A cuffless blood pressure estimation method and system

Abstract

The present application relates to the technical field of wearable device detection, and more particularly to a cuffless blood pressure estimation method and system, comprising the following steps: synchronously collecting electrocardiogram signals and photoplethysmogram signals, preprocessing the collected electrocardiogram signals and photoplethysmogram signals, delay embedding and state vector construction, reconstruction of nonlinear response operator, one-time cuff calibration and blood pressure sensitive feature recognition, online blood pressure estimation, universal boundary output gating, calibration drift monitoring and re-calibration triggering. The present application can fully exploit the nonlinear dynamics information of the cardiovascular system, realize automatic sparse recognition of blood pressure sensitive features, ensure the reliability of the output by introducing a universal boundary gating mechanism, has self-adaptive re-calibration capability, reduces the burden on the user of re-calibration operation while ensuring the continuity of the estimation accuracy, and is conducive to the long-term stable use of the device.

Description

Technical Field

[0001] This invention relates to the field of wearable device detection technology, and in particular to a cuffless blood pressure estimation method and system. Background Technology

[0002] With the widespread adoption of wearable devices, cuffless continuous blood pressure estimation technology based on photoplethysmography (PPG) and electrocardiogram (ECG) signals has become a research hotspot in the field of blood pressure monitoring due to its advantages of not requiring an inflatable cuff and its high wearability. Existing methods typically extract parameters from pulse wave transit time (PWTT) or pulse wave morphology characteristics to establish regression models for estimating systolic and diastolic blood pressure.

[0003] However, existing cuffless blood pressure estimation methods generally face two key problems:

[0004] First, the cardiovascular system is essentially a high-dimensional nonlinear dynamic system. Traditional methods based on PWTT or simple morphological features only extract a small amount of linear or quasi-linear information about the system, which cannot fully characterize the complex dynamic process of multiple factors such as vascular elasticity, peripheral resistance and neural regulation, resulting in a significant decrease in estimation accuracy when blood pressure fluctuates greatly.

[0005] Secondly, due to factors such as sensor contact status of wearable devices, changes in tissue perfusion, and vascular tension drift over time, the calibration parameters of the model are prone to shift. Existing methods lack effective model credibility judgment and adaptive calibration mechanisms, making it difficult to guarantee the reliability of estimations in long-term use.

[0006] To address the aforementioned issues, this invention proposes a cuffless blood pressure estimation method and system based on nonlinear response operator reconstruction and Lie-Jordan decomposition, taking into account the nonlinear dynamic reconstruction of the cardiovascular system. Summary of the Invention

[0007] The main objective of this invention is to overcome the shortcomings of the prior art and provide a cuffless blood pressure estimation method and system.

[0008] The technical solution adopted by this invention to achieve its technical objective is: a cuffless blood pressure estimation method, comprising the following steps:

[0009] Step S1: Synchronously acquire electrocardiogram (ECG) signals and photoplethysmography (PPG) pulse wave signals;

[0010] Step S2: Preprocess the acquired electrocardiogram signal and photoplethysmography (PPG) signal;

[0011] Step S3, Delay Embedding and State Vector Construction: The optimal delay time is determined using the first minimum mutual information method, and the minimum embedding dimension is determined using the pseudo nearest neighbor method or the Cao method. Based on the determined optimal delay time and embedding dimension, delay embedding is applied to the ECG signal and the photoplethysmography pulse wave signal respectively to construct a joint embedding state vector.

[0012] Step S4, Reconstruction of the nonlinear response operator: Divide the sequence of joint embedded state vectors into segments with a window length of 30-60 seconds, and use the Volterra series truncation approximation to reconstruct the nonlinear response operator in the joint embedded state vector sequence of each segment;

[0013] Step S5, disposable cuff calibration and blood pressure sensitive mode identification: while wearing the ECG sensor and photoplethysmography pulse wave sensor, use the cuff blood pressure monitor to measure the reference systolic blood pressure and reference diastolic blood pressure. With the reference systolic blood pressure and reference diastolic blood pressure as target variables and the real part, imaginary part and response amplitude coefficient of all intrinsic modes as candidate features, perform L1 regularized regression to automatically select the systolic blood pressure sensitive mode set and the diastolic blood pressure sensitive mode set.

[0014] Step S6, Online Blood Pressure Estimation: Perform Lie-Jordan decomposition on the nonlinear response operator and extract the real and imaginary parts of the eigenvalues ​​and response amplitude coefficients corresponding to the systolic and diastolic pressure sensitive mode sets. Substitute the extracted features into the frozen systolic and diastolic pressure calibration models to calculate the estimated systolic and diastolic blood pressure values.

[0015] Step S7, Universal Boundary Output Gating: Simultaneously calculate the steady-state entropy yield estimate, normalized acceleration response ratio, and slowest dissipation rate estimate for the current segment, and calculate the universal theoretical bound and calibration health index;

[0016] Step S8, Calibration Drift Monitoring and Recalibration Trigger: The device records the time series of the calibration health index during each online estimation. When the calibration health index shows a monotonically increasing trend within 24 hours and the p-value of the Kendall tau rank correlation coefficient test is less than 0.05, a recalibration reminder is triggered. When the calibration health index is always greater than 0.6 within 48 hours, recalibration is mandatory.

[0017] Preferably, in step S1: a wearable device or a medical-grade front-end is used to simultaneously collect the subject's single-lead electrocardiogram signal and photoplethysmography (PPG) wave signal;

[0018] The sampling rate of the electrocardiogram signal is not less than 256Hz, and the bandpass filtering range is 0.5-40Hz;

[0019] The sampling rate of the photoplethysmography signal is not less than 64Hz, the bandpass filtering range is 0.5-10Hz, and green light with a wavelength of 525nm or infrared light with a wavelength of 940nm is used as the light source.

[0020] The hardware synchronization error between the two channels is less than one sampling period;

[0021] A steady-state signal was continuously collected for no less than 30 seconds, during which the subject remained seated or lying down.

[0022] Preferably, step S4 includes the following sub-steps:

[0023] S4.1 Segmentation and Aggregation: The sequence of jointly embedded state vectors is segmented into segments with a window length of 30-60 seconds, and each segment contains 30-80 cardiac cycles;

[0024] S4.2 Volterra kernel estimation: The nonlinear response operator is reconstructed from the joint embedded state vector sequence in each segment using the Volterra series truncation approximation. The nonlinear response operator satisfies:

[0025] ;

[0026] in, It is an equivalent random perturbation vector;

[0027] The nonlinear response operator is approximated by truncation using a second-order Volterra series:

[0028] ;

[0029] When second-order accuracy is insufficient, the third-order Volterra series truncation approximation is used:

[0030] ;

[0031] in, It is a zero-order kernel. It is a first-order kernel. It is a second-order kernel. It is a third-order kernel; , , The regularization parameter is frozen in the independent calibration segment by estimation using the least squares method or regularized regression.

[0032] S4.3 Operator Consistency Verification: Apply the estimated nonlinear response operator F to the joint embedded state vector and calculate the one-step prediction residual. ;

[0033] For the predicted residual of the first step Perform the Ljung-Box whitening test. If the test fails, increase the order of the Volterra series or increase the embedding dimension until the whitening test is passed.

[0034] Preferably, in step S5, the disposable cuff calibration and blood pressure sensor identification are performed only once during initial use or recalibration, and include the following sub-steps:

[0035] S5.1 Calibration Data Acquisition: While the subjects wore the ECG sensor and photoplethysmography pulse wave sensor, reference systolic blood pressure and reference diastolic blood pressure were measured using a certified cuff blood pressure monitor.

[0036] The cuff measurement frequency is once every 1-2 minutes, lasting for 5-10 minutes, and a total of 3-10 sets of reference values ​​are obtained; ECG signals and photoplethysmography (PPG) signals are recorded simultaneously.

[0037] S5.2, Lie-Jordan Decomposition: Reconstruct the nonlinear response operator from the joint embedded state vector sequence within the calibration segment, and decompose the nonlinear response operator into Jordan parts. and Lie part :

[0038] ;

[0039] in, The symmetrical / dissipative component corresponds to an irreversible physiological dissipation process. This is the antisymmetric / coherent part, corresponding to the reversible dynamic subprocess;

[0040] right Perform spectral decomposition:

[0041] ;

[0042] in, For the first One orthogonal eigenmode The total number of intrinsic moduli. ; For the first Individual eigenmode The corresponding Jordan eigenvalues, For the first The attenuation rate of each eigenmode For the first The oscillation frequency of each eigenmode.

[0043] Record the real part of each eigenmode virtual part and response amplitude coefficient ;

[0044] S5.3 Sensitive modulus sparse selection: Using the reference systolic and diastolic blood pressure measured by the cuff as the target variables, and using all... The real part of each eigenmode virtual part and response amplitude coefficient As candidate features, L1 regularized regression is performed to automatically select the set of systolic blood pressure sensitive modes. and diastolic pressure sensitive mode set :

[0045] ;

[0046] ;

[0047] Among them, the set of systolic pressure sensitive modes The number of modules is 2-3, diastolic pressure sensitive module set The number of moduli is 2-3, and the sensitive moduli of the two do not completely overlap;

[0048] Freeze-throttle calibration model And diastolic pressure calibration model Stored in the device firmware:

[0049] ,in ;

[0050] ,in ;

[0051] in, and It is a sparse linear model or a low-order polynomial model.

[0052] Preferably, in step S6, during daily wear of the device, the following data is collected from a sliding window of ECG and photoplethysmography (PPG) signals every 10-30 seconds:

[0053] S6.1 Construct the joint embedded state vector and reconstruct the nonlinear response operator according to the methods in steps S3-S4;

[0054] S6.2 Perform Lie-Jordan decomposition on the nonlinear response operator to extract the set of systolic pressure sensitive modes. and diastolic pressure sensitive mode set The corresponding real part of the eigenvalue virtual part and response amplitude coefficient ;

[0055] S6.3 Substitute the extracted features into the frozen systolic blood pressure calibration model. And diastolic pressure calibration model Calculate the estimated systolic blood pressure value And diastolic blood pressure estimates ;

[0056] S6.4. Based on the residual standard deviation of the calibration segment leave-one-out cross-validation, calculate the confidence interval of the estimated value, and output it in the form of ( ) mmHg, of which This is the confidence interval for systolic blood pressure. This represents the confidence interval for diastolic blood pressure.

[0057] Preferably, in step S7: the universal boundary gating parameters of the current segment are calculated synchronously.

[0058] Calculate the steady-state entropy production rate estimate of the cardiovascular system ,in , Let be the state covariance matrix. The residual covariance matrix;

[0059] Calculate the normalized acceleration ratio ,in The maximum expansion rate of the positive part of the dynamic system;

[0060] Calculate the slowest dissipation rate estimate , ,Depend on The minimum eigenvalue real part is determined;

[0061] Upper limit of the universal theory of computation ;

[0062] Calculate the calibrated health index ;

[0063] Set the output suppression threshold ;

[0064] Output gating satisfy: If and only if ,Right now ;

[0065] when hour, Normal output systolic blood pressure estimate And diastolic blood pressure estimates and the corresponding confidence interval,

[0066] when hour, Normal output systolic blood pressure estimate And diastolic blood pressure estimates And the corresponding confidence interval, and mark the edge states;

[0067] when hour, Suppressing blood pressure readings triggers one of the following alternative actions:

[0068] (1) Output the last time The estimated blood pressure value at that time is marked as non-real-time / to be recalibrated; or

[0069] (2) Prompt the user to remain still and remeasure; or

[0070] (3) When If the value continues to rise, a recalibration alert will be triggered.

[0071] Preferably, in step S8: the device records the calibration health index during each online estimation. Time series;

[0072] When the calibrated health index shows a monotonically increasing trend over 24 hours, the Kendall tau rank correlation coefficient test... When the value is less than 0.05, a recalibration reminder is triggered;

[0073] If the calibration health index remains greater than 0.6 for a continuous period of 48 hours, recalibration is mandatory.

[0074] During recalibration, the original systolic pressure sensitive mode set... and diastolic pressure sensitive mode set Under the premise of keeping the systolic blood pressure calibration model constant, only update the systolic blood pressure calibration model. And diastolic pressure calibration model The weighting coefficients.

[0075] The present invention also provides a cuffless blood pressure estimation system, the system comprising:

[0076] The system includes a signal acquisition module, a signal preprocessing module, an embedded state construction module, a nonlinear response operator reconstruction module, a Lie-Jordan decomposition and spectral analysis module, a calibration module, an online blood pressure estimation module, a universal output gating module, a calibration drift monitoring and recalibration triggering module, a central processing unit, and a storage module.

[0077] The signal acquisition module is used to simultaneously acquire the subject's single-lead electrocardiogram signal and photoplethysmography pulse wave signal;

[0078] The signal preprocessing module is connected to the signal acquisition module and is used to preprocess the acquired electrocardiogram signal and photoplethysmography pulse wave signal.

[0079] The embedded state construction module is connected to the signal preprocessing module and is used to construct a joint embedded state vector based on the preprocessed electrocardiogram signal and photoplethysmography pulse wave signal.

[0080] The nonlinear response operator reconstruction module is connected to the embedded state construction module and is used to reconstruct the nonlinear response operator based on the sequence of the joint embedded state vector.

[0081] The Lie-Jordan decomposition and spectral analysis module is connected to the nonlinear response operator reconstruction module and is used to perform Lie-Jordan decomposition and spectral decomposition on the nonlinear response operator.

[0082] The calibration module is connected to the Lie-Jordan decomposition and spectral analysis module and is used to perform disposable cuff calibration and blood pressure sensitive model identification, and generate a frozen calibration model.

[0083] The online blood pressure estimation module is connected to the Lie-Jordan decomposition and spectral analysis module and the calibration module, and is used to estimate blood pressure in real time from a sliding window during daily wear of the device.

[0084] The universal bound output gating module is connected to the Lie-Jordan decomposition and spectral analysis module and the online blood pressure estimation module, and is used to gating the output of the blood pressure estimation value according to the universal bound theory.

[0085] The calibration drift monitoring and recalibration triggering module is connected to the universal output gating module and is used to monitor the time series of the calibration health index and trigger recalibration when the conditions are met.

[0086] The central processing unit is connected to the signal acquisition module, signal preprocessing module, embedded state construction module, nonlinear response operator reconstruction module, Lie-Jordan decomposition and spectral analysis module, calibration module, online blood pressure estimation module, universal output gating module, and calibration drift monitoring and recalibration triggering module, respectively, and is used to control the collaborative work of each module and data flow.

[0087] The storage module is connected to the central processing unit and is used to store calibration data, frozen calibration model parameters, calibration health index time series, and the user's historical blood pressure estimation results and corresponding confidence intervals.

[0088] Preferably, the signal acquisition module includes electrocardiogram electrodes, a photoplethysmography (PPG) sensor, and an analog-to-digital converter.

[0089] The ECG electrodes are configured as single leads, using an equivalent method of lead I or chest leads; the photoplethysmography (PPG) sensor includes a light-emitting diode (LED) and a photodiode. The LED is a green LED with a wavelength of 525nm or an infrared LED with a wavelength of 940nm, and the photodiode is used to receive light signals reflected or transmitted through tissue and convert them into electrical signals.

[0090] The analog-to-digital converter is used to convert the analog electrocardiogram (ECG) signal collected by the ECG electrodes and the analog photoplethysmography (PPG) signal output by the photodiode into digital signals. The sampling frequency of the ECG signal is configured to be no less than 256 Hz and the PPG signal sampling rate is no less than 64 Hz, and the hardware synchronization error between the two channels is less than one sampling period.

[0091] The signal acquisition module is a wrist-type integrated device, the ECG electrodes are placed on the surface of the device that contacts the skin of the wrist, and the photoplethysmography (PPG) sensor is placed at the bottom of the device where it contacts the skin of the wrist.

[0092] The signal acquisition module also includes a triaxial accelerometer for acquiring the subject's motion acceleration signals to identify and eliminate motion artifact segments.

[0093] Preferably, the nonlinear response operator reconstruction module includes:

[0094] The segmentation unit is used to segment the sequence of joint embedded state vectors into segments with a window length of 30-60 seconds, and each segment contains 30-80 cardiac cycles.

[0095] The Volterra kernel estimation unit is used to reconstruct the nonlinear response operator F from the joint embedded state vector sequence within each segment using the Volterra series truncation approximation, wherein the nonlinear response operator satisfies:

[0096] ;

[0097] in, It is an equivalent random perturbation vector;

[0098] The Volterra kernel estimation unit employs a second-order Volterra series truncation approximation:

[0099] ;

[0100] When second-order accuracy is insufficient, the third-order Volterra series truncation approximation is used:

[0101] ;

[0102] in, It is a zero-order kernel. It is a first-order kernel. It is a second-order kernel. It is a third-order kernel; , , The regularization parameter is frozen in the independent calibration segment by estimation using the least squares method or regularized regression.

[0103] The operator consistency verification unit is used to apply the estimated nonlinear response operator F to the joint embedded state vector and calculate the one-step prediction residual. The Ljung-Box whitening test is performed on the predicted residuals of the first step. If the test fails, a control signal is sent to the Volterra kernel estimation unit to increase the order of the Volterra series or to the embedded state construction module to increase the embedding dimension.

[0104] The working principle of this cuffless blood pressure estimation method and system is as follows:

[0105] The signal acquisition module synchronously acquires single-lead ECG signals and PEP signals from the subject through ECG electrodes and PEP sensors. After the signal preprocessing module removes baseline drift, detects feature points, and eliminates motion artifacts, the embedding state construction module uses the mutual information first minimum method to determine the optimal delay time and the pseudo nearest neighbor method or the Cao method to determine the minimum embedding dimension. Delay embedding is applied to the ECG signal and PEP signal respectively to construct a joint embedding state vector.

[0106] The nonlinear response operator reconstruction module uses the Volterra series truncation approximation to reconstruct the nonlinear response operator from the joint embedded state vector sequence in each segment, and the operator consistency verification unit ensures the operator accuracy through Ljung-Box whitening test;

[0107] The Lie-Jordan decomposition and spectral analysis module decomposes the nonlinear response operator into Jordan and Lie parts, performs spectral decomposition on the Jordan part, and extracts the real part, imaginary part, and response amplitude coefficient of each eigenmode.

[0108] During the calibration phase, the calibration module receives reference systolic and reference diastolic blood pressure measurements from the certified cuff blood pressure monitor through the calibration data interface unit. The sensitive mode sparse selection unit uses the real, imaginary parts, and response amplitude coefficients of all intrinsic modes as candidate features, and performs L1 regularized regression with reference systolic and reference diastolic blood pressure as target variables to automatically select the systolic and diastolic sensitive mode sets. The calibration model freezing unit then freezes and stores the systolic and diastolic calibration models in the storage module.

[0109] In the online estimation phase, the online blood pressure estimation module intercepts signal segments from the electrocardiogram signals and photoplethysmography signals during daily wearing through the sliding window unit, the real-time feature extraction unit extracts the real part, imaginary part and response amplitude coefficient of the eigenvalues corresponding to the systolic blood pressure sensitive mode set and the diastolic blood pressure sensitive mode set through the embedded state construction module, the nonlinear response operator reconstruction module and the Lie-Jordan decomposition and spectral analysis module, the blood pressure calculation unit substitutes the frozen systolic blood pressure calibration model and diastolic blood pressure calibration model to calculate the systolic blood pressure estimated value and the diastolic blood pressure estimated value, the confidence interval calculation unit calculates the confidence interval of the estimated value based on the residual standard deviation of the leave-one-out cross-validation of the calibration segment, and the output format unit organizes it into ( )mmHg format for output;

[0110] Meanwhile, the universal bound output gating module calculates the steady-state entropy production rate estimated value, the normalized acceleration response ratio, the slowest dissipation rate and the universal theoretical upper bound of the cardiovascular system, and calculates the calibrated health index, the comparison and judgment unit controls the output gating according to the health index: when r≤0.8, output normally and mark the health zone; when 0.8<r≤1.0, output normally and add a marginal state mark; when r>1.0, suppress the output of blood pressure values and trigger an alternative action;

[0111] The time series recording unit of the calibration drift monitoring and recalibration triggering module records the time series of the calibrated health index at each online estimation, the trend analysis unit performs the Kendall tau rank correlation coefficient test on the time series of the calibrated health index, when the health index shows a monotonically increasing trend within 24 hours and the Kendall tau rank correlation coefficient test's value is less than 0.05, the first triggering unit generates a suggested recalibration reminder; when the health index is always greater than 0.6 within 48 hours, the second triggering unit generates a forced recalibration signal, and the recalibration execution unit responds to the forced recalibration signal or the user-confirmed suggested recalibration reminder, and only updates the weight coefficients of the systolic blood pressure calibration model and the diastolic blood pressure calibration model on the premise that the original systolic blood pressure sensitive mode set and diastolic blood pressure sensitive mode set are fixed;

[0112] The central processing unit is respectively connected with the signal acquisition module, the signal preprocessing module, the embedded state construction module, the nonlinear response operator reconstruction module, the Lie-Jordan decomposition and spectral analysis module, the calibration module, the online blood pressure estimation module, the universal bound output gating module and the calibration drift monitoring and recalibration triggering module to control the cooperative operation and data flow of each module, and finally realizes cuff-less continuous blood pressure estimation based on electrocardiogram signals and photoplethysmography signals, and ensures the credibility of the estimation results and the long-term stability of the system through the universal bound theory and calibration drift monitoring.

[0113] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0114] This invention can fully exploit the nonlinear dynamic information of the cardiovascular system: by constructing a joint embedded state vector through delayed embedding and using Volterra series truncation to approximate the reconstruction of the nonlinear response operator, it can more completely capture the high-dimensional nonlinear dynamic characteristics of the cardiovascular system contained in the electrocardiogram signal and photoplethysmography pulse wave signal, overcome the shortcomings of insufficient information in traditional PWTT or morphological feature methods, and significantly improve the accuracy and robustness of blood pressure estimation.

[0115] This invention enables automatic sparse identification of blood pressure-sensitive features: by decomposing the nonlinear response operator through the Lie-Jordan decomposition and spectral analysis module, the intrinsic mode features reflecting the dissipation process and coherent kinetic subprocess are extracted, and L1 regularized regression is used to automatically screen the systolic blood pressure-sensitive mode set and the diastolic blood pressure-sensitive mode set, thereby realizing the sparse selection and personalized calibration of blood pressure-related features, effectively avoiding the interference of redundant features, and improving the generalization ability of the calibration model.

[0116] This invention introduces a universal gating mechanism to ensure output reliability: Through a universal output gating module, a calibration health index is calculated using steady-state entropy yield, normalized acceleration response ratio, and slowest dissipation rate, and the output of the blood pressure estimate is gated in real time. When the kinetic system deviates from the calibration state (calibrated health index exceeds the limit), the system automatically suppresses unreliable blood pressure values, effectively preventing erroneous estimates caused by sensor offset, sudden changes in physiological state, etc., from misleading users and ensuring safety during long-term use.

[0117] This invention features adaptive recalibration capability: through the calibration drift monitoring and recalibration triggering module, it continuously tracks the time series trend of the calibration health index. When a significant drift in the system dynamics is detected, it automatically triggers a recalibration reminder or forces recalibration. During recalibration, it only updates the weight coefficients of the calibration model while keeping the sensitive model set fixed. This ensures the continuity of estimation accuracy while reducing the burden of recalibration operations on users, which is beneficial for the long-term stable use of the equipment. Attached Figure Description

[0118] Figure 1 This is a flowchart illustrating the steps involved in a cuffless blood pressure estimation method.

[0119] Figure 2 This is a system framework diagram of a cuffless blood pressure estimation system. Detailed Implementation

[0120] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0121] In the description of this invention, it should be noted that when an element is referred to as being "fixed to" or "set on" another element, it can be directly on or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to or indirectly connected to the other element.

[0122] In the description of this invention, it should be noted that the terms "center," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. "Several" means one or more, unless otherwise explicitly specified.

[0123] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0124] Example 1:

[0125] Please see Figure 1 A cuffless blood pressure estimation method includes the following steps:

[0126] Step S1: Synchronously acquire electrocardiogram (ECG) signals and photoplethysmography (PPG) pulse wave signals:

[0127] Single-lead electrocardiogram signals of subjects are collected synchronously using wearable devices or medical-grade front-ends. Photoplethysmography (PPG) signal The electrocardiogram signal The sampling rate is not less than 256Hz, and the bandpass filtering range is 0.5-40Hz; the photoplethysmography signal The sampling rate is no less than 64Hz, the bandpass filter range is 0.5-10Hz, and green light with a wavelength of 525nm or infrared light with a wavelength of 940nm is used as the light source. The hardware synchronization error between the two channels is less than one sampling period. A steady-state signal of no less than 30 seconds needs to be continuously acquired, and the subject must remain seated or lying down during the acquisition period.

[0128] Step S2: Preprocess the acquired electrocardiogram (ECG) signal and photoplethysmography (PPG) signal:

[0129] The electrocardiogram signal acquired in step S1 Baseline drift removal was performed using median filtering or 0.5Hz high-pass filtering to detect the R-peak position and extract the RR interval sequence.

[0130] The photoplethysmography signal acquired in step S1 Baseline drift removal was performed, and the peak systolic position and dicrotic notch position were detected.

[0131] When the device integrates a three-axis accelerometer, it calculates the composite amplitude of the three axes of the accelerometer and discards continuous signal segments whose composite amplitude exceeds a predetermined threshold as motion artifact segments.

[0132] The heart rate is normalized along the time axis, and the signal of each cardiac cycle is resampled to a fixed number of points; at the same time, the original non-normalized signal is retained for subsequent operator reconstruction, and the normalized signal is only used to assist feature alignment.

[0133] Step S3, Delayed Embedding and State Vector Construction:

[0134] The optimal delay time is determined using the first minimum mutual information method. The minimum embedding dimension is determined using either the pseudo-nearest neighbor method or the Cao method. The embedding dimension The typical value range is 3 to 5.

[0135] Based on the determined optimal delay time and embedding dimension For electrocardiogram signals Photoplethysmography (PPG) signal Apply delayed embeddings separately to construct a joint embedding state vector. :

[0136] ;

[0137] in, For the first Each sampling time, , The sampling interval; the joint embedded state vector Total dimensions .

[0138] According to Takens' embedding theorem, in Under the condition that the effective dynamic dimension of the cardiovascular system is greater than that of the joint embedded state vector The evolution is homeomorphic to the original cardiovascular system state, i.e., jointly embedded state vectors. It encodes all observable information about cardiovascular dynamics.

[0139] Step S4, Nonlinear Response Operator Reconstruction:

[0140] S4.1 Segmentation and Aggregation:

[0141] Jointly embedding the state vector The sequence is segmented into windows of 30-60 seconds, with each segment containing 30-80 cardiac cycles. It is assumed that the cardiovascular state is approximately stable within each segment, with blood pressure changes of less than 5 mmHg.

[0142] S4.2, Volterra kernel estimation:

[0143] The nonlinear response operator is reconstructed from the joint embedded state vector sequence within each segment using the Volterra series truncation approximation. The nonlinear response operator satisfy:

[0144] ;

[0145] in, The equivalent random perturbation vector represents the sum of model residuals, sensor noise, and unmodeled physiological fluctuations.

[0146] The nonlinear response operator is approximated by truncation using a second-order Volterra series. :

[0147] ;

[0148] When second-order accuracy is insufficient, the third-order Volterra series truncation approximation is used:

[0149] ;

[0150] in, It is a zero-order kernel. It is a first-order kernel. It is a second-order kernel. It is a third-order kernel; , , The regularization parameter is frozen in the independent calibration segment by estimation using the least squares method or regularized regression.

[0151] S4.3 Operator Consistency Verification:

[0152] The estimated nonlinear response operator Applied to joint embedding state vector Calculate the one-step prediction residual:

[0153] ;

[0154] For the predicted residual of the first step Perform Ljung-Box whitening testing. If the test fails, increase the order of the Volterra series or increase the embedding dimension. Until it passes the whitening test.

[0155] Step S5: Disposable cuff calibration and blood pressure sensor identification:

[0156] This step is performed only once upon initial use or recalibration.

[0157] S5.1 Calibration Data Acquisition:

[0158] While subjects wore electrocardiogram (ECG) sensors and photoplethysmography (PPG) sensors, reference systolic blood pressure was measured using a certified cuff blood pressure monitor. and reference diastolic pressure The cuff should be used to measure values ​​every 1-2 minutes for 5-10 minutes, obtaining a total of 3-10 sets of reference values. Electrocardiogram (ECG) signals should be recorded simultaneously. Photoplethysmography (PPG) signal .

[0159] S5.2, Lie-Jordan decomposition:

[0160] Reconstructing the nonlinear response operator from the joint embedded state vector sequence within the calibration segment The nonlinear response operator Decomposed into Jordan part and Lie part :

[0161] ;

[0162] in, The symmetrical / dissipative component corresponds to irreversible physiological dissipation processes, including vasodilation and vasoconstriction and energy dissipation. The antisymmetric / coherent part corresponds to the reversible kinetic subprocesses, including pulse propagation, reflected waves, and heart rate variability oscillations.

[0163] right Perform spectral decomposition:

[0164] ;

[0165] in, For the first One orthogonal eigenmode The total number of intrinsic moduli. ; For the first Individual eigenmode The corresponding Jordan eigenvalues, For the first The attenuation rate of each eigenmode For the first The oscillation frequency of each eigenmode.

[0166] Record the real part of each eigenmode virtual part and response amplitude coefficient .

[0167] S5.3, Sensitive Mode Sparse Selection:

[0168] Reference systolic blood pressure measured by cuff and reference diastolic pressure As the target variable, with all The real part of each eigenmode virtual part and response amplitude coefficient As candidate features, construct feature vectors and execute... Regularized Regression:

[0169] ;

[0170] ;

[0171] in, , where are the regression weighting coefficients for systolic blood pressure. The regression weights for diastolic blood pressure are... and This is the regression bias term.

[0172] L1 regularization automatically compresses the weights of irrelevant intrinsic modes to zero, and the intrinsic modes corresponding to non-zero weights form a set of contraction-sensitive modes. The intrinsic modes corresponding to non-zero weights constitute the set of diastolic pressure sensitive modes. Shrinkage pressure sensitive mode set The number of modules is 2-3, diastolic pressure sensitive module set The number of moduli is 2-3, and the sensitive moduli of the two do not completely overlap.

[0173] Freeze-throttle calibration model And diastolic pressure calibration model Stored in the device firmware:

[0174] ,in ;

[0175] ,in ;

[0176] in, and It is a sparse linear model or a low-order polynomial model.

[0177] Step S6, Online Blood Pressure Estimation:

[0178] During routine use of the device, the following data is collected from a sliding window of ECG and photoplethysmography (PPG) signals every 10-30 seconds:

[0179] S6.1 Construct the joint embedded state vector and reconstruct the nonlinear response operator according to the methods in steps S3-S4. ;

[0180] S6.2, For the nonlinear response operator Perform Lie-Jordan decomposition to extract the set of systolic blood pressure sensitive modes. and diastolic pressure sensitive mode set The corresponding real part of the eigenvalue virtual part and response amplitude coefficient ;

[0181] S6.3 Substitute the extracted features into the frozen systolic blood pressure calibration model. And diastolic pressure calibration model Calculate the estimated systolic blood pressure value And diastolic blood pressure estimates ;

[0182] S6.4. Based on the residual standard deviation of the calibration segment leave-one-out cross-validation, calculate the confidence interval of the estimated value, and output it in the form of ( ) mmHg, of which This is the confidence interval for systolic blood pressure. This represents the confidence interval for diastolic blood pressure.

[0183] Step S7, Universal Output Gating:

[0184] Simultaneously calculate the universal boundary gating parameters of the current segment:

[0185] Calculate the steady-state entropy production rate estimate of the cardiovascular system ,in , Let be the state covariance matrix. The residual covariance matrix;

[0186] Calculate the normalized acceleration ratio ,in The maximum expansion rate of the positive part of the dynamic system;

[0187] Calculate the slowest dissipation rate estimate , ,Depend on The minimum eigenvalue real part is determined;

[0188] Upper limit of the universal theory of computation ;

[0189] Calculate the calibrated health index ,when At that time, it was a healthy zone. This is a warning zone;

[0190] Set the output suppression threshold ;

[0191] Output gating satisfy: If and only if ,Right now ;

[0192] when hour, Normal output systolic blood pressure estimate And diastolic blood pressure estimates The corresponding confidence interval indicates that the system is in the healthy dynamic region and the model is reliable;

[0193] when hour, Normal output systolic blood pressure estimate And diastolic blood pressure estimates The corresponding confidence intervals are also marked, and edge states are indicated to suggest that subjects remain still when approaching the nonlinear saturation boundary.

[0194] when hour, Suppressing blood pressure readings triggers one of the following alternative actions:

[0195] (1) Output the last time The estimated blood pressure value at that time is marked as non-real-time / to be recalibrated; or

[0196] (2) Prompt the user to remain still and remeasure; or

[0197] (3) When If the value continues to rise, a recalibration alert will be triggered.

[0198] Step S8, Calibration Drift Monitoring and Recalibration Trigger:

[0199] The device records the calibration health index during each online estimation. Time series;

[0200] When calibrating health index It exhibits a monotonically increasing trend over 24 hours, according to the Kendall-Tau rank correlation coefficient test. When the value is less than 0.05, a recalibration reminder is triggered;

[0201] When calibrating health index If the value remains greater than 0.6 for a continuous period of 48 hours, recalibration is mandatory.

[0202] During recalibration, the original systolic pressure sensitive mode set... and diastolic pressure sensitive mode set Under the premise of keeping the systolic blood pressure calibration model constant, only update the systolic blood pressure calibration model. And diastolic pressure calibration model The weighting coefficients.

[0203] The specific usage process of this cuffless blood pressure estimation method is divided into two main stages: a one-time calibration process and a daily online estimation process, as detailed below:

[0204] One-time calibration procedure (only performed upon initial use or when the system prompts for recalibration):

[0205] Subjects wore a wearable device integrating an electrocardiogram (ECG) sensor and a photoplethysmography (PPG) sensor, remaining seated or lying down. After activation, the hardware synchronously acquired single-lead ECG signals for at least 30 seconds. Photoplethysmography (PPG) signal Meanwhile, if the device integrates an accelerometer, it will simultaneously record triaxial acceleration signals for motion artifact removal.

[0206] While the device collects the aforementioned physiological signals, the operator uses a certified cuff blood pressure monitor to obtain the reference systolic blood pressure (S) and reference diastolic blood pressure (D) of 3-10 subjects within a 5-10 minute calibration window at a frequency of once every 1-2 minutes.

[0207] The device performs baseline drift removal, R-peak and systolic peak detection on the raw ECG and photoplethysmography (PPG) signals acquired in step S1, and removes motion artifacts using acceleration signals. Subsequently, the system uses the mutual information method and the Cao method to determine the optimal delay time, respectively. and embedding dimension The preprocessed signal is then subjected to delayed embedding to construct a joint embedding state vector. .

[0208] The system segments the state vector sequence with a window length of 30-60 seconds, and within each segment, it uses a second-order (or third-order if necessary) Volterra series to truncate and approximate the reconstruction of the nonlinear response operator. (satisfy The kernels of each order are estimated using the least squares method. Then, the system calculates the one-step prediction residuals and performs an Ljung-Box whitening test. If the test fails, the Volterra series order is automatically increased or the embedding dimension is increased. Until verification is passed, ensure that the operator It can accurately characterize current cardiovascular dynamics.

[0209] Operators within the calibration segment Perform Lie-Jordan decomposition to obtain the Jordan part. (Dissipative component) and Lie component (Conservative part). Regarding Jordan's part. Perform spectral decomposition and extract The real part of each eigenmode virtual part and response amplitude coefficient Then, using the reference systolic blood pressure measured in step S2... and diastolic blood pressure Using the features of all the above moduli as the target value, L1 regularized regression is performed to automatically select a set of 2-3 intrinsic moduli that are most sensitive to systolic blood pressure, forming a systolic blood pressure sensitive modulus set. The set of diastolic pressure sensitive modes consists of 2-3 intrinsic modes that are most sensitive to diastolic pressure. .

[0210] The system establishes a sparse linear model of systolic blood pressure based on the selected set of sensitive moduli. And diastolic pressure sparse linear model The two models (including the sensitivity index and weight coefficients) were then written into the device firmware as frozen calibration models, completing the calibration. Afterward, subjects no longer needed to carry the cuff blood pressure monitor during daily use.

[0211] II. Daily Online Estimation Process (Circular Execution Process During Daily Device Use):

[0212] Subjects wore the device in their daily lives. The device continuously collected electrocardiogram signals in a sliding window every 10-30 seconds. Photoplethysmography (PPG) signal It also performs preprocessing such as baseline drift removal, peak detection, and motion artifact removal in real time.

[0213] For the signal within the current sliding window, the system uses the delay time determined during the calibration phase. and embedding dimension Construct the joint embedded state vector under this window. The nonlinear response operator for the current window is reconstructed again using the Volterra series method. .

[0214] Operators newly reconstructed for the current window Lie-Jordan decomposition and spectral decomposition were performed to extract the set of systolic pressure sensitive modes selected during the calibration stage. and diastolic pressure sensitive mode set The corresponding real part of the current eigenvalue virtual part and response amplitude coefficient .

[0215] Substitute the current feature values ​​extracted in the above steps into the frozen calibration model. and In the middle, the current estimated systolic blood pressure value is calculated immediately. And diastolic blood pressure estimates Simultaneously, based on the standard deviation of the residuals from the leave-one-out cross-validation method during the calibration phase, the estimated systolic blood pressure value was calculated. And diastolic blood pressure estimates The confidence interval, and with ( Output in mmHg format.

[0216] While outputting blood pressure values, the system calculates the universal gating parameter (steady-state entropy yield) for the current window. Normalized accelerated response ratio Slowest dissipation rate and the upper limit of universal theory And calculate the calibrated health index. :

[0217] like If the condition is determined to be in the healthy zone, the above blood pressure value and confidence interval will be output normally.

[0218] like It outputs the normal blood pressure value and confidence interval, but adds a "marginal state" marker and prompts the subject to remain still.

[0219] like If the system determines that the area is in an alarm zone, it will suppress the output of the current blood pressure value and instead output the blood pressure value of the previous normal state (marked as non-real-time) or prompt the user to remain still and remeasure.

[0220] The system records the r-value and its timestamp for each online estimation. If within a consecutive 24-hour period... The value shows a monotonically increasing trend (Kendall tau test). If the r-value is greater than 0.6 within 48 hours, a "recalibration suggestion" reminder will be triggered; if the r-value remains greater than 0.6 within 48 hours, recalibration will be forcibly required. During recalibration, the system will automatically return to the "one-time calibration process," but will not recalibrate within the original sensitive module set. and Without changing the underlying model, only update the weight coefficients of the regression model.

[0221] Example 2:

[0222] Please see Figures 1-2 Based on the above embodiments, this invention also provides a cuffless blood pressure estimation system, which includes a signal acquisition module, a signal preprocessing module, an embedded state construction module, a nonlinear response operator reconstruction module, a Lie-Jordan decomposition and spectral analysis module, a calibration module, an online blood pressure estimation module, a universal output gating module, a calibration drift monitoring and recalibration triggering module, a central processing unit, and a storage module.

[0223] In this embodiment, the signal acquisition module is used to simultaneously acquire the subject's single-lead electrocardiogram signal. Photoplethysmography (PPG) signal .

[0224] The signal acquisition module includes ECG electrodes, a photoplethysmography (PPG) sensor, and an analog-to-digital converter. The ECG electrodes are configured as a single lead, using either lead I or a chest lead equivalent. The PPG sensor includes a light-emitting diode (LED) and a photodiode. The LED is a 525nm green LED or a 940nm infrared LED. The photodiode receives light signals reflected or transmitted through tissue and converts them into electrical signals.

[0225] The analog-to-digital converter (ADC) is used to convert the analog electrocardiogram (ECG) signal collected by the ECG electrodes and the analog photoplethysmography (PPG) signal output by the photodiode into digital signals. The sampling frequency of the ADC is configured such that the ECG signal sampling rate is not less than 256 Hz and the PPG signal sampling rate is not less than 64 Hz, and the hardware synchronization error between the two channels is less than one sampling period.

[0226] Furthermore, the signal acquisition module is a wrist-mounted integrated device, the ECG electrodes are disposed on the surface of the device that contacts the skin of the wrist, and the photoplethysmography (PPG) sensor is disposed at the bottom of the device where it contacts the skin of the wrist.

[0227] Furthermore, the signal acquisition module also includes a triaxial accelerometer for acquiring the motion acceleration signal of the subject in order to identify and eliminate motion artifact segments.

[0228] In this embodiment, the signal preprocessing module is connected to the signal acquisition module and is used to process the acquired electrocardiogram signals. Photoplethysmography (PPG) signal Preprocessing is performed.

[0229] The signal preprocessing module includes:

[0230] ECG preprocessing unit, used for processing ECG signals The ECG preprocessing unit performs baseline drift removal, R-peak location detection, and RR interval sequence extraction. The ECG preprocessing unit uses a median filter or a high-pass filter with a cutoff frequency of 0.5 Hz to remove baseline drift.

[0231] The photoplethysmography (PPG) preprocessing unit is used to process PPG signals. Baseline drift removal, peak systolic position detection, and dicrotic notch position detection were performed.

[0232] The motion artifact removal unit is connected to the triaxial accelerometer and is used to calculate the composite amplitude of the triaxial acceleration and remove continuous signal segments with composite amplitudes exceeding a predetermined threshold as motion artifact segments.

[0233] The time axis normalization unit is used to normalize the time axis according to the heart rate, resample the signal of each cardiac cycle to a fixed number of points, and retain the original non-normalized signal for operator reconstruction.

[0234] In this embodiment, the embedded state construction module is connected to the signal preprocessing module and is used to construct a joint embedded state vector based on the preprocessed electrocardiogram signal and photoplethysmography pulse wave signal.

[0235] The embedded state construction module includes:

[0236] The parameter determination unit is used to determine the optimal delay time using the first minimum mutual information method. The minimum embedding dimension d is determined using either the pseudo-nearest neighbor method or the Cao method. The typical value range is 3 to 5.

[0237] Joint embedding constructs are used to determine the optimal delay time. and embedding dimension For electrocardiogram signals Photoplethysmography (PPG) signal Apply delayed embeddings separately to construct a joint embedding state vector. :

[0238] ;

[0239] in, For the first Each sampling time, , The sampling interval; the joint embedded state vector Total dimensions .

[0240] In this embodiment, the nonlinear response operator reconstruction module is connected to the embedded state construction module, and is used to construct the joint embedded state vector. Sequence reconstruction nonlinear response operator .

[0241] The nonlinear response operator reconstruction module includes:

[0242] Segmentation units are used to combine the joint embedded state vector The sequence is segmented into windows of 30-60 seconds, with each segment containing 30-80 cardiac cycles.

[0243] The Volterra kernel estimation unit is used to reconstruct the nonlinear response operator F from the joint embedded state vector sequence within each segment using the Volterra series truncation approximation. satisfy:

[0244] ;

[0245] in, It is an equivalent random perturbation vector.

[0246] The Volterra kernel estimation unit employs a second-order Volterra series truncation approximation:

[0247] ;

[0248] When second-order accuracy is insufficient, the third-order Volterra series truncation approximation is used:

[0249] ;

[0250] in, It is a zero-order kernel. It is a first-order kernel. It is a second-order kernel. It is a third-order kernel; , , The regularization parameter is frozen in the independent calibration segment by estimation using the least squares method or regularized regression.

[0251] The operator consistency verification unit is used to verify the estimated nonlinear response operator. Applied to joint embedding state vector Calculate the one-step prediction residual And for the predicted residual in the first step Perform the Ljung-Box whitening test. If the test fails, send a control signal to the Volterra kernel estimation unit to increase the order of the Volterra series or to the embedded state construction module to increase the embedding dimension. .

[0252] In this embodiment, the Lie-Jordan decomposition and spectral analysis module is connected to the nonlinear response operator reconstruction module, and is used to reconstruct the nonlinear response operator. Perform Lie-Jordan decomposition and spectral decomposition.

[0253] The Lie-Jordan decomposition and spectral analysis module includes:

[0254] Operator decomposition unit, used to decompose nonlinear response operators Decomposed into Jordan part and Lie part :

[0255] ;

[0256] in, The symmetrical / dissipative part corresponds to an irreversible physiological dissipation process. This is the antisymmetric / coherent part, corresponding to the reversible kinetic subprocess.

[0257] Spectral decomposition unit, used for spectral decomposition Perform spectral decomposition:

[0258] ;

[0259] in, For the first One orthogonal eigenmode The total number of intrinsic moduli. ; For the first Individual eigenmode The corresponding Jordan eigenvalues, For the first The attenuation rate of each eigenmode For the first The oscillation frequency of each eigenmode.

[0260] The feature extraction unit is used to record the real part of each intrinsic modulus. virtual part and response amplitude coefficient .

[0261] In this embodiment, the calibration module is connected to the Lie-Jordan decomposition and spectral analysis module to perform disposable cuff calibration and blood pressure sensitive model identification, and to generate a frozen calibration model.

[0262] The calibration module includes:

[0263] The calibration data interface unit is used to receive reference systolic blood pressure measurements from a certified cuff blood pressure monitor. and reference diastolic pressure The cuff blood pressure monitor measures blood pressure every 1-2 minutes for 5-10 minutes, obtaining a total of 3-10 sets of reference values.

[0264] Sensitive modulus sparse selection unit for reference systolic pressure measured by cuff. and reference diastolic pressure As the target variable, with all The real part of each eigenmode virtual part and response amplitude coefficient As candidate features, L1 regularized regression is performed to automatically select the set of systolic blood pressure sensitive modes. and diastolic pressure sensitive mode set :

[0265] ;

[0266] ;

[0267] Among them, the set of systolic pressure sensitive modes The number of modules is 2-3, diastolic pressure sensitive module set The number of moduli is 2-3, and the sensitive moduli of the two do not completely overlap.

[0268] The calibration model freeze unit is used to freeze the systolic blood pressure calibration model. And diastolic pressure calibration model Freeze and store:

[0269] ,in ;

[0270] ,in ;

[0271] in, and The sparse linear model or low-order polynomial model determined by L1 regularized regression.

[0272] In this embodiment, the online blood pressure estimation module is connected to the Lie-Jordan decomposition and spectral analysis module and the calibration module, and is used to estimate blood pressure in real time from a sliding window during daily wear of the device.

[0273] The online blood pressure estimation module includes:

[0274] The sliding window unit is used to extract a signal segment of the current analysis window length from the electrocardiogram signal and photoplethysmography pulse wave signal every 10-30 seconds.

[0275] The real-time feature extraction unit is used to extract the set of systolic pressure sensitive modes within the current sliding window through the embedded state construction module, the nonlinear response operator reconstruction module, and the Lie-Jordan decomposition and spectral analysis module. and diastolic pressure sensitive mode set The corresponding real part of the eigenvalue virtual part and response amplitude coefficient .

[0276] The blood pressure calculation unit is used to input the features extracted by the real-time feature extraction unit into the frozen systolic blood pressure calibration model. And diastolic pressure calibration model Calculate the estimated systolic blood pressure value And diastolic blood pressure estimates .

[0277] The confidence interval calculation unit is used to calculate the confidence interval of the estimated value based on the residual standard deviation of the leave-one-out cross-validation method.

[0278] The output format unit is used to organize the outputs of the blood pressure calculation unit and the confidence interval calculation unit into ( Output in mmHg format.

[0279] In this embodiment, the universal boundary output gating module is connected to the Lie-Jordan decomposition and spectral analysis module and the online blood pressure estimation module, and is used to gate the output of the blood pressure estimation value according to the universal boundary theory.

[0280] The universal output gating module includes:

[0281] The entropy production rate calculation unit is used to calculate the estimated steady-state entropy production rate of the cardiovascular system. ,in , Let be the state covariance matrix. Let be the residual covariance matrix.

[0282] The acceleration response ratio calculation unit is used to calculate the normalized acceleration response ratio. ,in This represents the maximum expansion rate of the positive part of the dynamic system.

[0283] Dissipation rate estimation unit, used to estimate the slowest dissipation rate. , ,Depend on The minimum eigenvalue real part is determined.

[0284] Universal upper bound computation unit, used to calculate the universal theoretical upper bound. .

[0285] Health index calculation unit, used to calculate calibrated health index .

[0286] The comparison and determination unit is used to set the output suppression threshold. and determine the output gating : If and only if ,Right now ;

[0287] when When the system is determined to be in the healthy dynamic region, the output gate is activated. Control the online blood pressure estimation module to output the estimated systolic blood pressure value normally. And diastolic blood pressure estimates and the corresponding confidence interval;

[0288] when At that time, output gating Control the online blood pressure estimation module to output the estimated systolic blood pressure value normally. And diastolic blood pressure estimates The corresponding confidence interval is included, along with an output edge state marker.

[0289] when At that time, output gating The online blood pressure estimation module is controlled to suppress the output of blood pressure values ​​and trigger one of the following alternative actions:

[0290] (1) Output the last time Blood pressure estimates at that time are marked as non-real-time / to be recalibrated; or

[0291] (2) Generate a signal prompting the user to remain still and remeasure; or

[0292] (3) When calibrating health index When the value continues to rise, a recalibration reminder signal is generated.

[0293] In this embodiment, the calibration drift monitoring and recalibration triggering module is connected to the universal output gating module and is used to monitor the calibration health index. The time series is used to trigger recalibration when conditions are met.

[0294] The calibration drift monitoring and recalibration triggering module includes:

[0295] Time series recording unit, used to record calibrated health indices at each online estimation. The corresponding timestamps are used to generate a time series of calibrated health indices.

[0296] The trend analysis unit is used to perform Kendall-Tau rank correlation coefficient tests on the time series of the calibrated health index to determine the calibrated health index. Does it show a monotonous upward trend?

[0297] The first trigger unit is used when calibrating health indices. It shows a monotonically increasing trend over 24 hours and is tested by the Kendall-Tau rank correlation coefficient. When the value is less than 0.05, a recalibration reminder signal is generated.

[0298] The second trigger unit is used when calibrating health indices. If the value remains greater than 0.6 for 48 hours, a forced recalibration signal is generated.

[0299] The recalibration execution unit is used to respond to forced recalibration signals or user-confirmed recalibration reminders, within the existing systolic blood pressure sensitive module set. and diastolic pressure sensitive mode set Under the premise of keeping the systolic blood pressure calibration model constant, only update the systolic blood pressure calibration model. And diastolic pressure calibration model The weighting coefficients.

[0300] In this embodiment, the central processing unit is connected to the signal acquisition module, signal preprocessing module, embedded state construction module, nonlinear response operator reconstruction module, Lie-Jordan decomposition and spectral analysis module, calibration module, online blood pressure estimation module, universal output gating module, and calibration drift monitoring and recalibration triggering module, respectively, to control the collaborative work of each module and data flow.

[0301] The storage module is connected to the central processing unit and is used to store the following data:

[0302] Calibration data, including reference systolic blood pressure Reference diastolic blood pressure And synchronously recorded electrocardiogram signals and photoplethysmography (PPG) pulse wave signals;

[0303] Frozen calibration model parameters, including the systolic blood pressure sensitive model set. Diastolic pressure sensitive mode set Systolic blood pressure calibration model Weighting coefficients and diastolic blood pressure calibration model Weighting coefficients;

[0304] Calibrate the time series of health indices;

[0305] The user's historical blood pressure estimates and corresponding confidence intervals.

[0306] In this embodiment, the system is integrated into a wrist-worn wearable device, the ECG electrodes are disposed on the surface of the device housing that contacts the wrist, and the photoplethysmography (PPG) sensor is disposed at the bottom of the device.

[0307] Furthermore, the system also includes a display module for displaying blood pressure estimates, confidence intervals, edge status markers, insufficient signal quality prompts, and recalibration reminders.

[0308] Furthermore, the system also includes a wireless communication module for transmitting blood pressure estimates, confidence intervals, and calibrated health indices to external terminal devices or cloud servers.

[0309] The specific workflow of this cuffless blood pressure estimation system is a concrete mapping and data flow-driven process between the various physical modules of the above-mentioned method. Under the coordination of the central processing unit, the system operates in two modes: "calibration mode" and "online mode." Specifically:

[0310] System calibration mode workflow (upon initial power-on or upon receiving a recalibration command):

[0311] The operator uses the calibration data interface unit to input 3-10 sets of reference systolic blood pressure readings from the cuff blood pressure monitor. and diastolic blood pressure Input system. Simultaneously, the ECG electrodes, photoplethysmography (PPG) sensor, and analog-to-digital converter in the signal acquisition module synchronously acquire the subject's ECG signals. Photoplethysmography (PPG) signal The digital signal is collected by a triaxial accelerometer, which synchronously acquires the motion signal.

[0312] The ECG preprocessing unit and photoplethysmography (PPG) preprocessing unit of the signal preprocessing module filter and detect feature points on the original signal, respectively; the motion artifact removal unit uses acceleration data to remove motion interference segments; the time axis normalization unit completes cardiac cycle alignment and sends the processed clean signal to the embedded state construction module.

[0313] The parameter determination unit of the embedded state construction module calculates the optimal delay time and embedding dimension, and the joint embedding construction unit generates the joint embedded state vector. After the piecewise unit of the nonlinear response operator reconstruction module segments the vector, its Volterra kernel estimation unit constructs the nonlinear response operator through least squares estimation. The operator accuracy is ensured by the operator consistency verification unit through residual whitening test.

[0314] The Lie-Jordan decomposition and spectral analysis module for Decomposition and spectral analysis were performed to extract features from all intrinsic modes. The sensitive mode sparse selection unit of the calibration module, targeting reference blood pressure, was used to screen the systolic blood pressure sensitive mode set through L1 regression. and diastolic pressure sensitive mode set Subsequently, the calibration model freezing unit will generate a systolic blood pressure calibration model. And diastolic pressure calibration model The model parameters (including sensitive model indexes) are written to the storage module to complete the system calibration configuration.

[0315] Workflow in online estimation mode (daily wear phase after calibration):

[0316] During daily wear, the signal acquisition module continuously acquires physiological signals in a sliding window of 10-30 seconds. After being processed in real time by the signal preprocessing module, the signals are sent to the embedded state construction module and the nonlinear response operator reconstruction module to dynamically generate the nonlinear response operator F for the current window.

[0317] The Lie-Jordan decomposition and spectral analysis module decomposes the current operator, and its feature extraction unit reads only the frozen set of systolic pressure sensitive modes from the storage module. and diastolic pressure sensitive mode set And extract the current eigenvalues ​​(real part, imaginary part, amplitude coefficient) of these specific modes.

[0318] The blood pressure calculation unit of the online blood pressure estimation module reads the frozen systolic blood pressure calibration model from the storage module. And diastolic pressure calibration model Substitute the current feature values ​​into the model to calculate the estimated systolic blood pressure value. And diastolic blood pressure estimates Its confidence interval calculation unit calculates the confidence interval simultaneously, and the output format unit performs standardized encapsulation.

[0319] The universal output gating module calculates the entropy productivity, speedup response ratio, and calibration health index of the current window in real time. The comparison and determination unit is based on... The value determines the output strategy:

[0320] like The online blood pressure estimation module allows the display module to output blood pressure values ​​and confidence intervals. (with edge markers).

[0321] like If the gating is activated, the blood pressure value output will be suppressed, and alternative actions will be triggered (such as displaying historical values ​​or prompting for retesting).

[0322] The time-series recording unit of the calibration drift monitoring and recalibration triggering module continuously records the r-value for each estimation. The trend analysis unit performs a Kendall tau test. When an upward trend is met... When the value remains high (greater than 0.6 for 48 hours), the first trigger unit and the second trigger unit generate an alert signal or a forced command, which is then displayed to the user and the central processing unit is controlled to switch the system state back to calibration mode.

[0323] If the device integrates a wireless communication module, all the above-mentioned blood pressure estimates, confidence intervals, and calibrated health indices will be available. Both the event logs and other logs can be transmitted to external terminals or the cloud in real time for users or doctors to view and manage remotely.

[0324] The solution in this embodiment can be selectively combined with solutions in other embodiments.

[0325] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of this invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of this invention, or equivalent structural, procedural, or functional transformations made using the description and drawings of this invention, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of protection of this invention.

Claims

1. A cuffless blood pressure estimation method, characterized in that, Includes the following steps: Step S1: Synchronously acquire electrocardiogram (ECG) signals and photoplethysmography (PPG) pulse wave signals; Step S2: Preprocess the acquired electrocardiogram signal and photoplethysmography (PPG) signal; Step S3, Delay Embedding and State Vector Construction: The optimal delay time is determined using the first minimum mutual information method, and the minimum embedding dimension is determined using the pseudo nearest neighbor method or the Cao method. Based on the determined optimal delay time and embedding dimension, delay embedding is applied to the ECG signal and the photoplethysmography pulse wave signal respectively to construct a joint embedding state vector. Step S4, Reconstruction of the nonlinear response operator: Divide the sequence of joint embedded state vectors into segments with a window length of 30-60 seconds, and use the Volterra series truncation approximation to reconstruct the nonlinear response operator in the joint embedded state vector sequence of each segment; Step S5, disposable cuff calibration and blood pressure sensitive mode identification: while wearing the ECG sensor and photoplethysmography pulse wave sensor, use the cuff blood pressure monitor to measure the reference systolic blood pressure and reference diastolic blood pressure. With the reference systolic blood pressure and reference diastolic blood pressure as target variables and the real part, imaginary part and response amplitude coefficient of all intrinsic modes as candidate features, perform L1 regularized regression to automatically select the systolic blood pressure sensitive mode set and the diastolic blood pressure sensitive mode set. Step S6, Online Blood Pressure Estimation: Perform Lie-Jordan decomposition on the nonlinear response operator and extract the real and imaginary parts of the eigenvalues ​​and response amplitude coefficients corresponding to the systolic and diastolic pressure sensitive mode sets. Substitute the extracted features into the frozen systolic and diastolic pressure calibration models to calculate the estimated systolic and diastolic blood pressure values. Step S7, Universal Boundary Output Gating: Simultaneously calculate the steady-state entropy yield estimate, normalized acceleration response ratio, and slowest dissipation rate estimate for the current segment, and calculate the universal theoretical bound and calibration health index; Step S8, Calibration Drift Monitoring and Recalibration Trigger: The device records the time series of the calibration health index during each online estimation. When the calibration health index shows a monotonically increasing trend within 24 hours and the p-value of the Kendall tau rank correlation coefficient test is less than 0.05, a recalibration reminder is triggered. When the calibration health index is always greater than 0.6 within 48 hours, recalibration is mandatory.

2. The cuffless blood pressure estimation method according to claim 1, characterized in that, In step S1: wearable devices or medical-grade front-ends are used to simultaneously collect the subject's single-lead electrocardiogram signal and photoplethysmography (PPG) wave signal. The sampling rate of the electrocardiogram signal is not less than 256Hz, and the bandpass filtering range is 0.5-40Hz; The sampling rate of the photoplethysmography signal is not less than 64Hz, the bandpass filtering range is 0.5-10Hz, and green light with a wavelength of 525nm or infrared light with a wavelength of 940nm is used as the light source. The hardware synchronization error between the two channels is less than one sampling period; A steady-state signal was continuously collected for no less than 30 seconds, during which the subject remained seated or lying down.

3. The cuffless blood pressure estimation method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.1 Segmentation and Aggregation: The sequence of jointly embedded state vectors is segmented into segments with a window length of 30-60 seconds, and each segment contains 30-80 cardiac cycles; S4.2 Volterra kernel estimation: The nonlinear response operator is reconstructed from the joint embedded state vector sequence in each segment using the Volterra series truncation approximation. The nonlinear response operator satisfies: ； in, It is an equivalent random perturbation vector; The nonlinear response operator is approximated by truncation using a second-order Volterra series: ； When second-order accuracy is insufficient, the third-order Volterra series truncation approximation is used: ； in, It is a zero-order kernel. It is a first-order kernel. It is a second-order kernel. It is a third-order kernel; , , The regularization parameter is frozen in the independent calibration segment by estimation using the least squares method or regularized regression. S4.3 Operator Consistency Verification: Apply the estimated nonlinear response operator F to the joint embedded state vector and calculate the one-step prediction residual. ; For the predicted residual of the first step Perform the Ljung-Box whitening test. If the test fails, increase the order of the Volterra series or increase the embedding dimension until the whitening test is passed.

4. The cuffless blood pressure estimation method according to claim 1, characterized in that, In step S5, the disposable cuff calibration and blood pressure sensor identification are performed only once during initial use or recalibration, and include the following sub-steps: S5.1 Calibration Data Acquisition: While the subjects wore the ECG sensor and photoplethysmography pulse wave sensor, reference systolic blood pressure and reference diastolic blood pressure were measured using a certified cuff blood pressure monitor. The cuff measurement frequency is once every 1-2 minutes, lasting for 5-10 minutes, and a total of 3-10 sets of reference values ​​are obtained; ECG signals and photoplethysmography (PPG) signals are recorded simultaneously. S5.2, Lie-Jordan Decomposition: Reconstruct the nonlinear response operator from the joint embedded state vector sequence within the calibration segment, and decompose the nonlinear response operator into Jordan parts. and Lie part : ； in, The symmetrical / dissipative component corresponds to an irreversible physiological dissipation process. This is the antisymmetric / coherent part, corresponding to the reversible dynamic subprocess; right Perform spectral decomposition: ； in, For the first One orthogonal eigenmode The total number of intrinsic moduli. ; For the first Individual eigenmode The corresponding Jordan eigenvalues, For the first The attenuation rate of each eigenmode For the first The oscillation frequency of each eigenmode; Record the real part of each eigenmode virtual part and response amplitude coefficient ; S5.3 Sensitive modulus sparse selection: Using the reference systolic blood pressure and reference diastolic blood pressure measured by the cuff as the target variables, and using all... The real part of each eigenmode virtual part and response amplitude coefficient As candidate features, L1 regularized regression is performed to automatically select the set of systolic blood pressure sensitive modes. and diastolic pressure sensitive mode set : ； ； Among them, the set of systolic pressure sensitive modes The number of modules is 2-3, diastolic pressure sensitive module set The number of moduli is 2-3, and the sensitive moduli of the two do not completely overlap; Freeze-throttle calibration model And diastolic pressure calibration model Stored in the device firmware: ,in ; ,in ; in, and It is a sparse linear model or a low-order polynomial model.

5. The cuffless blood pressure estimation method according to claim 1, characterized in that, In step S6, during daily wear of the device, the following data is collected from a sliding window of ECG and photoplethysmography (PPG) signals every 10-30 seconds: S6.1 Construct the joint embedded state vector and reconstruct the nonlinear response operator according to the methods in steps S3-S4; S6.2 Perform Lie-Jordan decomposition on the nonlinear response operator to extract the set of systolic pressure sensitive modes. and diastolic pressure sensitive mode set The corresponding real part of the eigenvalue virtual part and response amplitude coefficient ; S6.3 Substitute the extracted features into the frozen systolic blood pressure calibration model. And diastolic pressure calibration model Calculate the estimated systolic blood pressure value And diastolic blood pressure estimates ; S6.

4. Based on the residual standard deviation of the calibration segment leave-one-out cross-validation, calculate the confidence interval of the estimated value, and output it in the form of ( ±CIS, ±CID) mmHg, where CIS is the confidence interval for systolic blood pressure and CID is the confidence interval for diastolic blood pressure.

6. The cuffless blood pressure estimation method according to claim 1, characterized in that, In step S7: the universal boundary gating parameters of the current segment are calculated synchronously. Calculate the steady-state entropy production rate estimate of the cardiovascular system ,in , Let be the state covariance matrix. The residual covariance matrix; Calculate the normalized acceleration ratio ,in The maximum expansion rate of the positive part of the dynamic system; Calculate the slowest dissipation rate estimate , ,Depend on The minimum eigenvalue real part is determined; Upper limit of computational universality theory ; Calculate the calibrated health index ; Set the output suppression threshold ; Output gating satisfy: If and only if ,Right now ; when hour, Normal output systolic blood pressure estimate And diastolic blood pressure estimates and the corresponding confidence interval, when hour, Normal output systolic blood pressure estimate And diastolic blood pressure estimates And the corresponding confidence interval, and mark the edge states; when hour, Suppressing blood pressure readings triggers one of the following alternative actions: (1) Output the last time The estimated blood pressure value at that time is marked as non-real-time / to be recalibrated; or (2) Prompt the user to remain still and remeasure; or (3) When If the value continues to rise, a recalibration alert will be triggered.

7. The cuffless blood pressure estimation method according to claim 1, characterized in that, In step S8: the device records the calibration health index during each online estimation. Time series; When the calibrated health index shows a monotonically increasing trend over 24 hours, the Kendall tau rank correlation coefficient test... When the value is less than 0.05, a recalibration reminder is triggered; If the calibration health index remains greater than 0.6 for a continuous period of 48 hours, recalibration is mandatory. During recalibration, the original systolic pressure sensitive mode set... and diastolic pressure sensitive mode set Under the premise of keeping the systolic blood pressure calibration model constant, only update the systolic blood pressure calibration model. And diastolic pressure calibration model The weighting coefficients.

8. A cuffless blood pressure estimation system for performing the cuffless blood pressure estimation method as described in any one of claims 1-7, characterized in that, The system includes: The signal acquisition module is used to simultaneously acquire the subject's single-lead electrocardiogram signal and photoplethysmography pulse wave signal; A signal preprocessing module, connected to the signal acquisition module, is used to preprocess the acquired electrocardiogram signals and photoplethysmography (PPG) pulse wave signals. An embedded state construction module, connected to the signal preprocessing module, is used to construct a joint embedded state vector based on the preprocessed electrocardiogram signal and photoplethysmography pulse wave signal. A nonlinear response operator reconstruction module, connected to the embedded state construction module, is used to reconstruct the nonlinear response operator based on the sequence of the joint embedded state vector; The Lie-Jordan decomposition and spectral analysis module is connected to the nonlinear response operator reconstruction module and is used to perform Lie-Jordan decomposition and spectral decomposition on the nonlinear response operator. The calibration module, connected to the Lie-Jordan decomposition and spectral analysis module, is used to perform disposable cuff calibration and blood pressure sensitive model identification, and generate a frozen calibration model; An online blood pressure estimation module, connected to the Lie-Jordan decomposition and spectral analysis module and the calibration module, is used to estimate blood pressure in real time from a sliding window during daily wear of the device. The universal boundary output gating module is connected to the Lie-Jordan decomposition and spectral analysis module and the online blood pressure estimation module, and is used to gating the output of the blood pressure estimation value according to the universal boundary theory. The calibration drift monitoring and recalibration triggering module is connected to the universal output gating module and is used to monitor the time series of calibration health index and trigger recalibration when conditions are met. The central processing unit is connected to each of the above modules and is used to control the collaborative work of the modules and the data flow. The storage module, connected to the central processing unit, is used to store calibration data, frozen calibration model parameters, calibration health index time series, and the user's historical blood pressure estimation results and corresponding confidence intervals.

9. The cuffless blood pressure estimation system according to claim 8, characterized in that, The signal acquisition module includes ECG electrodes, a photoplethysmography (PPG) sensor, and an analog-to-digital converter. The ECG electrodes are configured as single leads, using an equivalent method of lead I or chest leads; the photoplethysmography (PPG) sensor includes a light-emitting diode (LED) and a photodiode. The LED is a green LED with a wavelength of 525nm or an infrared LED with a wavelength of 940nm, and the photodiode is used to receive light signals reflected or transmitted through tissue and convert them into electrical signals. The analog-to-digital converter is used to convert the analog electrocardiogram (ECG) signal collected by the ECG electrodes and the analog photoplethysmography (PPG) signal output by the photodiode into digital signals. The sampling frequency of the ECG signal is configured to be no less than 256 Hz and the PPG signal sampling rate is no less than 64 Hz, and the hardware synchronization error between the two channels is less than one sampling period. The signal acquisition module is a wrist-type integrated device, the ECG electrodes are placed on the surface of the device that contacts the skin of the wrist, and the photoplethysmography (PPG) sensor is placed at the bottom of the device where it contacts the skin of the wrist. The signal acquisition module also includes a triaxial accelerometer for acquiring the subject's motion acceleration signals to identify and eliminate motion artifact segments.

10. The cuffless blood pressure estimation system according to claim 8, characterized in that, The nonlinear response operator reconstruction module includes: The segmentation unit is used to segment the sequence of joint embedded state vectors into segments with a window length of 30-60 seconds, and each segment contains 30-80 cardiac cycles. The Volterra kernel estimation unit is used to reconstruct the nonlinear response operator F from the joint embedded state vector sequence within each segment using the Volterra series truncation approximation, wherein the nonlinear response operator satisfies: ； in, It is an equivalent random perturbation vector; The Volterra kernel estimation unit employs a second-order Volterra series truncation approximation: ； When second-order accuracy is insufficient, the third-order Volterra series truncation approximation is used: ； in, It is a zero-order kernel. It is a first-order kernel. It is a second-order kernel. It is a third-order kernel; , , The regularization parameter is frozen in the independent calibration segment by estimation using the least squares method or regularized regression. The operator consistency verification unit is used to apply the estimated nonlinear response operator F to the joint embedded state vector and calculate the one-step prediction residual. The Ljung-Box whitening test is performed on the predicted residuals of the first step. If the test fails, a control signal is sent to the Volterra kernel estimation unit to increase the order of the Volterra series or to the embedded state construction module to increase the embedding dimension.

Images