Blood pressure measurement device, blood pressure measurement method, program

Abstract

This technology provides a way to estimate blood pressure using machine learning models with high accuracy, even with limited data. [Solution] A blood pressure measuring device that estimates blood pressure values ​​using a machine learning model comprises: an acquisition unit that acquires multiple types of waveform data measured from a subject by a sensor; and a blood pressure estimation unit that estimates blood pressure values ​​using the multiple types of waveform data and a machine learning model. The machine learning model comprises: multiple feature extraction layers that individually extract features from each of the multiple types of waveform data; multiple time aggregation layers that aggregate each of the multiple features output from the multiple feature extraction layers in the time direction; a feature aggregation layer that aggregates the multiple outputs from the multiple time aggregation layers into one; and a blood pressure estimation layer that estimates blood pressure values ​​based on the output of the feature aggregation layer.

Description

Technical Field

[0001] The present invention relates to a blood pressure measurement device that estimates blood pressure values using a machine learning model.

Background Art

[0002] Blood pressure measurement devices configured to estimate blood pressure values including diastolic blood pressure (DBP) and systolic blood pressure (SBP) by measuring biological signals with a sensor worn on a subject and inputting the obtained time-series waveform data into a machine learning model are known in the art.

[0003] For example, Non-Patent Document 1 discloses a method for estimating blood pressure values using feature extraction by a 1D-CNN (one-dimensional convolutional neural network ) from two types of waveform data: ECG (Electrocardiogram) and PPG (Photoplethysmogram). Specifically, ECG and PPG are converted into a single piece of waveform data by obtaining the difference between the two types of waveform data, and features of the converted waveform data are captured by CNN and LSTM (Long Short-Term Memory).

[0004] In addition, Patent Document 1 discloses a method for estimating blood pressure values, in which time-series data of pressure pulse waves measured by a cuff pressure sensor is decomposed into a plurality of pieces of waveform data, and a concatenation of the plurality of pieces of waveform data is input to a 1D-CNN to estimate a blood pressure value.

Prior Art Documents

Non-Patent Documents

[0005]

Non-Patent Document 1

[0006] [Patent Document 1] Japanese Patent Publication No. 2024-163785 [Overview of the project] [Problems that the invention aims to solve]

[0007] In the method proposed in Non-Patent Document 1, multiple types of waveform data are combined into one and input into a single CNN to extract features. However, in this process, the characteristics of each waveform data (for example, temporal information such as frequency characteristics, and information about unique waveform shapes) are lost. Therefore, high-precision estimation is not possible with small amounts of data, and a large amount of data is required to improve estimation accuracy. The method proposed in Patent Document 1 also has the same problem as the method proposed in Non-Patent Document 1, as it inputs multiple decomposed waveform data combined into one into a single CNN, and a large amount of data may be required to improve estimation accuracy.

[0008] The present invention aims to provide a technology that enables highly accurate blood pressure estimation using machine learning models, even with limited data. [Means for solving the problem]

[0009] This disclosure relates to a blood pressure measuring device that estimates blood pressure values ​​using a machine learning model, comprising: an acquisition unit that acquires multiple types of waveform data measured from a subject by a sensor; and a blood pressure estimation unit that estimates blood pressure values ​​using the multiple types of waveform data and a machine learning model, wherein the machine learning model comprises multiple feature extraction layers that individually extract features from each of the multiple types of waveform data, and each of the multiple features output from the multiple feature extraction layers The blood pressure measuring device includes a plurality of time aggregation layers that perform aggregation in the time direction, a feature aggregation layer that aggregates the plurality of outputs from the plurality of time aggregation layers into one, and a blood pressure estimation layer that estimates a blood pressure value based on the output of the feature aggregation layer.

[0010] In the machine learning model used in this invention, each waveform data is processed individually by multiple feature extraction layers, thereby appropriately extracting the unique features of each waveform data. Furthermore, since the feature quantities of each waveform data extracted by the multiple feature extraction layers are individually time-aggregated by multiple time-aggregation layers, the dimensionality in the time direction can be reduced while preserving the features of each waveform data. Therefore, it is expected that highly accurate blood pressure value estimation can be achieved from less data compared to conventional techniques.

[0011] The blood pressure measuring device further comprises a cuff for compressing the part of the subject to be measured and a sensor for detecting the pressure of the cuff, and the acquisition unit may acquire the multiple types of waveform data from the output signal of the sensor acquired while gradually pressurizing the cuff. With this configuration, multiple types of waveform data can be acquired with just one set of cuff and sensor, making it possible to estimate blood pressure values ​​with high accuracy while achieving miniaturization and cost reduction of the device.

[0012] Multiple types of waveform data obtained from the sensor's output signal may include pressure pulse waveform data and envelope data of the pressure pulse waveform. Furthermore, cuff pressure waveform data may also be included.

[0013] Each of the feature extraction layers of the machine learning model described above may be composed of a CNN with a different structure from the others. More specifically, each of the feature extraction layers of the machine learning model may be composed of a CNN having a structure corresponding to the waveform characteristics of the input waveform data. By designing the CNNs in advance in this way, it is possible to guide them to obtain the information to be extracted from each waveform data. Thus, the unique features of each waveform data can be appropriately extracted.

[0014] Each of the multiple time-aggregation layers in the machine learning model is preferably configured to aggregate the features input from the corresponding feature extraction layer in the time direction according to their time-varying characteristics. By designing the time-aggregation layers in this way, it is possible to reduce the time-direction dimensionality of each feature while appropriately preserving (or emphasizing) the time-varying characteristics of each waveform data. Each of the multiple time-aggregation layers in the machine learning model may be composed of a pooling layer, an attention mechanism, or an LSTM.

[0015] The blood pressure estimation layer of the machine learning model may be configured to output a single blood pressure value. Using a model with such a configuration, the model's output can be used directly as a blood pressure measurement, thus requiring a simpler processing program. Furthermore, when training the model, only the true blood pressure value needs to be provided as the training signal, which simplifies the preparation of training data and the learning process.

[0016] The blood pressure estimation layer of the machine learning model is configured to output a probability distribution of blood pressure values, and the blood pressure estimation unit may have a determination unit that determines the final blood pressure value to be output based on the probability distribution of blood pressure values ​​output from the machine learning model. By using a model with such a configuration, a probability distribution of blood pressure values ​​is output as the estimation result, so it can be expected that the reliability and accuracy can be improved compared to a model that outputs a single blood pressure value.

[0017] This disclosure relates to a blood pressure measurement method that estimates blood pressure values ​​using a machine learning model, comprising the steps of: acquiring multiple types of waveform data measured from a subject by a sensor; and estimating blood pressure values ​​using the multiple types of waveform data and a machine learning model. The learning model includes a blood pressure measurement method comprising: a plurality of feature extraction layers that individually extract features from each of the plurality of types of waveform data; a plurality of time aggregation layers that aggregate each of the plurality of features output from the plurality of feature extraction layers in the time direction; a feature aggregation layer that aggregates the plurality of outputs from the plurality of time aggregation layers into one; and a blood pressure estimation layer that estimates a blood pressure value based on the output of the feature aggregation layer.

[0018] This disclosure provides a program for causing the processor of a blood pressure measuring device that estimates blood pressure values ​​using a machine learning model to perform the steps of acquiring multiple types of waveform data measured from a subject by a sensor, and estimating blood pressure values ​​using the multiple types of waveform data and a machine learning model, wherein the machine learning model includes a program having a plurality of feature extraction layers that individually extract features from each of the multiple types of waveform data, a plurality of time aggregation layers that aggregate each of the plurality of features output from the plurality of feature extraction layers in the time direction, a feature aggregation layer that aggregates the plurality of outputs from the plurality of time aggregation layers into one, and a blood pressure estimation layer that estimates blood pressure values ​​based on the output of the feature aggregation layer.

[0019] The present invention may be understood as a blood pressure measuring device, blood pressure monitor, biological information acquisition device, etc., having at least a part of the above configuration. Alternatively, the present invention may be understood as a blood pressure measurement method, a blood pressure estimation method, or a program for implementing such a method, or a recording medium on which such a program is stored, including at least a part of the above processing. Furthermore, each of the above means and processing can be combined with each other as much as possible to constitute the present invention. [Effects of the Invention]

[0020] According to the present invention, in blood pressure estimation using a machine learning model, high-accuracy estimation can be performed even from a small amount of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] [Figure 1] FIG. 1 is a diagram explaining the basic configuration of a blood pressure measurement device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a diagram schematically showing an example of the hardware configuration of the blood pressure measurement device. [Figure 3] FIG. 3 is a block diagram showing an example of the functional configuration of the blood pressure measurement device. [Figure 4] FIG. 4 is a diagram showing an example of waveform data. [Figure 5] FIG. 5 is a flowchart showing the overall processing procedure of the blood pressure measurement method. [Figure 6] FIG. 6 is a schematic diagram for explaining learning processing of a machine learning model used for blood pressure estimation processing. [Figure 7] FIG. 7 is a diagram showing the structure of a machine learning model used in a second embodiment. [Figure 8] FIG. 8 is a diagram showing the configuration of a blood pressure estimation unit in a third embodiment. [Figure 9] FIG. 9 is a diagram showing a configuration example of a blood pressure estimation layer and a determination unit in the third embodiment. MODE FOR CARRYING OUT THE INVENTION

[0022] <Characteristics of Blood Pressure Measurement Device> Referring to FIG. 1, the basic configuration of a blood pressure measurement device according to an embodiment of the present invention will be described. As a configuration for blood pressure estimation, the blood pressure measurement device 1 includes: a waveform acquisition unit 20 that acquires a plurality of types of waveform data measured from a subject by a sensor; and a blood pressure estimation unit 30 that estimates a blood pressure value using the plurality of types of waveform data Da, Db, ···, Dx and a machine learning model M.

[0023] The machine learning model M consists of multiple feature extraction layers M1a, M1b, ..., M1x, multiple time-intensive layers M2a, M2b, ..., M2x, a feature-intensive layer M3, and a blood pressure estimation layer M4. The feature extraction layers M1a, M1b, ..., M1x and the time-intensive layers M2a, M2b, ..., M2x are provided for each waveform data Da, Db, ..., Dx.

[0024] When multiple types of waveform data Da, Db, ..., Dx acquired by the waveform acquisition unit 20 are input to the machine learning model M, multiple feature extraction layers M1a, M1b, ..., M1x individually extract feature quantities F1a, F1b, ..., F1x from each waveform data Da, Db, ..., Dx. Each feature quantity F1a, F1b, ..., F1x is, for example, data with a time dimension on the order of tens to thousands and a feature dimension on the order of tens to hundreds. The feature quantities F1a, F1b, ..., F1x output from the feature extraction layers M1a, M1b, ..., M1x are input to the time aggregation layers M2a, M2b, ..., M2x, respectively. The time-aggregation layers M2a, M2b, ..., M2x perform time-aggregation on each of the feature vectors F1a, F1b, ..., F1x. For example, the number of dimensions in the time direction is reduced to a few to a dozen or so dimensions. The time-aggregated feature vectors F2a, F2b, ..., F2x output from the time-aggregation layers M2a, M2b, ..., M2x are input to the feature-aggregation layer M3 and aggregated into one. The output F3 from the feature-aggregation layer M3 is then input to the blood pressure estimation layer M4, and the estimated blood pressure value is output from the machine learning model M. The output blood pressure value should preferably include DBP (diastolic blood pressure) and SBP (systolic blood pressure).

[0025] The multiple types of waveform data Da, Db, ..., Dx obtained through measurement generally have different waveform characteristics and time-varying characteristics (characteristics related to changes over time), and the features of the waveform data that should be focused on to obtain information effective for blood pressure estimation also differ. Therefore, if multiple types of waveform data are combined into one and input into a CNN (Convolutional Neural Network) as in conventional techniques, there is a possibility that the unique features of each waveform data may be lost. In contrast, the machine learning model M used in this embodiment can appropriately extract the unique features of each waveform data by processing each waveform data Da, Db, ..., Dx individually with multiple independent feature extraction layers M1a, M1b, ..., M1x. Furthermore, since the feature quantities F1a, F1b, ..., F1x of each waveform data extracted by multiple feature extraction layers M1a, M1b, ..., M1x are individually time-aggregated by multiple independent time-aggregation layers M2a, M2b, ..., M2x, it is possible to reduce the temporal dimensionality while preserving the features of each waveform data Da, Db, ..., Dx. In this way, by providing individual time-aggregation layers M2a, M2b, ..., M2x before the feature-aggregation layer M3, the temporal features of each waveform data Da, Db, ..., Dx can be appropriately transmitted to the feature-aggregation layer M3 and, consequently, to the blood pressure estimation layer M4, it is expected that highly accurate blood pressure estimation can be achieved even with a small amount of data.

[0026] <First Embodiment> (Configuration of a blood pressure measurement device) Figure 2 schematically shows an example of the hardware configuration of the blood pressure measuring device 1 according to this embodiment. The blood pressure measuring device 1 according to this embodiment includes a control unit 10, a storage unit 11, an oscillation circuit 121, a pump drive circuit 122, a valve drive circuit 123, a cuff 130, a pressure sensor 131, a pressurizing pump 132, an exhaust valve 133, an air tube 134, an operation switch 14, a display unit 15, and a power supply 16. The control unit 10 and the storage unit 11 are examples of the computer portion of the blood pressure measuring device 1.

[0027] The cuff 130 includes an air bag 1301 that contains air. The cuff 130 is equipped with a pressure sensor 131, a pressurizing pump 132, and an exhaust valve 133 via an air tube 134. The pressure sensor 131 is configured to detect the pressure (cuff pressure) inside the air bag 1301 of the cuff 130. The pressurizing pump 132 is configured to supply air into the air bag 1301. The exhaust valve 133 is provided as a boundary between the internal and external spaces of the air bag 1301 and is configured to be openable and closable. By closing the exhaust valve 133, air can be trapped inside the air bag 1301 and the pressure inside the air bag 1301 can be maintained. On the other hand, by opening the exhaust valve 133, air can be released from the air bag 1301 and the pressure can be reduced.

[0028] The control unit 10 includes a hardware processor such as a CPU, and is responsible for programs and various data The system is configured to perform information processing based on the data. The control unit 10 (CPU) is an example of the processor resources of the estimation device. The oscillation circuit 121 outputs a signal with an oscillation frequency corresponding to the output value of the pressure sensor 131 to the control unit 10. In this embodiment, the control unit 10 processes the signal output from the oscillation circuit 121 to obtain cuff pressure waveform and pressure pulse waveform data. The pump drive circuit 122 is configured to control the drive of the pressurizing pump 132 based on a control signal output from the control unit 10. The valve drive circuit 123 is configured to control the opening and closing of the exhaust valve 133 based on a control signal output from the control unit 10.

[0029] The storage unit 11 may be composed of, for example, a semiconductor memory. The storage unit 11 is an example of a memory resource of the estimation device. In this embodiment, the storage unit 11 stores various information such as the program 110 and model data 111. The program 110 includes a program that causes the CPU of the blood pressure measuring device 1 to execute information processing such as cuff pressurization control processing, pulse wave acquisition processing, blood pressure value estimation processing, and measurement result display processing. The program 110 includes a series of instructions for said information processing. Note that each processing may be composed of a different program module, or all processing may be executed with a single program. The model data 111 is a machine learning model. The model data 111 may be incorporated into the program 110. The storage unit 11 may appropriately hold other information such as the results of blood pressure measurement (cuff pressure, pulse wave data, estimated blood pressure, etc.).

[0030] The operation switch 14 is used to perform operations such as starting blood pressure measurement. The operation switch 14 may consist of at least one of a physical switch and a virtual switch. The display unit 15 is configured to display various information such as the blood pressure measurement results. The operation switch 14 and the display unit 15 may be integrally configured as a touch panel display. The power supply 16 is configured to supply power to each part, such as the control unit 10.

[0031] Regarding the specific hardware configuration of the blood pressure measurement device 1, components can be omitted, replaced, and added as appropriate depending on the embodiment. For example, the control unit 10 may include multiple hardware processors. Hardware processors may consist of microprocessors, FPGAs, DSPs, ASICs, etc. The blood pressure measurement device 1 may be equipped with a communication interface. For example, the blood pressure measurement device 1 may be configured to communicate data with a user terminal such as a smartphone by including a proximity wireless communication module. This allows the blood pressure measurement device 1 to acquire arbitrary data via the user terminal. At least one of the program 110 and model data 111 may be stored on a storage medium of an external computer such as a NAS. The blood pressure measurement device 1 may acquire at least one of the program 110 and model data 111 from an external computer. The blood pressure measurement device 1 may acquire data directly from an external computer, or it may acquire data indirectly from an external computer via a user terminal. The blood pressure measurement device 1 may be configured as a general blood pressure monitor, or as a wearable device such as a wristwatch.

[0032] Figure 3 is a block diagram showing an example of the functional configuration of the blood pressure measuring device 1 according to this embodiment. The same reference numerals are used for the same functional blocks as shown in Figure 1.

[0033] The blood pressure measuring device 1 has, as its main functions related to blood pressure measurement, a waveform acquisition unit 20, a blood pressure estimation unit 30, a pressurization control unit 40, and an information output unit 50. These functions are realized by the CPU of the control unit 10 reading and executing a program stored in the memory unit 11.

[0034] The waveform acquisition unit 20 receives the signal from the oscillation circuit 121. The waveform acquisition unit 20 detects the oscillation frequency of the signal input from the oscillation circuit 121 and converts the detected oscillation frequency into a pressure value signal. Convert. The waveform acquisition unit 20 includes an HPF (High Pass Filter) unit 21 that extracts and outputs a pressure pulse wave signal by processing the pressure value signal with an HPF (High Pass Filter), an LPF (Low Pass Filter) unit 22 that extracts and outputs a cuff pressure signal by processing the pressure value signal with an LPF (Low Pass Filter), and the envelope of the pressure pulse waveform. The system includes an envelope acquisition unit 23 that calculates and outputs the envelope. The time-series data of the pressure pulse waveform output from the HPF unit 21, the time-series data of the cuff pressure waveform output from the LPF unit 22, and the envelope data output from the envelope acquisition unit 23 are stored in the memory of the control unit 10. The sampling rate of the pressure pulse waveform and cuff pressure waveform is, for example, about 250 Hz, and the sampling rate of the envelope may be slower.

[0035] Figure 4 shows an example of waveform data. The top graph shows the pressure value signal. The horizontal axis is time, and the vertical axis is pressure, showing the output of the pressure sensor 131 when the cuff pressure is linearly increased. The pressure pulse wave signal is superimposed on the pressure value signal. The second graph from the top shows the cuff pressure waveform data extracted by the LPF unit 22, and the third graph shows the pressure pulse waveform data extracted by the HPF unit 21. The bottom graph is an example of envelope data obtained by the envelope acquisition unit 23. As the cuff pressure is increased, the amplitude of the pressure pulse waveform gradually increases, reaching a maximum value around the midpoint between DBP (diastolic blood pressure) and SBP (systolic blood pressure), and then gradually decreasing.

[0036] The pressure pulse waveform is data that reveals characteristics related to the pulse, and for example, the waveform shape, frequency amplitude, and phase characteristics for each beat are considered useful for estimating blood pressure values. The envelope of the pressure pulse waveform is data that reveals characteristics related to the characteristics and condition of the blood vessel, and for example, inflection points, rise speed, and fall speed are considered useful for estimating blood pressure values. The cuff pressure waveform is data that reveals characteristics related to the characteristics and condition of the cuff, and for example, the rise and slope (pressure rate) are considered useful for estimating blood pressure values.

[0037] The pressurization control unit 40 has the function of starting, stopping, and depressurizing the cuff 130 by sending control signals to the pump drive circuit 122 and the valve drive circuit 123 and controlling the operation of the pressurizing pump 132 and the exhaust valve 133.

[0038] The blood pressure estimation unit 30 estimates blood pressure values ​​using two types of waveform data acquired by the waveform acquisition unit 20: pressure pulse waveform data Da and envelope data Db, and a machine learning model M. The machine learning model M is composed of a feature extraction layer M1a and a time aggregation layer M2a for the pressure pulse waveform data Da, a feature extraction layer M1b and a time aggregation layer M2b for the envelope data Db, a feature aggregation layer M3, and a blood pressure estimation layer M4.

[0039] Feature extraction layers M1a and M1b each have independent Convolutional Neural Networks (CNNs). The CNNs in feature extraction layer M1a and M1b may have the same structure, but it is preferable that they have different structures. Specifically, the CNN in feature extraction layer M1a is preferably designed to have a structure that corresponds to the waveform characteristics (sampling interval, frequency characteristics, etc.) of the pulse waveform data Da so that desired information (information effective for estimating blood pressure values) can be obtained from the pulse waveform data Da. On the other hand, the CNN in feature extraction layer M1b is preferably designed to have a structure that corresponds to the waveform characteristics (sampling interval, frequency characteristics, etc.) of the envelope data Db so that desired information (information effective for estimating blood pressure values) can be obtained from the envelope data Db. For example, a downsampling layer may be added to the CNN in feature extraction layer M1a to match the sampling intervals of the pulse waveform data Da and the envelope data Db. Furthermore, the kernel size and number of filters of the CNN may be appropriately set according to the frequency characteristics of the waveform data (for example, whether there are many low-frequency components or many high-frequency components). In addition, an attention mechanism may be provided depending on the parts of the waveform data that deserve attention. In this way, by constructing the feature extraction layer M1a and feature extraction layer M1b as independent CNNs, the unique characteristics of the pulse waveform data Da and the envelope data Db can be determined. This allows for the appropriate extraction of specific features. Furthermore, since the input to the CNN in feature extraction layer M1a is limited to the pulse waveform data Da, and data with other waveform characteristics (i.e., data that could be considered noise) is not mixed in, the estimation accuracy and generalization ability of the CNN can be improved. For similar reasons, the estimation accuracy and generalization ability of the CNN in feature extraction layer M1b can also be improved.

[0040] The time aggregation layer M2a is configured to aggregate the feature quantity F1a input from the corresponding feature extraction layer M1a in the time direction according to its time-varying characteristics. For example, since feature quantity F1a is extracted from pulse waveform data Da, it is expected to include features related to pulsation such as waveform shape, amplitude, and phase for each beat. Therefore, it is desirable that time aggregation be performed in a way that extracts (emphasizes) the change points of these features. The time aggregation layer M2b is configured to aggregate the feature quantity F1b input from the corresponding feature extraction layer M1b in the time direction according to its time-varying characteristics. For example, since feature quantity F1b is extracted from envelope data Db, it is expected to include features such as inflection points, rise speed, and fall speed as time changes. Therefore, it is desirable that time aggregation be performed in a way that extracts (emphasizes) these features. By designing the time aggregation layers M2a and M2b in this way, it is possible to reduce the time-direction dimensionality of each feature quantity while appropriately preserving (or emphasizing) the time-varying characteristics of each waveform data. Each of the time-intensive layers M2a and M2b may be composed of a pooling layer, an attention mechanism, or an LSTM.

[0041] The feature aggregation layer M3 aggregates multiple outputs F2a and F2b from the time aggregation layers M2a and M2b into one. The feature aggregation layer M3 may be composed of models such as an attention mechanism, a Transformer, or an MLP-Mixer. By providing such a feature aggregation layer M3, it becomes possible to aggregate features while appropriately preserving features based on the waveform characteristics and time-varying characteristics of each waveform data Da and Db. Although not shown in the diagram, additional information other than the feature quantities F2a and F2b derived from the waveform data Da and Db may be provided as input data to the feature aggregation layer M3, and the feature quantities F2a and F2b and the additional information may be aggregated into one. As additional information, for example, the attribute information of the subject (gender, height, weight, age, race, lifestyle, etc.), measurement time, and measured body part may be used.

[0042] The blood pressure estimation layer M4 estimates blood pressure values ​​based on the output F3 of the feature aggregation layer M3. In this embodiment, the blood pressure estimation layer M4 is composed of fully connected layers and outputs a single blood pressure value. With this structure, the output of the machine learning model M can be used directly as the measured value (estimated value) of blood pressure, so a simple processing program is sufficient. Furthermore, when training the model M using machine learning, only the true value of the blood pressure needs to be provided as the training signal, which has the advantage of simplifying the preparation of training data and the learning process.

[0043] The information output unit 50 has the function of outputting information such as data acquired by the waveform acquisition unit 20 and blood pressure values ​​acquired by the blood pressure estimation unit 30 to the display unit 15.

[0044] In this embodiment, the functions shown in Figure 3 are implemented by a software program, but all or part of these functions may be replaced by circuits such as ASICs or FPGAs. Alternatively, the functions shown in Figure 3 may be implemented through collaboration with other computers connected via a network (such as a cloud server, a user's PC, or a smartphone).

[0045] (Operation of the blood pressure measuring device) Figure 5 is a flowchart showing the overall processing steps for the blood pressure measurement method using the blood pressure measurement device 1. The processing shown in Figure 5 is achieved when the CPU of the control unit 10 reads the program 110 from the storage unit 11 and executes it.

[0046] When measuring blood pressure, the subject will have a cuff 130 wrapped around the area to be measured beforehand. Now, we will explain an example where the upper arm is used as the measurement site, but the measurement site is not limited to the upper arm; it may also be the wrist or other parts of the body. Furthermore, we will explain assuming that the person being measured has made the necessary settings using the operation switch 14 and has given the instruction to start blood pressure measurement. In addition, when the blood pressure measuring device 1 receives the instruction to start blood pressure measurement, it performs the necessary initialization, such as opening the exhaust valve 133 and setting the cuff pressure to atmospheric pressure (initial pressure).

[0047] When blood pressure measurement is started, the pressurization control unit 40 starts pressurization control to pressurize the cuff 130 (step S50). During pressurization control, the cuff pressure is increased at a constant rate.

[0048] The pressure pulse wave and cuff pressure are acquired by the waveform acquisition unit 20 while gradually increasing the cuff pressure (step S51). The pressurization control unit 40 monitors the cuff pressure, and if the cuff pressure reaches a predetermined target value (Y in step S52), it stops pressurizing the cuff 130 at that point (step S53).

[0049] In step S54, the envelope acquisition unit 23 acquires the envelope of the pulse waveform acquired up to that point. Known methods can be used to generate the envelope. For example, the envelope may be obtained by simply connecting the peak positions of the pulses of each heartbeat, or the envelope may be obtained by fitting a curve to the sequence of peak positions of the pulses of each heartbeat.

[0050] In step S55, the blood pressure estimation unit 30 inputs the pressure pulse waveform data and envelope data into the machine learning model M to estimate the blood pressure value. The results obtained from the above measurement process (blood pressure value, heart rate, waveform data, etc.) are displayed on the display unit 15 (step S56).

[0051] (Training of machine learning model M) Figure 6 is a schematic diagram illustrating the training process of the machine learning model M used for blood pressure estimation. This training process can be performed by the control unit 10 of the blood pressure measurement device 1, but typically it is performed by a separate training device. The training device consists of a general-purpose computer equipped with a CPU, GPU, memory, storage, etc., and the training process described below is realized by loading the training program stored in storage into memory and executing the program using the CPU and GPU.

[0052] First, the learning device acquires training data, specifically pulse waveform data and blood pressure values ​​(true values ​​as the teacher signal) (step S60). At this time, it is advisable to prepare a large amount of pulse waveform data and blood pressure values ​​obtained from a large number of subjects.

[0053] The learning device first selects one pulse waveform data point and obtains the envelope from that pulse waveform data (step S61). Then, the pulse waveform data and the envelope data are input into the machine learning model M being trained to obtain an estimated blood pressure (step S62).

[0054] The learning device calculates the difference between the estimated blood pressure obtained in step S62 and the true value obtained as the training signal (step S63), and uses this difference (error) to modify the machine learning model M (step S64). As a result, the parameters of the feature extraction layers M1a, M1b, time aggregation layers M2a, M2b, feature aggregation layer M3, and blood pressure estimation layer M4 are updated.

[0055] The same process is repeated for each of the numerous pulse wave data points provided as training data. This allows the system to learn the correlation between pulse wave data, envelope data, and blood pressure values.

[0056] <Second Embodiment> In the first embodiment, blood pressure values ​​are obtained from two types of waveform data: pulse wave data and envelope data. In the first embodiment, blood pressure values ​​are estimated from three types of waveform data, including cuff pressure waveform data. The other configurations may be the same as those of the first embodiment.

[0057] Figure 7 shows the machine learning model M used in the second embodiment. The machine learning model M of this embodiment is configured to include a feature extraction layer M1a and a time aggregation layer M2a for pressure pulse waveform data Da, a feature extraction layer M1b and a time aggregation layer M2b for envelope data Db, a feature extraction layer M1c and a time aggregation layer M2c for cuff pressure waveform data Dc, a feature aggregation layer M3, and a blood pressure estimation layer M4.

[0058] The CNN in the feature extraction layer M1c should be designed to have a structure that corresponds to the waveform characteristics (sampling interval, frequency characteristics, etc.) of the cuff pressure waveform data Dc, so that desired information (information effective for estimating blood pressure values) can be obtained from the cuff pressure waveform data Dc. Furthermore, the time aggregation layer M2c should be configured to perform time-direction aggregation according to the time-varying characteristics of the feature quantities F1c extracted from the cuff pressure waveform data Dc.

[0059] The other configurations may be the same as those of the first embodiment. The configuration of the second embodiment also allows for highly accurate estimation of blood pressure values ​​from a small amount of data.

[0060] <Third Embodiment> In the first and second embodiments, the blood pressure estimation layer M4 was designed so that blood pressure values ​​were output from the machine learning model M. However, in the third embodiment, the blood pressure estimation layer M4 is designed so that a probability distribution of blood pressure values ​​is output from the machine learning model M. The other configurations may be the same as those of the first or second embodiment.

[0061] Figure 8 shows the configuration of the blood pressure estimation unit 30 in the third embodiment. The blood pressure estimation unit 30 in this embodiment includes a machine learning model M and a determination unit 31. The blood pressure estimation layer M4 of the machine learning model M is configured to output a probability distribution PD of blood pressure values, and the determination unit 31 determines the final blood pressure value to be output based on the probability distribution PD of blood pressure values.

[0062] Figure 9 shows an example configuration of the blood pressure estimation layer M4 and the determination unit 31. The blood pressure estimation layer M4 is composed of a CNN that performs multi-class classification, and multiple output nodes correspond to discretized blood pressure values. For example, blood pressure values ​​within a predetermined range are discretized in 0.5 mmHg increments to prepare k blood pressure value classes. The blood pressure estimation layer M4 then outputs the probabilities p1, p2, p3, ..., pk of the input data F3 being classified into each blood pressure value class from each output node. This output PD=[p1,p2,p3,...,pk] of the blood pressure estimation layer M4 can be considered as a pseudo-probability distribution of blood pressure values. The determination unit 31 calculates the expected value of the blood pressure value from the probability distribution PD and outputs the calculation result as the finally estimated blood pressure value.

[0063] When training the machine learning model M of this embodiment, it is necessary to provide a probability distribution of blood pressure values ​​as a training signal. For example, assuming that the probability distribution follows a Gaussian distribution, the probability distribution as a training signal may be defined by fitting the "true value of blood pressure" to the "mean" of the Gaussian distribution and the "variance of the true value of blood pressure in the entire training data" to the "variance" of the Gaussian distribution.

[0064] According to the configuration of this embodiment, since the probability distribution of blood pressure values ​​is output as the estimation result, it can be expected that reliability and accuracy can be further improved compared to the method of outputting a single blood pressure value as in the first and second embodiments.

[0065] <Other> The above embodiments are merely illustrative examples illustrating the configuration of the present invention. The specific form is not limited, and various modifications are possible within the scope of the technical concept. For example, in the above embodiment, the blood pressure estimation process was performed after the pressurization was stopped, but the blood pressure estimation may be performed at predetermined cycles during the process of pressurizing the cuff. Also, in the above embodiment, the pulse waveform and its envelope waveform data obtained from the output signal of the cuff pressure sensor were used, but the waveform data used for the blood pressure estimation process can be any type of waveform data, as long as it is measured from the subject by various sensors and contains some correlation with the blood pressure value. For example, ECG or PPG may be used, or waveform data of body movement measured by an acceleration sensor may be used. [Explanation of Symbols]

[0066] 1: Blood pressure measuring device 20: Waveform acquisition section 30: Blood pressure estimation unit M: Machine learning model M1a, M1b, M1c, M1x: Feature extraction layer M2a, M2b, M2c, M2x: Time-intensive layer M3: Feature aggregation layer M4: Blood pressure estimation layer 31: Decision-making section

Claims

1. In a blood pressure measurement device that estimates blood pressure values ​​using a machine learning model, An acquisition unit that acquires multiple types of waveform data measured from a subject by a sensor, It includes a blood pressure estimation unit that estimates blood pressure values ​​using the aforementioned multiple types of waveform data and a machine learning model, The aforementioned machine learning model, Multiple feature extraction layers that individually extract features from each of the aforementioned multiple types of waveform data, Multiple time aggregation layers perform time aggregation on each of the multiple feature quantities output from the multiple feature extraction layers, A feature aggregation layer that aggregates multiple outputs from the multiple time aggregation layers into one, A blood pressure estimation layer that estimates blood pressure values ​​based on the output of the feature aggregation layer, has Blood pressure measuring device.

2. A cuff for compressing the part of the subject to be measured, The system further includes a sensor for detecting the pressure of the cuff, The acquisition unit acquires the multiple types of waveform data from the output signal of the sensor acquired while gradually pressurizing the cuff. The blood pressure measuring device according to claim 1.

3. The aforementioned multiple types of waveform data include pressure pulse waveform data and envelope data of the pressure pulse waveform. The blood pressure measuring device according to claim 2.

4. The aforementioned multiple types of waveform data include pressure pulse waveform data, envelope data of the pressure pulse waveform, and cuff pressure waveform data. The blood pressure measuring device according to claim 2.

5. Each of the feature extraction layers of the aforementioned machine learning model is composed of CNNs with different structures. The blood pressure measuring device according to claim 1.

6. Each of the feature extraction layers of the aforementioned machine learning model is composed of a CNN having a structure corresponding to the waveform characteristics of the input waveform data. The blood pressure measuring device according to claim 1.

7. Each of the multiple time-aggregation layers in the machine learning model is configured to aggregate the features input from the corresponding feature extraction layer in the time direction according to their time-varying characteristics. The blood pressure measuring device according to claim 1.

8. Each of the multiple time-intensive layers in the aforementioned machine learning model is composed of a pooling layer, an attention mechanism, or an LSTM. The blood pressure measuring device according to claim 1.

9. The blood pressure estimation layer of the machine learning model is configured to output a single blood pressure value. The blood pressure measuring device according to claim 1.

10. The blood pressure estimation layer of the machine learning model is configured to output a probability distribution of blood pressure values. The blood pressure estimation unit includes a determination unit that determines the final blood pressure value to be output based on the probability distribution of blood pressure values ​​output from the machine learning model. The blood pressure measuring device according to claim 1.

11. In a blood pressure measurement method that estimates blood pressure values ​​using a machine learning model, The steps include acquiring multiple types of waveform data measured from the subject by a sensor, The process includes the step of estimating blood pressure values ​​using the aforementioned multiple types of waveform data and a machine learning model. The aforementioned machine learning model, Multiple feature extraction layers that individually extract features from each of the aforementioned multiple types of waveform data, Multiple time aggregation layers perform time aggregation on each of the multiple feature quantities output from the multiple feature extraction layers, A feature aggregation layer that aggregates multiple outputs from the multiple time aggregation layers into one, A blood pressure estimation layer that estimates blood pressure values ​​based on the output of the feature aggregation layer, has Blood pressure measurement method.

12. The processor of a blood pressure measuring device that estimates blood pressure values ​​using a machine learning model, The steps include acquiring multiple types of waveform data measured from the subject by a sensor, A program for performing the steps of estimating blood pressure values ​​using the aforementioned multiple types of waveform data and a machine learning model, The aforementioned machine learning model, Multiple feature extraction layers that individually extract features from each of the aforementioned multiple types of waveform data, Multiple time aggregation layers perform time aggregation on each of the multiple feature quantities output from the multiple feature extraction layers, A feature aggregation layer that aggregates multiple outputs from the multiple time aggregation layers into one, A blood pressure estimation layer that estimates blood pressure values ​​based on the output of the feature aggregation layer, A program that has a name.

Images