JP6573416B1 - Blood pressure estimation device, blood pressure estimation system, and blood pressure estimation program - Google Patents

Abstract

An object of the present invention is to estimate a blood pressure of a measurement subject. A blood pressure estimation device 100 represents a feature extraction network that converts input pulse wave data into a feature expression, and the feature expression converted by the feature extraction network represents a probability of a systolic blood pressure value and a diastolic blood pressure value. A blood pressure estimation network for converting into a probability vector, a training step execution means for executing a training step for causing the feature extraction network and the blood pressure estimation network to learn systolic blood pressure values and diastolic blood pressure values according to pulse wave data; The estimation is performed by inputting the pulse wave data of the measurement subject to the feature extraction network and the blood pressure estimation network that have been learned in the training step, and performing the estimation step of estimating the systolic blood pressure value and the diastolic blood pressure value of the measurement subject. Step execution means. [Selection] Figure 1

Description

  The present invention relates to a blood pressure estimation device, a blood pressure estimation system, and a blood pressure estimation program.

  The following blood pressure estimation apparatus is known. In this blood pressure estimation device, a method for estimating a subject's blood pressure from a pulse wave feature amount using a neural network learned based on a cuff sphygmomanometer measurement value and a pulse wave feature amount has been proposed. (For example, patent document 1).

JP 2016-16295 A

  In recent years, it has become possible to easily acquire a user's pulse wave with a wearable terminal worn by the user on a wrist or the like. For this reason, the method of estimating a blood pressure from the user's pulse wave acquired with the wearable terminal etc. using the conventional blood-pressure estimation apparatus can be considered. However, in the conventional technique, the feature quantity of the measurement subject's pulse wave is extracted, and when the extracted feature quantity is input, the estimated blood pressure is output, and is generated by support vector regression analysis based on teacher data. Was used to estimate the blood pressure of the measurement subject. When the blood pressure is estimated based on the pulse wave of the measurement subject, it is difficult to specify the blood pressure value uniformly from the pulse wave, and there is a possibility that the blood pressure varies within a certain range depending on the person. For this reason, it is expected that the estimation accuracy will be improved by estimating the blood pressure value corresponding to the pulse wave in consideration of the probability, but this point has not been studied in the past.

A blood pressure estimation apparatus according to the present invention is a blood pressure estimation apparatus that estimates blood pressure using a neural network, and includes a feature extraction network that converts input pulse wave data into a feature expression, and a feature that is converted by the feature extraction network. The blood pressure estimation network that converts the expression into a probability vector that represents the probabilities of systolic blood pressure value and diastolic blood pressure value, and the feature extraction network and blood pressure estimation network include the systolic blood pressure value and the diastolic blood pressure value according to the pulse wave data. The pulse wave data of the measurement subject is input to the training step execution means for executing the training step for learning, the feature extraction network and the blood pressure estimation network that have been learned by the training step, and the systolic blood pressure of the measurement subject Estimation step to estimate the blood pressure value and diastolic blood pressure value And a estimation step execution means for Rows, estimation step execution means obtains a probability vector representing the likelihood of the systolic blood pressure value and diastolic blood pressure values of the measured person, the measured person values became highest probability The systolic blood pressure value and the diastolic blood pressure value are estimated .
A blood pressure estimation system according to the present invention is a blood pressure estimation system that estimates blood pressure using a neural network, and includes a feature extraction network that converts input pulse wave data into a feature expression, and a feature that is converted by the feature extraction network. The blood pressure estimation network that converts the expression into a probability vector that represents the probabilities of systolic blood pressure value and diastolic blood pressure value, and the feature extraction network and blood pressure estimation network include the systolic blood pressure value and the diastolic blood pressure value according to the pulse wave data. The pulse wave data of the measurement subject is inputted to the training step execution means for executing the training step for learning, the feature extraction network learned by the training step and the blood pressure estimation network, and the systolic phase of the measurement subject Estimating blood pressure and diastolic blood pressure And a estimation step execution means for executing steps, estimation step execution means obtains a probability vector representing the likelihood of the systolic blood pressure value and diastolic blood pressure values of the measured person, measuring a value was the highest probability It is estimated as a systolic blood pressure value and a diastolic blood pressure value of the subject .
A blood pressure estimation program according to the present invention includes a feature extraction network that converts input pulse wave data into a feature expression, and a probability vector that represents the probability of systolic blood pressure value and diastolic blood pressure value from the feature expression converted by the feature extraction network. A blood pressure estimation program for estimating blood pressure using a neural network having a blood pressure estimation network that converts to a feature extraction network and a blood pressure estimation network, a systolic blood pressure value and a diastolic blood pressure value according to pulse wave data The measurement target person's pulse wave data is input to the training step execution procedure for executing the training step for learning, and the feature extraction network and blood pressure estimation network that have been learned by the training step, and the measurement subject's systole Estimate blood pressure and diastolic blood pressure That is executed and the estimated step execution procedure to perform the estimation step to the computer, the estimated step execution procedure obtains a probability vector representing the likelihood of the systolic blood pressure value and diastolic blood pressure values of the measured person, and highest probability The measured values are estimated as the systolic blood pressure value and the diastolic blood pressure value of the measurement subject .

  According to the present invention, in the training step, a pre-trained feature extraction network and blood pressure estimation network are used, and the pulse wave data of the measurement subject is converted into a probability vector representing the probabilities of the systolic blood pressure value and the diastolic blood pressure value. Since the conversion is performed, the blood pressure of the measurement subject can be estimated with high accuracy based on the converted probability vector.

1 is a block diagram illustrating a configuration of an embodiment of a blood pressure estimation device 100. FIG. It is the figure which showed typically the input data Ut in case the number of the sensor signals obtained from the wearable terminal is 1, and the length of each signal segment is n. It is the figure which showed typically the input data Ut in case the number of the sensor signals obtained from the wearable terminal is 3, and the length of each signal segment is n. It is the figure which showed typically the input data Ut in case the number of the sensor signals obtained from the wearable terminal is 4, and the length of each signal segment is n. It is the figure which showed the flow of the training step typically. It is the figure which showed the flow of the estimation step typically. It is a flowchart figure which shows the flow of a training step. It is a flowchart figure which shows the flow of an estimation step.

  FIG. 1 is a block diagram showing a configuration of an embodiment of a blood pressure estimation device 100 in the present embodiment. As the blood pressure estimation device 100, for example, a server device, a personal computer, a smartphone, a tablet terminal, or the like is used. FIG. 1 is a block diagram showing a configuration of an embodiment when a personal computer is used as blood pressure estimation apparatus 100 in the present embodiment.

  The blood pressure estimation device 100 includes an operation member 101, a control device 102, a storage medium 103, a display device 104, and a connection interface 105.

  The operation member 101 includes various devices operated by an operator of the blood pressure estimation device 100, such as a keyboard and a mouse.

  The control device 102 includes a CPU, a memory, and other peripheral circuits, and controls the entire blood pressure estimation device 100. In addition, the memory which comprises the control apparatus 102 is volatile memories, such as SDRAM, for example. This memory is used as a work memory for the CPU to expand the program when the program is executed and a buffer memory for temporarily recording data. For example, data read via the connection interface 102 is temporarily recorded in the buffer memory.

  The storage medium 103 is a storage medium for recording various data stored by the blood pressure estimation apparatus 100, data of a program to be executed by the control apparatus 102, and the like, for example, HDD (Hard Disk Drive) or SSD (Solid State). Drive) or the like is used. The program data recorded in the storage medium 103 is provided by being recorded on a recording medium such as a CD-ROM or DVD-ROM, or provided via a network, and stores the program data acquired by the operator. By installing in the medium 103, the control device 102 can execute the program. In the present embodiment, programs and various data used in the processing described below are recorded in the storage medium 103.

  The display device 104 is a liquid crystal monitor, for example, and displays various display data output from the control device 102.

  The connection interface 105 is an interface for the blood pressure estimation device 100 to communicate with an external device such as another device or a terminal through communication. For example, the communication interface 105 includes an interface for connecting to a communication line such as a LAN or the Internet, and an interface for short-distance wireless communication such as Bluetooth (registered trademark). The blood pressure estimation apparatus 100 can acquire data from a wearable terminal described later via the communication interface 105.

  The blood pressure estimation device 100 according to the present embodiment acquires pulse wave data from a device that can measure a user's pulse wave, and executes processing for estimating the user's blood pressure based on the acquired pulse wave data. . In the present embodiment, the device capable of measuring a user's pulse wave is, for example, a wearable terminal that is mounted on the user's wrist and has a pulse wave measurement sensor mounted on a portion that is in close contact with the user's body. Is assumed.

なお、式（１）において、ｗはウェアラブル端末から得られたセンサ信号の数であり、ｎは各信号セグメントの長さを表す。 In the present embodiment, a training step is executed in advance to capture output data from the wearable terminal and cause the neural network to learn a blood pressure value corresponding to the output signal from the wearable terminal. The output data from the wearable terminal used in the training step is expressed by the following equation (1) as a large set Ut of signal segments, for example.

In Equation (1), w is the number of sensor signals obtained from the wearable terminal, and n represents the length of each signal segment.

  Ut is configured to be input to a neural network, and segments obtained from the wearable terminal are configured into Ut segments as shown in FIGS.

  FIG. 2 shows a case where the wearable terminal includes one sensor, that is, the number w of sensor signals obtained from the wearable terminal is one. In FIG. 2, X represents a segment taken out from Sensor 1 included in the wearable terminal. When the length of the segment X is n, the Ut segment constitutes a 1 × n × 1 vector.

  FIG. 3 shows a case where the wearable terminal includes three sensors, that is, the number w of sensor signals obtained from the wearable terminal is three. In FIG. 3, X represents a segment extracted from Sensor 1 included in the wearable terminal, Y represents a segment extracted from Sensor 2 included in the wearable terminal, and Z represents a segment extracted from Sensor 3 included in the wearable terminal. Yes. When the lengths of the segment X, the segment Y, and the segment Z are n, the Ut segment constitutes a 1 × n × 3 vector.

  FIG. 4 shows a case where the wearable terminal includes four sensors, that is, the number w of sensor signals obtained from the wearable terminal is four. In FIG. 4, X represents a segment extracted from Sensor 1 included in the wearable terminal, Y represents a segment extracted from Sensor 2 included in the wearable terminal, and Z represents a segment extracted from Sensor 3 included in the wearable terminal. G represents a segment extracted from Sensor 4 included in the wearable terminal. When the lengths of the segment X, the segment Y, the segment Z, and the segment G are n, the U segment constitutes a 1 × n × 4 vector.

  In the present embodiment, when the training data Ut acquired in advance is input, the control device 102 executes the training step according to the flow shown in FIG. In FIG. 5, Ut shows the case where the number of sensor signals obtained from the wearable terminal is 3, and the length of each signal segment is n, as shown in FIG. As shown in FIG. 5, the control device 102 uses the data Ut as input data, and inputs the data Ut into a feature extraction network 5 a prepared in advance.

The feature extraction network 5a is Convolutional Layers in a known Convolutional Neural Network, and uses a feature mapping F (z || Ut) for mapping an input signal Ut to a feature expression z in the neural network. Ut is converted into a feature representation z. Here, the feature expression is also called a latent expression, and the feature expression z is expressed by the following equation (2). In the following equation (2), k is the length of the feature expression z.

The feature expression z converted by the feature extraction network 5a is input to the blood pressure estimation network 5b. The blood pressure estimation network 5b is a Fully Connected Layers in a well-known convolutional neural network (Fully Connected Neural Network), and is implemented using layers that are completely connected using the following equation (3).

The blood pressure estimation network 5b has an estimated mapping Q ((qs, qd) || z for mapping the feature expression z to a set of probability vectors qs and qd representing the respective probabilities of the systolic blood pressure value and the diastolic blood pressure value. ) To convert the feature expression z into a pair of probability vectors qs and qd. qs is a probability vector representing the systolic blood pressure value, and qd is a probability vector representing the diastolic blood pressure value. For example, the blood pressure estimation network 5b is normalized by the softmax function qi defined by the following equation (4), and converts the output vectors qs and qd into a probability vector whose sum is 1.

  The feature extraction network 5a and the blood pressure estimation network 5b repeat training while adjusting the network weight until the loss function L (ps, pd, qs, qd) is minimized. Note that ps and pd are probability vectors indicating actual systolic blood pressure values and diastolic blood pressure values with respect to the training data Ut. That is, for a person who has acquired training data Ut, actual systolic blood pressure values and diastolic blood pressure values are acquired using a known blood pressure monitor, and probability vectors indicating the values are set as ps and pd. Use it. As a result, the feature extraction network 5a and the blood pressure estimation network 5b can be learned until the difference between the blood pressure value estimated using the feature extraction network 5a and the blood pressure estimation network 5b and the actually measured blood pressure value is minimized. The blood pressure estimation accuracy in the estimation step described later can be improved.

Also, the network weight can be adjusted using any learning method such as, for example, gradient descent, stochastic gradient descent, or simulated annealing. For example, in this embodiment, the loss function L (ps, pd, qs, qd) is defined as in the following equation (5).

  When the training step described above is completed, it is possible to execute an estimation step for estimating the blood pressure of the user based on a signal input from the wearable terminal.

In the estimation step, the control device 102 inputs a signal U′t represented by the following equation (6) input from the wearable terminal to a network constituted by the feature extraction network 5a and the blood pressure estimation network 5b described above.

  As a result, a new segment feature representation z ′ and probability vectors qs ′ and qd ′ for the input signal signal U′t can be obtained. FIG. 6 is a diagram schematically showing the flow from when the input signal signal U′t is input until the probability vectors qs ′ and qd ′ are output.

Each entry of the output vector output by the blood pressure estimation network 5b represents the probability that the input segment corresponds to the systolic blood pressure value and the diastolic blood pressure value, respectively. For this reason, in this Embodiment, the value used as the highest probability is estimated as a systolic blood pressure and a diastolic blood pressure of a measuring subject using following formula (7), (8).

  For example, when a blood pressure value (˜49 mmHG, 50 mmHG—89 mmHG, 90 mmHG˜) is set with respect to the probability vector and qs = (0.3, 0.8, 0.0), the systolic blood pressure value Is estimated to be 50 mmHG to 89 mmHG with the highest probability calculated. When qd = (0.0, 0.44, 0.56), the diastolic blood pressure value is estimated to be 90 mmHG or higher for which the highest probability is calculated. In equations (7) and (8), probability vectors qs and qd are probability vectors representing the probabilities of estimation that the segment has systolic blood pressure value sbp ′ and diastolic blood pressure value dbp ′, respectively.

なお、式（９）において、確率ベクトルｐｓおよびｐｄは、セグメントが収縮期血圧値ｓｂｐおよび拡張期血圧値ｄｂｐをそれぞれ有する既知の確率を表す確率ベクトルである。 In the present embodiment, the following equation (9) is used to convert the blood pressure value (mmHG) into a vector.

In equation (9), probability vectors ps and pd are probability vectors representing known probabilities that a segment has systolic blood pressure value sbp and diastolic blood pressure value dbp, respectively.

  Through the above processing, the blood pressure of the user wearing the wearable terminal can be estimated with high accuracy based on the signal U′t input from the wearable terminal.

  FIG. 7 is a flowchart showing a flow of training steps in the present embodiment. The processing shown in FIG. 7 is executed by the control device 102 as a program to be activated when training data Ut acquired in advance is input.

  In step S10, the control device 102 inputs the input data Ut to the feature extraction network 5a. As a result, the feature extraction network 5a converts the input signal Ut into a feature expression z using the feature mapping F (z || Ut) described above. Then, it progresses to step S20.

  In step S20, the blood pressure estimation network 5b converts the feature expression z into a pair of probability vectors qs and qd using the estimated mapping Q ((qs, qd) || z). Then, it progresses to step S30.

  In step S30, the control device 102 calculates the loss function L (ps, pd, qs, qd) described above. Thereafter, the process proceeds to step S40.

  In step S40, the control device 102 determines whether or not the above-described loss function L has been minimized. When an affirmative determination is made in step S40, the weights at that time are adopted as the feature extraction network 5a and the blood pressure estimation network 5b, and the process is terminated. On the other hand, if a negative determination is made in step S40, the process proceeds to step S50.

  In step S50, as described above, the control apparatus 102 adjusts the weights of the feature extraction network 5a and the blood pressure estimation network 5b, and returns to step S10.

  FIG. 8 is a flowchart showing the flow of the estimation step in the present embodiment. The process shown in FIG. 8 is executed by the control device 102 as a program to be started when the measurement subject's data Ut ′ is input.

  In step S110, the control device 102 inputs the input data Ut ′ to the feature extraction network 5a. Thus, each signal segment of the input signal Ut ′ is converted into a feature representation z by the feature extraction network 5a using the above-described feature mapping F (z || Ut). Then, it progresses to step S120.

  In step S120, the blood pressure estimation network 5b converts the feature expression z into a pair of probability vectors qs ′ and qd ′ using the estimated mapping Q ((qs, qd) || z). Thereafter, the process proceeds to step S130.

  In step S130, as described above, the control device 102 calculates the systolic blood pressure value and the diastolic blood pressure value of the measurement subject from the probability vectors qs ′ and qd. Thereafter, the process proceeds to step S140.

  In step S140, the control apparatus 102 outputs and displays the calculated systolic blood pressure value and diastolic blood pressure value on the display device 104. Thereafter, the process ends.

According to the embodiment described above, the following operational effects can be obtained.
(1) A feature extraction network that converts input pulse wave data into a feature expression, and blood pressure estimation that converts the feature expression converted by the feature extraction network into a probability vector that represents the probabilities of systolic blood pressure values and diastolic blood pressure values The control device 102 executes a training step for causing the feature extraction network and the blood pressure estimation network to learn the systolic blood pressure value and the diastolic blood pressure value according to the pulse wave data, and learning is performed by the training step. The pulse wave data of the measurement subject is input to the completed feature extraction network and blood pressure estimation network, and an estimation step for estimating the systolic blood pressure value and the diastolic blood pressure value of the measurement subject is executed. Thus, the blood pressure of the measurement subject can be measured with high accuracy using the feature extraction network and the blood pressure estimation network that have been trained in advance.

(2) The control device 102 obtains a probability vector representing the probabilities of the measurement subject's systolic blood pressure value and diastolic blood pressure value, and uses the value having the highest probability as the measurement subject's systolic blood pressure value and diastolic phase. The blood pressure value was estimated. Thereby, the measurement accuracy of blood pressure can be improved.

(3) The control device 102 displays the estimated systolic blood pressure value and diastolic blood pressure value on the display device 104. Thereby, an operator such as a doctor can check the blood pressure value of the measurement subject on the display device 104.

(4) In the training step, the feature extraction network 5a and the blood pressure estimation network 5b repeat the training while adjusting the network weight until the loss function L (ps, pd, qs, qd) is minimized. As a result, blood pressure estimation accuracy based on pulse wave data can be improved.

-Modification-
Note that the blood pressure estimation system of the above-described embodiment can be modified as follows.
(1) In the above-described embodiment, the blood pressure estimation device 100 is a personal computer, and the example in which the control device 102 executes the above-described processing has been described. However, the terminal operated by the operator can be connected to the blood pressure estimation apparatus 100 via a communication line such as the Internet, and the pulse wave data from the wearable terminal is input by being transmitted from the operation terminal to the blood pressure estimation apparatus 100. May be. In this case, the blood pressure estimation result may be transmitted to the operation terminal and displayed on the operation terminal. Accordingly, a client server type or cloud type blood pressure estimation system can be constructed by connecting the operation terminal and the blood pressure estimation device 100 via a communication line, while the blood pressure estimation device 100 is used alone. Therefore, it can be used as a stand-alone device.

(2) In the above-described embodiment, the example in which the control device 102 executes the processing for the training step in the blood pressure estimation device 100 has been described. However, the processing for the training step may be executed by another device, and the data obtained by the training step may be recorded in the storage medium 103. In this case, the process for the training step in the blood pressure estimation device 100 is not necessary.

(3) In the embodiment described above, the feature extraction network 5a and the blood pressure estimation network 5b repeat the training while adjusting the network weight until the loss function L (ps, pd, qs, qd) is minimized. explained. However, the feature extraction network 5a and the blood pressure estimation network 5b may repeat training while adjusting the network weight until the loss function L (ps, pd, qs, qd) is equal to or less than a preset threshold value. .

  Note that the present invention is not limited to the configurations in the above-described embodiments as long as the characteristic functions of the present invention are not impaired. Moreover, it is good also as a structure which combined the above-mentioned embodiment and a some modification.

DESCRIPTION OF SYMBOLS 100 Blood pressure estimation apparatus 101 Operation member 102 Control apparatus 103 Storage medium 104 Display apparatus 105 Connection interface

Claims (9)

A blood pressure estimation device that estimates blood pressure using a neural network,
A feature extraction network that converts input pulse wave data into feature representations;
A blood pressure estimation network that converts the feature representation converted by the feature extraction network into a probability vector representing probabilities of systolic blood pressure values and diastolic blood pressure values;
Training step execution means for executing a training step for causing the feature extraction network and the blood pressure estimation network to learn systolic blood pressure values and diastolic blood pressure values according to pulse wave data;
An estimation step of inputting pulse wave data of a measurement subject to the feature extraction network and the blood pressure estimation network that have been learned in the training step, and estimating a systolic blood pressure value and a diastolic blood pressure value of the measurement subject. and a estimation step execution means for executing,
The estimation step executing means obtains a probability vector representing the probabilities of the measurement subject's systolic blood pressure value and diastolic blood pressure value, and uses the value having the highest probability as the measurement subject's systolic blood pressure value and dilation. A blood pressure estimation apparatus characterized by estimating the blood pressure value as a period . The blood pressure estimation device according to claim 1 ,
The blood pressure estimation apparatus further comprising display means for displaying on the display device the systolic blood pressure value and the diastolic blood pressure value of the measurement subject estimated by the estimation step execution means. The blood pressure estimation device according to claim 1 or 2 ,
The blood pressure estimation device, wherein the training step execution means repeats training while adjusting weights of the feature extraction network and the blood pressure estimation network until a loss function is minimized. A blood pressure estimation system that estimates blood pressure using a neural network,
A feature extraction network that converts input pulse wave data into feature representations;
A blood pressure estimation network that converts the feature representation converted by the feature extraction network into a probability vector representing probabilities of systolic blood pressure values and diastolic blood pressure values;
Training step execution means for executing a training step for causing the feature extraction network and the blood pressure estimation network to learn systolic blood pressure values and diastolic blood pressure values according to pulse wave data;
An estimation step of inputting pulse wave data of a measurement subject to the feature extraction network and the blood pressure estimation network that have been learned in the training step, and estimating a systolic blood pressure value and a diastolic blood pressure value of the measurement subject. and a estimation step execution means for executing,
The estimation step executing means obtains a probability vector representing the probabilities of the measurement subject's systolic blood pressure value and diastolic blood pressure value, and uses the value having the highest probability as the measurement subject's systolic blood pressure value and dilation. A blood pressure estimation system characterized in that the blood pressure is estimated as a blood pressure value during the period . The blood pressure estimation system according to claim 4 ,
The blood pressure estimation system further comprising display means for displaying a systolic blood pressure value and a diastolic blood pressure value of the measurement subject estimated by the estimation step executing means on a display device. The blood pressure estimation system according to claim 4 or 5 ,
The blood pressure estimation system, wherein the training step execution means repeats training while adjusting weights of the feature extraction network and the blood pressure estimation network until a loss function is minimized. A feature extraction network that converts input pulse wave data into a feature expression, and a blood pressure estimation network that converts the feature expression converted by the feature extraction network into a probability vector representing the probabilities of systolic blood pressure values and diastolic blood pressure values A blood pressure estimation program for estimating blood pressure using a neural network having
A training step execution procedure for executing a training step for causing the feature extraction network and the blood pressure estimation network to learn systolic blood pressure values and diastolic blood pressure values according to pulse wave data;
An estimation step of inputting pulse wave data of a measurement subject to the feature extraction network and the blood pressure estimation network that have been learned in the training step, and estimating a systolic blood pressure value and a diastolic blood pressure value of the measurement subject. and estimation step execution procedure of an execution cause the computer to execute,
The estimation step execution procedure obtains a probability vector representing the probabilities of the systolic blood pressure value and the diastolic blood pressure value of the measurement subject, and uses the value having the highest probability as the systolic blood pressure value and the dilation of the measurement subject. A blood pressure estimation program characterized in that the blood pressure is estimated as a period blood pressure value . In the blood pressure estimation program according to claim 7 ,
A blood pressure estimation program further comprising a display procedure for displaying a systolic blood pressure value and a diastolic blood pressure value of the measurement subject estimated in the estimation step execution procedure on a display device. In the blood pressure estimation program according to claim 7 or 8 ,
The training step execution procedure repeats training while adjusting weights of the feature extraction network and the blood pressure estimation network until a loss function is minimized.

Images