weight increase from the previous post-dialysis weight to the start of dialysis on the day of dialysis, weight increase from the dry weight on the day of dialysis, post-dialysis weight on the day of dialysis, systolic blood pressure and diastolic blood pressure before the start of dialysis on the day of dialysis, and whether or not a drop in blood pressure has occurred during dialysis in the past. The dialysis dataset is data acquired during each dialysis. Dialysis is generally performed about three times a week.

"Every unit time" in the above means every unit time from the start of dialysis. Preferably, "every unit time" is approximately every hour. One dialysis session for one patient generally takes about four hours, and for some patients, five hours. For this reason, each piece of data is obtained approximately one hour, two hours, three hours, four hours, and for some patients, approximately five hours from the start of dialysis.

In principle, dry weight can be calculated from the cardiothoracic ratio of each dialysis patient. The cardiothoracic ratio is calculated from the images taken of chest X-rays taken once or twice a month for each dialysis patient. The standard value for the cardiothoracic ratio is approximately 50%, so doctors determine the dry weight based on this standard value. However, for some dialysis patients, determining the amount of water removal during dialysis based on the calculated dry weight can cause a drop in blood pressure during dialysis. In such cases, doctors will make adjustments such as increasing the dry weight for each patient.

Regarding whether or not there has been a drop in blood pressure during dialysis in the past, a drop in blood pressure can be defined as a drop in systolic blood pressure of 20 mmHg or more from the previous measurement and a drop of 110 mmHg or less.

The profile dataset includes profiles of dialysis patients from which training data is obtained or prediction subjects, and profiles of the types of dialysis methods used by dialysis patients from which training data is obtained or prediction subjects. The profile data includes data that is basically unchanged or that changes on an annual basis, such as age groups.

The profile of the dialysis patient from whom training data is obtained, or the prediction subject, includes the time of day when dialysis is performed (morning or afternoon), the start date of dialysis, gender, and age group.
Age groups can be quantified, for example, into teens, twenties, thirties, forties, fifties, sixties, seventies, eighties, and nineties.
In principle, the time of day when dialysis is performed (morning or afternoon) is the same for each patient.
Types of dialysis procedures are hemodialysis (HD) therapy or hemodiafiltration (HDF) therapy.

For the various types of data mentioned above, if the data is continuous quantitative data (numeric data), the numerical value is used; if the data is qualitative data (non-continuous data) or an age group, a label value according to the type of quality is used.

The training of the artificial intelligence to build the assistance model is divided into two stages. In the first stage, the algorithm is initially trained using a training dataset obtained from multiple dialysis patients.

For initial training, (A) a dialysis dataset obtained from each of a number of dialysis patients, and (B) a profile dataset including a profile for each of the multiple dialysis patients and a profile for the dialysis device used by each of the patients are used. The above datasets (A) and (B) are linked to each patient. The above training datasets (A) and (B) are referred to as the initial training dataset.

Since each dialysis patient undergoes multiple dialysis sessions, the initial training uses all possible input data sets (A) and (B) obtained for each of the multiple patients.

In the initial training, the initial training data set is input to the input layer of the algorithm, and the target water removal volume set by the doctor for each patient's previous dialysis session is input to the output layer of the algorithm. The algorithm trained in this way is called the initial algorithm (or trained DCCN).

Here, it is preferable that the dialysis patients from which each data set used in the initial training is obtained are dialysis patients who have not experienced any events such as a drop in blood pressure, and whose doctor-determined target volume of water removal and actual volume of water removal are less than 100 mL.

The second stage of AI training is training to individualize the initial algorithm to suit the predicted subject. In this training, in order to individualize the initial algorithm to suit a specific predicted subject, (1) a dataset of multiple past dialysis sessions obtained from the specific predicted subject, and (2) a profile dataset including a profile of the specific predicted subject and a profile of the dialysis device used by the predicted subject are used. The above training datasets (1) and (2) are referred to as individual training datasets.

In training for individualization, all of the inputtable data sets (1) and (2) obtained for a specific prediction subject are used. In training for individualization, the individual training data set is input to the input layer of the initial algorithm, and the target water removal volume set by the doctor for each past dialysis session for the specific prediction subject is input to the output layer of the initial algorithm. An algorithm trained in this way is called an individualized algorithm (or P-DCCN).

The individualized algorithm constructed in this manner is used as a support model to assist in the advance prediction of the target water removal volume by dialysis on the day of dialysis for the specific prediction subject. By using the individualized algorithm, higher prediction accuracy can be obtained. Furthermore, the individualized algorithm can accurately predict the target water removal volume for the prediction subject even if the facility that obtained the initial training data is different from the facility that obtained the individualized training data. Furthermore, since the prediction subject receives dialysis therapy continuously, the individual training dataset is accumulated each time. The individualized algorithm is retrained with the added individual training dataset each time an individual training dataset for the prediction subject is added, thereby becoming a support model that is more suited to each prediction subject.

1-2. Training device for assistance model Fig. 1 shows the hardware configuration of a training device for an assistance model (hereinafter, simply referred to as a "training device") 10 that supports advance prediction of a target amount of water removal by dialysis on a dialysis day for a prediction subject.
The training device 10 may be connected to an input device 111 and an output device 112 .

In the training device 10, the processing unit (CPU) 101, memory 102, ROM (read only memory) 103, storage device 104, and interface 106 are connected to each other via a bus 109 so that they can communicate data with each other. The processing unit 101, memory 102, and ROM 103 function as the control unit 100 of the training device 10.

The processing unit 101 is the CPU of the training device 10 and is also called a calculation device. The processing unit 101 may work in cooperation with a GPU or MPU. The processing unit 101 works in cooperation with an operation system (OS) 1041 stored in the storage device 104 or ROM 103 to execute a training program 1042 (hereinafter also simply referred to as the "training program 1042") of an assistance model that supports advance prediction of the target water removal amount by dialysis on the day of dialysis for a prediction subject described below, and the computer functions as the training device 10.

The ROM 103 stores the training program 1042 executed by the processing unit 101 and data used therein. The ROM 103 stores the boot program executed by the processing unit 101 when the training device 10 is started up, as well as programs and settings related to the operation of the hardware of the training device 10.

The storage device 104 non-volatilely stores an operation system (OS) 1041, a training program 1042 for training the assistance model described below (hereinafter simply referred to as the "training program 1042"), and a model database 1043. The model database 1043 stores pre-training algorithms, or initial algorithms, and individualized algorithms. The model database 1043 can also store initial training data and individual training data.

The input device 111 is composed of a touch panel, a keyboard, a mouse, a pen tablet, a microphone, etc., and is used to input characters or voice to the training device 10. The input device 111 may be connected from outside the training device 10 or may be integrated with the training device 10.
The output device 112 is composed of, for example, a display, a printer, and the like.

The processing unit 101 may obtain application software and various settings necessary for controlling the training device 10 via a network instead of reading them from the ROM 103 or the storage device 104. The application program may be stored in a storage device of a server computer on the network, and the training device 10 may access this server computer to download a training program 1042 and store it in the ROM 103 or the storage device 104.

Also, an operation system that provides a graphical user interface environment, such as Windows (registered trademark) manufactured and sold by Microsoft Corporation in the United States or the open source Linux (registered trademark), is installed in the ROM 103 or the storage device 104. The training program runs on the operating system. In other words, the training device 10 can be a personal computer, etc.

2 shows a flow of processing performed by the training program 1042. The control unit 100 executes the training program 1042, whereby the computer functions as the training device 10.
The control unit 100 receives a processing start request inputted, for example, by an operator from the input device 111, executes the training program 1042, and starts the training processing.

In step S11, the control unit 100 reads the algorithm to be trained and the initial training data set from the model database 1043, inputs the initial training data set to the input layer of the algorithm, and inputs the target water removal volume to the output layer of the algorithm. The explanation of these inputs is given in 1-1 above.

In step S12, the control unit 100 trains the algorithm based on the data set input to the algorithm in step S11, and constructs an initial algorithm. The control unit 100 stores the initial algorithm in the storage device 104.

In step S13, the control unit 100 reads the initial algorithm and the individualized training data set from the storage device 104 in step S12, and inputs the individualized training data set for the specific prediction target into the input layer of the initial algorithm. The explanation of these inputs is as described in 1-1 above.

In step S14, the control unit 100 trains the initial algorithm based on the data set input to the initial algorithm, and constructs an individualization algorithm for each prediction target person.

In step S14, the control unit 100 accepts a command from the operator to start the verification process and verifies the individualization algorithm. The verification can be performed using the error between the target water removal volume set by the doctor for each prediction subject and the predicted water removal volume for each prediction subject predicted by the individualization algorithm as an index. For example, the index is the accuracy rate, the mean absolute percentage error (MAPE), and the mean absolute error (MAE). MAPE and MAE are calculated by the following formula.

The accuracy rate is the accuracy rate of the predicted water removal volume of each prediction target predicted by the individualization algorithm when the difference between the predicted water removal volume of each prediction target predicted by the individualization algorithm and the target water removal volume prescribed by the dialysis specialist is 250 ml or less (patients with a total water removal volume of less than 2350 ml in one dialysis session), or the difference between the predicted water removal volume of each prediction target predicted by the individualization algorithm and the target water removal volume prescribed by the dialysis specialist is within 10.7% (total water removal volume of 2350 ml or more in one dialysis session). The higher the accuracy rate, the smaller the error. The smaller the MAPE and MAE values, the smaller the error. Therefore, when the accuracy rate is higher than the standard value, or when MAPE or MAE is smaller than a specified standard value, it can be evaluated that the individualization algorithm is suitable for the prediction target. In this verification, if the accuracy rate is higher than the reference value, or if the MAPE, MAE, or RMSE is greater than a predetermined reference value, the operator reviews the initial training data set, adjusts the weights of each function in each neural network, and returns to step S11 to perform training again. Acceptable accuracy rates are, for example, 80% or more, 85% or more, 88% or more, and 90% or more. Acceptable MAPE is, for example, 15% or less, preferably 10% or less, and more preferably 8% or less.

2. Prediction of target water removal volume using support model 2-1. Overview A method for supporting the advance prediction of the target water removal volume by dialysis on the day of dialysis for a prediction subject (hereinafter, simply referred to as the "support method") can obtain the target water removal volume (e.g., mL) as an output value by inputting the analysis dataset of each prediction subject into the individualization algorithm constructed for each prediction subject in 1-3 above. The analysis dataset is (i) at least one past dialysis dataset immediately prior to the day of dialysis obtained from the prediction subject, and (ii) a profile dataset including a profile of the prediction subject and a profile of the dialysis device used by the prediction subject. The explanation of each dataset is incorporated herein by reference in 1-1 above.

Here, the number of sets of dialysis datasets is at least the most recent one going back from the date of dialysis, and preferably the most recent five going back from the date of dialysis. The dialysis dataset is acquired every time dialysis is performed, but since specimen test data is acquired about twice a month, one set of specimen test data is acquired for every five dialysis datasets acquired. The profile dataset does not change in principle, except for the age group.

2-2. Support device for supporting advance prediction of target water removal amount by dialysis Figure 3 shows the hardware configuration of a support device (hereinafter also simply referred to as "support device") 20 that supports advance prediction of a target water removal amount by dialysis on the day of dialysis for a prediction subject.
The assistance device 20 may be connected to an input device 211 and an output device 212 .
In the support device 20, a processing unit (CPU) 201, a memory 202, a ROM (read only memory) 203, a storage device 204, and an interface 206 are connected to each other via a bus 209 so as to be able to communicate data with each other. The processing unit 101, the memory 202, and the ROM 203 function as a control unit 200 of the support device 20.
Each component of the support device 20 is similar to the corresponding component of the training device 10 except for the configuration of the storage device 204 .

The storage device 204 non-volatilely stores an operation system (OS) 2041, a prediction support program 2042 (hereinafter simply referred to as the "support program 2042") that supports advance prediction of the target amount of water removal by dialysis on the day of dialysis for a prediction subject described below, and a model database 2043. The model database 2043 stores the individualization algorithm after training. The model database 2043 can also store analysis data.

4 shows a flow of processing performed by the assistance program 2042. The control unit 200 executes the assistance program 2042, causing the computer to function as the assistance device 20.
The control unit 200 receives a processing start request inputted, for example, by an operator from the input device 211, executes the assistance program 2042, and starts the assistance processing.

In step S21, the control unit 200 reads out the personalization algorithm corresponding to the person to be predicted and the analysis dataset of the prediction target from the model database 2043, and inputs the analysis dataset of the person to be predicted into the input layer of the personalization algorithm.

In step S22, the control unit 200 outputs the predicted value output from the individualization algorithm to the output device 212 as the target water removal amount for the predicted individual.

In step S23, the control unit 200 determines whether the predicted value for the prediction subject output from the individualization algorithm is greater than the dry weight of the prediction subject. For example, if the predicted value is 3% or more of the dry weight two days after the last dialysis, or if the predicted value is 5% or more of the dry weight three days after the last dialysis, it can be determined that the predicted value is greater than the dry weight.

If the control unit 200 determines in step S23 that the predicted value is greater than the dry weight (if "YES"), the control unit 200 proceeds to step S24 and outputs a warning to the output device 212 to prompt the user to review the dry weight. The warning may be, for example, a symbol such as an exclamation point, or a text message such as "Please check the dry weight."

If the control unit 200 determines in step S23 that the predicted value is not greater than the dry weight (if "NO"), it ends the process.
Steps S23 and S24 are optional processes.

3. Storage medium storing a program An embodiment of the present invention relates to a program product, such as a media drive, that stores the training program 1042, the assistance program 2042, and/or the retraining program. That is, the training program 1042, the assistance program 2042, and/or the retraining program may be stored in a media drive, such as a hard disk, a semiconductor memory device such as a flash memory, or an optical disk. The media drive may also be a computer, such as a server device. The format of recording the program to the media drive is not limited as long as each device can read the program. It is preferable that the recording to the media drive is non-volatile.

4. Modifications When the number of dialysis sessions of a prediction subject increases, the individualization algorithm may be retrained by adding a data set of the increased number of dialysis sessions to a new individual training data set, or by updating the individual training data set.

The individualization algorithm constructed by the method described in 1-1 above can also be used as a support model to assist in predicting the occurrence of a drop in blood pressure during dialysis by changing the data input to the output layer from the target amount of water removal set by the doctor to the presence or absence of a drop in blood pressure during dialysis in the past. In this case, the activation function uses a sigmoid function instead of ReLU or an identity function. The value output from the support model to assist in predicting the occurrence of a drop in blood pressure during dialysis is the probability that each prediction subject will experience a drop in blood pressure during dialysis.

5. Verification of the effect 5-1. Verification example 1
Dialysis datasets, specimen test datasets, and profile datasets were obtained from patients undergoing dialysis at Facility A, Facility H, Facility K, and Facility S. The breakdown is shown in Figure 5.

Next, for the patients shown in Figure 5, data from patients who underwent a small number of dialysis sessions and patients with missing data were excluded according to Figure 6, and ultimately data from 1,828 patients and 725,619 dialysis sessions was used for analysis. Furthermore, the data was divided into a learning (training) dataset, an inspection dataset, an adjustment dataset, and an evaluation dataset in the proportions shown in Figure 7, and the assistance algorithm was trained and its accuracy was evaluated.

The dialysis data set consisted of the day of the week on which dialysis was performed, temperature, weather, blood flow rate per unit time, dialysis fluid pressure per unit time, dialysis fluid temperature, set amount of water removed per unit time, target amount of water removed, actual amount of water removed per unit time, dry weight, previous post-dialysis weight, weight before the start of dialysis, weight gain from the date of the previous dialysis until before the start of the next dialysis, gain from dry weight, post-dialysis weight, systolic and diastolic blood pressure before the start of dialysis on the day of dialysis, and whether or not blood pressure drops occurred during dialysis in the past. A drop in blood pressure during dialysis was defined as a drop in systolic blood pressure of 20 mmHg or more from the previous measurement and a drop of 110 mmHg or less.

　Specimen test data included serum albumin concentration, serum BUN concentration, serum creatinine concentration, serum calcium concentration, corrected calcium concentration, serum inorganic phosphorus concentration, serum iron concentration, total iron binding capacity (TIBC), serum sodium concentration, serum potassium concentration, serum chloride concentration, white blood cell count, red blood cell count, hemoglobin concentration, hematocrit value, mean corpuscular hemoglobin (MCH), platelet count, and serum CRP concentration.

The profile of the predicted subject was the dialysis start date, gender, age group, and time of day when dialysis was performed (morning or afternoon), and the profile of the dialysis method was the type of dialysis method (HD or HDF).

5-1-2. Study steps and prediction accuracy The study history of this support model is shown in Figure 8. The prediction item is the occurrence or nonoccurrence of hypotension during dialysis.

In

steps

1 and 2, learning and validation were performed at the same facility. DCCN was used as the support model, and the training and validation datasets were randomly sampled data obtained at the same facility. These support models had an AUC (area under curve) of over 0.8, and the prediction accuracy was generally good. However, when a validation dataset obtained at a facility different from the training dataset was used in step 3, the AUC of the DCCN support model was 0.73. When DCCN was individualized using the individual training dataset of the prediction subject, the AUC increased to 0.79. Furthermore, when the amount of the training dataset was increased in step 4, the AUC became 0.91. In addition, the accuracy of predicting the occurrence of hypotension during dialysis, including the absence of events, was 0.91 AUC. For predicting event occurrence alone, the area under the precision-recall curve (AUPRC) was 0.68.

Figure 9 shows the prediction accuracy (RMSE) of the amount of water removed. The accuracy of the amount of water removed predicted by DCCN was compared with the accuracy of the amount of water removed predicted by other methods. The RMSE of the DCCN prediction was 183.36 mL, which was significantly better than other conventional prediction methods.

Next, the above dialysis dataset (including multiple dialysis data for one dialysis patient) was divided into the following four groups, excluding datasets with missing data, and the dataset of Group 1 and the corresponding specimen test dataset and profile dataset were used to train an initial algorithm, and an individualized algorithm was constructed for the dialysis datasets of Groups 1 to 4, and its accuracy was evaluated. The amount of water removed was predicted for Groups 1 to 4. The dialysis dataset of Group 1 contained approximately 400,000 pieces of data. The dialysis datasets of Groups 2 to 4 contained approximately 120,000 pieces of data.
Group 1: No events and [(physician-prescribed amount of fluid removed) - (actual amount of fluid removed)] < 100 ml (dialysis achieved fluid removal)
Group 2: No events and [(physician-prescribed amount of fluid removed) - (actual amount of fluid removed)] ≥ 100 ml (dialysis in which fluid removal was not achieved)
Group 3: Events occurred and [(physician-prescribed amount of fluid removed) - (actual amount of fluid removed)] < 100 ml (dialysis achieved fluid removal)
Group 4: No events and [(physician-prescribed amount of fluid removed) - (actual amount of fluid removed)] ≥ 100 ml (dialysis in which fluid removal was not achieved)
* An "event" indicates a drop in blood pressure during dialysis.

The difference between the target water removal volume set by the doctor and the water removal volume predicted by the assistance algorithm for the above four groups, and the MAPE are shown in Figure 10. Figure 11 shows the results of k-fold cross-validation in which each of groups 1 to 4 was divided into five groups.
The MAPE was less than 10.7% in all cases. Setting the amount of water removed from dialysis patients is influenced by factors such as the patient's condition, weather (the amount of insensible perspiration can easily change depending on whether or not there is a bowel movement, and temperature and humidity), the experience and knowledge of each staff member, and flow rate errors in the dialysis device itself, so even when prescribed by a specialist, there can be errors in the amount of water removed of more than a cup (250 ml, MAPE 10.7%), and so the results of this analysis were considered to be good.

5-2. Verification example 2
Next, the following dialysis datasets were used: blood flow rate per unit time, dialysis fluid pressure per unit time, dialysis fluid temperature on the day of dialysis, set hourly water removal volume per unit time, target total water removal volume determined by the doctor, actual water removal volume per unit time, dry weight on the day of dialysis, previous post-dialysis weight, weight before the start of dialysis on the day of dialysis, weight increase until the start of dialysis on the day of dialysis based on the previous post-dialysis weight, increase from dry weight on the day of dialysis, post-dialysis weight on the day of dialysis, systolic blood pressure and diastolic blood pressure before the start of dialysis on the day of dialysis, and whether or not a drop in blood pressure had occurred during dialysis in the past.The predictive accuracy of the trained support algorithm was examined using the following profile datasets: start date of dialysis, time period of dialysis, gender, and age group of each dialysis patient, and the type of dialysis method as the profile of the dialysis device.

　For comparison, in addition to the above dialysis dataset and profile dataset, a training assistance algorithm was constructed using a specimen test dataset (data consisted of 16 items: serum albumin concentration, serum BUN concentration, serum creatinine concentration, serum calcium concentration, serum inorganic phosphorus concentration, serum iron concentration, total iron binding capacity (TIBC), serum sodium concentration, serum potassium concentration, serum chloride concentration, white blood cell count, red blood cell count, hemoglobin concentration, hematocrit value, mean corpuscular hemoglobin (MCH), and platelet count) or a specimen test dataset (data consisted of 6 items: serum albumin concentration, serum BUN concentration, serum creatinine concentration, serum sodium concentration, hematocrit value, and total protein). The items entered as the analysis dataset were also the same as those in the training dataset for the assistance algorithm.

In addition, we obtained dialysis datasets, specimen test datasets, and profile datasets for each patient undergoing dialysis at sites A, H, and K.

In this study, dialysis in which the doctor-set volume of water removed was 1,000 ml or less or 5,000 ml or more was excluded from the calculation of accuracy rate, MAE, and MAPE. In addition, patients who had been on dialysis for less than one year were excluded from both the training dataset and the evaluation dataset.

The results are shown in Figure 12. Figure 12 (A) shows the prediction accuracy of the assistance algorithm constructed using a specimen test dataset of 16 items, a dialysis dataset, and a profile dataset. Figure 12 (B) shows the prediction accuracy of the assistance algorithm constructed using a specimen test dataset of 6 items, a dialysis dataset, and a profile dataset. Figure 12 (C) shows the prediction accuracy of the assistance algorithm constructed using the dialysis dataset and profile dataset. As is clear from Figure 12, the assistance algorithm constructed using the dialysis dataset and profile dataset shown in Figure 12 (C) showed high prediction accuracy in terms of accuracy rate, MAE, and MAPE. This result shows that the prediction accuracy of the assistance algorithm is improved when the specimen test dataset is not used.

20 Support device 200 Control unit

Claims

A method for supporting a prior prediction of a target water removal amount by dialysis on a dialysis implementation date for a prediction subject, comprising:
The method is assisted by artificial intelligence;
The artificial intelligence is
an algorithm comprising a first neural network structure and a second neural network structure different from the first neural network structure;
The model is initially trained using the following training datasets (A) and (B) obtained from multiple dialysis patients:
The initially trained initial algorithm is further trained using the following (1) and (2) obtained from the prediction target, thereby constructing the algorithm for each prediction target;
The artificial intelligence is configured to output the target water removal amount on the dialysis execution day based on the following analysis datasets (i) and (ii):
(A) a dialysis data set obtained from each of the plurality of dialysis patients;
(B) a profile data set including a profile of each of the plurality of dialysis patients and a profile of a dialysis machine used by each of the plurality of dialysis patients;
(1) a data set of multiple past dialysis sessions acquired from the subject;
(2) a profile dataset including a profile of the prediction subject and a profile of a dialysis device used by the prediction subject;
(i) at least one past dialysis data set immediately preceding the date of dialysis administration obtained from the subject;
(ii) a profile dataset including a profile of the predicted subject and a profile of a dialysis machine to be used by the predicted subject;
The dialysis data set in (A), (1) and (i) includes the blood flow rate per unit time, the pressure of the dialysis fluid per unit time, the temperature of the dialysis fluid on the day of dialysis, the amount of water removed per set unit time, the target total amount of water removed determined by the doctor, the actual amount of water removed per unit time, the dry weight on the day of dialysis, the previous value of the post-dialysis weight, the weight before the start of dialysis on the day of dialysis, the increase in weight from the previous value of the post-dialysis weight to the start of dialysis on the day of dialysis, the increase from the dry weight on the day of dialysis, the post-dialysis weight on the day of dialysis, the systolic blood pressure and the diastolic blood pressure before the start of dialysis on the day of dialysis, and whether or not a decrease in blood pressure has occurred during dialysis in the past.
The profile of each of the plurality of dialysis patients in (B) includes a dialysis start date, a dialysis time period, a gender, and an age group of each dialysis patient, and the profile of the dialysis device includes a type of dialysis method;
The profile of the predicted subject in (2) includes the dialysis start date, dialysis time zone, gender and age group of the predicted subject, and the profile of the dialysis device includes the type of dialysis method. The profile of the predicted subject in (ii) includes the dialysis start date, dialysis time zone, gender and age group of the predicted subject, and the profile of the dialysis device includes the type of dialysis method.
The method of claim 1.
The method of claim 1, wherein the first neural network has a feedforward neural network structure and the second neural network has a recurrent neural network structure.
The method of claim 3, wherein the feedforward neural network is a multi-layer perceptron and the recurrent neural network is a Long Short Term Memory (LSTM).
A support device for supporting a prediction of a target water removal amount by dialysis on a dialysis implementation date for a prediction subject,
The support device includes a control unit,
The control unit outputs the target water removal amount on the dialysis implementation day based on the following analysis data sets (i) and (ii),
(i) at least one past dialysis data set immediately preceding the date of dialysis administration obtained from the subject;
(ii) a profile dataset including a profile of the predicted subject and a profile of a dialysis machine to be used by the predicted subject;
The artificial intelligence is
an algorithm comprising a first neural network structure and a second neural network structure different from the first neural network structure;
The model is initially trained using the following training datasets (A) and (B) obtained from multiple dialysis patients:
The initially trained initial algorithm is further trained using the following (1) and (2) obtained from the prediction subject, thereby constructing the prediction subject for each prediction subject.
Support equipment:
(A) a dialysis dataset obtained from each of the plurality of dialysis patients;
(B) a profile data set including a profile of each of the plurality of dialysis patients and a profile of a dialysis machine used by each of the patients;
(1) a data set of multiple past dialysis sessions acquired from the subject;
(2) A profile dataset including a profile of the predicted subject and a profile of a dialysis device to be used by the predicted subject.
When the program is executed by a computer, the computer
A support program for supporting a prior prediction of a target water removal amount by dialysis on a dialysis implementation date for a prediction subject, the support program executing a step of outputting the target water removal amount on the dialysis implementation date based on the following analysis datasets (i) and (ii):
The artificial intelligence is
an algorithm comprising a first neural network structure and a second neural network structure different from the first neural network structure;
The model is initially trained using the following training datasets (A) and (B) obtained from multiple dialysis patients:
The initially trained initial algorithm is further trained using the following (1) and (2) obtained from the prediction subject, thereby constructing the prediction subject for each prediction subject.
Support programs:
(A) a dialysis data set obtained from each of the plurality of dialysis patients;
(B) a profile data set including a profile of each of the plurality of dialysis patients and a profile of a dialysis machine used by each of the patients;
(1) a data set of multiple past dialysis sessions acquired from the subject;
(2) a profile dataset including a profile of the prediction subject and a profile of a dialysis device used by the prediction subject;
(i) at least one past dialysis data set immediately preceding the date of dialysis administration obtained from the subject;
(ii) a profile dataset including a profile of the predicted subject and a profile of a dialysis machine to be used by the predicted subject;