WO2021114134A1

WO2021114134A1 - Method for blood pressure estimation

Info

Publication number: WO2021114134A1
Application number: PCT/CN2019/124557
Authority: WO
Inventors: Sato Shohei
Original assignee: Huawei Technologies Co., Ltd.
Priority date: 2019-12-11
Filing date: 2019-12-11
Publication date: 2021-06-17
Also published as: JP2022516820A

Abstract

A blood pressure (BP) estimation technique using machine learning, providing a method for blood pressure estimation,wherein the method includes obtaining a baseline BP value (13a) of a subject (S11), obtaining a photoplethysmogram (PPG) signal related to the subject (S12), and estimating a BP value of the subject by using the baseline BP value (13a), the PPG signal, and an estimation model (13b) which is configured to estimate a difference between the BP value and the baseline BP value (13a) based on the PPG signal (S14).

Description

METHOD FOR BLOOD PRESSURE ESTIMATION

TECHNICAL FIELD

The present disclosure relates to blood pressure (BP) estimation technique using machine learning, specifically to a method for BP estimation using photoplethysmogram (PPG) signals, and an apparatus such as a cuff-less BP monitor, a medical device, a personal BP measurement device, a mobile device such as a mobile phone, a smart phone, a communication terminal, a tablet device, a personal computer, a smart watch, a health care device, or the like.

BACKGROUND

Blood pressure (BP) is one of fundamental vital signs and is used to predict cardiovascular diseases such as atherosclerosis, ischemic disturbance, hypertension, arteritis, myocardial infarction, stroke, or the like. Frequent BP measurement is important for early diagnosis of such cardiovascular diseases. In order to facilitate the frequent BP measurement, cuff-less BP measurement using a photoplethysmogram (PPG) sensor is developed in place of traditional BP measurement requiring cuff's inflation and deflation which may cause an uncomfortable feeling for a subject. In addition, suggested is a new PPG measuring technique such as smartphone PPG using a LED (Light Emitting Diode) flush and a camera equipped within a smartphone, a method measuring pulse waves from a video image, or the like. Those cuff-less BP measuring techniques make it possible to achieve continuous BP monitoring.

The cuff-less BP measurement uses BP estimation based on PPG signals obtained by the PPG sensor. In regard to the BP estimation, for example, there exist pulse transition time (PPT) based approach and pulse waveform analysis (PWA) based approach. The PTT based approach considers that a pulse wave valocity (PWV) in an artery is proportional to stiffness of its vessel, and conducts the BP estimation based on the PWV which is represented as a pulse wave distance divided by the PTT. The principle used in the PTT based approach may be expressed by "Moens–Korteweg equation" . On the other hand, the PWA based approach uses "Windkessel model" which describes relationship between dilatation of elastic arteries such as an aorta and mean arterial pressure (MAP) by a differential equation. The relationship may be replaced by relationship between peripheral blood flow volumes and peripheral arterial pressures by analogy. The PPT based approach and the PWA based approach may be used in combination.

The BP may be defined by the product of total peripheral resistance (TPR) and cardiac output which is determined by a heart rate (HR) and a stroke volume. The BP fluctuation is configured by a long-term variations for example by age or season and a short-term fluctuation component. The short-term fluctuation component is affected by various factors such as sympathetic/parasympathetic nervous system, baroreceptor, or the like. Further, the BP may be affected by injection of epinephrine, postural change from lying to standing, white-coat hypertension, masked hypertension, and so on. These short-term effects should be eliminated to achieve adequate accuracy of the BP estimation for the assessment or management of hypertension.

In order to improve the accuracy of the BP estimation, several BP estimation techniques using machine learning are suggested in recent years. For example, M. Liu et al., "Cuffless Blood Pressure Estimation Based on Photoplethysmogramy Signal and Its Second Derivative" , International Journal of Computer Theory and Engineering, Vol. 9, No. 3, June 2017, 202-206 (hereinafter, Liu) teaches the PTT based approach using a Support Vector Regression (SVR) and describes that the SVR shows a better performance than Linear Regression and Neural Networks in the BP estimation. Y. Liang et al., "Hypertension Assessment via ECG and PPG Signals: An Evaluation Using MIMIC Database" , diagnostics 2018, 8, 65 (hereinafter, Liang) teaches the BP estimation using Logistic Regression, AdaBoost Tree, Bagged Tree and K Nearest Neighbors. S. Shimazaki et al., "Features Extraction for Cuffless Blood Pressure Estimation by Autoencoder from Photoplethysmogramy" , Conf. Proc., IEEE Engineering in Medicine and Biology Society, July, 2018, 2857-2860 (hereinafter, Shimazaki) teaches the BP estimation using machine learning based on features extracted by using Autoencoder. M. Kachuee et al., "Cuffless Blood Pressure Estimation Algorithms for Continuous Health-Care Monitoring" , IEEE transactions of Biomedical Engineering, Vol. 64, No. 5, April 2017 (hereinafter, Kachuee) teaches the BP estimation using Regularized Linear Regression, Decision Tree Regression, SVM, AdaBoost, and Random Forest Regression.

The BP estimation using machine learning based on optimal extracted features may reduce the above-mentioned short-term effects and improve the estimation accuracy of the BP. However, the accuracy of the existing BP estimation techniques is deteriorated due to the intrinsic differences in blood pressure that vary from person to person. Such intrinsic differences affect the common relationship between the BP and features and its effect is not negligible for improving the accuracy of BP estimation. In this regard, for example, baseline BP which may be defined by an 24 hours average of the BP or the lowest value of ambulatory blood pressure monitoring (ABPM) , gradually increases by age and/or a state of the subject because of increasing vascular resistance, arterial stiffness, arteriolar remodeling, resetting of pressure natriuretic relationship, cardiovascular diseases, and so on. Therefore, differences of the baseline BP may cause to reduce the accuracy of the BP estimation using the machine learning.

SUMMARY

Embodiments provide a method for BP estimation, and an apparatus having a processor and a PPG sensor and is used for BP estimation. The apparatus may be a cuff-less BP monitor, a medical device, a personal BP measurement device, or the like. The apparatus may also be a cuff-less BP monitor, a medical device, a personal BP measurement device, a mobile device such as a mobile phone, a smart phone, a communication terminal, a tablet device, a personal computer, a wearable device such as a smart watch, a health care device, or the like.

A first aspect of the embodiments provides a method for blood pressure estimation. In a first possible implementation form of the first aspect, the method comprising: obtaining a baseline blood pressure (BP) value of a subject; obtaining a photoplethysmogram (PPG) signal related to the subject; and estimating a BP value of the subject by using the baseline BP value, the PPG signal, and an estimation model which is configured to estimate a difference between the BP value and the baseline BP value based on the PPG signal.

According to the first possible implementation form of the first aspect, the estimation model estimates the difference between the BP value and the baseline BP value, and the baseline BP value of the subject is considered when estimating the BP value of the subject. Characteristics of the subject described by the baseline BP value is reflected on the estimated BP value, so that the accuracy of the BP estimation may be improved.

A second possible implementation form of the first aspect provides: the method according to the first possible implementation form of the first aspect, where the estimation model is built by machine learning using a training dataset which includes first data indicating a difference between each of given BP values and a given baseline BP value related to the given BP values.

According to the second possible implementation form of the first aspect, the training dataset consists of data indicating the difference between the given BP value and the corresponding given baseline BP value. This may eliminate at least a part of individuality of a person from the training dataset used for building the estimation model, and the improved accuracy of the BP estimation by the estimation model may be achieved.

A third possible implementation form of the first aspect provides: the method according to the second possible implementation form of the first aspect, where the training dataset further includes second data indicating at least one feature extracted from a given PPG signal related to the given BP values.

According to the third possible implementation form of the first aspect, the estimation model may be built based on a combination of the at least one feature extracted from the given PPG signal, the corresponding given BP values, and the corresponding difference between the given BP value and the given baseline BP value. The at least one feature may be at least one of a SDPTG (second derivative of photoplethysmogram) aging index, a RMSSD (root mean square of successive differences of peak-to-peak intervals in waveforms) , peak or valley timing of a PPG signal, peak timing in a first derivative of the PPG signal, peak timing in a second derivative of the PPG signal, or a combination thereof. Applying an optimal feature set as a portion of the training dataset may further improve the accuracy of the BP estimation.

A fourth possible implementation form of the first aspect provides: the method according to the second or third possible implementation form of the first aspect, where the training dataset includes third data indicating a difference between each of given heart rate (HR) values at the given BP values and a given HR value at the given baseline BP value.

According to the fourth possible implementation form of the first aspect, the estimation model which may consider a difference between an observed HR value and a HR value of the subject at the baseline BP value may be built by the machine learning. In a resting state of the subject, the HR may be usually proportional to the BP and also be one of optimal features associated with the BP. Therefore, considering the HR value at the baseline BP may further improve the accuracy of the BP estimation.

A fifth possible implementation form of the first aspect provides: the method according to the third possible implementation form of the first aspect, where the given PPG signal includes a plurality of PPG pulses, and the training dataset includes fourth data indicating a relationship between a feature of each PPG pulse and a corresponding feature of a PPG pulse related to the given baseline BP value.

According to the fifth possible implementation form of the first aspect, the estimation model which may consider a relationship between a feature extracted from observed PPG pulses and a feature of at least one PPG pulse corresponding to the baseline BP value may be built by the machine learning. This may eliminate at least a part of individuality of a person from the training data used for building the estimation model, and the improved accuracy of the BP estimation by the estimation model may be achieved.

A sixth possible implementation form of the first aspect provides: the method according to the fifth possible implementation form of the first aspect, where the relationship is a difference between said two features, a ratio of said two features or a predetermined function representing a mathematical relation between said two features. As mentioned above, a type of the relationship may be appropriately selected according to formality or a predetermined protocol of the two features.

A seventh possible implementation form of the first aspect provides: the method according to any one of the first to sixth possible implementation forms of the first aspect, where the estimating the BP value of the subject includes: estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) values of the subject.

According to the seventh possible implementation form of the first aspect, the subject may obtain the SBP and DBP values which may be useful for prediction and/or diagnosis of several cardiovascular diseases.

A eighth possible implementation form of the first aspect provides: the method according to any one of the first to seventh possible implementation forms of the first aspect, where the method further comprising: normalizing each PPG pulse waveform of the PPG signal based on a PPG pulse waveform corresponding to the baseline BP value.

According to the eighth possible implementation form of the first aspect, the observed PPG pulses are normalized based on the PPG pulse waveform corresponding to the baseline BP value, thus facilitating feature extraction from the PPG signal and improving the accuracy of the BP estimation.

A ninth possible implementation form of the first aspect provides: the method according to any one of the first to eighth possible implementation forms of the first aspect, where the obtaining the baseline BP value of the subject includes capturing an image of a medical checkup report, and reading a specific BP value from the image of the medical checkup report as the baseline BP value of the subject.

According to the ninth possible implementation form of the first aspect, the subject may use the medical checkup report to input the baseline BP value of the subject, thus improving usability.

A second aspect of the embodiments provides an apparatus for blood pressure estimation. In a first possible implementation form of the second aspect, the apparatus comprising a processor and a PPG sensor, where the processor is configured to: obtain a baseline blood pressure (BP) value of a subject; obtain a PPG signal related to the subject via the PPG sensor; and estimate a BP value of the subject by using the baseline BP value, the PPG signal, and an estimation model which is configured to estimate a difference between the BP value and the baseline BP value based on the PPG signal.

According to the first possible implementation form of the second aspect, the estimation model estimates the difference between the BP value and the baseline BP value, and the baseline BP value of the subject is considered when estimating the BP value of the subject. Characteristics of the subject described by the baseline BP value is reflected on the estimated BP value, so that the accuracy of the BP estimation may be improved.

A second possible implementation form of the second aspect provides: the apparatus according to the first possible implementation form of the second aspect, where the estimation model is built by machine learning using a training dataset which includes first data indicating a difference between each of given BP values and a given baseline BP value related to the given BP values.

According to the second possible implementation form of the second aspect, the training dataset consists of data indicating the difference between the given BP value and the corresponding given baseline BP value. This may eliminate at least a part of individuality of a person from the training data used for building the estimation model, and the improved accuracy of the BP estimation by the estimation model may be achieved.

A third possible implementation form of the second aspect provides: the apparatus according to the second possible implementation form of the second aspect, where the training dataset further includes second data indicating at least one feature extracted from a given PPG signal related to the given BP values.

According to the third possible implementation form of the second aspect, the estimation model may be built based on a combination of the at least one feature extracted from the given PPG signal, the corresponding given BP values, and the corresponding difference between the given BP value and the given baseline BP value. The at least one feature may be at least one of a SDPTG aging index, a RMSSD, peak or valley timing of a PPG signal, peak timing in a first derivative of the PPG signal, peak timing in a second derivative of the PPG signal, or a combination thereof. Applying an optimal feature set as a portion of the training dataset may further improve the accuracy of the BP estimation.

A fourth possible implementation form of the second aspect provides: the apparatus according to the second or third possible implementation form of the second aspect, where the training dataset includes third data indicating a difference between each of given HR values at the given BP values and a given HR value at the given baseline BP value.

According to the fourth possible implementation form of the second aspect, the estimation model which may consider a difference between an observed HR value and a HR value of the subject at the baseline BP value may be built by the machine learning. In a resting state of the subject, the HR may be usually proportional to the BP and also be one of optimal features associated with the BP. Therefore, considering the HR value at the baseline BP may further improve the accuracy of the BP estimation.

A fifth possible implementation form of the second aspect provides: the apparatus according to the third possible implementation form of the second aspect, where the given PPG signal includes a plurality of PPG pulses, and the training dataset includes fourth data indicating a relationship between a feature of each PPG pulse and a corresponding feature of a PPG pulse related to the given baseline BP value.

According to the fifth possible implementation form of the second aspect, the estimation model which may consider a relationship between a feature extracted from observed PPG pulses and a feature of at least one PPG pulse corresponding to the baseline BP value may be built by the machine learning. This may eliminate at least a part of individuality of a person from the training dataset used for building the estimation model, and the improved accuracy of the BP estimation by the estimation model may be achieved.

A sixth possible implementation form of the first aspect provides: the apparatus according to the fifth possible implementation form of the second aspect, where the relationship is a difference between said two features, a ratio of said two features or a predetermined function representing a mathematical relation between said two features. As mentioned above, a type of the relationship may be appropriately selected according to formality or a predetermined protocol of the two features.

A seventh possible implementation form of the second aspect provides: the apparatus according to any one of the first to sixth possible implementation forms of the second aspect, where when estimating the BP value of the subject, the processor is configured to estimate SBP and DBP values of the subject.

According to the seventh possible implementation form of the second aspect, the subject may obtain the SBP and DBP values which may be useful for prediction and/or diagnosis of several cardiovascular diseases.

A eighth possible implementation form of the second aspect provides: the apparatus according to any one of the first to seventh possible implementation forms of the second aspect, where the processor is further configure to: normalize each PPG pulse waveform of the PPG signal based on a PPG pulse waveform corresponding to the baseline BP value.

According to the eighth possible implementation form of the second aspect, the observed PPG pulses are normalized based on the PPG pulse waveform corresponding to the baseline BP value, thus facilitating feature extraction from the PPG signal and improving the accuracy of the BP estimation.

A ninth possible implementation form of the first aspect provides: the apparatus according to any one of the first to eighth possible implementation forms of the second aspect, where when obtaining the baseline BP value of the subject, the processor is configured to obtain an image of a medical checkup report, and read a specific BP value from the image of the medical checkup report as the baseline BP value of the subject.

According to the ninth possible implementation form of the second aspect, the subject may use the medical checkup report to input the baseline BP value of the subject, thus improving usability.

A third aspect of the embodiments provides a non-transitory computer readable storage medium storing a program to cause a computer to perform the method according to any one of the first to ninth implementation form of the first aspect. A fourth aspect of the embodiments provides a computer-readable program code to cause a computer to perform the method according to any one of the first to ninth implementation form of the first aspect.

According to any one of the third and sixth aspects, when building the estimation model, at least a part of individuality of a person may be eliminated from the training dataset and the improved accuracy of the BP estimation by the estimation model may be achieved.

A fifth aspect of the embodiments provides a method for building the estimation model according to any one of the first to ninth implementation form of the first aspect. A sixth aspect of the embodiments provides an apparatus for building the estimation model according to any one of the first to ninth implementation form of the first aspect. A seventh aspect of the embodiments provides a non-transitory computer readable storage medium storing a program to cause a computer to perform the method according to the fifth aspect. A eighth aspect of the embodiments provides a computer-readable program code to cause a computer to perform the method according to the fifth aspect.

According to any one of the fifth to eighth aspects, when building the estimation model, at least a part of individuality of a person may be eliminated from the training dataset and the improved accuracy of the BP estimation by the estimation model may be achieved.

BRIEF DESCRIPTION OF DRAWINGS

Fig. 1 is a schematic block diagram for describing a system according to an embodiment of the present disclosure,

Fig. 2 is a schematic block diagram for describing a BP measuring apparatus according to the embodiment of the present disclosure,

Fig. 3 is a schematic block diagram for describing a computing system according to an embodiment of the present disclosure,

Fig. 4 shows a first table for describing a training dataset according to the embodiment of the present disclosure,

Fig. 5 shows a second table for describing a training dataset according to a variation of the embodiment of the present disclosure,

Fig. 6 is a flowchart for describing a BP measurement operation according to the embodiment of the present disclosure,

Fig. 7 is a flowchart for describing a model building operation according to the first variation of the embodiment of the present disclosure.

DESCRIPTION OF EMBODIMENTS

The following describes technical solutions of the embodiments, referring to the accompanying drawings. It will be understood that the embodiments described below are not all but just some of embodiments relating to the present disclosure. It is to be noted that all other embodiments which may be derived by a person skilled in the art based on the embodiments described below without creative efforts shall fall within the protection scope of the present disclosure.

Fig. 1 is a schematic block diagram for describing a system according to an embodiment of the present disclosure. The system shown in Fig. 1 may include a BP measuring apparatus 10, a computing system 20, and a cuffed BP monitor 30. Optionally, this system may further include other healthcare monitors or sensors such as an electrocardiograph (ECG) sensor, a substitute of the ECG sensor, an impedance cardiography (ICG) sensor, at least one PPG sensor, a PPG sensor array, a pressure sensor, or the like, and the other healthcare monitors or sensors may be connected to the apparatus 10. For example, a device such as an acceleration sensor, a vibration sensor, a gyro sensor, a sheet-type pressure sensor, an ultrasonic sensor, a micro-wave sensor, a millimeter wave radar, or the like may be used as the substitute of the ECG sensor. The at least one PPG sensor may measure a PPG signal at a measuring point other than a point measured by a PPG sensor equipped within the apparatus 10. The pressure sensor may monitor a radial artery or the like by a tonometry method or manage contact pressure of the PPG sensor.

The apparatus 10 may be connected via a network NW with the system 20. The network NW may be a wireless or wired network such as a local area network (LAN) , a wide area network (WAN) , or a combination thereof. The apparatus 10 may be an apparatus such as a cuff-less BP monitor, a medical device, a personal BP measurement device, a mobile device such as a mobile phone, a smart phone, a communication terminal, a tablet device, a personal computer, a smart watch, a health care device, or the like. The system 20 may be a computer such as a personal computer, a workstation, a server computer, a distributed processing system, or the like. The monitor 30 may be a monitoring device with a cuff used for BP measurement.

The apparatus 10 is configured to perform cuff-less BP measurement and obtain BP values of a subject. The system 20 is configured to build an estimation model for estimating BP values based on PPG signals by machine learning. The built estimation model is provided to the apparatus 10 and used for the cuff-less BP measurement by the apparatus 10. The monitor 30 is configured to measure BP values of the subject, and the measurement results by the monitor 30 are used for regular calibration of the estimation model used by the apparatus 10. Optionally, the monitor 30 may be further configured to measure HR values of the subject.

Fig. 2 is a schematic block diagram for describing a BP measuring apparatus according to the embodiment of the present disclosure. As shown in Fig. 2, the apparatus 10 may include a PPG sensor 11, a processor 12 and a memory 13. The apparatus 10 may further include a display device such as a LCD (Liquid Crystal Display) device, a ELD (Electro-Luminescent Display) device, or the like. Optionally, the apparatus 10 may further include a capturing device such as a digital camera.

The PPG sensor 11 includes a light source such as a LED flash or the like, a light detecting device such as a CMOS (Complementary Metal Oxide Semiconductor) sensor or the like, and processing circuitry such as at least one CPU (Central Processing Unit) . Alternatively, the PPG sensor 11 may be a PPG measuring device by analyzing a captured video image of the subject to obtain PPG signals.

If the PPG sensor 11 is a transmission-type sensor, the PPG sensor 11 is configured to obtain PPG signals based on transmitted light through a body part of the subject such as a subject's finger or the like. If the PPG sensor 11 is a reflecting-type sensor, the PPG sensor 11 is configured to obtain PPG signals based on reflected light at an inner part of the subject. In the embodiment of the present disclosure, any PPG sensor type may be applicable to the PPG sensor 11.

The processor 12 is a hardware element for processing data, and may be processing circuitry such as at least one CPU, at least one FPGA (Field-Programmable Gate Array) , at least one GPU (Graphics Processing Unit) or the like. The memory 13 is a hardware element for storing data, and may be a storage device such as an SSD (Solid State Drive) , HDD (Hard Disk Drive) , RAM (Random Access Memory) , ROM (Read Only Memory) , a flash memory, a memory card or the like. The memory 13 may be a non-transitory computer readable storage medium.

For example, the processor 12 may be configured to obtain a baseline BP value 13a of a subject, obtain a PPG signal related to the subject via the PPG sensor 11, and estimate a BP value of the subject by using the baseline BP value 13a, the PPG signal, and an estimation model 13b. The estimation model 13b is configured to estimate a difference between the BP value and the baseline BP value 13a based on the PPG signal. Optionally, before estimating the BP value of the subject, the processor 12 may be configure to normalize each PPG pulse waveform of the PPG signal based on a PPG pulse waveform corresponding to the baseline BP value 13a. Further, the processor 12 may be configured to estimate SBP and DBP values of the subject.

The baseline BP value 13a may be inputted by the subject in advance and be stored in the memory 13. Optionally, the processor 12 may be configured to obtain an image of a medical checkup report of the subject and read a specific BP value from the image of the medical checkup report as the baseline BP value 13a. The medical checkup report is one example of an information medium including information of the baseline BP value 13a. The estimation model 13b may be built by the system 10, and be pre-stored in the memory 13 or received from the system 20 via the network NW.

Fig. 3 is a schematic block diagram for describing a computing system according to an embodiment of the present disclosure. As shown in Fig. 3, the system 20 may include a processor 21 and a memory 22. The processor 21 is a hardware element for processing data, and may be processing circuitry such as at least one CPU, at least one FPGA, at least one GPU or the like. The memory 22 is a hardware element for storing data, and may be a storage device such as an SSD, HDD, RAM, ROM, a flash memory, a memory card or the like. The memory 22 may be a non-transitory computer readable storage medium.

For example, the processor 21 may be configured to build an estimation model 22b by machine learning using a training dataset 22a which includes first data indicating a difference between each of given BP values and a given baseline BP value related to the given BP values.

In the embodiment of the present disclosure, the processor 21 may perform the machine learning based on a technique such as the Linear Regression, the Neural Networks, the Support Vector Regression, the Logistic Regression, the AdaBoost Tree, the Bagged Tree, the K Nearest Neighbors, the Decision Tree Regression, the Random Forest Regression, or the like.

For obtaining at least a part of the training dataset 22a, the Medical Information Mart for Intensive Care (MIMIC) database may be used, where the MIMIC database is a free-to-use database that contains tens of thousands of Intensive Care Unit (ICU) patients, and includes various types of data such as ABP data, ECG data, PPG data, or the like.

The training dataset 22a may include second data indicating at least one feature extracted from a given PPG signal related to the given BP values. The second data indicating the at least one feature may be obtained from a public database such as the MIMIC database, or may be extracted by using the Autoencoder.

For example, several types of features may be defined based on ECG, PPG, VPG (Velocity waveform of PPG signals) and APG (Acceleration waveform of PPG signals) signals, as PPG features. The VPG signal may be defined as the first order differential to the original PPG signal. The APG signal may be defined as the second order differential to the original PPG signal. One of the PPT features is the pulse arrival time (PAT) which may be extracted from ECG and PPG signals. Further, there exists PPG morphology features which may be extracted from PPG, VPG and APG signals. In addition, the SDPTG aging index, the RMSSD, peak or valley timing of a PPG signal, peak timing in the VPG signal, peak timing in the APG signal, or a combination thereof may be extracted as the PPG features applicable to the embodiment of the present disclosure.

Optionally, the training dataset 22a may include third data indicating a difference between each of given HR values at the given BP values and a given HR value at the given baseline BP value. For example, HR and BP values may be simultaneously measured by using a BP monitoring device with a HR measurement function. Therefore, if a database including pairs of HR and BP values is given, the relevant pairs of the HR and BP values may be used as the training dataset 22a for building the estimation model 22b. In addition, the training dataset 22a may include fourth data indicating a relationship between a feature of each PPG pulse and a corresponding feature of a PPG pulse related to the given baseline BP value. For example, the relationship may be a difference between two features, a ratio of two features or a predetermined function representing a mathematical relation between two features.

Fig. 4 shows a first table for describing a training dataset according to the embodiment of the present disclosure. As shown in Fig. 4, the first table includes datasets #01, #02, #11 and #12 relating to the same or similar subject, as one of examples. Each of datasets #01 and #02 consists of a baseline SBP value, a SBP value, a baseline DBP value, a DBP value, a baseline HR value, a HR value and features. Each of datasets #11 and #12 consists of a difference between the SBP value and the baseline SBP value, a difference between the DBP value and the baseline DBP value, a difference between the HR value and the baseline HR value, and the features. The baseline SBP/DBP value corresponds to a SBP/DBP value when a BP value is equal to the baseline BP value. The baseline HR value may be defined as a HR value when the BP value is equal to the baseline BP value. The datasets #01 and #02 correspond to the datasets #11 and #12, respectively.

If the datasets #1 and #2 are given as a source of the training dataset 22a, with regard to each of the datasets #1 and #2, the processor 21 may calculate the difference between the SBP value and the baseline SBP as follows: ΔSBP = SBP -baseline SBP. In the same manner, the processor 21 may calculate the difference between the DBP value and the baseline DBP as follows: ΔDBP = DBP -baseline DBP; and also calculate the difference between the HR value and the baseline HR as follows: ΔHR = HR -baseline HR. In an example of Fig. 4, the processor 21 sets the features in the datasets #01 and #02 to the datasets #11 and #12. This means that the processor 21 may directly use the features extracted the PPG signals when building the estimation model 22b. In the similar way, the processor 21 may prepare datasets including ΔSBP, ΔDBP, ΔHR and the corresponding features. Optionally, these datasets may be obtained from an external data source such as the MIMIC database.

The processor 21 builds the estimation model 22b based on the prepared datasets including ΔSBP, ΔDBP, ΔHR and the corresponding features. According to the example of Fig. 4, the prepared datasets mainly consist of the differences: ΔSBP, ΔDBP and ΔHR. This may eliminate at least a part of individuality of a person from the training dataset used for building the estimation model, and may contribute to achieve improved accuracy of BP estimation by the estimation model 22b.

Fig. 5 shows a second table for describing a training dataset according to a variation of the embodiment of the present disclosure. As shown in Fig. 5, the second table includes datasets #01, #02, #11 and #12 which are the same as the corresponding datasets in the first table of Fig. 4. In the example of Fig. 5, a record labeled by "BASELINE" is added to the table. This labeled record is added for indicating the features when the SBP value, the DBP value and the HR value equal to the baseline SBP value, the baseline DBP value and the baseline HR value, respectively. In the labeled record includes features of X, Y and Z which may be obtained by applying linear/nonlinear interpolation, linear/nonlinear extrapolation, statistic models, or the like to the given datasets such as the datasets #1 and #2. Further, some features used in a pulse wave analysis (PWA) approach may be extracted by synthesizing a normalized pulse waveform based on the PPG pulse at the baseline.

In the example of Fig. 5, the processor 21 may calculate ΔSBP, ΔDBP and ΔHR in the same manner as the example of Fig. 4. Further, the processor 21 may calculate a relationship between a feature A1 of the dataset #01 and a corresponding feature X of the labeled record. In regard to these features, the relationship is a difference between the corresponding two features. Thus, the processor 21 calculates the difference between the features A1 and X, and sets the difference to the dataset #11 corresponding to the dataset #01. Similarly, the processor 21 calculates a difference between the features A2 and X, and sets the difference to the dataset #12 corresponding to the dataset #02.

In the example of Fig. 5, features B1 and B2 in the second table correspond to a feature Y of the labeled record. In regard to these features, the relationship is a ratio of the corresponding two features. Thus, the processor 21 calculates a ratio of B1 and Y, and sets the ratio to the dataset #11. Similarly, the processor 21 calculates a ratio of B2 and Y, and sets the ratio to the dataset #12. Further, with respect to features C1 and C2 corresponding to a feature Z of the labeled record, the relationship is provided by a function F which may be a mathematical function. Regarding the features C1 and C2, the processor 21 calculates F (C1, Z) and F (C2, Z) , and sets these calculation results to the datasets #11 and #12, respectively. Optionally, these features may be obtained from the external data source such as the MIMIC database.

The processor 21 builds the estimation model 22b based on the prepared datasets including ΔSBP, ΔDBP, ΔHR and the relationship of the corresponding features. According to the example of Fig. 5, with regard to the features as well as ΔSBP, ΔDBP and ΔHR, at least a part of individuality of a person may be eliminated from the training dataset used for building the estimation model, so that improved accuracy of BP estimation by the estimation model 22b may be implemented.

Fig. 6 is a flowchart for describing a BP measurement operation according to the embodiment of the present disclosure. The BP measurement operation shown in Fig. 6 is mainly performed by the apparatus 10.

At step S11, the processor 12 of the apparatus 10 obtains a baseline BP value 13a of a subject. The baseline BP value may be inputted by the subject via an input interface such as a keyboard, a keypad, a touch panel, a touch pad, or the like. Optionally, the processor 12 may obtain the baseline BP value 13a from a BP monitoring device such as the cuffed BP monitor 30 or the like. In this case, the subject may measure the baseline BP value 13a by using the BP monitoring device, and transmit the baseline BP value 13a to the apparatus 10 via a wireless or wired communication network. The baseline BP value 13a is stored in the memory 13, and may be updated regularly.

At step S12, the processor 12 obtains a PPG signal related to the subject via the PPG sensor 11. Data indicating the PPG signal may be stored in the memory 13. Optionally, the processor 12 may further obtain other signals output from sensors (not shown) connected to the apparatus 10. For example, signals from an external ECG sensor or the like may be obtained by the processor 12.

At step S13, the processor 12 estimates a difference between the BP value and the baseline BP value 13a based on the estimation model 13b and the PPG signal. The estimation model 13b may be built by the system 20 and be pre-stored in the memory 13 of the apparatus 10. Optionally, the estimation model 13b may be received from the system 20 via the network NW. In this case, the estimation model 13b may be updated regularly and be provided to the apparatus 10.

Optionally, the estimation model 13b may be calibrated before estimating the BP value, or may not be calibrated. In this regard, Kachuee suggests that optional calibration of the BP estimation may improve BP estimation accuracy in the existing PPT based approach. In the embodiment of the present disclosure, the calibration may be performed by using practical BP values or ΔBP values output from the cuffed BP monitor 30. Optionally, the estimation model 13b may be calibrated by the system 20. In this case, the BP values and/or ΔBP values output from the cuffed BP monitor 30 may be provided to the system 20, and the system 20 may calibrate the estimation model 13b.

At step S14, the processor 12 calculates a BP value of the subject by using the obtained baseline BP value 13a, and the estimated difference between the BP value and the baseline BP value 13a by using the estimation model 13b. At step S15, the processor 12 displays the calculated BP value on the display device connected to the apparatus 10. After completing the step 15, the BP measurement operation shown in Fig. 6 is terminated.

Fig. 7 is a flowchart for describing a model building operation according to the first variation of the embodiment of the present disclosure. The model building operation is mainly performed by the system 20.

At step S21, the processor 21 of the system 20 extracts features used for building the estimation model 22b from the PPG signals. For example, the features of the PPG signals may be extracted by using the Autoencoder in the similar manner as Shimazaki. Optionally, the processor 21 may obtain the extracted features of the PPG signals from the public database such as the MIMIC database. In this step, several types of features may be extracted or obtained. For example, the PAT may be extracted from ECG and PPG signals, and the PPG morphology features may be extracted from PPG, VPG and APG signals. In addition, the processor 21 may extract the SDPTG aging index, the RMSSD, peak or valley timing of a PPG signal, peak timing in the VPG signal, peak timing in the APG signal, or a combination thereof.

At step S22, the processor 21 obtains BP values and baseline BP values related to the extracted features. The baseline BP values may be estimated based on the PPG signals and/or other signals such as the ECG signals and so on. At step S23, the processor 21 obtains HR values and baseline HR values related to the extracted features. According to the steps S22 and S23, the processor 21 may obtain source datasets such as the above-mentioned datasets #01 and #02 shown in Figs. 4 and 5. If the HR values are not obtained by the processor 22, the step S23 may be omitted.

At step S24, the processor 21 calculates a difference between each obtained BP value and the obtained baseline BP value. At step S25, the processor 21 calculates a difference between each obtained HR value and the obtained baseline HR value. At step S26, the processor 21 calculates a relationship between each extracted feature and a feature related to the baseline BP value. The feature related to the baseline BP value may be identified by applying an operation such as linear/nonlinear interpolation, linear/nonlinear extrapolation, statistic models or the like to the extracted features relevant to the same or similar subject. Optionally, if BP and HR related values at the baseline are not obtained, the step S26 may be omitted. In this case, the extracted features may be directly used for building the estimation model 22b.

At step S27, the processor 21 performs machine learning using the calculated differences and the calculated relationship as the training dataset 22a to build the estimation model 22b. Alternatively, if the step S26 is omitted, the processor 21 may perform machine learning using the calculated differences and the extracted features as the training dataset 22a. In the step S27, the machine learning may be based on a technique such as the Linear Regression, the Neural Networks, the Support Vector Regression, the Logistic Regression, the AdaBoost Tree, the Bagged Tree, the K Nearest Neighbors, the Decision Tree Regression, the Random Forest Regression, or the like.

At step S28, the processor 21 may perform calibration of the estimation model 22b based on practically measured BP values. The measured BP values may be obtained from a BP monitoring device such as the cuffed BP monitor 30 or the like. Although the calibration is performed after the step S28 once, the calibration may be performed periodically or in an arbitrary timing, and frequency of performing the calibration may be equal to or more than two. The built estimation model 22b may be provided to the apparatus 10 and be stored in the memory 13 as the estimation model 13b. After completing the step 28, the model building operation shown in Fig. 7 is terminated.

The foregoing disclosure merely discloses exemplary embodiments, and is not intended to limit the protection scope of the present invention. It will be appreciated by those skilled in the art that the foregoing embodiments and all or some of other embodiments and modifications which may be derived based on the scope of claims of the present invention will of course fall within the scope of the present invention.

Claims

A method for blood pressure estimation, comprising:

obtaining a baseline blood pressure (BP) value of a subject;

obtaining a photoplethysmogram (PPG) signal related to the subject; and

estimating a BP value of the subject by using the baseline BP value, the PPG signal, and an estimation model which is configured to estimate a difference between the BP value and the baseline BP value based on the PPG signal.
The method according to claim 1, wherein the estimation model is built by machine learning using a training dataset which includes first data indicating a difference between each of given BP values and a given baseline BP value related to the given BP values.
The method according to claim 2, wherein the training dataset further includes second data indicating at least one feature extracted from a given PPG signal related to the given BP values.
The method according to claim 2 or 3, wherein the training dataset includes third data indicating a difference between each of given heart rate (HR) values at the given BP values and a given HR value at the given baseline BP value.
The method according to claim 3, wherein the given PPG signal includes a plurality of PPG pulses, and the training dataset includes fourth data indicating a relationship between a feature of each PPG pulse and a corresponding feature of a PPG pulse related to the given baseline BP value.
The method according to claim 5, wherein the relationship is a difference between said two features, a ratio of said two features or a predetermined function representing a mathematical relation between said two features.
The method according to any one of claims 1 to 6, wherein the estimating the BP value of the subject includes: estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) values of the subject.
The method according to any one of claims 1 to 7, further comprising: normalizing each PPG pulse waveform of the PPG signal based on a PPG pulse waveform corresponding to the baseline BP value.
The method according to any one of claims 1 to 8, wherein the obtaining the baseline BP value of the subject includes: capturing an image of a medical checkup report, and reading a specific BP value from the image of the medical checkup report as the baseline BP value of the subject.
An apparatus for blood pressure estimation, the apparatus comprising a processor and a photoplethysmogram (PPG) sensor, wherein the processor is configured to:

obtain a baseline blood pressure (BP) value of a subject;

obtain a PPG signal related to the subject via the PPG sensor; and

estimate a BP value of the subject by using the baseline BP value, the PPG signal, and an estimation model which is configured to estimate a difference between the BP value and the baseline BP value based on the PPG signal.
The apparatus according to claim 10, wherein the estimation model is built by machine learning using a training dataset which includes first data indicating a difference between each of given BP values and a given baseline BP value related to the given BP values.
The apparatus according to claim 11, wherein the training dataset further includes second data indicating at least one feature extracted from a given PPG signal related to the given BP values.
The apparatus according to claim 11 or 12, wherein the training dataset includes third data indicating a difference between each of given heart rate (HR) values at the given BP values and a given HR value at the given baseline BP value.
The apparatus according to claim 12, wherein the given PPG signal includes a plurality of PPG pulses, and the training dataset includes fourth data indicating a relationship between a feature of each PPG pulse and a corresponding feature of a PPG pulse related to the given baseline BP value.
The apparatus according to claim 14, wherein the relationship is a difference between said two features, a ratio of said two features or a predetermined function representing a mathematical relation between said two features.
The apparatus according to any one of claims 10 to 15, wherein before estimating the BP value of the subject, the processor is configured to estimate systolic blood pressure (SBP) and diastolic blood pressure (DBP) values of the subject.
The apparatus according to any one of claims 10 to 16, wherein the processor is further configure to: normalize each PPG pulse waveform of the PPG signal based on a PPG pulse waveform corresponding to the baseline BP value.
The apparatus according to any one of claims 10 to 17, wherein when obtaining the baseline BP value of the subject, the processor is configured to obtain an image of a medical checkup report, and read a specific BP value from the image of the medical checkup report as the baseline BP value of the subject.
A non-transitory computer readable storage medium storing a program to cause a computer to perform the method for blood pressure estimation according to any one of claims 1 to 9.