Pregnancy state estimation device and pregnancy state estimation method

Abstract

The present invention is, even outside of a hospital, capable of inexpensively and conveniently identifying the state of a mother or a fetus. A pregnancy state estimation device according to the present invention comprises: an information acquisition unit that acquires an electrical activity record which has been derived from a mother and percutaneously measured; an inference unit that infers the pregnancy state of the mother by inputting the electrical activity record into a trained model which sets, as an input, the electrical activity record and sets, as an output, a pregnancy state including information related to at least one of the uterine contractions of the mother and the fetal heart rate in the mother; and an output unit that outputs the pregnancy state generated by the inference unit.

Description

Pregnancy status estimation device and pregnancy status estimation method

[0001] The present invention relates to a pregnancy status estimation device and a pregnancy status estimation method.

[0002] There are known techniques for measuring the mother's body during pregnancy and determining information about the mother and fetus based on the obtained data. For example, Patent Document 1 discloses a labor monitoring device that calculates the fetal heart rate based on signals from electrocardiogram electrodes attached to the fetal head after membrane rupture and an ultrasound transducer attached to the surface of the mother's abdomen, and calculates the mother's intrauterine pressure and the intensity of labor pains based on signals from an internal pressure transducer and a labor pain transducer inserted into the uterus after membrane rupture.

[0003] Japanese Patent Application Laid-Open No. 2006-223335

[0004] In addition to the labor monitoring device described in Patent Document 1, there are various other devices capable of obtaining information about the mother and fetus, such as fetal cardiotocography monitors and Doppler heart rate monitors. However, because these devices require specialized knowledge and skills to use and are expensive, they are intended to be installed in medical facilities such as hospitals and used by medical professionals such as doctors.

[0005] During pregnancy, there is a risk of premature birth and intrauterine fetal death, so attention must be paid to changes in the condition of the mother and fetus. However, when pregnant women are at home outside of a hospital, the only clues available are the pregnant woman's own subjective symptoms and fetal movements, and there are limited means of understanding the condition of the mother and fetus.

[0006] In order to detect early signs of premature birth or intrauterine fetal death and enable early therapeutic intervention, it is desirable to be able to monitor the condition of the mother or fetus inexpensively and easily outside of a hospital.

[0007] The present disclosure has been made in consideration of the above, and aims to provide a pregnancy status estimation device and a pregnancy status estimation method that enable the status of the mother or fetus to be grasped inexpensively and easily even outside a hospital.

[0008] The pregnancy status estimation device of the present invention comprises an information acquisition unit that acquires an electrical activity record measured transcutaneously from the mother, an inference unit that inputs the electrical activity record into a trained model that outputs a pregnancy status including information regarding at least one of the mother's uterine contractions and the fetal heart rate in the mother, and infers the pregnancy status of the mother by inputting the electrical activity record, and an output unit that outputs the pregnancy status generated by the inference unit.

[0009] The method for estimating pregnancy status according to the present invention comprises the steps of: acquiring a maternal electrical activity record measured transcutaneously; inputting the electrical activity record into a trained model that takes the electrical activity record as input and outputs a pregnancy status including information regarding at least one of the maternal uterine contractions and the fetal heart rate in the mother, thereby inferring the pregnancy status of the mother; and outputting the pregnancy status.

[0010] The pregnancy status estimation device of the present invention comprises an information acquisition unit that acquires electrical activity records measured transcutaneously from the mother, an inference unit that inputs the electrical activity records into a trained model that inputs the electrical activity records and outputs pregnancy status including information about intrauterine fetal movements within the mother, and infers the pregnancy status of the mother by inputting the electrical activity records, and an output unit that outputs the pregnancy status generated by the inference unit.

[0011] The method for estimating pregnancy status according to the present invention comprises the steps of acquiring a transcutaneously measured electrical activity record from the mother, inputting the electrical activity record into a trained model that takes the electrical activity record as input and outputs the pregnancy status including information on intrauterine fetal movements within the mother, thereby inferring the pregnancy status of the mother, and outputting the pregnancy status.

[0012] According to the present disclosure, it is possible to inexpensively and easily grasp the condition of the mother or fetus even outside a hospital.

[0013] FIG. 1 is a schematic block diagram showing a pregnancy status estimation device according to a first embodiment. FIG. 2 is an explanatory diagram showing a first trained model. FIG. 3 is an explanatory diagram showing a second trained model. FIG. 4 is an explanatory diagram explaining output of pregnancy status. FIG. 5 is an explanatory diagram explaining generation of a first trained model (machine learning). FIG. 6 is an explanatory diagram explaining generation of a second trained model (machine learning). FIG. 7 is a flowchart for explaining a pregnancy status estimation method according to an embodiment. FIG. 8 is an explanatory diagram showing a first configuration example of a pregnancy status estimation device according to an embodiment. FIG. 9 is an explanatory diagram showing a second configuration example of a pregnancy status estimation device according to an embodiment. FIG. 10 is a schematic block diagram showing a pregnancy status estimation device according to a second embodiment. FIG. 11 is an explanatory diagram showing a trained model according to the second embodiment. FIG. 12 is an explanatory diagram explaining evaluation information according to the second embodiment. FIG. 13 is an explanatory diagram explaining generation of a trained model (machine learning). FIG. 14 is a schematic diagram explaining generation of an individual trained model by fine tuning according to a modified example. FIG. 15 is a schematic diagram explaining inference processing using an individual trained model according to a modified example.

[0014] [First Embodiment] Hereinafter, embodiments of a pregnancy status estimation device and a pregnancy status estimation method according to the present invention will be described with reference to the drawings. Note that the present invention is not limited to these embodiments. Furthermore, the components in the following embodiments include those that are easily replaceable by those skilled in the art, or those that are substantially identical. The present invention is not limited to these embodiments, and when there are multiple embodiments, the present invention also includes configurations that combine the respective embodiments.

[0015] (Pregnancy Status Estimation Device) Fig. 1 is a schematic block diagram showing a pregnancy status estimation device 1 according to a first embodiment. As shown in Fig. 1, the pregnancy status estimation device 1 is a device that infers a pregnancy status 3 of a mother MO based on an electrical activity record 2 derived from the mother MO that is non-invasively and transcutaneously acquired from the mother MO of a pregnant woman, and outputs the obtained pregnancy status 3. The pregnancy status estimation device 1 is a device that can be used outside a medical facility such as a hospital, for example, in the home of a pregnant woman.

[0016] Pregnancy status 3 is information indicating the status of the maternal MO and / or the fetal FE in the maternal MO. In this specification, "pregnancy status" is a broad concept that encompasses the status of the maternal MO and the status of the fetal FE in the maternal MO. In the first embodiment, the maternal MO status included in pregnancy status 3 includes information about uterine contractions of the maternal MO. In the first embodiment, the fetal FE status included in pregnancy status 3 includes information about the fetal heart rate in the maternal MO. Pregnancy status 3 may include information related to signs of perinatal complications such as preterm birth, intrauterine fetal death, and placental abruption. Information about uterine contractions may include signs of preterm birth or placental abruption. Information about the fetal heart rate may include signs of intrauterine fetal death. By detecting signs of perinatal complications from pregnancy status 3, improvements in early medical intervention for pregnant women can be expected.

[0017] The electrical activity of the mother MO is acquired by a sensor 30 attached to the mother MO. The sensor 30 non-invasively measures the electrical activity of the mother MO. The sensor 30 is, for example, a sensor that detects electrical signals used in an electrocardiogram or an electromyogram. Such a sensor 30 transcutaneously measures electrical signals (changes in the electric field) generated in association with the biological activity of the mother MO using electrodes in contact with the mother MO. The pregnancy status estimation device 1 has an interface that communicates with the sensor 30 via a wired or wireless connection and receives measurement data from the sensor 30.

[0018] The pregnancy status estimation device 1 is configured as a computer including a processor such as a CPU (Central Processing Unit) and memories such as ROM (Read Only Memory) and RAM (Random Access Memory). The pregnancy status estimation device 1 includes an information acquisition unit 10, an inference unit 11, an evaluation unit 12, an output unit 13, and a storage unit 14. The information acquisition unit 10, the inference unit 11, the evaluation unit 12, and the output unit 13 are configured as functional blocks realized on a program 20 stored in the storage unit 14 by the processor executing the program 20. The program 20 is a program for causing a computer to function as the information acquisition unit 10, the inference unit 11, the evaluation unit 12, and the output unit 13. The information acquisition unit 10, the inference unit 11, the evaluation unit 12, and the output unit 13 may each be configured as dedicated hardware.

[0019] The information acquisition unit 10 acquires a transcutaneously measured electrical activity record 2 from the maternal MO. The information acquisition unit 10 acquires the electrical activity record 2 from a sensor 30 attached to the maternal MO. The electrical activity record 2 is time-series data of the electrical activity from the maternal MO measured by the sensor 30. The information acquisition unit 10 acquires the electrical activity record 2 from the sensor 30 at a predetermined sampling period and stores it in the memory unit 14.

[0020] The inference unit 11 generates a pregnancy status 3 of the maternal MO based on the electrical activity record 2 derived from the maternal MO acquired by the information acquisition unit 10. Specifically, the inference unit 11 inputs the electrical activity record 2 into a trained model 21 that outputs a pregnancy status including information on at least one of the maternal MO's uterine contractions and the fetal heart rate, thereby inferring the maternal MO's pregnancy status 3. The trained model 21 is generated by machine learning using training data including the maternal electrical activity record 2 as input and information on the uterine contractions and / or the fetal heart rate as output, and is pre-stored in the storage unit 14. The output pregnancy status 3 is numerical information related to at least one of the uterine contractions and the fetal heart rate. The inference unit 11 stores the pregnancy status 3 of the maternal MO, which is an inference result using the trained model 21, in the storage unit 14.

[0021] The evaluation unit 12 generates evaluation information 4 according to the content of the pregnancy status 3 generated by the inference unit 11. The evaluation unit 12 receives the pregnancy status 3 generated by the inference unit 11 as input, and outputs evaluation information 4 corresponding to the input pregnancy status 3. The evaluation information 4 is information indicating an evaluation of the status of the pregnant woman (mother MO) and fetus FE ascertained from the pregnancy status 3.

[0022] The evaluation unit 12 may generate the evaluation information 4 using a trained model for evaluation generated by machine learning, or may generate the evaluation information 4 by rule-based information processing. As an example of rule-based information processing, the evaluation unit 12 generates the evaluation information 4 according to the input pregnancy status 3 using evaluation data 22 that associates the content (numerical values) of the pregnancy status 3 with the evaluation information 4. The evaluation data 22 includes, for example, a table that associates a plurality of numerical ranges (thresholds) for classifying the pregnancy status 3 with the evaluation information 4 associated with each numerical range. The evaluation data 22 includes, for example, variables for inputting the values ​​of the pregnancy status 3 and a mathematical formula for determining the evaluation category of the pregnancy status 3. The evaluation data 22 is created in advance based on information obtained from the pregnancy status 3 and the knowledge of specialists in the field.

[0023] The output unit 13 outputs the pregnancy status 3 generated by the inference unit 11. The output unit 13 also outputs evaluation information 4 corresponding to the pregnancy status 3 in addition to the pregnancy status 3. The pregnancy status estimation device 1 can be connected to the display unit 31 via a wired or wireless connection via an interface (not shown). The display unit 31 may be a display device or a display device provided on an information processing terminal such as a smartphone. The output unit 13 presents this information to a pregnant woman or the like by outputting the pregnancy status 3 and the evaluation information 4 to the display unit 31. Users such as pregnant women and their families can understand the status of maternal MO and fetal FE from the pregnancy status 3 while at home or elsewhere outside a hospital. The evaluation information 4 allows users to understand the status of maternal MO and fetal FE without requiring medical expertise and to take appropriate action, such as visiting a medical institution, if necessary. This allows pregnant women to take action, such as seeking early medical care, based on objective information, even if they do not have any subjective symptoms (subjective information).

[0024] (Trained Model) In the first embodiment, the trained model 21 includes a first trained model 21A that outputs information 3A (see FIG. 2) related to uterine contractions of the mother MO, and a second trained model 21B that outputs information 3B (see FIG. 3) related to the fetal heart rate in the mother MO. The inference unit 11 infers a pregnancy status 3 that includes the information 3A related to uterine contractions of the mother MO and the information 3B related to the fetal heart rate in the mother MO.

[0025] FIG. 2 is an explanatory diagram showing a first trained model 21A. The first trained model 21A is generated in advance by machine learning so as to input an electrical activity record 2 derived from the mother MO and output information 3A related to the uterine contractions of the mother MO. As shown in FIG. 2, the inference unit 11 infers information 3A related to the uterine contractions of the mother MO by inputting the electrical activity record 2 derived from the mother MO acquired by the information acquisition unit 10 into the first trained model 21A. The electrical activity record 2 is time-series data of electrical signals measured by the sensor 30. In FIG. 2, the electrical activity record 2 is shown as a waveform graph in which the vertical axis indicates signal intensity and the horizontal axis indicates time (time u). The first trained model 21A outputs information 3A related to the uterine contractions of the mother MO as time-series data of the intensity of uterine contractions (electrical signals derived from uterine contractions). In FIG. 2, information 3A relating to uterine contractions of the mother MO is shown by a waveform graph in which the vertical axis indicates the strength of uterine contractions and the horizontal axis indicates time (time u).

[0026] The information 3A about uterine contractions includes at least one of information on the strength, duration, interval, period, frequency, and intensity waveform of uterine contractions. In the example shown in FIG. 2 , the information 3A about uterine contractions includes information on the strength H, duration L, and interval D of uterine contractions. The strength H of uterine contractions indicates the strength of a single uterine contraction. The duration L of a uterine contraction is the length of time from the start to the end of a single uterine contraction. The interval D of a uterine contraction is the time interval from one uterine contraction to the next. While FIG. 2 shows the strength waveform (time series data) of uterine contractions, the information 3A about uterine contractions may include only data on the strength H, duration L, and interval D of uterine contractions. In the first embodiment, the first trained model 21A outputs the strength waveform of uterine contractions including the strength H, duration L, and interval D of uterine contractions as the information 3A about uterine contractions. The information 3A about uterine contractions may include either the period or frequency of uterine contractions, or both. The period of a uterine contraction is the sum of the duration L and the interval D. Contraction frequency is the number of times a contraction occurs per period of time.

[0027] FIG. 3 is an explanatory diagram showing the second trained model 21B. The second trained model 21B is generated in advance by machine learning so as to input an electrical activity record 2 derived from the mother MO and output information 3B related to the fetal heartbeat in the mother MO. As shown in FIG. 3, the inference unit 11 infers information 3B related to the fetal heartbeat in the mother MO by inputting the electrical activity record 2 derived from the mother MO acquired by the information acquisition unit 10 into the second trained model 21B. The electrical activity record 2 is the same information as that input to the first trained model 21A. The second trained model 21B outputs the information 3B related to the fetal heartbeat as time series data of the intensity of the fetal heartbeat (electrical signal derived from the fetal heartbeat). In FIG. 3, the information 3B related to the fetal heartbeat is shown by a waveform graph in which the vertical axis indicates the intensity of the fetal heartbeat and the horizontal axis indicates time (time u).

[0028] The information 3B about the fetal heartbeat includes at least one of the following: fetal heartbeat, fetal heart rate, fetal heart sound, heartbeat interval, heart sound waveform, and fetal heartbeat findings. In the example shown in FIG. 3 , the information 3B about the fetal heartbeat in the mother MO includes at least one of the following: fetal heartbeat CB, fetal heart rate B, and fetal heart sound. The fetal heartbeat CB is time-series data of the fetal heartbeat intensity (data of the intensity change waveform). The fetal heart rate B is information about the number of beats per unit time, and is expressed, for example, in beats per minute (bpm). In FIG. 3 , "130," "132," "125," "127," and "132" indicate example values ​​of the fetal heart rate B calculated at each time point. The fetal heart sound is data that expresses the intensity change of the fetal heartbeat CB as sound. In the first embodiment, the second trained model 21B outputs a fetal heartbeat intensity waveform including a fetal heartbeat CB and a fetal heart rate B as information 3B related to the fetal heartbeat. The information 3B related to the fetal heartbeat may include one or more of a heartbeat interval, a phonocardiogram waveform, and fetal heartbeat findings. The heartbeat interval is information indicating whether the rhythm of the fetal heartbeat is irregular (arrhythmic) or irregular (arrhythmic). The phonocardiogram waveform is an intensity waveform of the fetal heartbeat. The fetal heartbeat findings are features that can be read from the heartbeat waveform (electrocardiogram waveform) or the phonocardiogram waveform (ultrasound waveform), such as blood flow velocity, blood flow velocity waveform, the presence or absence of tachycardia or bradycardia, and the degree of fluctuation in the heart rate (heartbeat interval).

[0029] In the first embodiment, the inference unit 11 may preprocess the electrical activity record 2 to be input to the trained model 21 (first trained model 21A, second trained model 21B). In this case, the trained model 21 is trained using training data of the electrical activity record 2 that has been preprocessed in the same way as the input data.

[0030] For example, the inference unit 11 may perform preprocessing on the electrical activity record 2, including filtering to remove predetermined signal components corresponding to the maternal MO's respiratory fluctuations. This allows fluctuations caused by the maternal MO's respiration to be removed from the electrical activity record 2, thereby improving the accuracy of estimating the pregnancy status 3. The filtering to remove predetermined signal components corresponding to the maternal MO's respiratory fluctuations is a frequency component removal process using a frequency-selective filter (e.g., a band-rejection filter). Respiratory fluctuations contained in the electrical activity record 2 generally fall within a frequency band of 0.08 Hz to 1 Hz. The frequency band to be removed is, for example, a range of 0.1 Hz to 0.5 Hz. By setting the frequency band to be removed within this range, the central signal component of the respiratory fluctuations can be effectively removed. When using a band-rejection filter, it is preferable to set the filter order to 2 or more, and it is particularly preferable to set the filter order to a predetermined value in the range of 2 to 4. This allows the signal components corresponding to the maternal MO's respiratory fluctuations to be effectively removed without excessively increasing the computational load.

[0031] Furthermore, for example, the inference unit 11 may perform preprocessing to extract information identifying at least one of a baseline, a P point, a Q point, an R point, an S point, and a T point from the electrical activity record 2, which is electrocardiogram data. Identifying the baseline means specifying the signal intensity indicating the baseline. Identifying the P point, the Q point, the R point, the S point, and the T point means specifying pairs of time information and signal intensity indicating the points. The inference unit 11 inputs the information identified from the electrocardiogram data to the trained model 21 as the preprocessed electrical activity record 2. For example, when all of the baseline, the P point, the Q point, the R point, the S point, and the T point are identified, the electrical activity record 2 becomes time-series data including the signal intensity of the baseline and pairs of time information and signal intensity indicating each of the P point, the Q point, the R point, the S point, and the T point. As a result, fluctuations between the P, Q, R, S, and T points in the electrocardiogram data are removed from the electrical activity record 2, and only components that characterize the electrocardiogram waveform can be extracted. In other words, the electrical activity record 2 after this preprocessing becomes feature-extracted data like a line graph in which the electrocardiogram waveform is constructed from a baseline and each point. Therefore, by removing noise components from the electrical activity record 2 and reducing the amount of data, the computational load in learning and inference can be effectively reduced. The inference unit 11 does not need to identify all feature points (P, Q, R, S, and T points), and may identify only some of the feature points.

[0032] 4 is an explanatory diagram illustrating the output of pregnancy status 3. As shown in FIG. 4, the data on pregnancy status 3 generated by the inference unit 11 is supplied to the evaluation unit 12 and the output unit 13.

[0033] The evaluation unit 12 generates evaluation information 4 based on the data on the pregnancy status 3 and evaluation data 22 pre-stored in the memory unit 14. The evaluation unit 12 generates evaluation information 4A based on uterine contractions based on information 3A related to uterine contractions of the mother MO. The evaluation unit 12 generates evaluation information 4B based on fetal heart rate based on information 3B related to the fetal heart rate in the mother MO. Furthermore, the evaluation unit 12 generates comprehensive evaluation information 4C based on both the information 3A related to uterine contractions and the information 3B related to the fetal heart rate. The generated evaluation information 4 includes the evaluation information 4A, evaluation information 4B, and comprehensive evaluation information 4C. The evaluation information 4 generated by the evaluation unit 12 is supplied to the output unit 13.

[0034] The output unit 13 outputs output data including the pregnancy status 3 generated by the inference unit 11 and the evaluation information 4 generated by the evaluation unit 12 to the display unit 31. The output unit 13 displays the pregnancy status 3 and the evaluation information 4 on the display unit 31, thereby presenting this information to the user.

[0035] 4, the output pregnancy status 3 includes information 3A related to uterine contractions of the mother MO and information 3B related to the fetal heart rate during the mother MO. The output evaluation information 4 is output, for example, in the form of a sentence (message) explaining the situation and advice grasped from the pregnancy status 3. In addition to the sentence, the evaluation information 4 may also include an indication of the degree of success of the pregnancy status 3, for example, in the form of a score (such as a score from 0 to 100) or a rank (such as A rank, B rank, C rank, etc., from most successful to least successful).

[0036] The evaluation information 4A based on uterine contractions includes, for example, an evaluation of the signs or possibility of preterm labor or placental abruption. The possibility of preterm labor can be evaluated based on the presence or absence of uterine contractions (true labor) that promote cervical ripening. True labor is characterized by strength H, duration L, and interval D. In true labor, the strength H and duration L of uterine contractions increase over time. In uterine contractions that do not promote cervical ripening (false labor), uterine contractions weaken or cease over time. Distinguishing between true labor and false labor requires continuous data acquisition over a certain period of time, but currently, no device capable of this determination exists outside of a hospital. Even within a hospital, time constraints make it difficult to obtain sufficient data. In the first embodiment, information 3A related to uterine contractions can be acquired and accumulated outside of a hospital, allowing evaluation indicators of signs of preterm labor to be obtained from the trends in the strength H and duration L of uterine contractions over time that characterize true labor. Placental abruption is a highly urgent condition, and early intervention improves the perinatal prognosis for both mother and child. A characteristic finding of placental abruption is known to be board-like hardness (a state in which the uterus is continuously hardened). Therefore, for example, an evaluation index for the possibility of placental abruption can be obtained using the strength H and duration L of uterine contractions, or a function that includes these as elements.

[0037] The evaluation information 4B based on the fetal heart rate includes, for example, an evaluation of the health status of the fetus FE and signs or possibility of intrauterine fetal death. As an example, the health status of the fetus FE can be evaluated based on the presence or frequency of transient tachycardia (a state in which the fetal heart rate B temporarily increases) within a certain period of time. If a predetermined number of transient tachycardia events are observed within the evaluation period based on the inferred pregnancy status 3, the state of the fetus FE is evaluated as good (reactive). If no transient tachycardia is observed, the state of the fetus FE is evaluated as abnormal (non-reactive). Abnormal values ​​can occur in a variety of cases, including cases in which there is no particular problem, and cases in which they are a sign of intrauterine fetal death. The comprehensive evaluation information 4C can include information such as an evaluation of the need to visit a medical institution based on the information on uterine contractions 3A and the information on the fetal heart rate 3B. This is expected to contribute to early medical intervention for important signs and improved treatment policies for perinatal complications, and will also reduce the burden on pregnant women and medical professionals by reducing unnecessary visits to the doctor when pregnancy status 3 is good.

[0038] (Generation of Trained Models) Fig. 5 is an explanatory diagram illustrating the generation (machine learning) of the first trained model 21A. Fig. 6 is an explanatory diagram illustrating the generation (machine learning) of the second trained model 21B. The learning algorithm of the trained models 21 (first trained model 21A, second trained model 21B) is not particularly limited, and may include, for example, one or a combination of a convolutional neural network (CNN), a recurrent neural network (RNN), a long short-term memory (LSTM), and an attention mechanism (including a transformer equipped with an attention mechanism).

[0039] 5 and 6 , the trained model 21 is generated by a learning unit ML, which is a computer that performs machine learning, performing machine learning of the trained model 40 using training data. The training data includes data actually measured from pregnant women who cooperated in creating the training data. The training data can be created from measurement data obtained using a dedicated device such as a fetal monitoring device in a hospital, for example.

[0040] As shown in FIG. 5 , the training data 41A used to generate the first trained model 21A includes a maternal electrical activity record 2 (time-series data) measured transcutaneously and a uterine contraction record 42 measured from the same mother. The uterine contraction record 42 is time-series data of the intensity of uterine contractions. The electrical activity record 2 and the uterine contraction record 42 included in the training data 41A are a pair of data measured at the same time t. The uterine contraction record 42 is information corresponding to the information 3A about uterine contractions generated by the inference unit 11. The learning unit ML uses the training data 41A to cause the training model 40 to acquire features for converting the maternal electrical activity record 2 measured transcutaneously into an output of the uterine contraction record 42 using deep learning. The features are weights or parameters for each node in each layer (output layer, input layer, hidden layer) constituting the training model 40. The training unit ML determines the features of the training model 40 by learning using multiple training data 41A. The learning model 40 whose features have been determined by machine learning is the first learned model 21A.

[0041] The training data 41A may use data from a specific period (number of weeks) during pregnancy, or data from the start of pregnancy for all weeks. By using uterine contraction data obtained from pregnant women for all weeks of pregnancy, including threatened miscarriage (from the start of pregnancy to less than 22 weeks), threatened preterm labor (from 22 weeks to less than 37 weeks), and placental abruption (22 weeks or later), as training data 41A, the output information 3A on maternal MO uterine contractions can be expected to contribute in various ways, such as early intervention from threatened miscarriage to threatened preterm labor, uterine rigidity in placental abruption, and prediction of the onset of labor. Using data from a specific period (number of weeks) as training data 41A enables machine learning specialized for a specific purpose, such as detecting signs of threatened preterm labor.

[0042] Furthermore, the teacher data 41A may be measurement data (so-called raw data) obtained using a dedicated device such as a fetal monitoring device, or may be data that has been annotated from the measurement data. Annotation is a process of adding information to the data (measurement data) used for learning. An example of annotation is identifying the time information for the start and end of uterine contractions in the measurement data. This allows for efficient learning of states in which effective uterine contractions are occurring and states in which effective uterine contractions are not occurring.

[0043] As shown in FIG. 6 , the teacher data 41B used to generate the second trained model 21B includes a maternal electrical activity record 2 (time-series data) measured transcutaneously and a fetal heart rate record 43 measured from the same mother. The fetal heart rate record 43 is time-series data of the fetal heart rate intensity. The teacher data 41B includes the maternal electrical activity record 2 and the fetal heart rate record 43, which are a pair of data measured at the same time t. The fetal heart rate record 43 is information corresponding to the fetal heart rate information 3B generated by the inference unit 11. The learning unit ML uses the teacher data 41B to cause the learning model 40 to acquire, using deep learning, feature values ​​for converting the maternal electrical activity record 2 measured transcutaneously into an output of the fetal heart rate record 43 when the maternal electrical activity record 2 is input. The learning unit ML determines the feature values ​​of the learning model 40 by learning using multiple pieces of teacher data 41B. The learning model 40, whose feature values ​​have been determined by machine learning, is the second trained model 21B.

[0044] The training data 41B may be data from a specific period (number of weeks), or data from all periods from the minimum number of weeks at which the fetal heartbeat can be detected by Doppler / echo. The minimum number of weeks at which the fetal heartbeat can be detected is usually about 5 weeks. By learning using data from the minimum number of weeks at which the heartbeat can be detected as the training data 41B, a second trained model 21B that can respond to changes in the heartbeat component due to the fetal growth process can be obtained.

[0045] As described above, when the inference unit 11 preprocesses the electrical activity records 2 to be input to the trained model 21, the training data used for training is also preprocessed in the same way. Therefore, the electrical activity records 2 used as training data may be subjected to a filter process to remove a predetermined signal component corresponding to the respiratory fluctuation of the maternal MO. Furthermore, the electrical activity records 2 used as training data may be electrocardiogram data that has been preprocessed to extract information identifying at least one of the baseline, P point, Q point, R point, S point, and T point.

[0046] Here, we will explain how to infer uterine contractions of the maternal MO and fetal heart rate from the electrical activity record 2 derived from the maternal MO measured transcutaneously. Human biological activities, such as muscle activity and brain activity, are caused by electrical stimulation. Therefore, the electrical signal (electrical activity record 2) measured transcutaneously from the maternal MO contains various signal components associated with biological activities, including signal components derived from uterine contractions and signal components derived from the fetal heart rate. The learning process shown in Figures 5 and 6 learns how to separate and extract signal components derived from uterine contractions and fetal heart rate components from the time-series data of the electrical activity record 2, which contains such various signal components. Note that the signal components derived from uterine contractions include not only electrical signals derived from uterine smooth muscle, but also electrical signals derived from other biological activities associated with uterine contractions (e.g., electrical signals generated from muscles other than uterine muscle due to maternal stiffness, electrocardiogram axis deviation due to changes in the diaphragm associated with uterine contractions, and changes in heart rate due to pain stimuli associated with uterine contractions). As an example of a technology for detecting a subject's disease or the like using machine learning from electrical signals acquired transcutaneously from the subject, the following non-patent document discloses a technology for detecting patients with arrhythmia using deep learning from the electrocardiogram signals of patients in sinus rhythm: Zachi I Attia, Peter A Noseworthy, Francisco Lopez-Jimenez, Samuel J Asirvatham, Abhishek J Deshmukh, Bernard J Gersh, et al., "An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction," Lancet. 2019 (published online Aug 1), Vol. 394, Issu. 10201, pp. 861-867, [online] URL: <https: / / doi.org / 10.1016 / S0140-6736(19)31721-0>

[0047] The memory unit 14 of the pregnancy status estimation device 1 shown in Figure 1 pre-stores the trained models 21 (first trained model 21A, second trained model 21B) generated in this manner.

[0048] (Pregnancy status estimation method) Next, a pregnancy status estimation method according to the first embodiment will be described. Fig. 7 is a flowchart for explaining the pregnancy status estimation method according to the first embodiment. The pregnancy status estimation method is also a method for operating the pregnancy status estimation device 1.

[0049] The pregnancy status estimation method includes a step S10 of acquiring an electrical activity record 2 measured transcutaneously from the mother MO, a step S20 of inputting the electrical activity record 2 into a trained model 21 that receives the electrical activity record 2 as an input and outputs a pregnancy status 3 including information on at least one of uterine contractions of the mother MO and a fetal heart rate in the mother MO, and inferring the pregnancy status 3 of the mother MO, and a step S40 of outputting the pregnancy status 3. The pregnancy status estimation method according to the first embodiment further includes a step S30 of generating evaluation information 4 according to the content of the pregnancy status 3 generated by the inference unit 11.

[0050] In step S10, the information acquisition unit 10 acquires the electrical activity record 2 measured transcutaneously from the maternal MO. The information acquisition unit 10 acquires the electrical activity record 2 from the sensor 30 attached to the maternal MO. After the electrical activity record 2 is measured by the sensor 30 attached to the maternal MO, the information acquisition unit 10 may acquire the electrical activity record 2 by reading data from a storage medium (e.g., a flash memory, an external server, etc.) that stores the electrical activity record 2.

[0051] In step S20, the inference unit 11 infers the pregnancy status 3 of the mother MO by inputting the electrical activity record 2 into the trained model 21. The inference unit 11 inputs the electrical activity record 2 acquired in step S10 into the first trained model 21A and the second trained model 21B and performs calculations based on the respective feature quantities. As a result, the inference unit 11 generates the pregnancy status 3 including information 3A regarding uterine contractions of the mother MO and information 3B regarding the fetal heart rate in the mother MO.

[0052] In step S30, the evaluation unit 12 generates evaluation information 4 based on the data of the pregnancy status 3 and the evaluation data 22. In the first embodiment, the evaluation unit 12 generates evaluation information 4A based on uterine contractions, evaluation information 4B based on fetal heart rate, and overall evaluation information 4C.

[0053] In step S40, the output unit 13 outputs output data including the pregnancy status 3 generated by the inference unit 11 and the evaluation information 4 generated by the evaluation unit 12 to the output destination device. In the first embodiment, the output destination device is the display unit 31. As a result, the pregnancy status 3 and the evaluation information 4 are displayed on the display unit 31.

[0054] (Configuration Example of Pregnancy Status Estimation Device) The pregnancy status estimation device 1 described above can be realized in various forms other than the example shown in FIG.

[0055] FIG. 8 is an explanatory diagram showing a first configuration example of the pregnancy status estimation device 1 according to the first embodiment. In the example of FIG. 8, the pregnancy status estimation device 1 is realized by a wearable device 100. The wearable device 100 is a small device that can be worn on a part of the body. The wearable device 100 has a band-like shape that can be worn on the wrist, ankle, upper arm, thigh, etc. FIG. 8 shows an example in which the wearable device 100 is a smart watch (a wristwatch-type information terminal worn on the wrist).

[0056] The wearable device 100 includes a control unit 102, a storage unit 14, a display unit 31, and a sensor 30. The control unit 102, the storage unit 14, the display unit 31, and the sensor 30 are incorporated into a single housing 101.

[0057] The control unit 102 is made up of a processor such as a CPU. The processor performs arithmetic processing in accordance with a program stored in the storage unit 14, thereby functioning as the control unit 102 that controls the wearable device 100. The storage unit 14 stores a program 20 for the pregnancy status estimation device 1. The information acquisition unit 10, inference unit 11, evaluation unit 12, and output unit 13 of the pregnancy status estimation device 1 are configured by the control unit 102 included in the wearable device 100 executing the program 20.

[0058] The storage unit 14 includes a non-volatile semiconductor memory and stores a program (operation system or firmware) for controlling the wearable device 100 and a program 20 (application program) for the pregnancy status estimation device 1. The storage unit 14 also stores the trained model 21 and evaluation data 22 shown in FIG. 1 .

[0059] The display unit 31 includes a self-luminous display device such as a liquid crystal display device or an organic electroluminescence (EL) display device. The display unit 31 displays various information of the wearable device 100 under the control of the control unit 102.

[0060] The sensor 30 transcutaneously measures electrical signals (changes in the electric field) generated in association with the wearer's biological activity using electrodes in contact with the wearer. The sensor 30 is, for example, an electrical heart rate sensor. In the example of FIG. 8 , the sensor 30 includes a first electrode provided on the inner peripheral surface of the housing 101 that contacts the wearer's wrist, and a second electrode provided on an operation unit 103 provided on the side of the housing 101. When the pregnant woman wears the wearable device 100 on the wrist of one hand (e.g., the left hand), the maternal body MO comes into contact with the first electrode. During measurement, the pregnant woman touches the operation unit 103 with her other hand (right hand), which brings the maternal body MO into contact with the second electrode. In this example, the sensor 30 can measure a signal corresponding to lead I of an electrocardiogram, which measures the potential difference between the left and right hands.

[0061] With this configuration, the information acquisition unit 10 acquires an electrical activity record 2 from a sensor 30 provided in the wearable device 100 that transcutaneously measures electrical activity. The generation of the pregnancy status 3 and evaluation information 4 by the inference unit 11 and the evaluation unit 12 is realized by calculation processing in which the control unit 102 provided in the wearable device 100 executes a program 20. The output unit 13 outputs the pregnancy status 3 and evaluation information 4 to a display unit 31 provided in the wearable device 100. The display unit 31 displays the pregnancy status 3 and evaluation information 4 supplied from the output unit 13. In this way, the pregnancy status estimation apparatus 1 can be realized by the wearable device 100. In other words, the wearable device 100 executes the program 20 to perform the pregnancy status estimation method according to the first embodiment.

[0062] 9 is an explanatory diagram showing a second configuration example of the pregnancy status estimation device 1 according to the first embodiment. In the example of FIG. 9, the pregnancy status estimation device 1 is realized by a server 200 connected to the network NW. The pregnancy status 3 and evaluation information 4 generated by the pregnancy status estimation device 1 are provided to the user in the form of a so-called cloud service.

[0063] The server 200 is a computer capable of communicating via a network NW. The server 200 includes an information acquisition unit 10, an inference unit 11, an evaluation unit 12, an output unit 13, and a storage unit 14. The information acquisition unit 10, the inference unit 11, the evaluation unit 12, and the output unit 13 of the pregnancy status estimation device 1 are realized on a program 20 (see FIG. 1 ) stored in the storage unit 14 by the server 200 executing the program 20. The server 200 may be configured as a single device or a group of devices (a server group). For example, the server 200 may be configured by a server device functioning as the information acquisition unit 10, a server device functioning as the inference unit 11, a server device functioning as the evaluation unit 12, and a server device functioning as the storage unit 14, and the pregnancy status estimation device 1 may be configured by mutual communication between the server devices. The output unit 13 is realized by the communication function of the server device.

[0064] In the example of FIG. 9 , a user, such as a pregnant woman or her family member, uses the pregnancy status estimation device 1 using an information communication terminal 210 that can communicate with the server 200 via a network NW. The information communication terminal 210 may be a smartphone, a tablet terminal, a PC (Personal Computer), or the like. The information communication terminal 210 includes a control unit 211, a communication unit 212, a display unit 213, and a storage unit 214. The storage unit 214 stores an application program for utilizing the functions of the pregnancy status estimation device 1 through communication with the server 200. The control unit 211 acquires, via the communication unit 212, an electrical activity record 2 derived from the maternal MO, which is measured transcutaneously, from the sensor 30 attached to the maternal MO. The control unit 211 executes the application program to establish communication with the server 200 via the communication unit 212 and transmits the electrical activity record 2 to the server 200.

[0065] In the server 200, the information acquisition unit 10 acquires the electrical activity record 2 transmitted from the information communication terminal 210. The inference unit 11 inputs the electrical activity record 2 into a trained model 21 stored in the memory unit 14 to infer a pregnancy status 3. The evaluation unit 12 generates evaluation information 4 based on the pregnancy status 3. The output unit 13 outputs the pregnancy status 3 and the evaluation information 4 to the information communication terminal 210 via the network NW.

[0066] The control unit 211 of the information communication terminal 210 receives the pregnancy status 3 and the evaluation information 4 from the server 200 via the communication unit 212. The control unit 211 displays the received pregnancy status 3 and evaluation information 4 on the display unit 213. In this way, the pregnancy status estimation device 1 is realized by the server 200 that provides cloud services. In other words, the server 200 executes the program 20 to perform the pregnancy status estimation method according to the first embodiment.

[0067] (Effects) As described above, the pregnancy status estimation device 1 according to the first embodiment includes an information acquisition unit 10 that acquires a transcutaneously measured electrical activity record 2 derived from the maternal MO; an inference unit 11 that inputs the electrical activity record 2 to a trained model 21 that outputs a pregnancy status including information related to at least one of the maternal MO's uterine contractions and the fetal heart rate; and an output unit 13 that outputs the pregnancy status 3 generated by the inference unit 11. This allows the pregnancy status 3 to be inferred and provided to a user simply by transcutaneously measuring the maternal MO's electrical activity record 2, without the need for an expensive, large, dedicated device such as a fetal monitoring device. The user can grasp the objective pregnancy status 3 generated by the inference process without requiring specialized knowledge or data interpretation work to interpret the electrical activity record 2. As a result, the status of the mother or fetus can be grasped inexpensively and easily, even outside of a hospital.

[0068] In the pregnancy status estimation device 1 according to the first embodiment, the trained model 21 includes a first trained model 21A that outputs information about uterine contractions of the mother MO and a second trained model 21B that outputs information about the fetal heart rate in the mother MO. The inference unit 11 infers a pregnancy status 3 that includes information about uterine contractions of the mother MO and information about the fetal heart rate in the mother MO. This allows the user to be provided with both information about uterine contractions of the mother MO and information about the fetal heart rate in the mother MO. As a result, the user can be motivated to engage in behavior that leads to early therapeutic intervention to prevent premature birth and intrauterine fetal death. Furthermore, by providing the first trained model 21A and the second trained model 21B separately, the learning process can be performed more easily than when both information about uterine contractions 3A and information about the fetal heart rate 3B are inferred using a single trained model.

[0069] The pregnancy status estimation device 1 according to the first embodiment further includes an evaluation unit 12 that generates evaluation information 4 according to the content of the pregnancy status 3 generated by the inference unit 11, and the output unit 13 outputs the evaluation information 4 according to the pregnancy status 3 in addition to the pregnancy status 3. This makes it possible to provide the user with evaluation information derived from the pregnancy status 3 in addition to objective information on the pregnancy status 3. This allows a user without specialized knowledge to more easily understand the status of the maternal MO and fetus FE.

[0070] In the pregnancy status estimation device 1 according to the first embodiment, the information 3A relating to the uterine contractions of the mother MO includes at least one of information on the strength, duration, interval, period, frequency, and intensity waveform of the uterine contractions, thereby making it possible to obtain objective data relating to the uterine contractions of the mother MO.

[0071] In the pregnancy status estimation device 1 according to the first embodiment, the information 3B relating to the fetal heartbeat in the mother MO includes at least one of the fetal heartbeat, fetal heart rate, fetal heart sound, heartbeat interval, cardiac waveform, and fetal heartbeat findings, thereby making it possible to obtain objective data relating to the fetal heartbeat in the mother MO.

[0072] In the pregnancy status estimation device 1 according to the first embodiment, the information acquisition unit 10, inference unit 11, and output unit 13 are configured by a control unit 102 included in a wearable device 100 that can be worn on a part of the body executing a program 20, and the information acquisition unit 10 acquires an electrical activity record 2 from a sensor 30 that transcutaneously measures electrical activity and is included in the wearable device 100, and the output unit 13 outputs a pregnancy status 3 to a display unit 31 included in the wearable device 100. In this way, the wearable device 100, which has sensing and display functions, can achieve everything from measuring the electrical activity record 2 to displaying the pregnancy status 3. When a wearable device 100 available on the market is used, it is possible to provide a pregnancy status estimation device 1 that is significantly cheaper than dedicated devices such as fetal monitoring devices and can be easily introduced outside of a hospital.

[0073] The pregnancy status estimation method according to the first embodiment includes a step S10 of acquiring a transcutaneously measured electrical activity record 2 derived from the maternal MO; a step S20 of inferring the pregnancy status 3 of the maternal MO by inputting the electrical activity record 2 into a trained model 21 that receives the electrical activity record 2 as an input and outputs a pregnancy status 3 including information on at least one of the uterine contractions of the maternal MO and the fetal heart rate in the maternal MO; and a step S40 of outputting the pregnancy status 3. This allows the pregnancy status 3 to be inferred from the transcutaneously measured electrical activity record 2 derived from the maternal MO and provided to the user without using an expensive, large, dedicated device such as a fetal monitoring device. The user can grasp the objective pregnancy status 3 generated by the inference process without requiring specialized knowledge to interpret the electrical activity record 2. As a result, the status of the mother or fetus can be grasped inexpensively and easily, even outside of a hospital.

[0074] Second Embodiment Next, a second embodiment will be described. In the second embodiment, an example will be described in which the inference unit 11 infers a pregnancy status 3 including information 3C (see FIG. 11 ) related to intrauterine movements of a fetus FE in a mother MO. In the second embodiment, descriptions of parts that are common to the first embodiment will be omitted.

[0075] Fig. 10 is a schematic block diagram showing a pregnancy status estimation device according to the second embodiment. As shown in Fig. 10, the pregnancy status estimation device 1A according to the second embodiment includes an information acquisition unit 10, an inference unit 11, an evaluation unit 12, an output unit 13, and a storage unit 14.

[0076] In the second embodiment, the storage unit 14 stores a trained model 121 that outputs a pregnancy status 3 including information 3C (see FIG. 11 ) related to the intrauterine movement of the fetus FE in the mother MO. In the example of FIG. 10 , the storage unit 14 also stores the trained model 21 according to the first embodiment.

[0077] In the second embodiment, the information acquisition unit 10 acquires the electrical activity record 2 via a sensor 30 that contacts one or more of the limbs of the mother MO to detect electrical signals. The sensor 30 detects electrical signals used in electrocardiograms or electromyograms. The sensor 30 is placed on the surface of one of the limbs of the mother MO, i.e., the right arm, left arm, right leg, or left leg. The sensor 30 transcutaneously measures electrical signals generated in association with the biological activity of the mother MO using electrodes that contact the mother MO. The placement location of the sensor 30 is not particularly limited, but in the case of the right arm or left arm, it may be anywhere from the shoulder to the fingertips. In the case of the right leg or left leg, it may be anywhere from the base of the leg to the toes. The sensor 30 may be placed on multiple limbs of the mother MO. In the second embodiment, the sensor 30 is a sensor built into a wearable sensor device or a multi-function wearable device (see FIG. 8 ). In the case of the wearable device 100 shown in FIG. 8 , the maternal MO comes into contact with the first electrode of the sensor 30 when the wearable device 100 is worn on the wrist of one hand (e.g., the left hand). During measurement, the pregnant woman touches the operation unit 103 with her other hand (right hand), bringing the maternal MO into contact with the second electrode. The pregnant woman may also touch the operation unit 103 to a body surface other than the limbs, such as the surface of the chest wall or abdominal wall. The sensor 30 measures the potential difference between the contact points of the first electrode and the second electrode. The information acquisition unit 10 acquires data measured while the pregnant woman wears the sensor 30 and goes about her daily life. Note that the sensor 30 may also be placed on a body surface other than the limbs of the maternal MO. For example, in the case of the wearable device 100 shown in FIG. 8 , the first electrode provided on the inner surface of the housing 101 is placed on the abdominal wall surface of the maternal MO, and the pregnant woman touches the operation unit 103 with either hand to bring the maternal MO into contact with the second electrode. This also makes it possible to obtain an electrical activity record 2.

[0078] In the second embodiment, the inference unit 11 infers the pregnancy status 3 using the trained model 121. That is, the inference unit 11 inputs the electrical activity record 2 into the trained model 121, which receives the electrical activity record 2 as input and outputs the pregnancy status 3 including information 3C related to the intrauterine movements of the fetus FE in the mother MO, to infer the pregnancy status 3 of the mother MO. The output pregnancy status 3 includes numerical information related to the intrauterine movements of the fetus FE in the mother MO. Note that in the example of FIG. 10 , the inference unit 11 infers information 3A related to uterine contractions of the mother MO using the first trained model 21A and infers information 3B related to the fetal heartbeat in the mother MO using the second trained model 21B. In the second embodiment, the pregnancy status 3 includes information 3A related to uterine contractions, information 3B related to the fetal heartbeat, and information 3C related to the intrauterine movements of the fetus FE. The pregnancy status 3 in the second embodiment does not necessarily include information 3A related to uterine contractions and information 3B related to the fetal heartbeat. As in the first embodiment, the inference unit 11 may perform preprocessing on the electrical activity record 2. The inference unit 11 may perform preprocessing on the electrical activity record 2, including filtering to remove predetermined signal components corresponding to respiratory fluctuations in the maternal MO. The inference unit 11 may perform preprocessing on the electrical activity record 2, which is electrocardiogram data, to add information identifying at least one of the baseline, P point, Q point, R point, S point, and T point.

[0079] 11 is an explanatory diagram showing a trained model 121 according to the second embodiment. The trained model 121 is generated in advance by machine learning so as to receive an electrical activity record 2 derived from the mother MO as input and to output information 3C relating to the intrauterine movement of the fetus FE in the mother MO. The electrical activity record 2 as training data may be preprocessed data.

[0080] In this specification, fetal movement refers to the movement (movement) of the fetus's body in the womb, and is a broad concept that includes fetal muscle tone, fetal movements, and breathing-like movements. Fetal movement particularly refers to temporary fetal movements, and does not include physiological activities of internal organs that are constantly occurring, such as the fetal heartbeat.

[0081] Information 3C regarding intrauterine movements of fetus FE in maternal MO includes at least one of information regarding muscle tone, fetal movements, and breathing-like movements of fetus FE. Muscle tone of fetus FE is the bending and stretching movement of a part of the body of fetus FE (trunk or limbs). Examples of muscle tone include the movement of the spine or limbs extending from a flexed position and returning to the original flexed position, and the opening and closing movement of the palms. Fetal movements of fetus FE are single or complex movements of the trunk or limbs of fetus FE, but refer to movements larger than muscle tone. Examples of fetal movements of fetus FE include movements such as changing the position of the body or kicking the uterine wall. Breathing-like movements of fetus FE are movements similar to breathing. Examples of breathing-like movements include intermittent movements of the diaphragm, abdominal wall, and thorax of fetus FE that last for several seconds to several tens of seconds.

[0082] In Figure 11, the trained model 121 outputs information 3C related to the intrauterine movements of the fetus FE in the mother MO as time-series data indicating the time points at which intrauterine movements occur. In Figure 11, the information 3C related to intrauterine movements is shown as a one-dimensional graph in which the horizontal axis represents time (time u) and the occurrence of intrauterine movements is indicated by the presence or absence of a bar plot. The length of the bar plot in the horizontal direction represents the duration from the start to the end of one intrauterine movement. A position where no bar plot is present indicates that no intrauterine movement has occurred. In the example of Figure 11, the information 3C related to intrauterine movements includes an interval E between intrauterine movements. The interval E between intrauterine movements is the time interval from one intrauterine movement to the next. The information 3C related to intrauterine movements may also include a duration G of intrauterine movements. The duration G of intrauterine movements is the duration from the start to the end of one intrauterine movement. The information 3C related to intrauterine movements may be time-series data including information on the intensity of intrauterine movements. In this case, information 3C regarding intrauterine movements can be represented as a two-dimensional graph in which the vertical axis represents the intensity of intrauterine movements and the horizontal axis represents time (time u). The trained model 121 may generate information 3C regarding intrauterine movements that distinguishes between muscle tension, fetal movements, and respiratory-like movements, or may generate information 3C regarding intrauterine movements without distinguishing between muscle tension, fetal movements, and respiratory-like movements.

[0083] 12 is an explanatory diagram illustrating evaluation information according to the second embodiment. The evaluation unit 12 generates evaluation information 4 based on data on pregnancy status 3 and evaluation data 22 (see FIG. 10 ) pre-stored in the storage unit 14. In the second embodiment, the evaluation unit 12 generates evaluation information 4D based on intrauterine movement based on information 3C related to intrauterine movement of fetus FE.

[0084] The evaluation information 4D based on intrauterine movement includes, for example, an evaluation of the frequency of intrauterine movement (the length of the interval E between intrauterine movements). There is a correlation between the frequency of intrauterine movement of the fetus FE and the health of the fetus FE, and a decrease in the frequency or absence of intrauterine movement (so-called decreased fetal movement) is a sign or possibility indicator of intrauterine fetal death or placental abruption. The evaluation information 4D based on intrauterine movement includes, for example, an evaluation of whether the interval E between intrauterine movements reaches a predetermined criterion (i.e., an evaluation of whether fetal movement is decreasing). The predetermined criterion is, for example, a relative criterion such that the interval E is N% or more (N>100) compared to the average value over the most recent M hours. The frequency of intrauterine movement per certain time period may be calculated from the interval E, and the frequency of intrauterine movement may be compared with the criterion. The evaluation information 4D based on intrauterine movement may include, for example, an evaluation of whether the duration G of intrauterine movement exceeds a predetermined criterion (i.e., an evaluation of the length of one intrauterine movement).

[0085] In the second embodiment, information 3C relating to intrauterine movements is acquired daily and continuously by a wearable sensor 30, and evaluation information 4D based on the intrauterine movements is generated by an evaluation unit 12. The information 3C relating to intrauterine movements and the evaluation information 4D based on the intrauterine movements allow doctors and other medical professionals to obtain objective data and evaluations over a relatively long period of time to determine whether or not there is a decrease in fetal movements.

[0086] 12 , the evaluation unit 12 generates comprehensive evaluation information 4C based on information 3C related to intrauterine movements of the fetus FE as well as information 3A related to uterine contractions and information 3B related to the fetal heart rate shown in the first embodiment. The evaluation information 4 generated by the evaluation unit 12 is supplied to the output unit 13. The output unit 13 outputs output data including the pregnancy status 3 generated by the inference unit 11 and the evaluation information 4 generated by the evaluation unit 12 to the display unit 31.

[0087] 13 is an explanatory diagram illustrating generation (machine learning) of the trained model 121. In the second embodiment, the training data 141 used in the machine learning of the trained model 40 is created from measurement data obtained using dedicated devices such as a fetal monitoring device and an ultrasound examination device (echo, Doppler) in a hospital, for example.

[0088] The teacher data 141 used to generate the trained model 121 includes a maternal electrical activity record 2 (time-series data) measured transcutaneously and a fetal movement record 142 of the fetus FE measured from the same mother. The fetal movement record 142 is time-series data of the fetal movement of the fetus FE in the mother MO. The teacher data 141 includes the electrical activity record 2 and the fetal movement record 142, which are a pair of data measured at the same time t. The fetal movement record 142 is obtained by applying a probe of an ultrasound examination device or an ultrasound transducer of a fetal monitoring device to the abdominal wall surface of the mother MO to measure an acoustic signal and extracting signal components derived from the fetus FE through arithmetic processing. The fetal movement record 142 is information equivalent to the information 3C regarding the fetal FE's intrauterine movement generated by the inference unit 11. The learning unit ML uses teacher data 141 to cause the learning model 40 to acquire, by deep learning, feature quantities for converting a maternal electrical activity record 2 measured transcutaneously into an output of an intrauterine movement record 142. The learning unit ML determines the feature quantities of the learning model 40 by performing learning using a plurality of teacher data 141. The learning model 40 whose feature quantities have been determined by machine learning is the trained model 121.

[0089] Note that the intrauterine movements of the fetus FE are accompanied by bioelectric signals that are much larger than the fetal heartbeat described in the first embodiment. The bioelectric signals derived from intrauterine movements, estimated as information 3C related to intrauterine movements, can be indirectly measured from the fetus FE via the mother MO. The bioelectric signals derived from intrauterine movements include not only electrical signals due to fetal movements themselves, but also maternal biological activity associated with the occurrence of fetal movements (such as electrical signals derived from the rectus abdominis and biceps brachii muscles due to the mother stiffening), and this maternal biological activity is not due to conscious or unconscious activity. The trained model 121 acquires, through machine learning, a process for separating and extracting signal components derived from intrauterine movements of the fetus FE from the electrical activity record 2 of the mother MO. The generated trained model 121 is pre-stored in the memory unit 14 of the pregnancy status estimation device 1 shown in FIG. 10.

[0090] (Pregnancy status estimation method) Next, a description will be given of a pregnancy status estimation method according to the second embodiment. The process flow of the pregnancy status estimation method according to the second embodiment is the same as that shown in the flowchart of Fig. 7, with only the process contents of some steps being different.

[0091] In step S10, the information acquisition unit 10 acquires an electrical activity record 2 derived from the maternal MO measured transcutaneously. The information acquisition unit 10 acquires the electrical activity record 2 directly from the sensor 30 attached to the maternal MO or via a storage medium that records measurement data from the sensor 30. The information acquisition unit 10 acquires the electrical activity record 2 using the wearable sensor 30. In each measurement, the pregnant woman being measured measures the electrical activity record 2, for example, in a semi-sitting position and in a resting state for approximately 20 to 40 minutes. The measurement may be performed, for example, when the pregnant woman feels "abdominal distension" or "fetal movement is weak," but may also be performed periodically, for example, at a specific time each day.

[0092] In step S20, the inference unit 11 infers the pregnancy status 3 of the mother MO by inputting the electrical activity record 2 into the trained model 21. In the second embodiment, the inference unit 11 inputs the electrical activity record 2 acquired in step S10 into the trained model 121, which receives the electrical activity record 2 as input and outputs the pregnancy status 3 including information 3C related to intrauterine movements of the fetus FE in the mother MO, and performs calculations based on the features. As a result, the inference unit 11 infers the pregnancy status 3 including the information 3C related to intrauterine movements of the fetus FE. In the second embodiment, the inference unit 11 further inputs the electrical activity record 2 acquired in step S10 into the first trained model 21A and the second trained model 21B, and performs calculations based on the respective features. As a result, the inference unit 11 generates the pregnancy status 3 including information 3A related to uterine contractions of the mother MO and information 3B related to fetal heartbeats in the mother MO, in addition to information 3C related to intrauterine movements.

[0093] In step S30, the evaluation unit 12 generates evaluation information 4 based on the data of the pregnancy status 3 and the evaluation data 22. In the second embodiment, the evaluation unit 12 generates evaluation information 4D based on intrauterine movement, evaluation information 4A based on uterine contractions, evaluation information 4B based on fetal heart rate, and overall evaluation information 4C.

[0094] In step S40, the output unit 13 outputs output data including the pregnancy status 3 generated by the inference unit 11 and the evaluation information 4 generated by the evaluation unit 12 to an output destination device such as the display unit 31. As a result, the pregnancy status 3 and the evaluation information 4 are displayed on the display unit 31.

[0095] (Configuration example of pregnancy status estimation device) As with the first embodiment, the pregnancy status estimation device 1A according to the second embodiment can be realized in various forms other than the example shown in Fig. 10. The pregnancy status estimation device 1A according to the second embodiment can be realized, for example, by the wearable device 100 shown in Fig. 8. The pregnancy status estimation device 1A according to the second embodiment can also be realized, for example, by the server 200 shown in Fig. 9, and can provide the pregnancy status 3 and evaluation information 4 in the form of a cloud service.

[0096] (Effects) As described above, the pregnancy status estimation device 1A according to the second embodiment includes an information acquisition unit 10 that acquires a transcutaneously measured electrical activity record 2 derived from the mother MO, an inference unit 11 that inputs the electrical activity record 2 to a trained model 121 that receives as input the electrical activity record 2 and outputs a pregnancy status 3 including information 3C related to intrauterine movements of the fetus FE in the mother MO, and infers the pregnancy status 3 of the mother MO by inputting the electrical activity record 2 to the trained model 121, and an output unit 13 that outputs the pregnancy status 3 generated by the inference unit 11. This allows the pregnancy status 3 to be inferred and provided to the user simply by transcutaneously measuring the electrical activity record 2 derived from the mother MO, without using expensive, large, dedicated devices such as a fetal monitoring device or an ultrasound examination device. The user can grasp the objective pregnancy status 3 generated by the inference process without needing specialized knowledge or data interpretation work to interpret the electrical activity record 2. As a result, the status of the fetus FE can be grasped inexpensively and easily, even outside of a hospital.

[0097] Intrauterine movements of fetus FE can also be measured outside of a hospital, for example, by fixing a dedicated sensor to the surface of the pregnant woman's abdominal wall using an ultrasound examination device. However, such sensors and data processing devices are expensive and difficult to introduce, and it is difficult to continue daily life or perform long-term measurements while the sensor is fixed to the abdominal wall, making it difficult to obtain sufficient measurement data outside of a hospital. For this reason, in the past, intrauterine movements of fetus FE were mostly determined by interviewing pregnant women visiting hospitals for regular outpatient visits. However, the information obtained was based on the pregnant woman's subjective impression of decreased fetal movement, and was therefore not useful as an indicator for identifying signs of intrauterine fetal death or placental abruption. According to the second embodiment, doctors and others can obtain objective data for determining whether or not fetal movement has decreased using information 3C regarding intrauterine movement, which is expected to contribute to early medical intervention and improved treatment strategies for intrauterine fetal death or placental abruption.

[0098] In the pregnancy status estimation device 1A according to the second embodiment, information 3C relating to intrauterine movements of the fetus FE in the mother MO includes information on at least one of muscle tone, fetal movements, and breathing-like movements of the fetus FE, thereby making it possible to obtain objective data that serves as an indicator of the health of the fetus FE.

[0099] In the pregnancy status estimation device 1A according to the second embodiment, the information acquisition unit 10 acquires an electrical activity record 2 via a sensor 30 that detects electrical signals by contacting one or more of the limbs of the mother MO. This allows the electrical activity record 2 to be acquired simply by attaching a wearable sensor 30 to an arm or leg, unlike conventional methods that involve fixing a dedicated sensor to the surface of the abdominal wall to assess the status of the fetus FE. For example, in the case of the sensor 30 provided on the wearable device 100 shown in FIG. 8 , as described above, the first electrode of the sensor 30 worn on the wrist of one hand contacts the mother MO. Therefore, measurement can be performed by the pregnant woman touching the operation unit 103 with her other hand or by bringing the operation unit 103 into contact with the surface of the chest wall or abdominal wall. Because the pregnant woman only needs to touch the operation unit 103, she can remain fully clothed. Even when the operation unit 103 is to be brought into contact with the chest wall or abdominal wall, measurement can be easily performed by simply inserting her hand inside her clothing. Therefore, there are few restrictions on the posture, movements, clothing, etc. of the pregnant woman wearing the sensor 30 during measurement, so continuous and frequent measurements can be performed in daily life without impairing the quality of life (QOL) of the pregnant woman, which increases the possibility of detecting important signs related to intrauterine fetal death or placental abruption.

[0100] Other effects of the second embodiment are similar to those of the first embodiment.

[0101] Although the embodiments of the present invention have been described above, the embodiments are not limited to the contents of these embodiments. Furthermore, the above-described components include those that can be easily imagined by a person skilled in the art, those that are substantially the same, and those that are within the scope of what is called equivalents. Furthermore, the above-described components can be combined as appropriate. Furthermore, various omissions, substitutions, or modifications of the components can be made without departing from the spirit of the above-described embodiments.

[0102] (Fine Tuning) For example, the trained model 21 (121) used in the first and second embodiments may be fine-tuned based on additional information regarding the pregnancy status 3 of the mother MO or fetus FE. Fig. 14 is a schematic diagram illustrating the generation of an individual trained model by fine tuning according to a modified example. Fig. 15 is a schematic diagram illustrating the inference process using an individual trained model according to a modified example.

[0103] As shown in Figure 14, the trained model 21 includes a plurality of individual trained models 321 that have been fine-tuned using individual teacher data 341 of electrical activity records 2 and pregnancy status 3 classified based on additional information 5 for a pre-trained pre-trained model 320. The individual trained models 321 are generated by fine-tuning the pre-trained model 320, which is a trained model 40 pre-trained using the pre-trained teacher data 340, using the individual teacher data 341. Hereinafter, the trained model fine-tuned using the individual teacher data 341 will be referred to as the individual trained model 321.

[0104] The additional information 5 is information other than the electrical activity record 2 that relates to the pregnancy status 3 of the mother MO or fetus FE. The additional information 5 includes at least one of the number of weeks of pregnancy, estimated fetal weight, maternal BMI (Body Mass Index), maternal height, maternal weight, and maternal age. The additional information 5 classifies the teacher data of the electrical activity record 2 and the pregnancy status 3 into one of multiple groups. Individual teacher data 341, which is a group of teacher data belonging to one of the groups in the additional information 5, is created for each group.

[0105] The learning model 40 is machine-learned (pre-trained) using pre-training data 340 that is not grouped (randomly includes data from all groups), thereby generating a pre-trained model 320. Part or all of the generated pre-trained model 320 is trained by fine-tuning using individual training data 341. As a result, an individual trained model 321 is generated for each group of additional information 5. The number of data points in the individual training data 341 may be smaller than the number of data points in the pre-training data 340.

[0106] The method for classifying (grouping) the individual training data 341 is not particularly limited. An example will be described in which the additional information 5 is the number of weeks of pregnancy. The additional information 5 on the number of weeks of pregnancy may be classified into three groups, for example, early pregnancy (before 13 weeks and 6 days), mid-pregnancy (14 weeks and 0 days to 27 weeks and 6 days), and late pregnancy (after 28 weeks and 0 days), depending on the number of weeks of pregnancy. The individual training data 341 is composed of training data obtained from maternal MOs whose number of weeks of pregnancy belongs to one of the groups. Then, by fine-tuning using the individual training data 341 for early pregnancy, the individual training data 341 for mid-pregnancy, and the individual training data 341 for late pregnancy, an individual trained model 321 for early pregnancy, an individual trained model 321 for mid-pregnancy, and an individual trained model 321 for late pregnancy are separately created.

[0107] 15 , the information acquisition unit 10 acquires additional information 5 related to the pregnancy status 3 of the mother MO or fetus FE in addition to the electrical activity record 2. The additional information 5 (or information required to calculate the additional information 5) may be input (e.g., transmitted from an information communication terminal) by the user, such as the mother MO or a doctor, to the pregnancy status estimation device 1, or may be acquired by the information acquisition unit 10 from electronic medical record information, etc. The number of weeks of pregnancy, for example, can be calculated from the date on which the mother MO's pregnancy began and the current date, and the date on which the mother MO's pregnancy began may be acquired by the above-mentioned method.

[0108] The inference unit 11 selects an individual trained model 321 to input the electrical activity record 2 based on the additional information 5 acquired by the information acquisition unit 10. The inference unit 11 inputs the electrical activity record 2 into the selected individual trained model 321, thereby generating a pregnancy status 3 of the mother MO and supplying it to the output unit 13.

[0109] This selects the individual trained model 321 to be used for inference depending on which of the three groups (early, mid, or late pregnancy) the maternal MO's gestational age falls into. As a result, the accuracy of inference according to the number of weeks of pregnancy can be improved. The frequency of uterine contractions and the frequency and intensity of fetal movement tend to increase as the number of weeks of pregnancy advances, and fine-tuning according to the number of weeks of pregnancy is effective in improving inference accuracy. Furthermore, electrical activity records 2 such as electrocardiograms vary from the statistical average depending on the subject's physique (maternal BMI, maternal height, maternal weight). Fine-tuning according to maternal BMI, maternal height, maternal weight, etc. can address statistical bias in the input data, and is therefore effective in improving inference accuracy. As the number of weeks of pregnancy advances and the estimated fetal weight increases, the frequency and intensity of fetal movement tend to increase. Therefore, like the number of weeks of pregnancy, fine-tuning according to the estimated fetal weight is also effective in improving inference accuracy. Furthermore, as maternal age increases, the strength of uterine contractions is more likely to weaken, which puts older mothers at risk of experiencing weak labor and labor arrest. Age-based fine-tuning makes it possible to take into account the degree of weakening of uterine contractions, thereby contributing to improved inference accuracy.

[0110] There is no particular limitation on the number of groups into which the additional information 5 is classified. The number of weeks of pregnancy may also be classified into two or four or more groups, or each week of pregnancy may be a separate group. The same applies to the other additional information 5. In addition to the above-mentioned number of weeks of pregnancy, estimated fetal weight, maternal BMI, maternal height, maternal weight, and maternal age, the additional information 5 may also include pregnancy / parity status, race, nationality, lifestyle history (smoking, drinking, occupational history), pregnancy history (natural conception, artificial insemination, in vitro fertilization), medical history, obstetric complications, oral medications, allergy history, family history, presence or absence of edema, and urine test results (urine protein, urine sugar).

[0111] 14 shows an example in which fine tuning is performed on the pre-trained model 320 based on the additional information 5, but the individual trained model 321 may be created (without fine tuning) by separately performing machine learning on the learning model 40 based on the additional information 5. In other words, the individual training data 341 classified based on the additional information 5 may be used to perform machine learning on the learning model 40, thereby separately creating trained models 21 for each classification based on the additional information 5.

[0112] 1, 1A... pregnancy status estimation device, 2... electrical activity record, 3... pregnancy status, 3A... information on uterine contractions, 3B... information on fetal heart rate, 3C... information on intrauterine movement, 4... evaluation information, 4A... evaluation information based on uterine contractions, 4B... evaluation information based on fetal heart rate, 4C... overall evaluation information, 4D... evaluation information based on intrauterine movement, 5... additional information, 10... information acquisition unit, 11... inference unit, 12... evaluation unit, 13... output unit, 14... memory unit, 20... program, 21... trained model, 21A... first trained model, 21B... second trained model, 22... evaluation data, 30... sensor, 31... display unit, 40... training model, 41A, 41B... training data, 42... child Uterine contraction recording, 43...fetal heart rate recording, 100...wearable device, 101...housing, 102...control unit, 103...operation unit, 121...trained model, 141...teaching data, 142...intrauterine movement recording, 200...server, 210...information and communication terminal, 211...control unit, 212...communication unit, 213...display unit, 214...memory unit, 320...pre-trained model, 321...individual trained model, 340...pre-teaching data, 341...individual teaching data, B...fetal heart rate, D...interval between uterine contractions, E...interval between fetal movements, FE...fetus, G...duration of fetal movements, H...strength of uterine contractions, L...duration of uterine contractions, ML...learning unit, MO...mother, NW...network.

Claims

1. A pregnancy status estimation device comprising: an information acquisition unit that acquires a record of electrical activity from the mother measured transcutaneously; an inference unit that inputs the electrical activity record into a trained model that outputs a pregnancy status including information regarding at least one of the mother's uterine contractions and the fetal heart rate in the mother, and infers the pregnancy status of the mother; and an output unit that outputs the pregnancy status generated by the inference unit.

2. The pregnancy status estimation device of claim 1, wherein the trained models include a first trained model that outputs information regarding uterine contractions of the mother and a second trained model that outputs information regarding the fetal heart rate in the mother, and the inference unit infers the pregnancy status including information regarding uterine contractions of the mother and information regarding the fetal heart rate in the mother, respectively.

3. A pregnancy status estimation device as described in claim 1 or 2, further comprising an evaluation unit that generates evaluation information according to the content of the pregnancy status generated by the inference unit, and the output unit outputs the evaluation information according to the pregnancy status in addition to the pregnancy status.

4. The pregnancy status estimation device according to claim 1 or 2, wherein the information relating to the maternal uterine contractions includes at least one of information regarding the strength, duration, interval, period, frequency and intensity waveform of the uterine contractions.

5. A pregnancy status estimation device as described in claim 1 or 2, wherein the information regarding the fetal heartbeat in the mother's body includes at least any one of information regarding the fetal heartbeat, fetal heart rate, fetal heart sounds, heartbeat intervals, heart sound waveforms, and fetal heartbeat findings.

6. A pregnancy status estimation device as described in claim 1 or 2, wherein the information acquisition unit, the inference unit and the output unit are configured by a control unit provided in a wearable device that can be worn on a part of the body executing a program, the information acquisition unit acquires the electrical activity record from a sensor provided in the wearable device that measures electrical activity transcutaneously, and the output unit outputs the pregnancy status to a display unit provided in the wearable device.

7. A method for estimating pregnancy status comprising: a step of acquiring a maternal electrical activity record measured transcutaneously; a step of inferring the pregnancy status of the mother by inputting the electrical activity record into a trained model that takes the electrical activity record as input and outputs the pregnancy status including information regarding at least one of the maternal uterine contractions and the fetal heart rate in the mother; and a step of outputting the pregnancy status.

8. A pregnancy status estimation device comprising: an information acquisition unit that acquires a record of electrical activity from the mother measured transcutaneously; an inference unit that inputs the electrical activity record into a trained model that outputs a pregnancy status including information regarding intrauterine movement of the fetus in the mother's body, and infers the pregnancy status of the mother; and an output unit that outputs the pregnancy status generated by the inference unit.

9. The pregnancy status estimation device according to claim 8, wherein the information relating to intrauterine movements of the fetus in the mother's body includes at least one of information on muscle tone, fetal movements, and respiratory-like movements of the fetus.

10. The pregnancy status estimation device according to claim 8, wherein the information acquisition unit acquires the electrical activity record via a sensor that contacts one or more of the mother's limbs to detect electrical signals.

11. A pregnancy status estimation device as described in any one of claims 8 to 10, further comprising an evaluation unit that generates evaluation information according to the content of the pregnancy status generated by the inference unit, and the output unit outputs the evaluation information according to the pregnancy status in addition to the pregnancy status.

12. A method for estimating pregnancy status, comprising: a step of acquiring a maternal electrical activity record measured transcutaneously; a step of inferring the pregnancy status of the mother by inputting the electrical activity record into a trained model that takes the electrical activity record as input and outputs the pregnancy status including information about intrauterine movement of the fetus in the mother; and a step of outputting the pregnancy status.

Images