JP6909078B2 - Disease onset prediction device, disease onset prediction method and program - Google Patents

Description

The present invention relates to a disease onset prediction device, a disease onset prediction method and a program.

Conventional medical device devices issue alerts based on simple judgments such as the patient's biometric data falling below a certain value. In recent years, techniques for predicting diseases by analyzing biological data of patients in real time have been studied. For example, Patent Document 1 discloses a disease prediction device that obtains the relationship between two parameters indicating the biological data of a patient and compares it with a statistical value to analyze a sign of a disease. Further, Patent Document 2 discloses a method of monitoring a person's health condition by modeling a person's biological data and analyzing the difference between an estimated value obtained from the model and an actual measured value. ..

JP-A-2016-040693 Special Table 2009-517188

When a disease develops, biometric data often show signs for hours. If such signs can be caught, it will be possible to prevent the outbreak of the disease by accurately predicting the onset of the disease for the patient and communicating it to doctors, nurses, etc. in the medical field.

However, the technique disclosed in Patent Document 1 only shows that the patient has changed by comparing with the statistical values, and cannot accurately capture the signs appearing before the onset of a specific disease. Further, the method disclosed in Patent Document 2 only identifies whether the actual measured value is in a healthy state or not, and does not predict the onset of a specific disease.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a disease onset prediction device, a disease onset prediction method, and a program capable of accurately predicting the onset of a specific disease in advance.

In order to achieve the above object, the disease onset prediction device according to the first aspect of the present invention is
A biometric data acquisition means for acquiring biometric data to be predicted and biometric data for constructing a prediction model,
An onset information acquisition means for acquiring onset information indicating whether or not a patient has developed a disease corresponding to the biological data for constructing the prediction model, and an onset information acquisition means.
When the onset information indicates that the disease has developed, a positive sample is obtained from the biometric data for constructing the prediction model from the biometric data from before the reference period at the onset time of the onset of the disease to the onset time. A sample extraction means for extracting biological data and extracting biological data as a negative sample from biological data before the reference period at the time of onset.
The positive sample is used as training data for developing the disease within the period from the reference time, which is the time point before the reference period at the time of onset, to the time after the reference period , and the negative sample is used from the reference time to the above. A predictive model construction means for constructing a predictive model by using it as training data that does not develop the disease within the period up to after the reference period.
Said time obtaining the prediction target biometric data and the reference time, on the basis the prediction target biometric data to said predictive model, the patient corresponding to the biometric data of the prediction target, the reference from the time before Symbol reference It is provided with an onset prediction means for predicting whether or not the disease will develop within the period up to after the period.
In the sample extraction means, the result of performing a predetermined calculation on the extracted biological data to be a positive sample is defined as the positive sample, and the result of performing a predetermined calculation on the biological data to be a negative sample is defined as the negative sample. ..

When the onset information indicates that the patient has developed the disease, the sample extraction means may use the biological data for constructing the prediction model from before the total period of the reference period and the predetermined period at the time of onset. A plurality of biological data up to the time of onset and corresponding to the predetermined period are extracted so as to partially overlap the predetermined period to obtain the positive sample.
You may do so.

It is provided with an action information acquisition means for acquiring action information indicating the content of an action performed on a patient corresponding to the biological data for constructing the prediction model.
Based on the action information, the sample extraction means preferentially extracts from the biological data for constructing the prediction model to obtain the positive sample or the negative sample.
You may do so.

When the action information indicates that the action information was an action for avoiding the onset of the disease, the sample extraction means means that the patient corresponding to the biological data for constructing the prediction model developed the disease. Assuming that the onset information indicates that the disease has occurred, the biometric data serving as the positive sample is extracted.
You may do so.

The action information acquisition means further acquires the prediction target action information representing the content of the action performed on the patient corresponding to the biometric data of the prediction target predicted to develop the disease by the onset prediction means.
The biometric data acquisition means newly acquires data corresponding to the biometric data of the prediction target of the patient in which the act is performed as the biometric data for constructing the prediction model.
When the sample extraction means indicates that the prediction target action information is not an action for avoiding the onset of the disease, the newly acquired biological data for constructing a prediction model is used as another living body. The negative sample is extracted with priority over the data.
You may do so.

When the action information indicates that the action information was not an action for avoiding the onset of the disease and the onset information indicates that the disease has developed, the biometric data of the prediction target. Based on, set the conditions for outputting cautionary information indicating that actions should be taken to avoid the onset of the disease.
When the prediction target action information indicates that the action was not an action for avoiding the onset of the disease, the caution information output that outputs the caution information based on the conditions set by the prediction model construction means. Equipped with more means
You may do so.

When the onset prediction means predicts that the disease will develop, the causative components predicted to develop the disease are classified by a predetermined method based on the biological data of the prediction target, and output as alert information. Further equipped with alert information output means,
You may do so.

In order to achieve the above object, the disease onset prediction method according to the second aspect of the present invention is
A biometric data acquisition step for acquiring biometric data to be predicted and biometric data for building a predictive model,
An onset information acquisition step for acquiring onset information indicating whether or not a patient has developed a disease corresponding to the biological data for constructing the prediction model, and an onset information acquisition step.
When the onset information indicates that the disease has developed, a positive sample is obtained from the biometric data for constructing the prediction model from the biometric data from before the reference period at the onset time of the onset of the disease to the onset time. A sample extraction step of extracting biometric data to be a negative sample from biometric data prior to the reference period at the time of onset.
The positive sample is used as training data for developing the disease within the period from the reference time, which is the time point before the reference period at the time of onset, to the time after the reference period , and the negative sample is used from the reference time to the above. A predictive model building step to build a predictive model by using it as training data that does not develop the disease within the period up to after the reference period.
Said time obtaining the prediction target biometric data and the reference time, on the basis the prediction target biometric data to said predictive model, the patient corresponding to the biometric data of the prediction target, the reference from the time before Symbol reference It is provided with an onset prediction step for predicting whether or not the disease will develop within the period up to after the period.
In the sample extraction step, the result of performing a predetermined calculation on the extracted biological data to be a positive sample is defined as the positive sample, and the result of performing a predetermined calculation on the biological data to be a negative sample is defined as the negative sample. ..

In order to achieve the above object, the program according to the third aspect of the present invention is
On the computer
A biometric data acquisition step for acquiring biometric data to be predicted and biometric data for building a predictive model,
An onset information acquisition step for acquiring onset information indicating whether or not a patient has developed a disease corresponding to the biological data for constructing the prediction model, and an onset information acquisition step.
When the onset information indicates that the disease has developed, a positive sample is obtained from the biometric data for constructing the prediction model from the biometric data from before the reference period at the onset time of the onset of the disease to the onset time. A sample extraction step of extracting biometric data to be a negative sample from biometric data prior to the reference period at the time of onset.
The positive sample is used as training data for developing the disease within the period from the reference time, which is the time point before the reference period at the time of onset, to the time after the reference period , and the negative sample is used from the reference time to the above. A predictive model building step to build a predictive model by using it as training data that does not develop the disease within the period up to after the reference period.
Said time obtaining the prediction target biometric data and the reference time, on the basis the prediction target biometric data to said predictive model, the patient corresponding to the biometric data of the prediction target, the reference from the time before Symbol reference The onset prediction step of predicting whether or not the disease will develop within the period up to the period after the period is executed.
In the sample extraction step, the result of performing a predetermined calculation on the extracted biological data to be a positive sample is defined as the positive sample, and the result of performing a predetermined calculation on the biological data to be a negative sample is defined as the negative sample. ..

According to the present invention, it is possible to provide a disease onset prediction device, a disease onset prediction method and a program capable of accurately predicting the onset of a specific disease in advance.

It is a figure which shows the outline of the alert system which concerns on embodiment of this invention. It is a functional block diagram of the alert system which concerns on embodiment of this invention. It is a figure which shows the sample extraction method of the alert system which concerns on embodiment of this invention. It is a figure which shows an example of the extracted sample of the alert system which concerns on embodiment of this invention. It is a hardware block diagram of the alert system which concerns on embodiment of this invention. It is a flowchart of the prediction model construction process of the alert system which concerns on embodiment of this invention. It is a figure which shows an example of the principal component extracted in the alert cause classification condition setting process of the alert system which concerns on embodiment of this invention. It is a figure which shows an example of the principal component score given in the alert cause classification condition setting process of the alert system which concerns on embodiment of this invention. It is a figure which shows an example of the sample which was given the classification in the alert cause classification condition setting process of the alert system which concerns on embodiment of this invention. It is a flowchart of the prediction model reconstruction process of the alert system which concerns on embodiment of this invention. It is a figure which shows an example of the setting value in the prediction model reconstruction processing of the alert system which concerns on embodiment of this invention. It is a flowchart of the alert processing of the alert system which concerns on embodiment of this invention. It is a figure which shows an example of the output screen of the alert information displayed on the mobile terminal which concerns on embodiment of this invention. It is a figure which shows an example of the action selection screen displayed on the mobile terminal which concerns on embodiment of this invention. It is a figure which shows (a) an example, and (b) a figure which shows another example for demonstrating the effect of the alert system which concerns on embodiment of this invention.

(Embodiment 1)
Hereinafter, an embodiment in which the disease onset prediction device of the present invention is applied to an alert system for notifying a patient's disease prediction will be described with reference to the drawings.

The alert system 1 of the present embodiment is the disease onset prediction device of the present invention, which predicts that a patient's disease will develop within a reference period and notifies the doctor in charge. One or more specific diseases are set in advance in the alert system 1. Then, the alert system 1 predicts whether or not each of the set diseases will develop in the target patient, and if it predicts that it will develop, notifies the doctor to that effect. The administrator of the alert system 1 sets a reference period in the alert system 1 in advance for each disease, for example, 3 hours for sepsis, 5 hours for cardiovascular disease, etc., in consideration of the characteristics of the disease. ..

The alert system 1 receives the patient's biometric data measured by the medical device 2 described later in real time via the hospital network 5, analyzes the received biometric data, and predicts whether or not the patient will develop a disease. .. Further, when the alert system 1 predicts that a patient will develop a disease, the alert system 1 analyzes the biometric data that led to the prediction and generates alert information, and the generated alert information is transmitted to the mobile terminal 4 described later via the hospital network 5. Send to. Further, the alert system 1 acquires the operation information performed by the doctor from the mobile terminal 4 described later and reflects it in the updated content of the prediction model described later.

As shown in FIG. 1, the alert system 1 is communicably connected to the medical device 2, the hospital information system 3, and the mobile terminal 4 via the hospital network 5.

The medical device 2 is a device for measuring biological data of a patient, and is typically a device for measuring vital signs of a patient in an intensive care unit. In addition, the medical device 2 may be a device that measures all biological data of a patient, such as a respirator and an infusion pump. The patient's biological data is mainly quantified information on the patient's health condition, and in addition to so-called vital signs such as blood pressure, pulse rate, respiratory rate, and body temperature, platelet count, bilirubin, etc. It includes data such as mean arterial pressure and urine volume. The medical device 2 measures biological data at regular intervals (for example, 1 second) and transmits it to the alert system 1 and the hospital information system 3 via the hospital network 5.

The hospital information system 3 is an information system constructed in the hospital, and includes subsystems such as an electronic medical record system, an examination information system, a drug system, and a patient information system. The hospital information system 3 stores basic patient information, administered drugs, electronic medical records, and the like. Further, the hospital information system 3 receives real-time biological data of the patient from the medical device 2 via the hospital network 5, and stores the measured time in an internal storage device. In response to the transmission request from the alert system 1, the hospital information system 3 provides the basic information of the patient, the administered drug, the information such as the electronic medical record, and the stored past biometric data to the alert system via the hospital network 5. Send to 1.

Further, the medical staff of the hospital can access the hospital information system 3 through various devices (not shown) and refer to the stored information. Further, the hospital information system 3 receives operation information performed by a doctor from a mobile terminal 4 described later. As a result, the medical staff can visually recognize the operation information of the doctor.

The mobile terminal 4 is a terminal for a doctor to carry, such as a smartphone, a business mobile terminal, a tablet, a PDA (Personal Digital Assistant), or a mobile phone. The mobile terminal 4 displays the alert information received from the alert system 1 on the screen so that the doctor can visually recognize the alert information.

Further, the mobile terminal 4 receives an operation by a doctor and transmits operation information to the alert system 1 and the hospital information system 3 via the hospital network 5. As a result, the medical staff acquires the operation information of the doctor via the hospital information system 3. Further, the alert system 1 acquires the operation information of the doctor and reflects it in the update content of the prediction model described later.

Next, the function of the alert system 1 will be described with reference to FIG. The alert system 1 has a biological data acquisition unit 11, an onset information acquisition unit 12, a sample extraction unit 13, and a prediction model as functional units for performing prediction model construction processing, prediction model reconstruction processing, and alert processing, which will be described later. It includes a construction unit 14, an onset prediction unit 15, an alert information output unit 16, and an action information acquisition unit 17. These functions are realized by the CPU (Central Processing Unit) of the control unit, which will be described later, reading the program and executing the processing. It should be noted that the alert system 1 is described separately below before and during operation. The operation here means that the alert system 1 is actually used for the patient to be predicted. Therefore, "before operation" means that it is a preparatory stage before actually using it.

The biometric data acquisition unit 11 requests the hospital information system 3 for biometric data acquired in the past from a large number of patients for machine learning before or during operation, and as a response, the biometric data acquisition unit 11 requests the hospital information system 3 via the hospital network 5. Receive data. In addition, the biometric data acquisition unit 11 receives biometric data of the patient to be predicted, which is transmitted in real time from the medical device 2 via the hospital network 5 for predicting the onset during operation.

The onset information acquisition unit 12 indicates whether or not the patient has developed a disease based on the past biometric data acquired by the biometric data acquisition unit 11 from the hospital information system 3 for machine learning before and during operation. Acquire onset information. Specifically, the onset information acquisition unit 12 stores in advance a criterion for determining whether or not a patient has developed a specific disease from biological data, and refers to this to determine whether or not the patient has developed a specific disease. judge. It should be noted that the onset information judged by the medical staff (mainly doctors) by the medical examination is input to the hospital information system 3 via the device and transmitted to the alert system 1 via the hospital network 5 instead of judging from the biological data. You may. In this way, the onset information acquisition unit 12 may acquire the onset information by receiving from the hospital information system 3 via the hospital network 5. The onset information includes information such as the ID of the onset patient, the name of the onset disease, and the date and time of onset.

The sample extraction unit 13 extracts biometric data as a sample for machine learning before and during operation from the past biometric data acquired by the biometric data acquisition unit 11 from the hospital information system 3. The alert system 1 uses the extracted sample as machine learning training data for building and updating predictive models used to predict the onset of disease. In machine learning, a sample in which the alert system 1 learns to develop a disease within a reference period is a positive sample, and a sample in which the alert system 1 learns not to develop a disease within a reference period is a negative sample. The sample extraction unit 13 separates the positive sample and the negative sample from the past biological data and extracts each of them. In addition, this reference period is called an alert target period, and in the following, it will be 180 minutes.

Specifically, the sample extraction unit 13 determines whether the past biometric data acquired by the biometric data acquisition unit 11 is the biometric data of the patient who has developed the disease, based on the onset information acquired by the onset information acquisition unit 12. Recognize whether the data is biometric data of a patient who has not developed the disease. Then, basically, the sample extraction unit 13 extracts the biological data that becomes a positive sample from the patient who has developed the disease, and extracts the biological data that becomes the negative sample from the patient who has not developed the disease. However, even if it is the biometric data of a patient who has developed the disease, the biometric data before the alert target period from the onset often does not show any sign of the onset, so it is extracted as a negative sample. The sample extraction unit 13 extracts a sample of biological data acquired by the biological data acquisition unit 11 at regular intervals for a predetermined period (for example, 30 minutes).

The sample extraction unit 13 applies the extracted biological data as a sample to a predetermined function for a predetermined period (for example, 30 minutes) to perform a calculation. This function may be, for example, a function for calculating the standard deviation, the average, the maximum, the minimum, etc. from one type of biometric data, or a function for calculating the standard deviation, the average, the maximum, the minimum, etc. from a plurality of types of biometric data. As described above, the calculation result by each function is referred to as a feature amount as a feature characterizing the patient's condition for 30 minutes, and the predetermined period of 30 minutes described above is referred to as a feature amount window period.

A method in which the sample extraction unit 13 extracts a sample from the biological data of a patient who has developed a disease will be described in detail with reference to FIG. From the time of onset of this, sampling intervals are repeated and extracted repeatedly. Specifically, the alert system 1 is based on the biological data of the patient who developed the disease, from the sample candidates (arrow 101 in FIG. 3) in which the feature amount window period is collected starting from the time when the disease occurred. Biometric data as a positive sample is extracted from the sample candidates (arrow 107 in FIG. 3) in which the feature amount window period is collected starting from before the above-mentioned alert target period for predicting the disease. In addition, a sample candidate (arrow 108 in FIG. 3) that collects the feature amount window period starting from the time retroactive from the time of onset of the disease by the alert target period and the non-target period, and the feature amount window period starting from the past. Negative samples are extracted from the sample candidates (arrows 109, 110, 111 in FIG. 3) in which the minutes are collected.

In such an extraction method, since the biological data that can be the target of the positive sample is the biological data of the patient who developed the disease for a limited period of time, the positive sample is generally better than the negative sample. The number that can be obtained is small. If these are used as training data for machine learning at the same ratio, there is a possibility that the prediction of the alert system 1 will be biased to the prediction that the disease will not develop because there are more negative samples. Therefore, the alert system 1 extracts all the sample candidates in the range shown in FIG. 3 as candidates for positive samples, and on the other hand, randomly selects negative samples from the sample candidates that can be targeted so that the same number of samples are obtained. It is desirable to extract. However, when the number of positive samples is large, positive samples may be randomly selected from the candidates for positive samples that can be targeted. In addition, the negative sample biometric data obtained from the patient who developed the disease is an effective sample that characterizes the changes that occurred during the pre-symptomatic alert target period, especially when compared to the positive sample. Therefore, the biometric data of the patient who has developed the disease may be extracted with priority over the biometric data of the patient who has not developed the disease.

The sample extraction unit 13 applies the above-mentioned function to the biological data to be a positive sample obtained in this way, performs a calculation, and calculates a feature amount to obtain a positive sample. In addition, a negative sample is obtained by applying the above-mentioned function to the biological data to be a negative sample, performing a calculation, and calculating a feature amount.

FIG. 4 shows an example of the data converted into the feature amount by the sample extraction unit 13. The average blood pressure, the standard deviation of the pulse rate, etc. are the features. In this example, a plurality of positive samples (1st to 3rd records) and negative samples (4th to 6th records) are obtained from the same patient.

Returning to FIG. 2, the prediction model building unit 14 builds a prediction model used by the alert system 1 to predict the onset of a patient's disease. Specifically, the prediction model construction unit 14 selects from prediction models such as a support vector machine, a decision tree model, and a logistic regression model based on the positive sample and the negative sample extracted by the sample extraction unit 13. Evaluation indexes, evaluation methods, etc. for selecting a prediction model are preset as parameters for building a prediction model. The prediction model construction unit 14 selects a prediction model based on the prediction model construction parameters, and determines the model parameters to be given to the prediction model. Then, the prediction score is calculated for each sample based on the model parameters given to the prediction model such as the rules, mathematical formulas, and coefficients of the selected prediction model. In this way, the prediction model building unit 14 builds the prediction model. In addition, the prediction model building unit 14 may build a different prediction model for each disease. Since prediction models such as support vector machines, decision tree models, and logistic regression models are well known, detailed explanations are omitted here.

Further, the prediction model construction unit 14 determines a threshold value for alert output. The threshold value for alert output is whether or not the prediction score (corresponding to the probability of developing a disease within the alert target period) calculated by applying the prediction model by the onset prediction unit 15 described later is the score for which the alert should be output. It is a threshold value for determining. Before the operation, the prediction model construction unit 14 may determine the threshold value for alert output by referring to the preset parameters, or may use the median value of the prediction score as the threshold value. Alternatively, due to business requirements, a threshold may be set so that, for example, alert notification is less than once an hour. In addition, the prediction model construction unit 14 acquires the output result of the alert from the hospital information system 3 via the hospital network 5 during operation, and the case where the alert does not occur even though the alert is notified (false positive). Predictive models are reconstructed based on samples such as (called samples) and cases that develop even though alerts are not notified (called false negative samples), and according to the predicted score at the time of reconstructing the predicted model. Update the threshold for alert output. Alternatively, the prediction model construction unit 14 may update only the threshold value for alert output.

Further, the prediction model construction unit 14 sets the alert cause classification condition. The alert cause classification condition is a condition used by the alert information output unit 16 described later in order to include the classification of the feature amount in the alert information based on the main component causing the alert. Specifically, the prediction model construction unit 14 extracts principal components by applying principal component analysis using each feature amount of the positive sample as an input, and determines the score for each extracted principal component for each positive sample. Then, the prediction model construction unit 14 performs clustering based on the determined score, and sets the condition for assigning the classification as the alert cause classification condition. The main component is a component related to the onset of a disease such as "respiratory fluctuation" and "blood pressure". The classification is a type that characterizes a positive sample, such as "the value tends to decrease" and "the value fluctuates greatly". Further, clustering is a basic data analysis method for dividing a set to be classified into subsets, and the details thereof are omitted in the present specification. The alert cause classification condition does not necessarily have to be based on principal component analysis. For example, depending on the number and nature of the features, the factors may be identified by visualizing the average of the features of the positive sample and the negative sample and the features of the patient to be predicted.

The onset prediction unit 15 analyzes the biological data transmitted from the medical device 2 via the hospital network 5 according to the prediction model constructed by the prediction model construction unit 14, and the target patient develops a disease within the alert target period. Calculate the predicted score that represents the probability of doing so.

When the prediction score calculated by the onset prediction unit 15 is equal to or higher than the threshold value for alert output, the alert information output unit 16 creates alert information based on the alert cause classification condition set by the prediction model construction unit 14. It is transmitted to the mobile terminal 4 via the hospital network 5. The alert information includes the principal component that is the cause of predicting the onset of the disease and the information indicating the assigned classification (see FIG. 13). In addition, the alert information may include a feature amount related to the principal component that is the cause of predicting the onset of the disease or biometric data corresponding to the feature amount.

The action information acquisition unit 17 sees the alert information displayed on the mobile terminal 4 and receives the content of operating the mobile terminal 4 (hereinafter referred to as action information) via the hospital network 5. The action information is a type of medical practice selected from options such as "medication", "continuation of monitoring", and "no problem". These action information is also transmitted from the mobile terminal 4 to the hospital information system 3 via the hospital network 5. When the doctor is not near the patient, medical staff such as nurses can access the hospital information system 3 to obtain the action information selected by the doctor and how to perform medical treatment for the patient. grasp. The contents of the above options are as follows. "Dosing" means performing medical treatment for a patient. "Continued monitoring" means to continuously monitor the patient's condition. "No problem" means that there is no need to continue monitoring. It should be noted that these options are merely examples, and any content may be used as long as the object and function of the invention are satisfied. In addition, the action information includes a patient ID that identifies the patient and the date and time when the alert was issued. Then, the action information acquisition unit 17 stores the action information acquired from the mobile terminal 4, and when performing machine learning during operation, acquires the action information used for learning from the stored action information.

The caution condition setting unit 18 analyzes the onset information acquired by the onset information acquisition unit 12 from the patient whose action information acquired by the action information acquisition unit 17 is "no problem", and sets the action to "no problem". Set the caution condition for outputting the caution information that calls attention to it. Specifically, the caution condition setting unit 18 compares the biometric data of the patients who subsequently developed the disease with the biometric data of the patients who did not develop the disease among the patients who were the targets of the action of "no problem". Then, the conditions requiring attention are set by machine learning such as a decision tree. In this case, instead of machine learning, rule-based conditions may be set.

Further, the caution condition setting unit 18 determines a threshold value for caution information. The threshold value for the caution information is a threshold value for the caution information output unit 19, which will be described later, to determine whether or not the caution score calculated by applying the caution condition is a score for outputting the caution information. .. The caution condition setting unit 18 may determine the threshold value for the caution information by referring to the preset parameters, or may use the median value of the caution score as the threshold value.

The attention-requiring information output unit 19 acquires the action information of "no problem" by the action information acquisition unit 17, and calculates the caution-requiring score based on the caution-required condition set by the caution-required condition setting unit 18. Then, when the calculated caution score is equal to or higher than the threshold value for the caution information, the caution information output unit 19 outputs the caution information to the mobile terminal 4 via the hospital network 5. The information requiring attention to be output includes information on matching conditions requiring attention.

Next, the hardware configuration of the alert system 1 will be described with reference to FIG. The alert system 1 includes a control unit 20, a storage unit 30, and a communication unit 40.

The control unit 20 includes a CPU 21, a RAM (Random Access Memory) 22, and a ROM (Read Only Memory) 23.

The CPU 21 realizes each of the above-mentioned functional units by reading the program stored in the storage unit 30 into the RAM 22 and executing the program.

The RAM 22 includes a volatile memory and is used as a work area of the CPU 21.

The ROM 23 includes a non-volatile memory such as a flash memory and an organic memory, and stores a control program for the basic operation of the alert system 1 executed by the CPU 21, a BIOS (Basic Input Output System), and the like.

The storage unit 30 includes a hard disk drive, a flash memory, and the like, and stores various information and programs. The storage unit 30 stores various information such as a sample extracted by the sample extraction unit 13, a prediction model constructed by the prediction model construction unit 14, a threshold value for alert output, and alert cause classification conditions. In addition, the storage unit 30 also stores information on the disease to be predicted, a standard for determining whether or not a specific disease has developed from biological data, and the like.

The communication unit 40 is communicably connected to the hospital network 5 and transmits and receives various information to and from the hospital network 5.

Next, various processes executed by the alert system 1 will be specifically described with reference to the drawings. Since the alert system 1 needs to build a prediction model before starting operation, it executes a prediction model construction process. The prediction model construction process is a process corresponding to machine learning before operation.

When the alert system 1 starts the prediction model construction process in response to a remote command execution instruction from a terminal (not shown) used by the system administrator, the biometric data acquisition unit 11 acquires biometric data as shown in FIG. (Step S11). This biometric data is the biometric data of the patient before operation stored in the hospital information system 3. This step S11 is a biometric data acquisition step of acquiring the biometric data of the patient, and in this step, the biometric data acquisition unit 11 functions as a biometric data acquisition means.

Next, the onset information acquisition unit 12 acquires the onset information of the patient whose biological data was acquired in step S11 (step S12). Specifically, the onset information acquisition unit 12 determines whether or not each patient has developed the disease to be predicted from the biological data based on the determination criteria stored in the storage unit 30. Further, as described above, the onset information acquisition unit 12 may acquire what the doctor has determined on the onset of the disease of each patient. This step S12 is an onset information acquisition step for acquiring onset information indicating whether or not the patient has developed a disease, and in this step, the onset information acquisition unit 12 functions as an onset information acquisition means.

Next, the sample extraction unit 13 extracts a sample from the biological data acquired in step S11 (step S13). The sample extraction unit 13 extracts a positive sample and a negative sample, respectively, by the method described with reference to FIG. For example, 50 positive samples and 50 negative samples are extracted. Then, as shown in FIG. 4, the sample extraction unit 13 stores the positive sample and the negative sample obtained by converting the biological data into the feature amount in the storage unit 30. This step S13 is a sample extraction step for extracting a sample, and in this step, the sample extraction unit 13 functions as a sample extraction means.

Next, the prediction model building unit 14 builds a prediction model based on the positive sample and the negative sample extracted by the sample extraction unit 13 (step S14). As described above, the prediction model construction unit 14 selects the prediction model based on the parameters for model selection stored in the storage unit 30. Then, the prediction model construction unit 14 calculates the prediction score for each sample extracted in step S13 based on the model parameters given to the prediction model such as the rules, mathematical formulas, and coefficients of the selected prediction model. The prediction model construction unit 14 stores in the storage unit 30 model parameters to be given to the prediction model such as rules, mathematical formulas, and coefficients of the selected prediction model, and a sample with a prediction score. This step S14 is a prediction model construction step for constructing a prediction model, and in this step, the prediction model construction unit 14 functions as a prediction model construction means.

Next, the prediction model construction unit 14 determines the threshold value for alert output (step S15). For example, when the median prediction score calculated in step S14 is 0.5, the prediction model construction unit 14 determines the threshold value for alert output to 0.5, which is the median prediction score. Then, the prediction model construction unit 14 stores the determined threshold value for alert output in the storage unit 30.

Subsequently, the prediction model construction unit 14 sets the alert cause classification condition (step S16). Here, a specific procedure for setting the alert cause classification condition will be described with reference to the drawings. The prediction model construction unit 14 applies principal component analysis to the acquired positive sample to extract the principal components. Then, the main components are extracted from the main components having a high contribution rate that contributes to the prediction of the onset of the disease, and the main components are extracted until the preset cumulative contribution rate is reached.

For example, FIG. 7 shows an example in which the main component is extracted from the positive samples (first to third records) shown in FIG. 4, assuming that the cumulative contribution rate is set to 80%. In this example, the contribution rate of the principal component 1 is 50%, the contribution rate of the principal component 2 is 20%, the contribution rate of the principal component 3 is 10%, and the cumulative contribution rate of the three principal components (that is, the total contribution rate). ) Is 80%, so these three principal components are extracted. In addition, each principal component has a weight for each feature amount, and it is clarified to what extent each feature amount is related to which principal component. For example, in the example of FIG. 7, the principal component 1 has a weight of 0.8 for the average blood pressure and 0.2 for the standard deviation.

Next, the prediction model construction unit 14 performs principal component analysis on the features of each positive sample. Then, the prediction model construction unit 14 extracts the main components and gives a score for each main component of the positive sample as shown in FIG. Then, the prediction model construction unit 14 performs clustering based on the score for each principal component, and sets a condition for assigning a classification for each positive sample as shown in FIG. 9 as an alert cause classification condition. The number of classifications to be assigned may be set as a parameter in advance. Then, the prediction model construction unit 14 stores the set alert cause classification condition in the storage unit 30.

As described above, the alert system 1 constructs a prediction model, determines a threshold value for alert output, and executes a prediction model construction process for setting alert cause classification conditions.

Next, a prediction model reconstruction process for reconstructing the prediction model by machine learning and updating the threshold value for alert output and the alert cause classification condition during operation will be described with reference to FIG. The alert system 1 executes the prediction model reconstruction process periodically, for example, once a day in a nighttime batch process. The prediction model reconstruction process is a process corresponding to machine learning during operation.

When the alert system 1 starts the prediction model reconstruction process by automatically starting the batch process or the like, the biometric data acquisition unit 11 acquires the biometric data from the hospital information system 3 via the hospital network 5 (step S21). This biometric data is biometric data acquired by the medical device 2 from the patient during operation and stored in the hospital information system 3.

Next, the onset information acquisition unit 12 acquires the onset information of the patient whose biological data was acquired in step S21 (step S22). Specifically, the onset information acquisition unit 12 determines whether or not each patient has developed the disease to be predicted from the biological data based on the determination criteria stored in the storage unit 30. In addition, the onset information acquisition unit 12 may acquire what the doctor has determined on the onset of the disease of each patient.

Next, the action information acquisition unit 17 acquires the action information for the operating patient stored in the storage unit 30 (step S23). As described above, the action information is the information selected from "medication", "continuation of monitoring", and "no problem", the patient ID, and the date and time when the alert was issued. In step S23, the action information acquisition unit 17 functions as an action information acquisition means for acquiring action information representing the content of the action performed on the predicted target patient predicted to develop the disease by the onset prediction means.

Next, the sample extraction unit 13 extracts a sample from the biological data acquired in step S21 (step S24). The sample extraction unit 13 extracts, for example, 25 positive samples and 25 negative samples by the method described with reference to FIG. Then, the sample extraction unit 13 extracts 25 positive samples and 25 negative samples from the 50 positive samples and negative samples extracted in step S11 of the above-mentioned prediction model construction process in step S24. Update to. Then, the sample extraction unit 13 stores the positive sample and the negative sample for which the feature amount is calculated from the biological data as shown in FIG. 4 in the storage unit 30. However, unlike the prediction model construction process, the sample extraction unit 13 preferentially extracts a sample based on the action information acquired by the action information acquisition unit 17. In the above-mentioned example, the sample extraction unit 13 updates 50% of the samples extracted in step S11 of the prediction model construction process in this step 24, but the update rate is set as a parameter in advance. May be stored in the storage unit 30. Also, instead of updating the sample, a new sample may be added.

This preferential sample extraction method will be specifically described with reference to FIG. A false-negative sample is a sample in which a patient who is predicted not to develop a disease within the alert target period actually develops a disease within the alert target period. In this case, the sample extraction unit 13 extracts the biological data of the patient from the time when the alert is issued to the time after the alert target period as a positive sample, and sets the priority to 1. Here, the priority means that the sample having a large numerical value is preferentially extracted, and the sample having the priority 2 has a larger number of samples than the sample having the priority 1, for example, about twice the number of samples. Extract to. In the sample not set in FIG. 11, the priority is set to 0.

A false-positive sample is a sample in which a patient who is predicted to develop a disease within the alert target period does not actually develop the disease within the alert target period. In this case, the alert system 1 outputs an alert. Subsequent actions selected by the doctor will change the treatment and priority of the samples to be extracted. First, when "dosing" is selected, the sample is treated as a positive sample and the priority is set to 2. This means that since it is considered that the onset of the disease could be avoided by the action of "medication", it is treated as if it actually occurred. If "continue monitoring" is selected, it is treated as a negative sample and the priority is set to 1. If "No problem" is selected, it is treated as a negative sample and the priority is set to 2. The sample extraction unit 13 extracts a sample according to the priority set in this way. In addition, instead of the above-mentioned priority, the ratio of sampling as a sample may be set. However, as described above, it is necessary to maintain a balance between the number of positive samples and the number of negative samples. Therefore, in reality, the sample extraction unit 13 extracts the sample according to the priority or ratio setting within the range in which this balance can be maintained.

Returning to FIG. 10, the prediction model building unit 14 builds a prediction model based on the acquired information, and reconstructs the prediction model stored in the storage unit 30 (step S25). The method for constructing the prediction model is the same as in step S14 of the prediction model construction process. The prediction model construction unit 14 reconstructs the prediction model by replacing it with the constructed prediction model. The prediction model construction unit 14 may update the prediction model by comparing the constructed prediction model with the prediction model stored in the storage unit 30 and updating the difference.

Next, the prediction model construction unit 14 updates the threshold value for alert output (step S26). Specifically, when there are many false positive samples, the prediction model construction unit 14 considers that it is necessary to tighten the alert output conditions and reduce the number of alert outputs, so that the threshold value for alert output is increased. .. On the contrary, when there are many false negative samples, the prediction model construction unit 14 reduces the threshold value for alert output.

Next, the prediction model construction unit 14 updates the alert cause classification condition (step S27). The basic processing procedure is the same as step S16 of the prediction model construction process shown in FIG. The prediction model construction unit 14 updates the alert cause classification condition by setting the alert cause classification condition based on the biological data, action information, etc. acquired from the patient in operation and replacing it with the set alert cause classification condition. do. The prediction model construction unit 14 may update the alert cause classification condition by comparing the set alert cause classification condition with the alert cause classification condition stored in the storage unit 30 and updating the difference. ..

Next, the caution condition setting unit 18 sets or updates the caution condition (step S28). The caution condition setting unit 18 extracts the biometric data obtained by the biometric data acquisition unit 11 in step S21 and the action information acquired by the action information acquisition unit 17 in step S23 is “no problem”. do. Then, the caution condition setting unit 18 calculates the feature amount based on the extracted biological data. Then, the caution condition setting unit 18 sets the caution condition for the characteristic amount of the patient who developed the disease within the reference period and the characteristic amount of the patient who did not develop the disease by machine learning such as a decision tree. In this case, instead of machine learning, rule-based conditions may be set. Specifically, the caution condition setting unit 18 sets the caution condition for defining the caution score as 0.9 for the condition (rule) that the average blood pressure is 140 mmHg or more, for example. Then, the caution condition setting unit 18 stores the set caution condition in the storage unit 30. Further, when the caution condition is already stored in the storage unit 30, the caution condition setting unit 18 updates the caution condition by replacing it with the set caution condition. The caution condition setting unit 18 may update the caution condition by comparing the set caution condition with the caution condition stored in the storage unit 30 and updating the difference.

Next, the caution condition setting unit 18 determines the threshold value for the caution information (step S29). The caution condition setting unit 18 determines a threshold value for caution information from preset parameters. Further, the caution condition setting unit 18 may calculate a caution score for the patient feature amount calculated in step S28 and use the median value thereof as a threshold value. Then, the caution condition setting unit 18 stores the determined threshold value for the caution information in the storage unit 30.

As described above, the alert system 1 reconstructs the prediction model, updates the threshold for alert output and the alert cause classification condition, sets or updates the caution condition, and sets or updates the threshold for caution information. Execute the prediction model reconstruction process to determine. Through these processes, the alert system 1 can improve the accuracy of predicting the onset of a disease during operation.

Next, an alert process in which the alert system 1 predicts a patient's disease and outputs an alert will be described with reference to FIG. The alert system 1 identifies the target patient and the target disease, and executes this alert process on a regular basis, for example, every 5 minutes. This execution interval is preferably about the same as the sampling interval described above.

When the alert system 1 starts the alert process, the biometric data acquisition unit 11 acquires the biometric data in real time by receiving the biometric data transmitted from the medical device 2 (step S301). The biological data acquisition unit 11 stores the received biological data in the storage unit 30 and holds the received biological data for at least 30 minutes (feature amount window time) or more. Then, the biometric data acquisition unit 11 calculates the feature amount for each patient based on the biometric data for 30 minutes including the retained biometric data and the received biometric data.

Next, the onset prediction unit 15 builds a prediction model (constructed in step S14 of the prediction model construction process or step S25 of the prediction model reconstruction process) on the feature amount calculated by the biological data acquisition unit 11 in step S301 for each patient. The prediction model reconstructed in (step S302) is applied to calculate a prediction score representing the probability of developing a disease within the alert target period (step S302). Hereinafter, up to step S310, it is a process to be executed for each patient.

Next, in the onset prediction unit 15, the prediction score calculated by the onset prediction unit 15 in step S302 is determined in the threshold value for alert output (determined in step S15 of the prediction model construction process, or in step S26 of the prediction model reconstruction process. It is determined whether or not it is equal to or higher than the updated alert output threshold value (step S303). This step S303 is an onset prediction step of predicting whether or not the predicted target patient develops the disease within the reference period, and in this step, the onset prediction unit 15 functions as an onset prediction means.

When the onset prediction unit 15 determines that the prediction score is equal to or higher than the threshold value (step S303: Yes), the alert information output unit 16 generates alert information (step S304). Specifically, the alert information output unit 16 is based on the alert cause classification condition (the alert cause classification condition constructed in step S16 of the prediction model construction process or updated in step S27 of the prediction model reconstruction process). Calculate the main component score of the data and assign a classification. Then, the alert information output unit 16 creates alert information including the calculated principal component score and the assigned classification. In addition, the alert information may include a feature amount or biometric data related to the calculated principal component.

On the other hand, when the onset prediction unit 15 determines that the prediction score is not equal to or higher than the threshold value (step S303: No), the control unit 20 of the alert system 1 determines whether or not all the prediction target persons have been processed (step). S311). When the control unit 20 determines that all the prediction target persons have been processed (step S311: Yes), the control unit 20 ends the alert processing. On the other hand, when the control unit 20 determines that all the prediction target patients are not processed (step S311: No), the onset prediction unit 15 calculates the prediction score of the next prediction target patient in step S302. Return to processing. In this example, the data of each patient is sequentially processed, but the control unit 20 may process the data of a plurality of patients in parallel by executing the plurality of processes in a distributed manner. good.

Subsequently, the alert information output unit 16 transmits the generated alert information to the mobile terminal 4 via the hospital network 5 (step S305). The mobile terminal 4 notifies the doctor of the receipt of the alert information by emitting a notification sound, and displays the received alert information on the screen. On the screen displayed by the mobile terminal 4, for example, as shown in FIG. 13, the target disease, the predicted score, the principal component score, the assigned classification, and the like are visually displayed. The mobile terminal 4 generates such a screen based on the received alert information. Further, the mobile terminal 4 may display the received alert information by generating and transmitting the alert information to be displayed on the screen by the alert information output unit. That is, the alert system 1 may perform the process of converting the information such as the target disease, the predicted score, the main component score, and the assigned classification included in the alert information into the information to be visually displayed, or the mobile phone. The terminal 4 may go. Further, the alert system 1 may perform some conversion processing, and the mobile terminal 4 may perform other conversion processing. Further, the mobile terminal 4 may display a feature amount or biological data related to the principal component on the screen. In these steps S304 and S305, the alert information output unit 16 applies the principal component analysis to the biometric data of the predicted target patient predicted to develop the disease by the onset prediction means or the feature amount representing the characteristics of the biometric data. As a result, it functions as an alert information output means for outputting the alert information to which the classification is given.

Further, as shown in FIG. 14, the mobile terminal 4 displays action options such as “medication”, “continuation of monitoring”, and “no problem”. The doctor grasps the patient's situation from the information displayed on the mobile terminal 4 and decides the action to be taken. Then, the doctor operates the mobile terminal 4 to select one of the displayed options. The mobile terminal 4 transmits action information representing the selected action to the alert system 1 and the hospital information system 3 via the hospital network 5.

The action information acquisition unit 17 acquires the action information by receiving the real-time action information transmitted by the mobile terminal 4 via the hospital network 5 (step S306). This action information is the action information of the prediction target, and is referred to as the prediction target action information below. The action information acquisition unit 17 stores the acquired prediction target action information in the storage unit 30 in association with the target patient ID and the date and time when the alert is issued for reference in machine learning.

Next, the caution information output unit 19 determines whether or not the prediction target action information acquired by the action information acquisition unit 17 is “no problem” (step S307). When the attention-requiring information output unit 19 determines that the action information is "no problem" (step S307: Yes), the caution-requiring information output unit 19 calculates the caution-requiring score (step S308). Specifically, the caution information output unit 19 analyzes the patient feature amount calculated by the biological data acquisition unit 11 in step S301 according to the caution condition set or updated in step S28 of the prediction model reconstruction process. , Calculate the caution score. For example, when there is a caution condition that the caution score is defined as 0.9 for the condition (rule) that the average blood pressure is 140 mmHg or more, the caution information output unit 19 has an average blood pressure of 143 mmHg. The attention score of a patient is calculated as 0.9 because it meets this condition.

On the other hand, when the caution information output unit 19 determines that the prediction target action information is not "no problem" (step S307: No), whether or not the control unit 20 of the alert system 1 has processed all the prediction target persons. Proceed to the process of step S311 for determining whether or not.

Next, the attention-requiring information output unit 19 determines whether or not the caution-requiring score is equal to or greater than the threshold value for the caution-requiring information determined in step S29 of the prediction model reconstruction process (step S309). For example, when the threshold value is 0.5 and the calculated attention-requiring score is 0.9, the attention-requiring information output unit 19 determines that the caution-requiring score is equal to or higher than the threshold value. When the caution information output unit 19 determines that the caution score is equal to or higher than the threshold value (step S309: Yes), the caution information output unit 19 generates the caution information and transmits it to the mobile terminal 4 via the hospital network 5 (step S310). Then, the process returns to the process of step S301 in which the biometric data acquisition unit 11 acquires the biometric data. On the other hand, when the attention-requiring information output unit 19 determines that the caution-requiring score is not equal to or higher than the threshold value (step S309: No), the control unit 20 of the alert system 1 determines whether or not all the prediction target persons have been processed. The process proceeds to step S311. In step S310, the attention-requiring information output unit 19 functions as a caution-requiring information output means for outputting the caution-requiring information.

The caution information includes information such as matching rules and caution scores. Then, the mobile terminal 4 displays the received information requiring attention. Further, similarly to the alert information, the alert system 1 may perform the process of converting the caution information into the information for screen display, or the mobile terminal 4 may perform the process. Further, the alert system 1 may perform some conversion processing, and the mobile terminal 4 may perform other conversion processing.

In this way, the alert system 1 applies a prediction model to predict whether or not the target patient will develop a specific disease. Then, when it is predicted that a disease will develop, alert information is transmitted to the mobile terminal 4 via the hospital network 5. Further, the alert system 1 acquires the action information input to the mobile terminal 4, and if the action is "no problem", determines whether or not the rule for transmitting the cautionary information is met. Then, when it is determined that the rule is met, the alert system 1 transmits the caution information to the mobile terminal 4 via the hospital network 5.

According to the alert system 1 of the present embodiment, the onset of the disease can be predicted from several hours in advance and notified to the doctor. In fact, for some illnesses, signs of onset often appear hours before. For example, as in patient A shown in FIG. 15A, when the blood pressure tends to decrease several hours before the onset, the sample extraction unit 13 of the alert system 1 sets the alert target period to several hours or more. As a result, the blood pressure during the feature window period shown in FIG. 15 (a) is extracted as biological data as a positive sample. Then, the alert system 1 calculates the slope of blood pressure in this range as a feature amount, learns the feature amount as a positive sample, and constructs a prediction model based on the slope of blood pressure. Thereby, the onset of the disease can be predicted for the patient who shows the same signs as the patient A. Similarly, as in patient B shown in FIG. 15 (b), when the fluctuation of blood pressure several hours before the onset is large, the variance or standard deviation of blood pressure in the range of the feature amount window period is calculated and learned as the feature amount. By doing so, a predictive model based on blood pressure variance or standard deviation is constructed. Thereby, the onset of the disease can be predicted for the patient who shows the same signs as the patient B.

The alert system 1 of the present embodiment is a biometric data of a patient who has developed a disease, which is a positive sample from the biometric data from the time of the onset of the disease to the time before the reference period of the onset of the disease. Is obtained, and biometric data that is a negative sample is extracted from the biometric data before the reference period of onset and used for learning. As a result, the alert system 1 can learn from the same patient by comparing the characteristics exhibited by the biometric data from the time of onset of the disease to before the reference period with the characteristics exhibited by the biometric data before the reference period. It can be done, and it is possible to catch the signs that the onset will occur.

Further, the alert system 1 is biological data for constructing a prediction model, which is biological data from before the total period of the alert target period and the non-target period at the time of onset to the time of onset, and corresponds to the feature window period. Multiple data are extracted so that a part of the feature window period overlaps to obtain a positive sample. As a result, the signs of the onset of the disease can be grasped in a complete form, and the onset of the disease can be accurately predicted. In addition, in order to avoid acquiring the biometric data as a positive sample and the biometric data as a negative sample in duplicate, it is preferable that the above-mentioned non-target period is the same as the feature amount window period.

The alert system 1 of the present embodiment reflects the action information selected by the doctor in the sample extraction method by the sample extraction unit 13. As a result, the alert system 1 can use the content of the response to the patient to improve the accuracy of machine learning. In particular, in the alert system 1, even if the extracted sample is a false positive sample in which the patient predicted to develop the disease did not develop the disease within the alert target period, the doctor receives the alert and "doses". If you select an action, treat it as a positive sample instead of a negative sample. As a result, even if a patient does not actually develop the disease, if it is considered that the disease has developed without medical treatment for the patient, a sign of developing the disease can be obtained from the biometric data of the patient. can do.

Further, in the alert system 1, the extracted sample is a false positive sample in which the patient predicted to develop the disease did not develop the disease within the alert target period, and the doctor received the alert and "no problem" or If the "Continue monitoring" action is selected, it is preferentially extracted as a negative sample. As a result, the alert system 1 can preferentially learn the biometric data that has made an erroneous judgment as a judgment to be corrected.

The alert system 1 of the present embodiment classifies the predicted causative component of developing a disease by applying the alert cause classification condition, and transmits it as alert information to the mobile terminal 4 via the hospital network 5. This allows the physician to obtain information that will guide them in considering how to respond to the target patient.

When the action for the patient who issued the alert information is "no problem", the alert system 1 of the present embodiment generates caution information and transmits it to the mobile terminal 4 via the hospital network 5. This gives the doctor the opportunity to notice a misjudgment.

The biological data of the present embodiment may include information on medication to the patient, treatment, environmental information, and the like. As a result, it is possible to more accurately determine the prediction of the onset, the generation of alert information, the generation of information requiring attention, and the like. For example, by applying the medication information to the classification of the alert information, the doctor who sees the alert information can know what kind of medication should be taken.

In the present embodiment, an example in which there are only a positive sample and a negative sample as samples has been illustrated, but the scope of the present invention is not limited to this, and for example, a sample that is neither a positive sample nor a negative sample may be applied. For example, a sample that is neither positive nor negative may be applied, that is, a sample for learning that it is not possible to specify whether the sample is positive or negative. As a result, the disease onset prediction device can learn about the feature amount that cannot be judged to be positive or negative, so that a prediction model with ambiguity can be constructed.

In the present embodiment, the alert system 1 calculates a prediction score and compares it with a threshold value to predict whether or not the patient to be predicted develops a disease. However, the scope of the present invention is not limited to this, and as long as it predicts whether or not the patient to be predicted will develop the disease, the prediction score is not calculated, but for example, the prediction that the patient to be predicted will develop is predicted. 1. Whether or not the patient who is the target of prediction will develop the disease may be directly calculated, such as 0 for the prediction that the disease will not occur.

In the present embodiment, mutual communication between the alert system 1, the medical device 2, the hospital information system 3, and the mobile terminal 4 is performed via the hospital network 5. However, the scope of the present invention is not limited to this, and each may directly communicate with each other by a dedicated line or the like. Further, the alert system 1 may be built in or integrated with the medical device 2 or the hospital information system 3.

This embodiment illustrates the operation of the alert system 1 in a hospital. However, the scope of the present invention is not limited to this, and for example, if a wearable sensor can be used to acquire necessary biometric data, it may be operated outside the hospital such as at home.

The alert system 1 can be realized by using an ordinary computer without using a dedicated device. For example, the alert system 1 that executes the above-mentioned processing may be configured by installing the program in the computer from a recording medium in which the program for executing any of the above is executed in the computer. Further, one alert system 1 may be configured by operating a plurality of computers in cooperation with each other.

Also, the method for supplying the program to the computer is arbitrary. For example, it may be supplied via a communication line, a communication network, a communication system, or the like.

When the OS (Operation System) provides a part of the above-mentioned functions, the part other than the functions provided by the OS may be provided by a program.

1 ... Alert system, 2 ... Medical equipment, 3 ... Hospital information system, 4 ... Mobile terminal, 5 ... Hospital network, 11 ... Biological data acquisition unit, 12 ... Onset information acquisition unit, 13 ... Sample extraction unit, 14 ... Prediction model Construction unit, 15 ... Onset prediction unit, 16 ... Alert information output unit, 17 ... Action information acquisition unit, 18 ... Caution condition setting unit, 19 ... Caution information output unit, 20 ... Control unit, 21 ... CPU, 22 ... RAM, 23 ... ROM, 30 ... Storage unit, 40 ... Communication unit, 101, 102, 103, 104, 105, 106, 107 ... Arrows indicating the target period of positive samples, 108, 109, 110, 111 ... Negative samples Arrow representing the target period

Claims (9)

A biometric data acquisition means for acquiring biometric data to be predicted and biometric data for constructing a prediction model,
An onset information acquisition means for acquiring onset information indicating whether or not a patient has developed a disease corresponding to the biological data for constructing the prediction model, and an onset information acquisition means.
When the onset information indicates that the disease has developed, a positive sample is obtained from the biometric data for constructing the prediction model from the biometric data from before the reference period at the onset time of the onset of the disease to the onset time. A sample extraction means for extracting biological data and extracting biological data as a negative sample from biological data before the reference period at the time of onset.
The positive sample is used as training data for developing the disease within the period from the reference time, which is the time point before the reference period at the time of onset, to the time after the reference period , and the negative sample is used from the reference time to the above. A predictive model construction means for constructing a predictive model by using it as training data that does not develop the disease within the period up to after the reference period.
Said time obtaining the prediction target biometric data and the reference time, on the basis the prediction target biometric data to said predictive model, the patient corresponding to the biometric data of the prediction target, the reference from the time before Symbol reference It is provided with an onset prediction means for predicting whether or not the disease will develop within the period up to after the period.
In the sample extraction means, the result of performing a predetermined calculation on the extracted biological data to be a positive sample is defined as the positive sample, and the result of performing a predetermined calculation on the biological data to be a negative sample is defined as the negative sample. ,
Disease onset prediction device. When the onset information indicates that the patient has developed the disease, the sample extraction means may use the biological data for constructing the prediction model from before the total period of the reference period and the predetermined period at the time of onset. A plurality of biological data up to the time of onset and corresponding to the predetermined period are extracted so as to partially overlap the predetermined period to obtain the positive sample.
The disease onset prediction device according to claim 1. It is provided with an action information acquisition means for acquiring action information indicating the content of an action performed on a patient corresponding to the biological data for constructing the prediction model.
Based on the action information, the sample extraction means preferentially extracts from the biological data for constructing the prediction model to obtain the positive sample or the negative sample.
The disease onset prediction device according to claim 1 or 2. When the action information indicates that the action information was an action for avoiding the onset of the disease, the sample extraction means means that the patient corresponding to the biological data for constructing the prediction model developed the disease. Assuming that the onset information indicates that the disease has occurred, the biometric data serving as the positive sample is extracted.
The disease onset prediction device according to claim 3. The action information acquisition means further acquires the prediction target action information representing the content of the action performed on the patient corresponding to the biometric data of the prediction target predicted to develop the disease by the onset prediction means.
The biometric data acquisition means newly acquires data corresponding to the biometric data of the prediction target of the patient in which the act is performed as the biometric data for constructing the prediction model.
When the sample extraction means indicates that the prediction target action information is not an action for avoiding the onset of the disease, the newly acquired biological data for constructing a prediction model is used as another living body. The negative sample is extracted with priority over the data.
The disease onset prediction device according to claim 3 or 4. When the action information indicates that the action information was not an action for avoiding the onset of the disease and the onset information indicates that the disease has developed, the biometric data of the prediction target. Based on, set the conditions for outputting cautionary information indicating that actions should be taken to avoid the onset of the disease.
When the prediction target action information indicates that the action was not an action for avoiding the onset of the disease, the caution information output that outputs the caution information based on the conditions set by the prediction model construction means. Equipped with more means
The disease onset prediction device according to claim 5. When the onset prediction means predicts that the disease will develop, the causative components predicted to develop the disease are classified by a predetermined method based on the biological data of the prediction target, and output as alert information. Further equipped with alert information output means,
The disease onset prediction device according to any one of claims 1 to 6, characterized in that. A biometric data acquisition step for acquiring biometric data to be predicted and biometric data for building a predictive model,
An onset information acquisition step for acquiring onset information indicating whether or not a patient has developed a disease corresponding to the biological data for constructing the prediction model, and an onset information acquisition step.
When the onset information indicates that the disease has developed, a positive sample is obtained from the biometric data for constructing the prediction model from the biometric data from before the reference period at the onset time of the onset of the disease to the onset time. A sample extraction step of extracting biometric data to be a negative sample from biometric data prior to the reference period at the time of onset.
The positive sample is used as training data for developing the disease within the period from the reference time, which is the time point before the reference period at the time of onset, to the time after the reference period , and the negative sample is used from the reference time to the above. A predictive model building step to build a predictive model by using it as training data that does not develop the disease within the period up to after the reference period.
Said time obtaining the prediction target biometric data and the reference time, on the basis the prediction target biometric data to said predictive model, the patient corresponding to the biometric data of the prediction target, the reference from the time before Symbol reference It is provided with an onset prediction step for predicting whether or not the disease will develop within the period up to after the period.
In the sample extraction step, the result of performing a predetermined calculation on the extracted biological data to be a positive sample is defined as the positive sample, and the result of performing a predetermined calculation on the biological data to be a negative sample is defined as the negative sample. ,
Disease onset prediction method. On the computer
A biometric data acquisition step for acquiring biometric data to be predicted and biometric data for building a predictive model,
An onset information acquisition step for acquiring onset information indicating whether or not a patient has developed a disease corresponding to the biological data for constructing the prediction model, and an onset information acquisition step.
When the onset information indicates that the disease has developed, a positive sample is obtained from the biometric data for constructing the prediction model from the biometric data from before the reference period at the onset time of the onset of the disease to the onset time. A sample extraction step of extracting biometric data to be a negative sample from biometric data prior to the reference period at the time of onset.
The positive sample is used as training data for developing the disease within the period from the reference time, which is the time point before the reference period at the time of onset, to the time after the reference period , and the negative sample is used from the reference time to the above. A predictive model building step to build a predictive model by using it as training data that does not develop the disease within the period up to after the reference period.
Said time obtaining the prediction target biometric data and the reference time, on the basis the prediction target biometric data to said predictive model, the patient corresponding to the biometric data of the prediction target, the reference from the time before Symbol reference It is a program for executing the onset prediction step for predicting whether or not the disease will develop within the period up to the period after the period.
In the sample extraction step, the result of performing a predetermined calculation on the extracted biological data to be a positive sample is defined as the positive sample, and the result of performing a predetermined calculation on the biological data to be a negative sample is defined as the negative sample. ,
program.

Images