Pregnancy-induced hypertension prediction support system, program, and method

Abstract

To support prediction of an onset of hypertensive disorder in pregnancy (HDP) from health checkup data of a prenatal checkup having received by a pregnant woman.SOLUTION: An onset prediction system comprises: a storage part for storing a discriminator that has learned so as to predict transitions among a plurality of states with different incidence of HDP based on prenatal checkup data for learning for every pregnant woman for learning and presence / absence of onset of HDP; an acquisition part for acquiring checked checkup data obtained by checkups that a prediction object pregnant woman has received during pregnancy; a discrimination part for discriminating a state that the pregnant woman may have during a remaining period of the pregnancy out of a plurality of states based on the health checkup data received by the pregnant woman and the discriminator; and a calculation part for calculating health checkup data during the remaining period of the pregnant woman predicted based on a state within the remaining period of the pregnant woman. The discriminator has leaned using a hidden Markov model with an internal state of the pregnant woman used as a hidden state and health checkup data for every internal state used as an output variable of the hidden state.SELECTED DRAWING: Figure 15

Description

[Technical Field]

[0001] The present invention relates to a system, a program, and a method for supporting prediction of the onset of pregnancy-induced hypertension. [Background technology]

[0002] Hypertensive disorder in pregnancy (HDP) is a type of perinatal complication that develops in approximately 5% of pregnant women. HDP causes various disorders, including liver and kidney dysfunction in the mother and intrauterine growth restriction in the baby. HDP develops suddenly and irreversibly and rapidly worsens. When it becomes severe, it can cause eclampsia, a type of maternal convulsion, and placental abruption. In other words, HDP is significantly associated with maternal and neonatal deaths and permanent disabilities in newborns. Therefore, early detection of HDP onset and early initiation of treatment are essential.

[0003] Proposals relating to HDP have been made in the past (see, for example, Patent Documents 1 to 4). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Special Publication No. 2015-525870 [Patent Document 2] Special Publication No. 2015-519107 [Patent Document 3] Japanese Patent Application Laid-Open No. 2002-122592 [Patent Document 4] Special Publication No. 2018-536170 Summary of the Invention [Problem to be solved by the invention]

[0005] The present invention aims to provide an onset prediction support system, an onset prediction support program, and an onset prediction support method that do not require invasive testing and that use the health check data of pregnant women obtained through general prenatal checkups to help predict the onset of HDP during pregnancy. [Means for solving the problem]

[0006] The HDP onset prediction support system of the present invention supports prediction of whether a pregnant woman will develop HDP by predicting the state the pregnant woman can be in during pregnancy from among multiple states with different incidence rates of preeclampsia.The HDP onset prediction support system includes: a memory unit that stores a classifier trained to predict transitions between multiple states based on training pregnancy health checkup data for each training pregnant woman and the presence or absence of HDP; an acquisition unit that acquires health checkup data obtained at health checkups attended by the pregnant woman who is the target of prediction during pregnancy; a discrimination unit that discriminates the state the pregnant woman can be in from among multiple states based on the pregnant woman's health checkup data and the classifier; and a calculation unit that calculates the predicted health checkup data for the pregnant woman's remaining period based on the pregnant woman's state during the remaining period.The classifier is characterized in that it is trained using a hidden Markov model in which the state is a hidden state and the health checkup data for each state is an output variable of the state. [Effects of the Invention]

[0007] The present invention uses medical checkup data of a pregnant woman obtained during prenatal checkups to assist in predicting the onset of HDP during pregnancy. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a schematic diagram showing an overview of a method for supporting prediction of HDP onset according to the present invention. FIG. [Figure 2] FIG. 1 is a schematic diagram showing the relationship between the hidden state of a pregnant woman and the output variable in a classifier (hidden Markov model) used in the above method. [Figure 3] 1 is a functional block diagram showing an embodiment of an HDP onset prediction support system according to the present invention. [Figure 4] FIG. 2 is a schematic diagram showing an example of output probabilities (diastolic blood pressure) for each hidden state stored in a storage unit included in the system. [Figure 5] FIG. 10 is a schematic diagram showing an example of output probability (systolic blood pressure) for each hidden state stored in the storage unit. [Figure 6] FIG. 2 is a schematic diagram showing an example of output probability (urinary protein) for each hidden state stored in the storage unit. [Figure 7] FIG. 10 is a schematic diagram showing an example of average values ​​of output variables for each hidden state stored in the storage unit. [Figure 8] FIG. 4 is a schematic diagram showing an example of state transition probability of the classifier. [Figure 9] FIG. 4 is a schematic diagram showing an example of state transition probability of the classifier. [Figure 10] FIG. 4 is a schematic diagram showing an example of the state transition probability stored in the storage unit. [Figure 11] FIG. 4 is a schematic diagram showing an example of the state transition probability stored in the storage unit. [Figure 12] 10 is a schematic diagram illustrating an example of an evaluation amount for each of the classifier candidates. FIG. [Figure 13] FIG. 2 is a schematic diagram showing an example of an ROC curve of the classifier. [Figure 14] FIG. 2 is a schematic diagram showing an example of an output variable predicted by the classifier. [Figure 15] FIG. 10 is a schematic diagram showing an example of output from the onset prediction support system including the classifier. [Figure 16] FIG. 10 is a schematic diagram showing an example of a hidden state predicted by the classifier. [Figure 17] FIG. 10 is a schematic diagram showing an example of a hidden state predicted by the classifier. [Figure 18] FIG. 10 is a schematic diagram showing an example of a hidden state predicted by the classifier. [Figure 19] FIG. 10 is a schematic diagram showing an example of a hidden state predicted by the classifier. [Figure 20] FIG. 10 is a schematic diagram showing an example of a hidden state predicted by the classifier. [Figure 21]FIG. 10 is a schematic diagram showing an example of a hidden state predicted by the classifier. DETAILED DESCRIPTION OF THE INVENTION

[0009] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of a system, a program, and a method for supporting prediction of pregnancy-induced hypertension (hereinafter referred to as "HDP") according to the present invention will be described below with reference to the drawings.

[0010] Overview of the onset prediction support method FIG. 1 is a schematic diagram showing an overview of the HDP onset prediction support method according to the present invention (hereinafter referred to as "the method"). The figure shows the relationship between the number of weeks of pregnancy and blood pressure. The figure shows, with solid circles, blood pressures (examples of medical checkup data) obtained at prenatal checkups attended by pregnant women before the onset of symptoms. The figure also shows, with dotted circles, blood pressures predicted by the classifier of the present invention based on the blood pressures obtained at prenatal checkups. The figure shows that pregnant women whose predicted blood pressures exceed the diagnostic line, which requires a doctor's diagnosis, are in the onset group, and pregnant women whose predicted blood pressures do not exceed the diagnostic line are in the non-onset group.

[0011] In this method, a classifier predicts the medical checkup data for the remaining period of a pregnant woman's pregnancy based on the medical checkup data the pregnant woman has received during her pregnancy. That is, this method predicts the medical checkup data for the pregnant woman for each week of pregnancy from the N+1th week of pregnancy onwards (for example, up to the 40th week) based on the medical checkup data obtained from the prenatal checkups the pregnant woman received up to the Nth week of pregnancy. A doctor who uses this method predicts whether or not the pregnant woman will develop HDP based on the medical checkup data predicted by this method. As a result, the doctor can take measures such as treatment for pregnant women who are predicted to develop HDP before the onset of the disease occurs.

[0012] FIG. 2 is a schematic diagram showing the relationship between the hidden state (internal state) of a pregnant woman and the output variable in the classifier used in this method. This method uses a classifier that is statistically machine-learned using a Hidden Markov Model (HMM).

[0013] Here, HMM is a learning model that can handle time-series data. HMM is a learning model in which a hidden state represented by a Markov process is observed through an output variable (response variable). HMM can handle missing data without relying on imputation, etc. HMM can handle the checkup data of each pregnant woman in a unified manner, even if the timing of the prenatal checkup varies slightly for each pregnant woman. Compared to other learning models such as recurrent neural networks (RNNs) that can handle time-series data, HMM has the advantages of having a high-speed learning algorithm and not requiring teacher labels except for the output variable. From this perspective, the inventors decided to use a classifier trained using HMM.

[0014] The classifier used in this method uses the number of weeks of pregnancy as a covariate for the state transition probability so that the transition of the pregnant woman's state according to the progression of the number of weeks of pregnancy (time) is reflected in the classification results. In other words, the classifier is an extension of the conventional HMM to a model (Markov dependent mixture model) that allows covariates dependent on the number of weeks of pregnancy (time) to be included in the parameters of the state transition matrix.

[0015] FIG. 2 is a schematic diagram showing the relationship between the hidden state, which is the internal state of a pregnant woman, and the output variable in the classifier (hidden Markov model) used in this method. The figure shows how the hidden state, which is the internal state of a pregnant woman, transitions from x1 to x2, and from x2 to x3, as the number of weeks of pregnancy progresses. The figure shows that the state transition probability from hidden state x1 to x2 is P 12 The state transition probability from hidden state x2 to x3 is P 23 This figure shows that the blood pressure and urinary protein (y1, y2, y3), which are the medical checkup data of pregnant women for each week of pregnancy, can be observed from the hidden state (x1, x2, x3) for each week of pregnancy.

[0016] In this embodiment, as described below, the method determines the current state of a pregnant woman from among 14 hidden states (hereinafter referred to as "states") from state 1 to state 14, which have different incidence rates of HDP, based on the medical checkup data the pregnant woman has undergone. Of the 14 states, state 1 has the lowest incidence rate of HDP, and state 14 has the highest incidence rate of HDP.

[0017] This method predicts the future state of a pregnant woman based on the determined current state of the pregnant woman (the state determined from the medical checkup data she has already received) and the state transition probability. This method predicts the future medical checkup data of the pregnant woman based on the predicted future state of the pregnant woman. A doctor who uses this method refers to the predicted future medical checkup data of the pregnant woman to determine whether the pregnant woman will develop HDP or the predicted timing of onset (number of weeks of pregnancy), and starts treatment for the pregnant woman as necessary.

[0018] ●Configuration of the onset prediction support system● FIG. 3 is a functional block diagram showing an embodiment of an HDP onset prediction support system (hereinafter referred to as "this system") according to the present invention.

[0019] This system is realized by an information processing device such as a personal computer. This system executes a HDP onset prediction support program (hereinafter referred to as "this program") according to the present invention. This program running on this system cooperates with the hardware resources of this system to realize this method.

[0020] The hardware resources of this system are, for example, processors such as a CPU (Central Processing Unit), MPU (Micro Processing Unit), DSP (Digital Signal Processor), etc. The processor executes instructions written in this program to realize each of the means provided by this system, which will be described later.

[0021] When an information processing device (not shown) different from the present system executes the present program, the information processing device functions in the same way as the present system and realizes the present method.

[0022] The system 1 includes a storage unit 2, a medical checkup data acquisition unit 3, a condition determination unit 4, an output variable calculation unit 5, and an output unit 6.

[0023] The storage unit 2 stores the program (classifier) ​​and information used by the system 1 to implement the method. The storage unit 2 is, for example, a semiconductor memory element such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), a RAM (Random Access Memory), or a flash memory. Details of the information used by the system 1 to implement the method will be described later.

[0024] The medical checkup data acquisition unit 3 acquires medical checkup data for each gestational week of the pregnant woman to be predicted. The medical checkup data in this embodiment includes systolic blood pressure, diastolic blood pressure, and urinary protein. In the present invention, the medical checkup data used to predict the onset of HDP includes at least one of systolic blood pressure, diastolic blood pressure, and urinary protein.

[0025] The medical checkup data acquiring unit 3 reads, for example, medical checkup data that a user (such as a doctor) of the present system 1 inputs on a screen displayed on a display (not shown) of the present system 1, and stores the data in the storage unit 2. Alternatively, the medical checkup data acquiring unit 3 receives medical checkup data from an information processing device (not shown) connected via a communication network, and stores the data in the storage unit 2.

[0026] The state determination unit 4 determines the state that the pregnant woman may be in within the remaining period of pregnancy from among a plurality of states based on the medical checkup data of the pregnant woman and a classifier. Details of the state determination will be described later.

[0027] The output variable calculation unit 5 calculates the health checkup data for each week of pregnancy within the predicted remaining period of the pregnant woman based on the state of the pregnant woman within the remaining period. The output variable calculation unit 5 calculates the predicted value of the health checkup data based on the states that the pregnant woman can take within the remaining period of the pregnant woman and the output probability of each output variable in each state. Details of the calculation of the health checkup data for each week of pregnancy of the pregnant woman will be described later.

[0028] The output unit 6 outputs the state determined by the state determination unit 4 and the medical checkup data calculated by the output variable calculation unit 5. The output unit 6 outputs the state and medical checkup data in various ways, such as storing the state in the memory unit 2, displaying the state on a display (not shown) of the present system 1, or transmitting the state to an information processing device connected to the present system 1 via a communication network.

[0029] ●Information stored in the memory unit● As described above, the memory unit 2 stores information used by the present system 1 to realize the present method. The information stored in the memory unit 2 includes information constituting the classifier and information for each pregnant woman to be predicted. The information constituting the classifier includes the state transition probability between each state for each week of pregnancy and the output probability for each state of the health check data (diastolic blood pressure, systolic blood pressure, and urinary protein). The information for each pregnant woman to be predicted includes the health check data (diastolic blood pressure, systolic blood pressure, and urinary protein) that the pregnant woman has undergone for each week of pregnancy.

[0030] Output probability 4 to 6 are schematic diagrams showing examples of output probabilities stored in the storage unit 2. FIG.

[0031] FIG. 4 shows an example of the output probability of diastolic blood pressure for each state. The figure shows the states (state 1 to state 14) and the possible diastolic blood pressures (d1, d2, . . . , d m ) output probability (P 1(d1) ,P 1(d2) ,···,P 14(dm) ) are stored in the storage unit 2 in association with each other. For example, in the state 1, the output probability of the diastolic blood pressure d1 is P 1(d1)In other words, the figure shows that the diastolic pressure in state 1 is 1(d1) indicates that it is d1.

[0032] Here, in this embodiment, "information J and information K are stored in the storage unit 2 in association with each other" means that the system 1 can read out information K (or information J) from the storage unit 2 by referring to the storage unit 2 based on information J (or information K). That is, for example, the system 1 can read out the output probability P 1(d1) can be read out.

[0033] FIG. 5 shows an example of the output probability of systolic blood pressure for each state. The figure shows the states (state 1 to state 14) and the possible systolic blood pressures (s1, s2, . . . , s n ) output probability (P 1(s1) ,P 1(s2) ,···,P 14(sn) ) are stored in the storage unit 2 in association with each other. For example, in the state 1, the output probability of the systolic blood pressure s1 is P 1(s1) In other words, the figure shows that the systolic blood pressure in state 1 is 1(s1) This indicates that it is s1.

[0034] FIG. 6 shows an example of the output probability of urinary protein for each state. The figure shows the states (state 1 to state 14) and the possible urinary proteins (p1, p2, . . . , p r ) output probability (P 1(p1) ,P 1(p2) ,···,P 14(pr) ) are stored in the storage unit 2 in association with each other. For example, in the state 1, the output probability of the urinary protein p1 is P 1(p1) In other words, the figure shows that in state 1, urinary protein has an output probability P 1(p1) indicates that it is p1.

[0035] In this way, the system 1 can specify the output probability for each piece of medical checkup data that can be obtained at the time of the determined state based on the determined state.

[0036] FIG. 7 is a schematic diagram showing an example of the average value of the medical checkup data for each condition. The figure shows that the state and the average value for each medical checkup data are stored in association with each other in the storage unit 2. The figure shows that, for example, the average value of the diastolic blood pressure in state 1 is 57 mmHg. This average value of the diastolic blood pressure is calculated based on the output probability of the diastolic blood pressure that can be taken for each state shown in FIG. 4. That is, for example, the average value of the diastolic blood pressure in state 1 is Average value = d1 × P 1(d1) +d2×P 1(d2) +···+d m ×P 1(dm) It is calculated as follows.

[0037] The method for calculating the average values ​​of systolic blood pressure and urinary protein for each state shown in the figure is the same as the method for calculating the average value of diastolic blood pressure described above.

[0038] ●State transition probability 8 and 9 are schematic diagrams showing examples of state transition probabilities of classifiers in the present invention.

[0039] FIG. 8 shows an example of state transition probabilities at the 25th week of pregnancy. FIG. 9 shows an example of state transition probabilities at the 35th week of pregnancy. In the figure, rectangles represent each state. In the figure, the letters inside the rectangles representing states, for example, ST5 represents state 5. The numbers in parentheses inside the rectangles, from left to right on the page, represent the average diastolic blood pressure, average systolic blood pressure, and average urinary protein in that state (see Figure 7). In the figure, arrows connecting rectangles indicate the transition from the state of the rectangle at the origin of the arrow to the state of the rectangle at the end of the arrow, and the numbers attached to the arrows are the state transition probabilities of that transition.

[0040] That is, for example, Figure 8 shows that the state to which State 5 transitions is either State 1, State 5 (no change), State 12, or State 13, and that the state transition probability of a transition to State 1 is 0.02, the state transition probability of a transition to State 5 is 0.91, the state transition probability of a transition to State 12 is 0.04, and the state transition probability of a transition to State 13 is 0.03.

[0041] Also, for example, Figure 9 shows that the state to which State 5 transitions is either State 1, State 5 (no change), State 12, State 13, or State 14, and that the state transition probability of a transition to State 1 is 0.01, the state transition probability of a transition to State 5 is 0.85, the state transition probability of a transition to State 12 is 0.06, the state transition probability of a transition to State 13 is 0.07, and the state transition probability of a transition to State 14 is 0.01.

[0042] 10 and 11 are schematic diagrams showing examples of state transition probabilities of classifiers stored in the storage unit 2. In FIG.

[0043] FIG. 10 shows an example of state transition probabilities at the 25th week of pregnancy shown in FIG. FIG. 11 shows an example of state transition probabilities at the 35th week of pregnancy shown in FIG. The figures show that the transition source state and the state transition probability from the transition source state for each transition destination state are stored in the storage unit 2 in association with each other.

[0044] ●Example of a classifier● Below is an example of a classifier created by the inventors.

[0045] ●Number of states The classifier was selected from multiple classifier candidates. Each of the multiple classifier candidates has a different number of states as a hyperparameter. That is, the inventors treated the number of states as a hyperparameter and had seven classifier candidates with numbers of states ranging from 10 to 16 learn training data. The training data included medical checkup data of training pregnant women (4,038 people) and information on whether each training pregnant woman developed HDP. Training was performed using statistical machine learning using an HMM. The HMM was trained as a model that predicts transitions between states (the most likely transitions between states). From the classifier candidates, the inventors selected the candidate that was thought to have the highest accuracy in predicting the onset of HDP.

[0046] FIG. 12 is a schematic diagram showing an example of the evaluation amount of each classifier candidate. The figure shows the relationship between the number of states and the evaluation value. In the figure, AIC indicates the evaluation value of Akaike Information Criterion, and BIC indicates the evaluation value of Bayes Information Criterion. The figure shows that the AIC measure decreases with increasing number of states. The figure shows that the BIC evaluation value decreases as the number of states increases, and then increases when the number of states is 15 or more. Therefore, the inventors selected a candidate classifier with 14 states from among the classifier candidates.

[0047] ●Prediction accuracy of the classifier FIG. 13 is a schematic diagram showing an example of an ROC (Receiver Operating Characteristic) curve of a classifier selected by the present inventors from among candidate classifiers. In the figure, graph A is the ROC curve for the training data (health checkup data for 3,384 pregnant women throughout their entire pregnancy), graph B is the ROC curve for the validation data (health checkup data for 847 pregnant women throughout their entire pregnancy), and graph C is the ROC curve for the validation data (health checkup data for 847 pregnant women up to the 31st week of pregnancy). For each of graphs A, B, and C, the same discriminator was used to estimate (predict) the most likely transition from the health checkup data for each pregnant woman, and the highest-risk state (the state with the highest incidence of HDP) was used as the score.

[0048] The figure shows the AUC (Area Under the Curve) for each of graphs A, B, and C as follows: Graph A: AUC=0.842 Graph B: AUC=0.841 Graph C: AUC=0.725

[0049] The figure shows that the AUC for each of graphs A and B is greater than 0.84, indicating good accuracy, and that both have similar levels of accuracy. On the other hand, the figure shows that although the AUC of graph C is inferior to the AUC of graphs A and B, it is possible to predict the onset of HDP during the remainder of the pregnancy from health checkup data prior to the 31st week of pregnancy.

[0050] ● Classifier prediction value FIG. 14 is a schematic diagram showing an example of an output variable predicted by a classifier. The figure shows blood pressure for each week of pregnancy predicted by a classifier based on the medical checkup data of pregnant women. In the figure, the solid line graph on the top of the page shows the median systolic blood pressure (average of predicted values), and the solid line graph on the bottom of the page shows the median diastolic blood pressure. In the figure, the line surrounding the graph of the median systolic blood pressure indicates the range of predicted values ​​with a 95% credible interval for systolic blood pressure for each gestational age. In the figure, the line surrounding the graph of the median diastolic blood pressure indicates the range of predicted values ​​with a 95% credible interval for diastolic blood pressure for each gestational age. In the figure, the circle displayed within the line surrounding the graph of the median systolic blood pressure indicates the actual measured value of systolic blood pressure included in the medical checkup data of pregnant women.In the figure, the triangle displayed within the line surrounding the graph of the median diastolic blood pressure indicates the actual measured value of diastolic blood pressure included in the medical checkup data of pregnant women.

[0051] FIG. 15 is a schematic diagram showing an example of the output of the prediction result of the present system 1. The figure shows a prediction result output screen on which predicted values ​​of medical checkup data for the remaining period of a pregnant woman's pregnancy are displayed. This screen is displayed on a display (not shown) of the system 1 by the output unit 6. The predicted values ​​of the medical checkup data displayed on this screen are calculated by the output variable calculation unit 5 based on the state of each pregnancy week determined by the state determination unit 4.

[0052] The figure shows predicted values ​​for the pregnant woman's condition and medical checkup data for each week of pregnancy from the 34th week onwards, based on the medical checkup data of the pregnant woman up to the 33rd week of pregnancy. Based on the predicted values ​​shown in the figure, the user (doctor) of this system 1 determines the possibility of the pregnant woman developing HDP and takes measures before the onset of the disease, if necessary.

[0053] Example of prediction by classifier 16 to 21 are schematic diagrams showing examples of the state of each pregnant woman predicted by the classifier. Figures 16-18 show the states predicted by the classifier based on the medical checkup data of pregnant women who developed HDP, while Figures 19-21 show the states predicted by the classifier based on the medical checkup data of pregnant women who did not develop HDP.

[0054] The figures show the state of each pregnancy week predicted by the classifier for each pregnant woman being predicted. That is, Figures 16 to 21 show prediction examples for each of six pregnant women. In the figures, the pregnancy weeks are displayed in 5-week increments for the sake of convenience, but in reality, the predictions are made for each week, for example, from week 1 to week 45.

[0055] In the figure, "●" indicates the number of weeks of pregnancy for which medical checkup data for the pregnant woman being predicted exists. That is, for example, for the pregnant woman shown in Figure 16, the classifier predicts the condition of the pregnant woman from the 21st week of pregnancy onwards based on the medical checkup data up to the 20th week of pregnancy.

[0056] In addition, when predicting the condition of a pregnant woman from the 21st week of pregnancy onwards based on medical checkup data up to the 20th week of pregnancy, the classifier also estimates the pregnant woman's past condition before the 20th week of pregnancy. That is, as described above, the classifier in the present invention is trained using a hidden Markov model. That is, as shown in FIG. 2 , for example, the condition of a pregnant woman at the 21st week of pregnancy is determined based on the condition of the pregnant woman at the 20th week of pregnancy and the state transition probability. Therefore, the classifier also estimates the condition of a pregnant woman up to the 20th week of pregnancy based on medical checkup data up to the 20th week of pregnancy.

[0057] Here, the prediction by the classifier is made based on the medical checkup data of the pregnant woman to be predicted. Therefore, the prediction period (weeks of pregnancy) of the classifier shown in the figures varies for each pregnant woman. That is, for example, for the pregnant woman shown in FIG. 16, the classifier predicts the condition from the 21st week of pregnancy onwards based on the medical checkup data up to the 20th week of pregnancy. Also, for example, for the pregnant woman shown in FIG. 17, the classifier predicts the condition from the 26th week of pregnancy onwards based on the medical checkup data up to the 25th week of pregnancy.

[0058] As described above, predictions by the classifier are made based on medical checkup data of the pregnant woman being predicted. For example, for the pregnant woman shown in FIG. 16, the figure shows that the state of the pregnant woman from the 21st week of pregnancy onward is predicted based on medical checkup data up to the 20th week of pregnancy. For the same pregnant woman, if medical checkup data up to the 25th week of pregnancy exists five weeks later, the classifier predicts the state of the pregnant woman from the 26th week of pregnancy onward based on the medical checkup data up to the 25th week of pregnancy. In this case, the prediction result for the 26th week of pregnancy onward based on the medical checkup data up to the 25th week of pregnancy may differ from the prediction result for the 26th week of pregnancy onward based on the medical checkup data up to the 20th week of pregnancy. In other words, the more medical checkup data a pregnant woman has, the higher the accuracy of the prediction, i.e., the accuracy of the prediction as to whether the pregnant woman will develop HDP.

[0059] In the figure, the blackened areas indicate the states predicted by the classifier. Figure 16 shows that, based on the medical checkup data of a pregnant woman up to the 20th week of pregnancy, the classifier estimated that the pregnant woman would be in state 8 up to the 10th week of pregnancy, state 11 from the 11th to the 20th week of pregnancy, and predicted that the pregnant woman would be in state 12 from the 21st to the 30th week of pregnancy and state 13 from the 31st to the 45th week of pregnancy. Figure 17 shows that, based on the medical checkup data of a pregnant woman up to the 25th week of pregnancy, the classifier estimated that the pregnant woman's condition up to the 15th week of pregnancy was 12, and that the condition from the 16th to 25th weeks was 8, and predicted that the condition from the 26th to 30th weeks was 13, and that the condition from the 31st to 45th weeks was 14. Figure 18 shows that, based on the medical checkup data of a pregnant woman up to the 15th week of pregnancy, the classifier estimated that the pregnant woman would be in state 13 up to the 15th week of pregnancy, and predicted that the pregnant woman would be in state 13 from the 16th week to the 45th week of pregnancy.

[0060] In Figures 16 to 18, the thick lines arranged vertically indicate the gestational age at which the pregnant woman developed HDP. For example, for the pregnant woman shown in Figure 16, the figure shows that she developed HDP at the 35th week of pregnancy. In other words, the figure shows a superimposed display of the estimated / predicted state of the pregnant woman at each gestational age based on her medical checkup data up to the 20th week of pregnancy, and the gestational age at which she actually developed HDP.

[0061] Figure 19 shows that pregnant women who were estimated or predicted to remain in state 2 throughout their pregnancies did not actually develop HDP. Figure 20 shows that pregnant women who were estimated or predicted to remain in condition 12 throughout their pregnancies did not actually develop HDP. FIG. 21 shows that pregnant women whose condition was estimated or predicted to gradually progress to a state with a high incidence of HDP as the pregnancy progressed did not actually develop HDP.

[0062] ●Summary● According to the embodiment described above, the present system 1 predicts the medical checkup data for the remaining period of a pregnant woman's pregnancy using a classifier trained using a hidden Markov model in which the state (internal state) of the pregnant woman is defined as a hidden state and the medical checkup data for each state is defined as an output variable of the state. A user (doctor) of the present system 1 determines the possibility of the pregnant woman developing HDP by referring to the predicted values ​​of the medical checkup data output by the present system 1, and takes measures before the onset of HDP, if necessary. In other words, the present system 1 uses the medical checkup data of the pregnant woman obtained during prenatal checkups to help predict the onset of HDP during the pregnant woman's pregnancy.

[0063] As explained above, the present system 1 predicts the future physical condition of a pregnant woman from the fluctuation patterns in the transition of her medical checkup data (blood pressure and blood data) over time during her pregnancy. The predictions made by the present system 1 are based on the medical checkup data accumulated at each prenatal checkup for each pregnant woman, and do not depend on the experience of the doctor examining the pregnant woman.

[0064] Here, the health checkup data used when training the classifier of this system 1 and when making predictions in this system 1 is obtained from prenatal checkups that all pregnant women in Japan undergo regularly during their pregnancy. In other words, the health checkup data required by this system 1 is not information unique to this system, but rather longitudinal information that has been accumulated in Japan over many years. In other words, the health checkup data used by this system 1 is dense and well-organized data obtained from prenatal checkups that all pregnant women undergo at the same gestational week. The classifier of this system 1, which trains using this dense and well-organized data, has high discrimination accuracy.

[0065] Furthermore, this system 1 uses information obtained from existing prenatal checkup tests, and does not require new test items for prenatal checkups or invasive tests.

[0066] Furthermore, this system 1 uses information obtained from the prenatal checkup tests that are uniformly implemented in Japan, and therefore can be used in medical institutions in Japan and other countries that perform tests similar to those used in prenatal checkups in Japan.

[0067] In this way, this system 1 can assist doctors by accurately predicting the onset of HDP based on medical checkup data obtained during prenatal checkups, which have previously only been used as sequential data for pregnancy management.

[0068] ●Features of this system, this program, and this method● The features of the system, program, and method described above are summarized below.

[0069] ●Features of this system This system is A support system for predicting whether a pregnant woman will develop pregnancy-induced hypertension by predicting a state that the pregnant woman may be in during pregnancy from among a plurality of states having different rates of development of pregnancy-induced hypertension (for example, the system 1 shown in FIG. 3 ), a storage unit (e.g., storage unit 2 in FIG. 3) that stores a classifier trained to predict transitions between the plurality of states (e.g., state 1 to state 14) based on the training pregnancy medical checkup data for each training pregnant woman and the presence or absence of the onset of pregnancy-induced hypertension syndrome; An acquisition unit (e.g., medical examination data acquisition unit 3 in FIG. 3) that acquires medical examination data (e.g., diastolic blood pressure, systolic blood pressure, urinary protein) that the pregnant woman who is the target of prediction has undergone during pregnancy; A discrimination unit (e.g., state discrimination unit 4 in FIG. 3) that discriminates the state that the pregnant woman may be in within the remaining period of pregnancy from among the plurality of states based on the medical examination data of the pregnant woman that has been taken and the discriminator; a calculation unit (for example, the output variable calculation unit 5 of FIG. 3 ) that calculates predicted medical examination data for the remaining period of the pregnant woman based on the state of the pregnant woman for the remaining period of the pregnant woman; and the classifier is trained using a hidden Markov model in which the states are hidden states and the medical examination data for each state is an output variable of the state; It is characterized by:

[0070] In this system, the classifier is selected from a plurality of classifier candidates trained using the medical examination data of the training pregnant woman, the number of states as hyperparameters of each of the plurality of candidates is different; It can also be something like this.

[0071] In this system, The storage unit state transition probabilities between a plurality of said states; Remember, The determination unit determining the current condition of the pregnant woman based on the medical examination data of the pregnant woman; determining the state that the pregnant woman may take within the remaining period based on the current state of the pregnant woman and the state transition probability; It can also be something like this.

[0072] In this system, the classifier is trained using the medical examination data for training obtained by having the training pregnant woman undergo medical examinations at each week of pregnancy, The determination unit determines the state of the pregnant woman for each pregnancy week within the remaining term. It can also be something like this.

[0073] In this system, The storage unit an output probability of the health checkup data for each of the plurality of states; Remember, the calculation unit calculates the predicted health checkup data based on the states that the pregnant woman may take within the remaining period and the output probability. It can also be something like this.

[0074] In this system, the calculation unit calculates the predicted medical checkup data for each pregnancy week within the remaining period of the pregnant woman; It can also be something like this.

[0075] In this system, The acquisition unit acquires the medical examination data of the pregnant woman for each pregnancy week. It can also be something like this.

[0076] In this system, The medical checkup data includes at least one of systolic blood pressure, diastolic blood pressure, and urinary protein. It can also be something like this.

[0077] ●Features of this program This program causes a computer to function as this system. It is characterized by:

[0078] ●Features of this method The method comprises: A method for supporting prediction of pregnancy-induced hypertension, which is executed by an onset prediction support system that supports prediction of whether a pregnant woman will develop pregnancy-induced hypertension by predicting a state that the pregnant woman may take during pregnancy from a plurality of states (e.g., state 1 to state 14) having different rates of onset of pregnancy-induced hypertension, The onset prediction support system (for example, the onset prediction support system 1 of FIG. 3) a storage unit (e.g., storage unit 2 in FIG. 3 ) that stores a classifier trained to predict transitions between the plurality of states based on the training pregnancy medical checkup data for each training pregnant woman and the presence or absence of the onset of pregnancy-induced hypertension syndrome; Equipped with The onset prediction support system comprises: An acquisition step of acquiring medical checkup data obtained from medical checkups that the pregnant woman who is the target of prediction has undergone during pregnancy; a determining step of determining the state that the pregnant woman may be in within the remaining period of pregnancy from among the plurality of states based on the medical examination data of the pregnant woman that has been taken and the classifier; a calculation step of calculating predicted medical examination data for the remaining period of the pregnant woman based on the condition of the pregnant woman for the remaining period of the pregnancy; and the classifier is trained using a hidden Markov model in which the states are hidden states and the medical examination data for each state is an output variable of the state; It is characterized by: [Explanation of symbols]

[0079] 1. Onset prediction support system 2 Storage section 3. Health checkup data acquisition section 4. Status detection section 5. Output variable calculation section 6 Output section

Claims

1. A support system for predicting whether a pregnant woman will develop pregnancy-induced hypertension by predicting a state that the pregnant woman may be in during pregnancy from among a plurality of states having different rates of development of pregnancy-induced hypertension, a storage unit that stores a classifier that has been trained to predict transitions between the plurality of states based on the training health checkup data during pregnancy for each training pregnant woman and the presence or absence of the onset of pregnancy-induced hypertension syndrome; an acquisition unit that acquires medical checkup data obtained from medical checkups that the pregnant woman who is the target of prediction has undergone during pregnancy; a discrimination unit that discriminates the state that the pregnant woman may be in within the remaining period of pregnancy from among the plurality of states based on the medical examination data of the pregnant woman that has been taken and the discriminator; a calculation unit that calculates predicted medical examination data for the remaining period of the pregnant woman based on the condition of the pregnant woman for the remaining period of the pregnancy; and the classifier is trained using a hidden Markov model in which the states are hidden states and the medical examination data for each state is an output variable of the state; The storage unit an output probability of the health checkup data for each of the plurality of states; Remember, the calculation unit calculates the predicted health checkup data based on the states that the pregnant woman may take within the remaining period and the output probability. A support system for predicting the onset of pregnancy-induced hypertension.

2. the classifier is selected from a plurality of classifier candidates trained using the medical examination data of the training pregnant woman, the number of states as hyperparameters of each of the plurality of candidates is different; The support system for predicting the onset of pregnancy-induced hypertension according to claim 1.

3. The storage unit state transition probabilities between a plurality of said states; Remember, The determination unit determining the current condition of the pregnant woman based on the medical examination data of the pregnant woman; determining the state that the pregnant woman may take within the remaining period based on the current state of the pregnant woman and the state transition probability; 3. The support system for predicting the onset of pregnancy-induced hypertension according to claim 1 or 2.

4. the classifier is trained using the medical examination data for training obtained from medical examinations taken by the training pregnant woman at each week of pregnancy, The determination unit determines the state of the pregnant woman for each pregnancy week within the remaining term. The support system for predicting the onset of pregnancy-induced hypertension according to any one of claims 1 to 3.

5. the calculation unit calculates the predicted medical checkup data for each pregnancy week within the remaining period of the pregnant woman; The support system for predicting the onset of pregnancy-induced hypertension according to claim 1.

6. The acquisition unit acquires the medical checkup data obtained at a medical checkup taken by the pregnant woman for each week of pregnancy. The support system for predicting the onset of pregnancy-induced hypertension according to claim 5.

7. The medical checkup data includes at least one of systolic blood pressure, diastolic blood pressure, and urinary protein. The support system for predicting the onset of pregnancy-induced hypertension according to claim 1.

8. A computer is caused to function as the pregnancy-induced hypertension onset prediction support system according to claim 1. A support program for predicting the onset of pregnancy-induced hypertension.

9. A method for supporting prediction of pregnancy-induced hypertension, which is executed by an onset prediction support system that supports prediction of whether a pregnant woman will develop pregnancy-induced hypertension by predicting a state that the pregnant woman may take during pregnancy from a plurality of states having different rates of onset of pregnancy-induced hypertension, The onset prediction support system comprises: a storage unit that stores a classifier that has been trained to predict transitions between the plurality of states based on the training health checkup data during pregnancy for each training pregnant woman and the presence or absence of the onset of pregnancy-induced hypertension syndrome; Equipped with the classifier is trained using a hidden Markov model in which the states are hidden states and the medical examination data for each state is an output variable of the state; The storage unit an output probability of the health checkup data for each of the plurality of states; Remember, The onset prediction support system comprises: An acquisition step of acquiring medical checkup data obtained from medical checkups that the pregnant woman who is the target of prediction has undergone during pregnancy; a determining step of determining the state that the pregnant woman may be in within the remaining period of pregnancy from among the plurality of states based on the medical examination data of the pregnant woman that has been taken and the classifier; a calculation step of calculating predicted medical checkup data for the remaining period of the pregnant woman based on the states that the pregnant woman may take within the remaining period of the pregnant woman and the output probability; consisting of A method for supporting prediction of the onset of pregnancy-induced hypertension.

Images