JP2018175850A - Data collection device and data collection method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a data collection technique for improving HDP onset prediction accuracy.SOLUTION: One embodiment of the invention relates to a data collection apparatus comprising: a data acquisition function for acquiring data related to a pregnant woman from each of plural data source devices; an HDP onset prediction function for executing HDP onset prediction of the pregnant woman from the pieces of data acquired based on an HDP onset prediction model; and a result notification function for notifying of a result of the HDP onset prediction. The data acquisition function acquires the pieces of data at plural time points in a data acquisition period since any period of before and after pregnancy of the pregnant woman to any period of before and after birth.SELECTED DRAWING: Figure 1

Description

  The present invention relates generally to medical diagnostic techniques, and more particularly to a data collection apparatus and method for predicting the onset or absence of pregnancy and hypertension for noninvasively predicting onset or abnormality detection of pregnancy and hypertension.

  Pregnancy related diseases have traditionally been difficult to predict. The onset of pregnancy related diseases is not only due to genetic factors but also to environmental factors in a complex way, but it is difficult to grasp environmental factors. For this reason, there are many pregnancy related diseases for which a preventive method has not been established, for which a precursor or a cause has not yet been identified.

  In obstetrics and medical care at the present time, it is mainly based on examination information acquired at the time of medical examination, medical history listening and / or prediction of onset by medical examination information, and it was difficult to obtain continuous living environment information such as at home. In the research on disease onset, conducting surveys using questionnaires has been the main method in the past as a method of acquiring information on the living environment of pregnant women. However, this method has the problem of accuracy because the memory of the pregnant woman is incorrect and there is an error in the response (recollection bias), and the implementation interval is early and there is the problem of acquisition frequency etc. once in half a year. The Therefore, it was an insufficient method to grasp the actual condition of the life of the pregnant woman and clarify the causal relationship with the disease. In the future, in order to precisely predict the onset of the disease, it is necessary to obtain continuous environmental information other than at the time of consultation and analyze them in an integrated manner at the time of consultation, but these information have not been Because it was difficult to obtain, it has not been used to predict the onset of pregnancy related diseases.

  Gestational hypertension syndrome (HDP: Hypertensive Disorders of Pregnancy) is a typical perinatal medical disease that is associated with the life of mothers and infants causing maternal death, fetal failure, fetal death, permanent placental exfoliation, etc. It is a multifactorial disease that develops due to a complex interaction of environmental and environmental factors. In addition, it is known that pregnant women suffering from HDP have a high risk of developing lifestyle-related diseases thereafter.

  As an epidemiological HDP onset study based on the living environment of a pregnant woman, for example, there is a report that BMI (Body Mass Index) before pregnancy is measured and the height of the BMI concerned increases the risk of HDP (non-patent literature) 1). However, in this study, the prediction accuracy is 3.4 to 8.8 in odds ratio because it is based on single parameter and measurement results at one time point, and it can not be said that the accuracy is clinically sufficient. . Also, the evaluation index is only the odds ratio, and it is not possible to predict the final HDP of the pregnant woman.

"" Pregnancy Body Mass Index and the Occurrence of Severe Hypertension Disorders of Pregnancy "", Lisa M. Bodnar, et al., Epidemiology, Volume 18, Number 2, March 2007 (http://journals.lww.com/epidem/ Abstract / 2007/03000 / Prepregnancy_Body_Mass_Index_and_the_Occurrence_of.11.aspx)

  In view of the above-mentioned problems, it is an object of the present invention to provide a data collection technique for improving the accuracy of predicting the onset of HDP.

  In order to solve the above-mentioned subject, one mode of the present invention predicts HDP onset of the above-mentioned pregnant woman from the data acquired based on a data acquisition function which acquires data about a pregnant woman from a plurality of data source devices, and HDP onset prediction model. The function has an HDP onset prediction function to be executed and a result notification function to notify the result of the HDP onset prediction, and the data acquisition function is from any time before or after pregnancy of the pregnant woman to any time from before to after delivery The present invention relates to a data collection apparatus that acquires the data at a plurality of time points in a data acquisition period up to

  According to the present invention, it is possible to provide a data collection technique for improving the accuracy in predicting the onset of HDP.

FIG. 1 is a block diagram showing a system configuration according to an embodiment of the present invention. FIG. 2 is a block diagram showing a functional configuration of a measured value transfer apparatus according to an embodiment of the present invention. FIG. 3 is a block diagram showing a functional configuration of the physical condition registration device according to one embodiment of the present invention. FIG. 4 is a block diagram showing a functional configuration of a weather information acquisition apparatus according to an embodiment of the present invention. FIG. 5 is a block diagram showing a functional configuration of a survey form registration device according to an embodiment of the present invention. FIG. 6 is a block diagram showing a functional configuration of the medical care record registration device according to one embodiment of the present invention. FIG. 7 is a block diagram showing a functional configuration of a data collection device according to an embodiment of the present invention. FIG. 8 is a flowchart showing data collection processing in the first embodiment of the present invention. FIG. 9 is a flowchart showing HDP onset prediction model generation processing in the third embodiment of the present invention. FIG. 10 is a table showing combinations of feature amounts selected in Example 3 in which HDP onset / presence prediction was performed, and a graph showing prediction accuracy of HDP onset / absence when using a combination of these feature amounts. FIG. 11 is a flowchart showing HDP onset prediction model generation processing in the fourth embodiment of the present invention. FIG. 12 is a table showing combinations of feature amounts selected in Example 4 in which HDP onset / presence prediction was performed, and a graph comparing the prediction accuracy of HDP onset / absence in Example 3 and Example 4. FIG. 13 is a flowchart showing HDP onset prediction model generation processing in the sixth embodiment of the present invention. FIG. 14 is a flowchart showing HDP onset prediction model generation processing in the seventh embodiment of the present invention. FIG. 15 is a flowchart showing HDP onset date prediction model generation processing in the eighth embodiment of the present invention. FIG. 16 is a flowchart showing HDP onset risk prediction model generation processing in the ninth embodiment of the present invention.

  Hereinafter, embodiments of the present invention will be described based on the drawings.

  In the embodiments described below, a data acquisition device for predicting the onset of HDP is disclosed. Since the prediction accuracy is low in the conventional HDP onset prediction using a single index such as BMI, the data collection device according to the following embodiment improves the prediction accuracy by combining data of the living environment of the pregnant woman. Let In addition, the data collection device is objectively obtained by acquiring data to be acquired not at measurement values at one time but at multiple time points, for example, continuously or periodically, and eliminating recall bias from the acquired data as much as possible. Use the following data.

Specifically, according to the data collection system according to the following example, in order to clarify the environmental factor of the onset of pregnancy related diseases by objectively acquiring the living environment of the pregnant woman more objectively and more frequently, From before, various data such as the following are acquired.
1) Daily data of measuring devices such as blood pressure meter, body weight body composition meter, activity meter, thermometer 2) Daily mood, sleep, bowel movement, nausea, pain, uterine contraction, palpitation, interval interval, meal contents, medication contents Physical condition data 3) Daily living environment data such as temperature, humidity, barometric pressure, tide, etc. 4) Eating habits, sleeping conditions, regular questionnaire data on lifestyle habits such as smoking 5) Pregnant women Periodic medical record data such as a medical examination By acquiring such data, many kinds of data can be used, and prediction accuracy can be improved more than conventional HDP onset prediction based on a single index. Also, with regard to the above 1) to 3), during or before pregnancy, data are obtained frequently, such as several times a day or several times a week, or continuously during the period, and further, 1), 3) above. And objective data acquisition, and elimination or reduction of recall bias by shortening the interval which the pregnant woman should memorize in daily registration by 2) can be realized. Of the above data, the data items 1) to 3) are defined as a life log.

  As a result, in the present invention, in order to generate an HDP onset prediction model, it is possible to acquire the life log, the questionnaire, and the medical record, and enrich the environment information of the pregnant woman. That is, the present invention can be used as a daily life log during pregnancy or after delivery of a pregnant woman, for example, in a blood pressure monitor, body composition monitor, activity meter, thermometer, in order to non-invasively predict the onset of HDP or detect abnormalities. Objective data measured by etc., daily mood, sleep, bowel movement, nausea, pain, contraction of the uterine contraction, palpitation, fetal movement, diet, medication, etc. while acquiring data on the physical condition of the pregnant woman, temperature of the residence of the pregnant woman, weather, By acquiring data on one or more data items out of objective data such as humidity, barometric pressure, tide and the like and analyzing them in combination, it is possible to realize more accurate HDP onset prediction.

  First, referring to FIG. 1, a data collection system according to an embodiment of the present invention will be described. FIG. 1 is a block diagram showing a system configuration according to an embodiment of the present invention.

  As shown in FIG. 1, the data collection system 10 includes a measurement device 20, a measurement value transfer device 30, a physical condition registration device 40, a weather information acquisition device 50, a survey form registration device 60, a medical record registration device 70, and a data collection device. It has 100. The illustrated system configuration is merely an example, and some devices may be omitted or other devices may be added. For example, the data collection system 10 includes one or both of the measurement device 20 or the physical condition registration device 40, the medical record registration device 70 and the data collection device 100, and optionally, the measurement value transfer device 30, weather information acquisition The device 50 and / or the questionnaire registration device 60 may be included.

  The measuring device 20 includes a sphygmomanometer, a body weight body composition meter, an activity meter, a thermometer, and the like, and measures physiological data such as the blood pressure of a pregnant woman.

  The measurement value transfer device 30 transfers data between the measurement device 20 and the data acquisition device 100. The measurement device 20 itself may include the function of the measurement value transfer device 30, and the measurement value may be directly transmitted to the data collection device 100.

  The physical condition registration device 40 is a device for a pregnant woman to register daily physical condition information, and can be implemented, for example, in a smartphone and an application. Specifically, the pregnant woman inputs his physical condition information into the physical condition registration device 40. Here, the physical condition information may include, but is not limited to, data items such as, for example, mood, sleep, bowel movement, nausea, pain, uterine contractions, palpitations, intervals between meals, food contents, medication contents, etc. .

  The meteorological information acquisition apparatus 50 acquires meteorological information in a residence of a pregnant woman from a database (not shown) of meteorological information that is open to the public. Here, the weather information includes, for example, data items such as weather, precipitation probability, temperature, humidity, wind speed, wind direction, barometric pressure, tide, etc. for each predetermined period (one hour unit, one day unit, etc.) Although it is good, it is not limited to these.

  The questionnaire registration device 60 is a device for registering information of a regular questionnaire from a pregnant woman. Specifically, the person in charge of input registers the information on the survey form acquired from the pregnant woman in the survey form registration device 60. Here, information on the questionnaire includes, for example, the family structure of a pregnant woman and a family member, medical history of a pregnant woman and a family, pregnancy and delivery history, diet, sleep status, medication status, smoking and drinking status, exercise status, work status, mental status It may also include data items such as, but not limited to, specific situations.

  The medical care record registration device 70 is a device for registering a medical care record at the time of delivery of a pregnant woman. In addition, medical care information such as a regular pregnancy checkup of a pregnant woman may optionally be further registered in the medical care record registration device 70.

  The data collection device 100 is one of the measurement device 20, the measurement value transfer device 30, the physical condition registration device 40, the weather information acquisition device 50, the survey form registration device 60, and the medical record registration device 70, as described in detail below. Data on the above-mentioned data items are collected from the network via the network 200, HDP occurrence prediction model is generated, and prediction is performed. The network 200 may be composed of a plurality of networks instead of a single network.

  Each device described above may typically be realized by an information processing device having a communication function such as a computer, a smartphone, a tablet, or a server. For example, a processor mounted on an information processing apparatus executes various functions and processes to be described later by processing and executing data and programs stored in a memory device. However, each device is not limited to any specific hardware configuration, and may be realized by an appropriate hardware configuration.

  Next, with reference to FIG. 2, a measurement value transfer apparatus according to an embodiment of the present invention will be described. FIG. 2 is a block diagram showing a functional configuration of a measured value transfer apparatus according to an embodiment of the present invention.

  As shown in FIG. 2, the measurement value transfer device 30 has a data acquisition function 31, a data transfer function 32, a measurement device authentication function 33, and a user authentication function 34.

  The data acquisition function 31 communicates with one or more measuring devices 20 to obtain data and measurements from the measuring devices 20. Communication between the measurement value transfer device 30 and the measurement device 20 is not limited to the following, and Bluetooth (registered trademark), Wi-Fi, NFC, ZigBee, USB connection, mobile communication (3G, LTE, 4G , 5G, etc.). For example, the data items acquired from the measuring device 20 are not limited to the following, and may be physiological data on a pregnant woman such as a sphygmomanometer, a body weight body composition meter, an activity meter, and a thermometer.

  The data transfer function 32 transfers the acquired data and measurement values to the data collection device 100 via the network 200. The data may be obtained frequently or continuously several times a day or several times a week during or before pregnancy, and may be transferred to the data collection device 100.

  The measuring device authentication function 33 authenticates the measuring device 20 at the communication destination.

  The user authentication function 34 authenticates the user of the measurement value transfer device 30 such as a pregnant woman.

  Next, referring to FIG. 3, a physical condition registration device according to an embodiment of the present invention will be described. FIG. 3 is a block diagram showing a functional configuration of the physical condition registration device according to one embodiment of the present invention.

  As shown in FIG. 3, the physical condition registration device 40 has a data registration function 41, a data transfer function 42 and a user authentication function 43.

  The data registration function 41 registers input content from a user such as a pregnant woman as data. For example, data items input from a pregnant woman are not limited to the following, and physical condition data of the pregnant woman such as daily mood, sleep, bowel movement, nausea, pain, uterine contraction, palpitations, intervals of feces, food content, medication contents, etc. It may be

  The data transfer function 42 transfers the registered data to the data collection device 100. The data may be obtained frequently or continuously several times a day or several times a week during or before pregnancy, and may be transferred to the data collection device 100.

  The user authentication function 43 authenticates the user of the physical condition registration device such as a pregnant woman.

  Next, a weather information acquiring apparatus according to an embodiment of the present invention will be described with reference to FIG. FIG. 4 is a block diagram showing a functional configuration of a weather information acquisition apparatus according to an embodiment of the present invention.

  As shown in FIG. 4, the weather information acquisition apparatus 50 has a data acquisition function 51 and a data transfer function 52.

  The data acquisition function 51 acquires weather information from the weather information database. The weather information database includes, for example, a database that is open to the public. For example, the weather information is not limited to the following, and may be living environment data such as the temperature, weather, humidity, barometric pressure, tide, etc. of the residence of the pregnant woman.

  The data transfer function 52 transfers the acquired data to the data collection device 100. The data may be obtained frequently or continuously several times a day or several times a week during or before pregnancy, and may be transferred to the data collection device 100.

  Next, with reference to FIG. 5, a survey form registration device according to an embodiment of the present invention will be described. FIG. 5 is a block diagram showing a functional configuration of a survey form registration device according to an embodiment of the present invention.

  As shown in FIG. 5, the survey form registration device 60 has a data registration function 61 and a data transfer function 62.

  The data registration function 61 registers data input by the person in charge of inputting the survey form information. For example, information on questionnaires regarding pregnant women is not limited to the following: family structure of pregnant women and their families, medical history of pregnant women and families, pregnancy delivery history, diet, sleep status, medication status, smoking drinking status, exercise It may be data items on the lifestyle habits of pregnant women and their families such as situations, work situations, and mental situations.

  The data transfer function 62 transfers the registered data to the data collection device 100.

  Next, a medical record registration apparatus according to an embodiment of the present invention will be described with reference to FIG. FIG. 6 is a block diagram showing a functional configuration of the medical care record registration device according to one embodiment of the present invention.

  As shown in FIG. 6, the medical care record registration device 70 has a data registration function 71 and a data transfer function 72.

  The data registration function 71 registers data input by the person in charge of medical record entry. For example, the medical record is not limited to the following, and may be a medical record such as a pregnant woman medical checkup recorded by a doctor or the like.

  The data transfer function 72 transfers the registered data to the data collection device 100.

  Next, with reference to FIG. 7, a data collection device according to an embodiment of the present invention will be described. FIG. 7 is a block diagram showing a functional configuration of a data collection device according to an embodiment of the present invention.

  As shown in FIG. 7, the data collection device 100 has a data acquisition function 110, a data storage function 120, an HDP onset prediction model generation function 130, an HDP onset prediction function 140, and a result notification function 150.

  The data acquisition function 110 acquires data on a pregnant woman from a plurality of data source devices. Further, the data acquisition function 110 acquires data on a pregnant woman at a plurality of time points in a data acquisition period from any time before and after pregnancy of the pregnant woman to any time before and after delivery. That is, from the data source device such as the measurement device 20, the measurement value transfer device 30, the physical condition registration device 40, the weather information acquisition device 50, the survey sheet registration device 60 and / or the medical record registration device 70 described above. Data of one or more data items regarding a pregnant woman are acquired at a plurality of time points, for example, periodically, in a data acquisition period from any time before or after pregnancy of the pregnant woman to any time before or after delivery.

  Here, data relating to a pregnant woman are, for example, age at pregnancy, blood pressure, pulse, body weight, BMI, body fat rate, skeletal muscle rate, number of steps, calories, body temperature, mood, sleep time, bedtime, wake time, bedtime Number of waking up, quality of sleep, number of bowel movements, softness of the bowel movement, uterine contraction, head pain, tooth pain, upper abdominal pain, lower abdominal pain, lower back pain, nausea, vomiting, palpitations, fetal movement interval Food content, medication content, past HDP history, pregnancy status, temperature, room temperature, humidity, barometric pressure, weather conditions, tides and one or more data items of age.

  The data storage function 120 stores the acquired data on the pregnant woman.

  The HDP onset prediction model generation function 130 generates and holds an HDP onset prediction model. For example, the HDP onset prediction model generation function 130 generates and updates the HDP onset prediction model based on data on pregnant women acquired from a plurality of pregnant women including a pregnant woman who developed HDP and a pregnant woman who did not develop HDP. Good. Specifically, the HDP onset prediction model may be generated according to a model derivation method as described in detail below.

  The HDP onset prediction function 140 executes HDP onset prediction of the pregnant woman from the acquired data on the pregnant woman based on the HDP onset prediction model. Specifically, the HDP onset prediction function 140 inputs the acquired data on the pregnant woman into the HDP onset prediction model, and predicts the presence or absence of the HDP onset, the HDP onset date prediction or the HDP onset of the pregnant woman based on the output result of the HDP onset prediction model. Risk calculation may be performed. That is, in predicting HDP onset, the HDP onset prediction function 140 predicts whether the pregnant woman has a possibility of HDP onset (presence or absence of HDP onset), and predicts the date of HDP onset of the pregnant woman (HDP onset date prediction), Alternatively, the risk of developing HDP may be calculated for the pregnant woman (calculation of risk of developing HDP). The HDP onset prediction function 140 may execute HDP onset prediction, HDP onset date prediction or HDP onset risk calculation of the pregnant woman based on different data items depending on the number of weeks of the pregnant woman.

  The result notification function 150 notifies pregnant women, medical personnel, and the like of the result of the HDP prediction performed by the HDP prediction function 140.

Note that in the devices described above, each may be a single device, or one device may have some or all of the functions of another device.
Example 1
Next, with reference to FIG. 8, data collection processing in the data collection device 100 according to an embodiment of the present invention will be described. FIG. 8 is a flowchart illustrating data collection processing according to an embodiment of the present invention.

  As shown in FIG. 8, in step S100, the data collection device 100 includes the measurement device 20, the measurement value transfer device 30, the physical condition registration device 40, the weather information acquisition device 50, the survey form registration device 60, and / or medical treatment described above. Data is acquired from all or part of a data source device such as the record registration device 70, and the acquired data is stored. Specifically, the data collection device 100 acquires data at a plurality of time points, such as periodically, in a data acquisition period from any time before and after pregnancy of a pregnant woman to any time before and after delivery.

  First, a pregnant woman uses the measuring device 20 from any time before or during pregnancy to any time after birth or postpartum. The measuring device 20 includes, for example, a sphygmomanometer, a body weight body composition meter, an activity meter, a thermometer, and the like. The sphygmomanometer measures systolic blood pressure, diastolic blood pressure, pulse rate, room temperature and the like. The body weight / body composition meter measures body weight, BMI, body fat percentage, skeletal muscle percentage and the like. The activity meter measures the number of steps and calories consumed. The thermometer measures the temperature and so on. It is assumed that the measurement by the measuring device 20 is performed twice in the morning and evening for the sphygmomanometer and once in the other morning for the sphygmomanometer, but it is also possible to measure at arbitrary intervals. The measurement value transfer device 30 acquires measurement data from each of the measurement devices 20 and transfers the measurement data to the data collection device 100.

  Further, the pregnant woman registers the physical condition of the pregnant woman in the physical condition registration device 40. Although the case where the application for physical condition registration and a web page are used for a smart phone as the physical condition registration apparatus 40 is assumed, it is not the limitation. As physical condition information, mood, bedtime, wake up time, sleep time, bedtime, wake up time, frequency of waking up at bedtime, quality of sleep, frequency of bowel movements, softness of stool, total time for feeling nausea, vomiting Number of times, air vomiting number of times, head pain, teeth pain, lower back pain, upper abdominal pain, lower abdominal pain, painless contractions number of contractions, number of painful contractions number of contractions, palpitation, interval interval, Although it is assumed that one or more data items of meal content and medication content are registered, it is not limited thereto. The registration period of physical condition information is assumed to be from any time before or during pregnancy to any time after birth or postpartum. In addition, the physical condition registration interval assumes that the main data items are once a day, but for some data items such as the contents of meals and medications, it is assumed that inputs are made multiple times a day. These data items can also be measured at arbitrary intervals, not daily.

  Also, the weather information acquisition device 50 collects daily weather information. For example, it can be obtained from the website of the Meteorological Agency. Data items to be acquired assume date, precipitation, average temperature, maximum temperature, minimum temperature, total solar radiation, average cloud amount, snowfall depth, deepest snowfall, maximum instantaneous wind speed, tide and / or age. Furthermore, temperature, precipitation, wind direction, wind speed, sunshine duration, humidity, barometric pressure, general conditions of weather are assumed every hour, but it is not limited to these. Also, the unit of data is not limited to one day or time, but can be weekly, monthly, or yearly.

  In addition, the survey form registration device 60 acquires information of a regular survey form entered by a pregnant woman. The information on the questionnaire includes data on the family structure of pregnant women and their families, medical history of pregnant women and families, pregnancy birth history, diet, sleep status, medication status, smoking and drinking status, exercise status, work status, mental status, etc. It includes items but is not limited to them. As for the frequency of acquisition, the frequency of questionnaires is assumed.

  Further, the medical care record registration device 70 acquires records such as test values and findings when a pregnant woman receives a medical checkup for a pregnant woman and the like. As for the frequency of acquisition, the frequency of consultation is assumed.

  In step s101, the data collection device 100 detects and removes abnormal values from the acquired data, and interpolates missing values. For example, when a pregnant woman whose daily systolic blood pressure is 120 mmHg is 60 mmHg for only one day, the data collection device 100 regards the measured value as not a real value but a measurement error and removes the measured value. Do. As an example of this removal criterion, the data collection device 100 calculates the mean μ and the standard deviation σ of the previous systolic blood pressure of the pregnant woman, and when a measured value is dissociated by nσ or more from the mean μ, There are methods such as removing the measured value (n is an arbitrary numerical value), etc., but it is not limited thereto. Further, with regard to the missing value, there are a linear interpolation or spline interpolation using points before and after, a method of complementing using a model fitted to a measurement point other than the missing value, and the like, but it is not limited thereto.

  In step s102, the data collection device 100 predicts the occurrence of data items having a large difference in data distribution between the HDP onset group and the other groups, using the data of a pregnant woman who has already given birth and the birth status has been determined. Extract as feature quantity of model. As a definition of the size of the difference in data distribution, for example, the indicator “the probability that the average value of the data is not the same is large” or “the probability that the distribution of the data is not the same is large” can be used Although there is no limitation to them. This is useful with methods such as Student's t-test, Welch's t-test, multivariate regression, multivariate logistic regression, chi-square test, Fisher's exact test, Lasso regression, Ridge regression, Elastic Net, etc. There are a method of limiting the data items, a method of limiting the data items from the contribution rate of the principal component analysis, and the like, but it is not limited thereto. The feature amount refers to a data item used as a variable of a model when creating any prediction model, such as blood pressure and body temperature.

  In steps s103, s105, and s107, the data collection device 100 generates an HDP onset prediction model using the extracted feature quantities. Specifically, in step s103, the data collection device 100 generates an HDP onset prediction model as a classification model for predicting whether to become an HDP. Further, in step s105, the data collection device 100 generates an HDP onset day prediction model as a regression model for predicting the HDP onset date of each pregnant woman. In addition, the data collection device 100 generates an HDP onset risk calculation model for calculating the risk of HDP onset in s107. The regression model includes, but is not limited to, linear regression model, logistic regression model, regression model by SVM, regression model by multi-layered neural network, and the like. Further, in the classification model, there are linear classification, logistic regression, Bayesian network, SVM, k-neighbor method, random forest, classification model by multilayer neural network, etc., but it is not limited thereto. The HDP risk calculation model includes, but is not limited to, odds ratio, contribution risk, relative risk, and the like. Further, steps s103, s105, s107 may be processed integrally without being separated from s102.

  In steps s104, s106 and s108, the data collection device 100 inputs the data of the pregnant woman to be predicted into the generated HDP onset prediction model, the HDP onset prediction model and the risk calculation model of HDP onset, respectively, and predicts the HDP onset. Run. Specifically, in step s104, the data collection device 100 predicts the presence or absence of future HDP onset from the data of the pregnant woman to be predicted using the HDP onset prediction model. Further, in step s106, the data collection device 100 predicts the date of HDP onset from the data of the pregnant woman to be predicted using the HDP onset day prediction model. Further, in step s108, the data collection device 100 calculates the HDP risk from the data of the pregnant woman to be predicted using the HDP onset risk calculation model. In addition, the HDP presence / absence prediction is a final prediction of HDP onset / presence, prediction of not developing HDP in the next n1 weeks (n1 is any numerical value), prediction of developing HDP in the next n2 weeks (n2 is arbitrary numerical value), Including, but not limited to.

In step s109, the data collection device 100 notifies the pregnant woman or medical institution of the HDP onset prediction result, the HDP onset date prediction result and / or the HDP onset risk calculation result, and ends the data collection processing.
(Example 2)
In Example 1 described above, data items used to generate and / or enter the HDP onset prediction model, HDP onset prediction model or HDP onset risk calculation model include gestational age, blood pressure, pulse, weight , BMI, body fat rate, skeletal muscle rate, number of steps, calories burned, body temperature, mood, sleep time, bedtime, bedtime, wake up time, number of waking up at bedtime, quality of sleep, number of bowel movements, softness of bowel movement, uterus Contraction, head pain, tooth pain, upper abdominal pain, lower abdominal pain, lower back pain, nausea, vomiting, palpitations, interval between meals, meal contents, medication contents, past HDP history, pregnancy status (such as multiple pregnancy) Temperature, room temperature, humidity, barometric pressure, weather conditions, tides and age may be limited to one or more (except in the case of BMI alone). This makes it possible to generate and use a model efficiently with a smaller number of data.
(Example 3)
FIG. 9 is a flowchart showing HDP onset prediction model generation processing according to an embodiment of the present invention. Regarding Example 2 described above, in the HDP onset prediction model generation process described later, the magnitude of the difference in data distribution between the HDP onset group and the non-HDP group is used to stably predict high HDP onset in a plurality of gestational weeks An HDP onset prediction model is generated that shows accuracy. The HDP onset prediction model generation process can be executed in steps s102 and s103 of FIG. 8 described above, and steps s102-1 to s8 described later correspond to the feature value extraction process of step s102, and step s103- 1 to 9 correspond to the generation processing of the HDP presence / absence prediction model in step s103. In Example 3, the HDP onset prediction model generation function 130 divides the acquired data into the HDP group and the non-HDP group, and the difference in data distribution between the two groups of data items is large (ie, distinguish, discriminate, A feature value used to generate an HDP onset prediction model by reducing the number of data items by extracting data items (such as separating) and further selecting one of the highly correlated data items in each data item. Extract In addition, the HDP onset prediction model generation function 130 selects an HDP onset prediction model with high prediction accuracy by sequentially increasing the number of feature amounts from the feature amount having a large difference in data distribution between the HDP group and the non-HDP group. The above is an example of a method of extracting feature amounts used in the HDP onset prediction model and a selection method of the HDP onset prediction model, and the HDP onset prediction generation model 130 is not limited to this. For example, there is also a method of extracting feature quantities by a multi-layered neural network or Lasso regression to generate an HDP onset prediction model.

  As shown in FIG. 9, in step s102-1, the data collection device 100 processes some data items of the acquired data items. Specifically, the data collection device 100 is not limited to the following for some data items, but the difference with the data of several days ago, the average and variance for several days, and the derivative thereof that generated the approximate model of the measurement value , Derive second-order derivatives, etc.

  First, the feature quantity extraction process will be described.

  In step s102-2, the data collection device 100 combines the original data and the data generated in s102-1 and extracts data of a pregnant woman who has already given birth for model generation.

  In step s102-3, the data collection device 100 presupposes that the data items useful for predicting the occurrence of HDP change with gestational weeks, and p-values between the HDP group and the non-HDP group for each data item are made pregnant. Start a loop to calculate every week number. For example, when extracting the feature quantity in each week up to N to N + n weeks, execute the in-loop process with t1 = N × 7 (days) as parameter t1 indicating the number of pregnancy days, and then execute as t1 = (N + 1) × 7 Finally, it is executed as t1 = (N + n) × 7.

  In step s102-4, the data collection device 100 selects data whose period is from (t1-a) day to (t1 + b) day from the data extracted in s102-2 at the beginning of in-loop processing, and those data Divide the data into a group in which the pregnant women linked to the data became HDP and a group in which they were not. a and b are arbitrary numerical values, for example, a = 0 and b = 6 in the case of generating a model on a weekly basis. Further, by enlarging a = 28 or the like, it is possible to extract subsequent data items by using long-term data after b days from data 28 days before t1. Thereafter, in step s102-4, the data collection device 100 calculates ap value using, for example, Welch's t-test.

  In step s102-5, the data collection device 100 calculates the p value of each data item by the amount of t1, and ends the loop.

  In step s102-6, the data collection device 100 selects one or more arbitrary week combinations listN1 out of N to N + n weeks, and calculates an average value of p values of each data item in listN1. Do. In averaging p values, weights may be given for each gestational week. Note that listN1 may be all weeks of N to N + n weeks.

  In step s102-7, the number of data items is narrowed by putting together data items having high correlation. Here, the correlation coefficient of each data item is calculated, and those of c or more are classified as the same cluster, and the item with the lowest average value of p values in each cluster is left except the others, Reduce the number of data items to the number of clusters. Here, c is an arbitrary numerical value, for example, c = 0.7. As a cluster configuration method, it is assumed that the mutual data items in the same cluster are all d or more, but it is not limited thereto.

  In step s102-8, the data collection device 100 excludes data items whose average value of p values is d or more. Here, d is an arbitrary numerical value, for example, d = 0.05. However, the present invention is not limited to this, and any appropriate discrimination model (for example, a statistical model, a learning model, etc.) capable of separating the HDP group and the non-HDP group may be applied. The data items remaining above are rearranged in ascending order of the average value of p values, and these are used as feature amounts.

  Next, a method of generating an HDP onset prediction model will be described. Since a model using all of the feature quantities generated above is not necessarily optimal, it is considered to gradually increase the number of feature quantities to generate a model and select a model with high prediction accuracy.

  In step s103-1, the data collection device 100 increases the number of feature quantities for model generation from 1 and stores the list of feature quantity names in the parameter f, and executes the subsequent in-loop processing. In addition, the combination of the used feature-value is shown in FIG.

  In step s103-2, the data collection device 100 extracts one or more arbitrary weeks out of N to N + n weeks, generates a combination list N2 of those converted to gestational days, and generates a pregnancy to the parameter t2 Store a list of days and execute the in-loop processing of the latter stage. Note that the week used in the generation of listN2 may be all weeks from N to N + n weeks, and may be the same as N1 described above.

  In step s103-3, the data collection device 100 generates learning data and evaluation data as preparation for model generation. Specifically, the data collection device 100 extracts data of the period (t2-a) to (t2 + b), and first divides the data into the HDP group D1 and the non-HDP group D2. Further, the data collection apparatus 100 further divides D1 into D1a and D1b, divides D2 into D2a and D2b, and sets learning data as D1a + D2a, where e is a ratio of learning data to evaluation data as e. The data is D1b + D2b. Here, e may be any numerical value, for example, e = 0.5. It is also possible to adjust the number of pregnant women used as D2, for example the number of pregnant women of D2 can be up to twice the number of pregnant women of D1.

  In step s103-4, the data collection device 100 learns a model by limiting the feature amount to the feature amount name list f. For example, a support vector machine (SVM) is used as a prediction model at this time, and learning is performed using data for learning. Since there can be a large number of onset prediction models, this onset prediction model is M (f, t2) for distinction.

  In step s103-5, the data collection device 100 uses the evaluation data to evaluate the accuracy with which the model can predict the HDP. As an evaluation method, for example, an ROC (Receiver Operating Characteristic) curve is prepared, which has an F value, sensitivity, specificity, or a false positive rate on the horizontal axis and sensitivity on the vertical axis, and the area under the curve is AUC Calculate Area Under the Curve. The higher the F value, sensitivity, specificity, or AUC, the better the prediction accuracy. The data collection device 100 records the evaluated F value, sensitivity, specificity, or AUC together with the prediction model M (f, t2). The evaluation method is not limited to this, and it is also possible to combine the value of the HDP risk as calculated in the same manner as in s102-1 to s103-9 in s107 of FIG.

  In step s103-6, the data collection device 100 generates a model for t2 and terminates the loop.

  In step s103-7, the data collection device 100 calculates the F value in the list N2, the sensitivity, the specificity, or the average value of AUC. In averaging of the F value, sensitivity, specificity or AUC, weighting may be performed every t2.

  In step s103-8, the data collection device 100 calculates the F value, the sensitivity, the specificity, or the average value of AUC by f, and ends the loop.

  In step s103-9, the data collection device 100 selects f having the highest F value, sensitivity, specificity, or average value of AUC as an optimal combination (model) of feature quantities. Note that this selection method is not limited to this, and the value of the HDP risk that is calculated by the same method as s102-1 to s103-9 in s107 of FIG. Or in combination with AUC.

  As described above, it is possible to generate an HDP onset prediction model that stably shows high HDP onset prediction accuracy in a plurality of gestational weeks. For example, when listN1 and listN2 are [24 × 7, 25 × 7, 26 × 7], blood pressure, BMI, age, poor sleep quality and painless uterine contraction are used. On the other hand, when listN1 and listN2 are [27 × 7, 28 × 7, 29 × 7], blood pressure is used.

  Although the process for generating the HDP onset / prediction model is described in the above-described embodiment, it is apparent that the HDP onset day prediction model and the HDP onset risk calculation model can be generated by the same process.

Further, as understood from the above-described embodiment, the HDP onset prediction model generation function 130 sequentially generates features from HDP group and non-HDP group from the feature amount having a large difference in data distribution when generating the model from the feature amount. A model with high prediction accuracy may be selected by increasing the quantity number. In these figures, when the number of feature quantities used in the model is taken on the horizontal axis, and the average value of F values at weeks 24, 25 and 26 is confirmed, the implementation result as shown in FIG. 10 is obtained. As the number of feature quantities increases, the F value tends to be high, and it can be seen that the accuracy of predicting the onset of HDP is enhanced. Thereby, it can be confirmed that the prediction accuracy is improved by combining a plurality of data. In addition, it can be understood that the average value of the F value is higher in the case of eight than in the case where the number of feature amounts is nine or more, and the loop of the parameter f in step s103-1 is effective. This suggests that it is possible to find a more optimal combination of data by selecting what kind of data to combine. In the case of using the optimum combination of feature quantities (eight feature quantities) selected in Example 3, the F value indicated by the prediction model for the data for each week is 0.62 in the data for 24 weeks and 0 for 25 weeks. It was 0.50 at .52, 26 weeks. In addition, the odds ratio for the HDP indicated by the prediction model for the data for each week is 33 for the data for 24 weeks, 15 for the 25th week, 11 for the 26th week, and the odds ratio for 3.4 to 8.8 shown in Non-Patent Document 1. It can be seen that the prediction of the onset of HDP is improved by using the data collection system 10 of the present invention.
(Example 4)
FIG. 11 is a flowchart showing HDP onset prediction model generation processing according to an embodiment of the present invention. With regard to Example 2 described above, in the HDP onset prediction model generation process described later, an HDP onset prediction model is generated that stably shows the HDP onset prediction accuracy higher than that of Example 3 in a plurality of gestational weeks. The HDP onset prediction model generation process can be executed in steps s102 and s103 of FIG. 8 described above, and steps s102-1 to 8 described later correspond to the feature quantity extraction process of step s102, and the steps The steps s103-1 'to 10' correspond to the process of generating the HDP onset / prediction model in step s103. Also, steps s102-1 to 7 'are equal to s102-1 to 7 described in the third embodiment. In the fourth embodiment, the HDP onset prediction model generation function 130 sequentially reduces feature amounts with small absolute values of weights in the HDP onset prediction model from feature amounts with large differences in data distribution between the HDP group and the non-HDP group. , Select HDP onset prediction model with high prediction accuracy. In addition, the combination of the used feature-value is shown in FIG.

  As described above, it is possible to generate an HDP onset prediction model that stably shows high HDP onset prediction accuracy more stably than Example 3 in a plurality of gestational weeks. For example, when listN1 and listN2 are [24 × 7, 25 × 7, 26 × 7], blood pressure, lower back pain, painless uterine contraction, BMI, number of occurrences, age are used. On the other hand, when listN1 and listN2 are [27 × 7, 28 × 7, 29 × 7], blood pressure and lower back pain are used.

For the models generated in Example 3 and Example 4, when the average of the F values at weeks 24, 25 and 26 is confirmed with the feature quantity number on the horizontal axis, the implementation as shown in FIG. 12 is performed. The result is obtained. It can be confirmed that a model is generated in which the F value of the model generated in Example 4 is higher than that in Example 3 and the prediction accuracy is higher. When the optimum combination of feature amounts selected in Example 4 (6 feature amounts) is used, the F value indicated by the prediction model for the data for each week is 0.55 in the data for 24 weeks, and 0 for 25 weeks. The average value was 0.56 at 0.58 at .57 and 26 weeks. Also, the odds ratio for the HDP indicated by the prediction model for each week of data was 17, in the 24 week data, 45 in the 25th week, 48 in the 26th week, and the average value was 37. The average value of the F value and the odds ratio in Example 4 exceeds the average value of 0.55 of the F value in Example 3 and the average value of the odds ratio 20 in Example 3, respectively, and using the data collection system of the present invention It can be seen that the prediction of the onset of HDP is improved.
(Example 5)
The HDP onset prediction model generation function 130 and the HDP onset prediction function 140 of the data collection device 10 according to the present invention can set a target disease other than HDP. For example, gestational diabetes, SGA (small-for-gestational age), fetal growth failure, stillbirth, threatened abortion, premature birth, premature birth, premature rupture, premature amniotic fluid, excessive amniotic fluid, severe amniotic fluid, eclampsia, acute pregnancy fatty liver, HELLP syndrome, It is possible to set one or more of atopy, mental disease of the child, cardiovascular disease of the child, metabolic disorder of the child, and other diseases related to the pregnant woman and the fetus, and these onset prediction, onset date prediction or onset risk Calculations can be made. It is also possible to predict full-term delivery.

Thus, according to the present invention, in addition to conventional questionnaires and medical records, daily life logs can be collected noninvasively during pregnancy and before and after pregnancy, and it is possible to predict the presence of HDP in HDP presence / absence prediction. It is possible to diversify data items, increase data acquisition period and number of acquisition points, and minimize data objectivity and recall bias. Furthermore, by combining and analyzing these obtained various data, it is possible to predict the onset of HDP, leading to prior prevention and prompt response. In addition, it is possible to predict daily life logs and detect abnormalities, and prevent the early detection of symptoms and seriousness by prompting a consultation as needed without waiting for a pregnancy checkup once every several weeks. Leads to quick action.
(Example 6)
FIG. 13 is a flowchart showing an optional onset prediction model generation process according to an embodiment of the present invention. As described later, by using the disease onset prediction model generation process of the sixth embodiment, a disease onset prediction model indicating a disease onset prediction score higher than that of the third embodiment is generated stably at a plurality of gestational weeks. The disease onset prediction model generation process can be executed in steps s102 and s103 of FIG. 8 described above, and steps s1022-1 to s1022-6 described later correspond to the feature value extraction process of step s102, and the steps Although s1032-1 to s1032-11 correspond to the generation processing of the disease onset / prediction prediction model of step s103, they can not be distinguished strictly between steps s102 and s103, and overlap partially. In the sixth embodiment, the disease onset prediction model generation function 130 selects a feature amount having the highest prediction score in the disease onset prediction model from all feature amounts in which the difference in data distribution is large between the disease group and the normal group. A disease onset prediction model with a high prediction score is selected by sequentially adding feature amounts that further increase the prediction score from the remaining feature amounts. Although various indexes can be designated as the prediction score, for example, an F value can be designated.

  In FIG. 13, steps s1022-1 to s1022-5 are equal to steps s102-1 to s102-5 described in the third embodiment. Step s1022-6 is equivalent to step s102-7 of the third embodiment. Up to this point, in each data item, calculate the p value to divide the disease onset group and the normal group, compare the correlation between each data item, and narrow down the data items to one out of high correlation. It is possible to reduce the number of items.

  Next, select one or more weeks out of any week from N to N + n (for example, 16, 20, 24, 28, 32 weeks) and create a disease prediction model in the data for those weeks. Think. In step s-1032-1, a loop is started by sequentially inserting the days (for example, 16 × 7, 20 × 7, 24 × 7, 28 × 7, 32 × 7) corresponding to the number of weeks selected for the parameter t2 . In step s1032-2, an average of p values of each data item before and after t2 at t is calculated. For example, Δt = 7. In step s-1032-3, in the number of weeks for which a disease prediction model is to be generated, thereby selecting each data item group in which each data item is stable and the difference in data distribution between the two groups is large, and use them as feature values. . Also, in order to calculate the final optimal feature value group, an empty array variable called f_best is prepared, and a variable called best_score is prepared.

  In step s1032-4, a loop is started in which all the feature quantities are sequentially substituted into the variable f and the subsequent processing is executed, and the feature having the highest prediction score from the feature quantity group in the variable f and the array variable f_best Search for groups of quantities, ie combinations of data items. For example, when there are four feature amount groups {fa, fb, fc, fd}, first execute steps s-1032-5 to s-1032-8 as f = fa, and then f = fb Execute the same process and repeat it to fd.

  In step s1032-5, if the variable f is included in f_best, the process moves to the next variable f so that feature amounts already included in the array variable f_best are not searched redundantly. For example, when fa is processed as a variable f and fa is included in f_best, the process of fa is ended, and the following process is performed from step s1032-4 with fb as a variable f.

  In step s1032-6, the data collection device 100 generates learning data and evaluation data as preparation for model generation. Specifically, the data collection device 100 extracts data of the period (t2-a) to (t2 + b), and first divides the data into the disease group D1 and the normal group D2. Further, the data collection apparatus 100 further divides D1 into D1a and D1b, divides D2 into D2a and D2b, and sets learning data as D1a + D2a, where e is a ratio of learning data to evaluation data as e. The data is D1b + D2b. Here, e may be any numerical value less than 1, for example, e = 0.5. It is also possible to adjust the number of pregnant women used as D2, for example the number of pregnant women of D2 can be up to twice the number of pregnant women of D1.

  In step s1032-7, the data collection device 100 learns a model by limiting the feature amount to f_best + f. That is, when f = fc and f_best has two feature quantities fa and fb, the prediction model is learned with three feature quantities fa, fb and fc. For example, a support vector machine (SVM) is used as a prediction model at this time, and learning is performed using data for learning. Since there are a large number of prediction models for the onset eventually, this onset prediction model is M (f_best + f, t2) for discrimination.

  In step s1032-8, the data collection device 100 uses the data for evaluation to evaluate a prediction score that allows the model to predict a disease. As an evaluation method, for example, an ROC (Receiver Operating Characteristic) curve is prepared, which has an F value, sensitivity, specificity, or a false positive rate on the horizontal axis and sensitivity on the vertical axis, and the area under the curve is AUC Calculate Area Under the Curve. The higher the F value, sensitivity, specificity, or AUC, the better the prediction score. If the estimated score thus evaluated is higher than the best_score at this time, the data collection device 100 sets the array variable f_best 'to f_best' = f_best + f, further changes the variable best_score 'to the current prediction score, and recursively performs step s-1032. Execute from -4. At this time, f_best 'is used in place of f_best and best_score' is used in place of best_score in steps s-1032-4 and later. That is, using the new f_best 'and best_score' updated this time, the prediction score of the pattern in which one of all the feature quantities is added as a variable f 'is evaluated, and if there is a higher one, the search is further recursively performed Do. If there is no higher prediction score through all f ', end the recursive loop and substitute the array variable f_best' and the variable best_score 'at that time into the array variable f_best before the recursive call and the variable best_score to update . The evaluation method is not limited to this.

  The step s-1032-9 corresponds to the end of the loop of the variable f which started the loop in the step s-1032-4 and is repeated until the f ends all the feature quantities.

  In step s-1032-10, the prediction model M (f_best, t2) that has become the best_score is selected as the optimum prediction model.

  Step s-1032-11 corresponds to t2 at which the loop is started in step s-1032-1 and is repeated until t2 ends the entire period for which a prediction model is desired to be created.

From the above, it is possible to create a predictive model that is searchable.
(Example 7)
FIG. 14 is a flowchart showing HDP onset prediction model generation processing according to an embodiment of the present invention. Compared to the third embodiment described above, in the process described later, using data of a plurality of weeks, not only the magnitude of the difference between the data distribution of the HDP group and the non-HDP group but also the time series change of the data distribution It is possible to generate a more sophisticated prediction model by using the magnitude of the difference of For the definition of the magnitude of the difference in time series change of data distribution, for example, while the data distribution of certain data items in each of the HDP group and the non-HDP group changes according to the number of gestational weeks, There is an indication that “the probability that can not be said to be the same is large” is available, but it is not limited thereto. For example, in the non-HDP group, certain data items decrease according to the number of gestational weeks, whereas in the HDP group, no decrease is observed. This is useful with methods such as Student's t-test, Welch's t-test, multivariate regression, multivariate logistic regression, chi-square test, Fisher's exact test, Lasso regression, Ridge regression, Elastic Net, etc. There are a method of limiting the data items, a method of limiting the data items from the contribution ratio of the principal component analysis, etc.

The HDP onset prediction model generation process can be executed in steps s102 and s103 of FIG. 8 described above, and steps s202-1 to s7 described later perform the feature value extraction process of step s102, and step s203-1 6 perform the generation processing of the HDP onset / prediction prediction model in step s103.
In the fifth embodiment, the HDP onset prediction model generation function 130 divides the acquired data into the HDP group and the non-HDP group, and the data distribution differs between the two groups among the data items (that is, distinguish, discriminate, separate Etc.) but extract confounding factors including multiple weekly data and static data (age, non-fertility BMI, height, medical history, past labor history, etc.) and adjust their effects. ,Extract. Further, the number of data items is reduced by selecting one of the data items having high correlation in each data item, thereby extracting the feature value used for generating the HDP onset prediction model. In addition, the HDP onset prediction model generation function 130 extracts a predetermined number of high-order feature amounts among feature amounts having different data distributions in the HDP group and the non-HDP group, and uses the extracted feature amounts to calculate the HDP onset prediction model. Create and identify HDP onset prediction model with the best performing feature value. Next, the HDP onset prediction model generation function 130 creates a two-variable prediction model combining the remaining feature amounts excluding the feature amount and the specified feature amounts, and including these models, an optimal model Choose The HDP onset prediction model generation function 130 recursively executes the above-described series of processes, and finally identifies the HDP onset prediction model. The above is an example of a method of extracting feature quantities used in the HDP onset prediction model and a selection method of the HDP onset prediction model, and the HDP onset prediction model generation function 130 is not limited thereto. For example, there is also a method of extracting feature quantities by multi-layer neural network, Lasso regression, etc., and creating an HDP onset prediction model.

  As shown in FIG. 14, in step s202-1, the data collection device 100 processes some data items of the acquired data items. Specifically, the data collection device 100 is not limited to the following for some data items, but the difference with the data of several days ago, the average and variance for several days, and the derivative of the approximation model of the measurement value created , Derive second-order derivatives, etc.

  First, the feature quantity extraction process will be described.

  In step s202-2, the data collection device 100 combines the original data and the data created in s202-1 and extracts pregnancy data and time-lapse data of a pregnant woman who has already given birth, for model creation. Pregnancy data are data of static data items not depending on gestational age, and include gestational age, non-fertility BMI, non-fertility weight, height, postpartum history, medical history, family history and the like. Temporal data is a list of data obtained by aggregating measurement values of a specific period during pregnancy. For example, one week can be specified as a specific period, user ID, blood pressure of 16 weeks, pulse of 16 weeks, The calculated values of each pregnant woman are described in the lateral direction such as body temperature of 16 weeks, ..., blood pressure of 17 weeks, pulse of 17 weeks, body temperature of 17 weeks, and so on. In addition, an arbitrary period can be designated to a specific period. Moreover, as a method of tabulating measurement values in a specific period, any method such as average value or median value can be specified.

  In step s202-3, with respect to the above pregnancy data, data of a group in which a pregnant woman linked to the data becomes HDP and a group of a group in which the pregnant woman does not become HDP are extracted. Thereafter, the data collection device 100 extracts data items having different data distributions between the two groups as a confounding data item using the method described in s102. Here, for example, data items in which the p value is less than a (for example, a = 0.01) in multivariate logistic regression are extracted. Furthermore, a seasonal term may be included in the confounding data item. The present invention is not limited to this, and any appropriate discrimination model (for example, a statistical model, a learning model, etc.) capable of separating the HDP group and the non-HDP group may be applied.

  In step s202-4, the correlation coefficient between those confounding data items is calculated, and those having a correlation of b (for example, b = 0.75) or more are divided into the same cluster, leaving the item with the lowest p value in each cluster. Reduce the number of confounding data items to the number of clusters by excluding the others. As a cluster configuration method, it is assumed that the mutual data items in the same cluster are all b or more, but it is not limited thereto. In some cases, s202-3 may be executed after s202-4 is executed, or s202-3 and s202-54 may be repeatedly executed.

  Next, in step s202-5, the data collection device 100 adds a confounding data item to each data item one by one with respect to pregnancy data and time-lapse data to calculate p value and odds ratio by logistic regression, p By selecting a data item whose value is less than c (for example, c = 0.01), a data item that divides the two groups into significant after excluding the influence of the confounding factor is specified.

  In step s202-6, the correlation coefficient between those data items is calculated, and those having a correlation of d (for example, d = 0.75) or more are divided into the same cluster, leaving the item with the lowest p value in each cluster Reduce the number of data items to the number of clusters by excluding the others. The cluster configuration method is the same as in the case of s202-4.

  Next, in step s202-7, the data collection device 100 rearranges the data items remaining above in ascending order of p value, and uses these as feature amounts for creating an HDP onset prediction model. In addition, information such as data item name, odds ratio, p value, and regression coefficient β is stored as a feature amount table.

  Next, a method of creating an HDP onset prediction model will be described.

  Since a model using all of the feature quantities created above is not necessarily optimal, it is considered to create a model by gradually increasing the number of feature quantities and select a model with high prediction accuracy.

  In step s203-1, the data collection device 100 assumes that the risk of the subsequent onset of HDP changes according to the progress of pregnancy, and constructs a prediction model of the onset of the HDP in fixed time units. A certain unit of time may be, for example, one week, but it is not limited thereto. Here, it is assumed that a prediction model is built at each week (t week) of 16 weeks to 36 weeks of pregnancy.

  In step s203-2, the data collection device 100 extracts pregnancy data and time-lapse data before t week from the feature quantities identified in s202-7 in the pregnancy data and time-lapse data of s202-2.

  In step s203-3, the data collection device 100 creates learning data and evaluation data as preparation for model creation. Specifically, the data acquisition device 100 is first divided into the HDP group D1 and the non-HDP group D2. Further, the data collection apparatus 100 further divides D1 into D1a and D1b, divides D2 into D2a and D2b, and sets learning data as D1a + D2a, where e is a ratio of learning data to evaluation data as e. The data is D1b + D2b. Here, e may be any numerical value, for example, e = 0.5. It is also possible to adjust the number of pregnant women used as D2, for example the number of pregnant women of D2 can be up to twice the number of pregnant women of D1.

  In step s203-4, the data collection device 100 may oversample D1a in the learning data D1a and D2a, and / or undersample D2a to improve the prediction performance.

  In step s203-5, the data collection device 100 selects feature quantities by forward stepwise selection using the learning data D1a and D2a, and learns an HDP onset prediction model Mt at t week time point. For example, a support vector machine (SVM) is used as a prediction model at this time, and learning is performed using data for learning. For example, for the first 100 feature quantities (pregnancy data and time-lapse data), the data collection apparatus 100 uses only one feature quantity to predict the HDP onset provisional model Mt (1) to Mt (100) at time t. Create Then, the data collection apparatus 100 identifies Mt (A) (using the feature amount A) with the highest performance, using the evaluation data D1b and D2b. As the performance score, for example, area under the curve (AUC), f-score, harmonic mean of sensitivity and specificity, etc. may be used. Next, using the learning data D1a and D2a, the data collection device 100 selects one feature quantity from the remaining 99 feature quantities excluding the feature quantity, and causes the two variables HDP onset to be combined with A. Predictive models Mt (A, 1) to Mt (A, 99) are created, and the optimal data Mt (A, B) including Mt (A) is selected using the evaluation data D1b and D2b. The data acquisition device 100 recursively executes the series of processes described above, and finally identifies the HDP onset prediction model Mt (A, B, C,...).

  The HDP onset prediction model M may be generated based on SVM, linear logistic regression, neural network, random forest, and the like. In addition, as a method of selecting the optimal feature amount, not only forward step selection but backward stepwise selection or bidirectional stepwise selection may be used. Also, after using all data items and reducing dimensions by PCA, MDS, NMDS, etc., apply forward / backward / bidirectional stepwise selection to N feature values of PC1, PC2, ..., PCN. It is also good.

In step s203-6, the data collection device 100 creates models at all points in time for t, and ends the loop.
(Example 8)
FIG. 15 is a flowchart showing HDP onset date prediction model creation processing according to another embodiment of the present invention. By using multiple weeks of data, it is possible to generate a higher-performance prediction model by using not only the difference in data distribution between HDP group and non-HDP group but also the difference in time series change of data distribution . The HDP onset date prediction model creation processing can be executed in steps s102 and s105 of FIG. 8 described above, and the above-described steps s202-1 to s7 perform the feature amount extraction processing of step s102, and the below-described step s205 -1 to 6 perform generation processing of the HDP onset day prediction model in step s105. Steps s202-1 to 7 in FIG. 15 are the same as steps s202-1 to 7 in FIG. 14, and steps s205-1 to s4 and s205-6 are steps s203-1 to s203 and s203-6 in FIG. Since it is the same as, its explanation is omitted.

  In step s205-5 of FIG. 15, the data collection device 100 selects feature quantities by forward stepwise selection using the learning data D1a and D2a, and learns an HDP onset date prediction model Mt at t week time point. For example, a support vector machine (SVM) is used as a prediction model at this time, and learning is performed using data for learning. For example, for the first 100 feature quantities (pregnancy data and time-lapse data), the data collection apparatus 100 uses only one feature quantity to predict the date of onset of HDP at time t, and the provisional models Mt (1) to Mt (100 (100) Create). Then, the data collection apparatus 100 identifies Mt (A) (using the feature amount A) with the highest performance, using the evaluation data D1b and D2b. As a performance score, for example, MSE (mean squared error) or R2 (decision coefficient) may be used. Next, using the learning data D1a and D2a, the data collection device 100 selects one feature quantity from the remaining 99 feature quantities excluding the feature quantity, and causes the two variables HDP onset to be combined with A. A daily prediction model Mt (A, 1) to Mt (A, 99) is created, and an optimal model Mt (A, B) including Mt (A) is selected using the evaluation data D1b and D2b. The data acquisition device 100 recursively executes the above-described series of processes, and finally identifies the HDP onset date prediction model Mt (A, B, C,...).

The HDP onset day prediction model M may be generated based on SVM, linear regression, neural network, or the like. In addition, as a method of selecting the optimal feature amount, not only forward step selection but backward stepwise selection or bidirectional stepwise selection may be used. Also, after using all data items and reducing dimensions by PCA, MDS, NMDS, etc., apply forward / backward / bidirectional stepwise selection to N feature values of PC1, PC2, ..., PCN. It is also good.
(Example 9)
FIG. 16 is a flowchart showing HDP risk prediction model creation processing according to another embodiment of the present invention. By using multiple weeks of data, it is possible to generate a higher-performance prediction model by using not only the difference in data distribution between HDP group and non-HDP group but also the difference in time series change of data distribution . The feature quantity extraction / model creation process can be executed in steps s102 and s107 of FIG. 8 described above, and the above-described steps s202-1 to s7 correspond to the feature quantity extraction process of step s102, and Steps s207-1 to s6 correspond to the process of generating an HDP onset risk prediction model in step s107. Steps s202-1 to 7 in FIG. 16 are the same as steps s102-1 to 7 in FIG. 14, and steps s207-1 to s4 and s207-6 are steps s203-1 to 4, s203-6 in FIG. Since it is the same as, its explanation is omitted.

  In step s207-5 of FIG. 16, the data collection device 100 creates the HDP onset risk prediction model Mt in the week t by the same method as s203-5, but the model used at this time includes multivariate logistic regression and the like. Do.

  The data collection device 100 of FIG. 8 can calculate the probability p of becoming an HDP by applying feature amount data to the HDP risk prediction model Mt1 in a pregnant woman who has not given birth weeks t1 in s108. .

  Although the embodiments of the present invention have been described in detail, the present invention is not limited to the specific embodiments described above, and various modifications may be made within the scope of the subject matter of the present invention described in the claims.・ Change is possible.

DESCRIPTION OF SYMBOLS 10 Data acquisition system 20 Measurement apparatus 30 Measurement value transfer apparatus 40 Physical condition registration apparatus 50 Weather information acquisition apparatus 60 Questionnaire registration apparatus 70 Medical record registration apparatus 100 Data acquisition apparatus 200 Network

Claims (7)

A data acquisition function for acquiring data on pregnant women from multiple data source devices;
An HDP onset prediction function that executes HDP onset prediction of the pregnant woman from the acquired data based on the HDP onset prediction model;
A result notification function for notifying the result of the HDP prediction;
Have
The data acquisition device acquires the data at a plurality of time points in a data acquisition period from any time before or after pregnancy of the pregnant woman to any time before or after delivery of the pregnant woman.   The data on the pregnant woman include pregnancy age, blood pressure, pulse, body weight, BMI, body fat rate, skeletal muscle rate, number of steps, calorie consumption, body temperature, mood, sleeping time, bedtime, wake up time, number of times during sleeping , Quality of sleep, frequency of bowel movements, softness of bowel movements, uterine contraction, head pain, tooth pain, upper abdominal pain, lower abdominal pain, lower back pain, nausea, vomiting, palpitations, interval of feces, food contents The data collection device according to claim 1, comprising one or more data items of medication content, past HDP history, pregnancy status, temperature, room temperature, humidity, barometric pressure, weather conditions, tides and age.   The data on the pregnant woman include pregnancy age, blood pressure, pulse, body weight, BMI, body fat rate, skeletal muscle rate, number of steps, calorie consumption, body temperature, mood, sleeping time, bedtime, wake up time, number of times during sleeping , Quality of sleep, frequency of bowel movements, softness of bowel movements, uterine contraction, head pain, tooth pain, upper abdominal pain, lower abdominal pain, lower back pain, nausea, vomiting, palpitations, interval of feces, food contents The data collection device according to claim 1 or 2, comprising a combination of two or more data items of medication content, past HDP history, pregnancy status, temperature, room temperature, humidity, barometric pressure, weather conditions, tide and age.   The HDP onset prediction function uses the acquired data, and in the HDP onset prediction model, the magnitude of the difference in data distribution between the HDP group and the non-HDP group and / or the magnitude of the difference in time series change in data distribution The data collection device according to any one of claims 1 to 3, wherein the prediction of the onset of HDP, the prediction of the onset date of HDP, or the calculation of the risk of developing HDP is performed by using the above.   The data collection according to any one of claims 1 to 4, further comprising an HDP onset prediction model generation function of generating the HDP onset prediction model by sequentially increasing the number of feature quantities separating the HDP group and the non-HDP group. apparatus.   The data collection according to any one of claims 1 to 4, further comprising an HDP onset prediction model generation function of generating the HDP onset prediction model by sequentially reducing the number of feature quantities separating the HDP group and the non-HDP group. apparatus. Obtaining data about a pregnant woman from a plurality of data source devices;
Performing HDP onset prediction of the pregnant woman from the acquired data based on the HDP onset prediction model;
Notifying the result of the HDP prediction.
Have
The step of performing HDP prediction of the pregnant woman includes acquiring the data at a plurality of time points in a data acquisition period from any time before and after pregnancy of the pregnant woman to any time before and after delivery. Data collection method.

Images