Blood pressure measuring method

Abstract

The invention discloses a blood pressure measuring method, and belongs to the technical field of medical health monitoring. Comprising the following steps: synchronously acquiring an ECG signal and a PPG signal, and preprocessing the ECG signal and the PPG signal; performing denoising and feature extraction on the ECG signal and the PPG signal; a weighted kernel function regression model fused with multi-modal features is constructed, and a blood pressure predicted value is generated in combination with random forest feature selection and a piecewise linear model dynamic switching mechanism; the model weight is updated based on a recursive least square method, dual-mode calibration is carried out by adopting an event triggering and period triggering mixed mechanism, and a final blood pressure value is obtained through global fine tuning. Compared with the prior art, the method has the beneficial effect that the accurate prediction of the blood pressure can be realized by combining the regression analysis principle with the individual physiological characteristics.

Description

Technical Field

[0001] The present invention relates to the field of medical health monitoring technology, and more specifically, to a blood pressure measurement method. Background Art

[0002] As a core physiological parameter for evaluating the function of the cardiovascular system, accurate monitoring of blood pressure is of great significance for the diagnosis of hypertension, early warning of cardiovascular disease risks, and evaluation of treatment effects. The limitations of traditional blood pressure measurement methods that rely on cuff-type sphygmomanometers are becoming increasingly prominent. First, non-continuous single measurements cannot capture dynamic changes in blood pressure such as circadian rhythms, exercise, or mood swings, and abnormal dynamic blood pressure (such as nocturnal hypertension) is closely related to the risk of cardiovascular events; second, repeated inflation and deflation of the cuff leads to poor comfort, which is especially unsuitable for long-term monitoring (such as ICU patients or sleep studies); finally, the large size and complex operation of the device lead to insufficient portability, making it difficult to meet the needs of daily health management or wearable devices.

[0003] In recent years, non-invasive blood pressure monitoring technology based on photoplethysmography (PPG) and electrocardiogram (ECG) signals has gradually become a research hotspot. In related technologies, for example, Chinese patent CN118697311A provides an intelligent blood pressure prediction method and system based on human PPG signals, which uses training sets and test sets to train the initial blood pressure prediction network TCN model, deploys the trained blood pressure prediction network TCN model to the wearable carrier, and uses photoplethysmography technology to collect human PPG signals from human tissues, and then normalizes and processes them before transmitting them to the blood pressure prediction network TCN model for blood pressure prediction; another example is Chinese patent CN118161139A, which discloses a blood pressure measurement method based on PPG signals, including obtaining PPG signals. and physiological information, segment the PPG signal to obtain multiple pulse waves and extract signal features, preset multiple regression models, and input each pulse wave into the regression model corresponding to the pulse wave classification according to the pulse wave classification to obtain the predicted blood pressure value; for example, Chinese patent CN117100239A provides a blood pressure continuous measurement method based on PPG and its multi-order differential signals, pre-processes the original PPG signal, eliminates noise and useless signals, segments the PPG signal and converts it into a time series signal, then uses a convolutional neural network for feature extraction, and uses a long short-term memory network to automatically estimate blood pressure. These methods have made progress in feature extraction, signal processing and model building, but still have the following significant defects: (1) Single signal analysis dimension: The synergistic complementarity of multimodal signals has not been fully explored, and the advantages of multi-signal fusion have not been fully utilized; (2) Poor individual adaptability: Existing methods lack consideration of individual differences (such as age, gender and weight, etc.) and the physiological state of the same individual at different times. Static parameter correction is difficult to cover dynamic physiological changes between and within individuals, resulting in insufficient universality of the prediction model; (3) The model is complex and inefficient: The model is highly complex and computationally intensive, making it difficult to achieve real-time monitoring on mobile devices; (4) Weak anti-interference ability: The signal noise interference problem has not been effectively solved. Although the existing PPG signal processing technology uses filtering algorithms, it has limited effect on suppressing common interference such as motion artifacts and photoelectric sensor offset. In particular, the signal quality is easily affected in dynamic monitoring scenarios, resulting in distortion of key feature extraction and affecting the stability and reliability of the measurement. Summary of the invention

[0004] 1. Technical issues to be solved In order to solve the problem in the prior art of how to suppress the interference of signal noise on key feature extraction and improve the accuracy and stability of blood pressure prediction, the present invention provides a blood pressure measurement method, which can realize accurate prediction of blood pressure through regression analysis principle combined with individual physiological characteristics.

[0005] 2. Technical solution The purpose of the present invention is achieved through the following technical solutions.

[0006] The content of this application is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this application is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0007] Some embodiments of the present application provide a blood pressure measurement method, comprising the following steps: synchronously acquiring an ECG signal and a PPG signal, preprocessing the ECG signal and the PPG signal, and aligning the timestamps of the ECG signal and the PPG signal using a synchronous trigger mechanism; The ECG and PPG signals are denoised based on discrete wavelet transform, the starting point of the ECG signal R peak and the PPG signal main peak are aligned through the dynamic time warping algorithm, and the features of the ECG and PPG signals are extracted; A weighted kernel function regression model integrating multimodal features was constructed, and the blood pressure prediction value was generated by combining random forest feature selection with the dynamic switching mechanism of the piecewise linear model. The model weights are updated based on the recursive least squares method, and a hybrid mechanism of event triggering and period triggering is used for dual-mode calibration. The final blood pressure value is obtained through global fine-tuning.

[0008] Furthermore, the synchronization trigger mechanism includes: controlling the ADC sampling clocks of the ECG signal and the PPG signal through a field programmable gate array to achieve hardware-level synchronization; The PPG signal acquisition is started with the R peak of the ECG signal as the trigger signal, and a unified timestamp and synchronization flag are added to each frame of the signal.

[0009] Furthermore, the ECG signal preprocessing process includes: using a preamplifier to amplify the signal; eliminating baseline drift and high-frequency noise through a bandpass filter, and combining a notch filter to suppress power frequency interference; converting the collected ECG signal from an analog signal to a digital signal through an analog-to-digital converter.

[0010] Furthermore, the discrete wavelet transform process includes: selecting the db4 wavelet basis, performing 5-layer decomposition on the ECG signal and the PPG signal, respectively, calculating the median absolute deviation for the approximation coefficient and detail coefficient of each layer, and estimating the noise variance; The optimal threshold is determined based on the SUREShrink risk function, the soft threshold function is used for coefficient processing, and the denoised signal is reconstructed by inverse discrete wavelet transform.

[0011] Furthermore, feature extraction of ECG signals includes: obtaining the R wave position by R wave peak detection, calculating the average RR interval, average heart rate, and heart rate variability; The process of obtaining the R wave position includes: taking the derivative of the filtered ECG signal, squaring the derivative, and then smoothing the signal using a moving window integral to obtain a curve representing the QRS complex; Set the dynamic threshold, detect the peak value in the curve, and determine the R wave position of the ECG signal.

[0012] Furthermore, the average RR interval is obtained The expression is: ; in, is the time interval between two adjacent R waves, in seconds; Get average heart rate The expression is: .

[0013] Furthermore, feature extraction of PPG signals includes obtaining the average pulse rate and pulse wave transmission time; Get average pulse rate The process is expressed as: ; is the total number of PPG cycles, in seconds; is the serial number of the PPG cycle; The process of obtaining the pulse wave transmission time is expressed as: ; in, is the time point of the R wave in the ECG signal, It is the time point at which the corresponding pulse wave arrives in the PPG signal.

[0014] Furthermore, feature extraction of the PPG signal also includes obtaining a rising interval and a falling interval; Get the slope of the rising interval The process is expressed as: ; is the amplitude difference from the lowest point to the highest point in a single PPG cycle, For the time difference.

[0015] Get the slope of the descending interval The process is expressed as: ; is the amplitude difference from the highest point to the PPG end point in a single PPG cycle, For the time difference.

[0016] Furthermore, the dynamic switching mechanism of the piecewise linear model includes: when the activity state indicator MAI=0, the linear weight vector in the static state and the bias term in the static state are adopted; When MAI=1, switch to the linear weight vector in motion state and the bias term in motion state.

[0017] Furthermore, the dual-mode calibration mechanism includes an initial calibration and a periodic calibration; the initial calibration initializes the model weights based on the mean of more than two standard blood pressure measurements in a resting state; Periodic calibration triggers the update of model parameters when the preset cumulative running time is reached.

[0018] 3. Beneficial effects Compared with the prior art, the advantages of the present invention are: (1) The physiological relevance of blood pressure prediction is significantly improved by fusion of multi-scale features of ECG and PPG signals (RR interval, HRV, PAT and pulse waveform slope) and nonlinear coupling modeling of individual physiological parameters (age, gender and BMI). Compared with the single signal model, the prediction error of the fusion model is reduced, solving the problem of prediction bias caused by signal noise interference or individual differences in traditional methods. (2) Through the dual-mode calibration mechanism (initial parameter calibration + residual triggered dynamic update) and Top-5 feature screening (based on random forest), the model's adaptability to subjects across age, gender, and BMI gradients is significantly enhanced; by combining individual difference information, the method is applicable to subjects of different ages, genders, and weights, improving individual adaptability and the universality of the prediction model; (3) Through piecewise linear model conversion (adaptive switching between static and moving states) and feature dimension compression, the model computational complexity is reduced and power consumption is reduced; the lightweight model design reduces computational complexity, enabling real-time blood pressure monitoring on mobile devices to meet daily health management needs; (4) Through the 5-layer db4 wavelet decomposition combined with SUREShrink threshold denoising, ECG baseline drift and PPG motion artifacts are effectively suppressed; combined with dynamic time warping (DTW) alignment, the problem of feature offset caused by physiological delay or noise is solved; the optimized signal processing algorithm effectively reduces noise interference, improves the stability and reliability of measurement, and enhances the practicality of the method; (5) Through the online recursive least squares (RLS) method and the dual-cycle calibration strategy, the model's adaptability to long-term physiological changes (such as changes in exercise habits and weight fluctuations) is significantly improved. Through the incremental learning mechanism, historical data are used to dynamically update the model parameters to adapt to long-term physiological changes, further improving the model's applicability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of the steps of a blood pressure measurement method according to an embodiment of the present invention; Figure 2 Schematic diagram of a blood pressure measurement method according to an embodiment of the present invention; Figure 3 FIG. 4 is a schematic diagram of an adaptive dual-mode calibration process in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0021] Combination Figures 1 to 3 A blood pressure measurement method of the present invention comprises the following steps: Signal acquisition and processing: synchronously acquire ECG and PPG signals, pre-process them, and use a synchronous trigger mechanism to align the timestamps of ECG and PPG signals; Denoising and multimodal feature extraction: Denoise the ECG and PPG signals based on discrete wavelet transform, align the starting point of the ECG signal R peak with the PPG signal main peak through the dynamic time warping algorithm, and extract features from the ECG and PPG signals; Modeling and personalized regression: Construct a weighted kernel function regression model that integrates multimodal features, combine random forest feature selection with the dynamic switching mechanism of the piecewise linear model, and generate blood pressure prediction values; Incremental learning and dynamic calibration: The model weights are updated based on the recursive least squares method, and a hybrid mechanism of event triggering and period triggering is used for dual-mode calibration. The final blood pressure value is obtained through global fine-tuning.

[0022] In a specific embodiment, the specific process of the blood pressure measurement method is as follows: S1. Signal acquisition and processing Collect multimodal signals and physiological information, and preprocess the collected ECG signals and PPG signals.

[0023] Specifically, the multimodal signal includes an ECG signal (electrocardiogram signal) and a PPG signal (photoplethysmography signal).

[0024] In this embodiment, ECG signals are captured by ECG sensors. ECG sensors use dry electrodes or wet electrodes. ECG signals are cardiac electrical activity signals that are used to reflect the myocardial depolarization and repolarization process. ECG (electrocardiogram) is a technology that uses an electrocardiograph to record the electrical activity change graph generated by each cardiac cycle from the body surface.

[0025] The choice of wet electrodes and dry electrodes is aimed at different scenarios. Wet electrodes (WetElectrodes) reduce skin impedance through conductive gel and directly contact the skin surface to collect ECG signals. They have a high signal-to-noise ratio suitable for clinical high-precision measurements, but long-term wearing may cause skin irritation and are more suitable for short-term use. Dry electrodes (DryElectrodes) use metal (such as stainless steel or gold plating) or conductive polymer materials to directly contact the skin. They do not require conductive gel and rely on capacitive coupling or contact impedance matching to collect signals. They are highly comfortable and suitable for long-term continuous monitoring.

[0026] Specifically, in the process of collecting ECG signals, the ECG signals are preprocessed, and the process includes pre-amplification, noise suppression and digitization.

[0027] The collected ECG signal is amplified by using a preamplifier with a gain of 1000 times and a bandwidth of 0.05Hz~150Hz to complete the preamplification.

[0028] More specifically, noise suppression includes baseline drift elimination, power frequency interference suppression, and high-frequency noise filtering. Baseline drift and high-frequency noise are eliminated through a bandpass filter, and power frequency interference is suppressed through a 50Hz notch filter.

[0029] Finally, the acquired ECG signal is converted from analog to digital at a sampling rate of 500 Hz by a 16-bit analog-to-digital converter (ADC) to ensure accurate detection of the R wave (the steep peak in the ECG that represents ventricular depolarization).

[0030] In this embodiment, the PPG signal (photoplethysmography signal) is captured by a photoelectric sensor. The photoelectric sensor uses a green LED (wavelength 530nm) to illuminate the skin, and detects the change in reflected light intensity through a photodiode, reflecting the periodic fluctuation of blood vessel volume with the heartbeat.

[0031] Specifically, in the process of collecting PPG signals, the PPG signals are preprocessed, and the process includes photoelectric conversion, dynamic gain control, noise suppression and digitization.

[0032] Photoelectric conversion uses a transimpedance amplifier (TIA) to convert photocurrent into a voltage signal. Dynamic gain control uses automatic gain control (AGC) to adapt to signal amplitude differences of different skin colors and subcutaneous tissue thickness. Noise suppression includes ambient light interference filtering and motion artifact suppression.

[0033] Specifically, a 0.5Hz~10Hz bandpass filter is used to eliminate ambient light fluctuations, and the acceleration sensor data is combined to dynamically compensate for motion interference.

[0034] The collected PPG signals were converted from analog signals to digital signals through a 16-bit analog-to-digital converter (ADC) at a sampling rate of 100 Hz.

[0035] In a specific embodiment, in order to ensure that the timestamps of the ECG signal and the PPG signal are aligned, a synchronous trigger mechanism is used in the process of collecting the ECG signal and the PPG signal. The hardware-level synchronization of the ECG signal and the PPG signal acquisition is achieved through a field programmable gate array (FPGA, Field-Programmable Gate Array), and the ADC sampling clock of the ECG signal and the PPG signal is controlled to ensure that the timestamp synchronization error is less than 1 millisecond.

[0036] The R wave peak is detected in real time in the ECG signal and used as a trigger signal to start the acquisition of the PPG signal to ensure that the ECG signal and the PPG signal are phase-aligned.

[0037] Specifically, a unified timestamp and synchronization flag are added to each frame of ECG signal and PPG signal for subsequent calculation of pulse wave transmission time (PAT, PulseArrivalTime, i.e., the time difference from the R wave peak to the PPG main wave peak). The PPG main wave peak corresponds to the vasodilation caused by cardiac ejection.

[0038] Among them, the R wave peak is the point with the largest positive amplitude in the QRS complex. The QRS complex is the core waveform in the electrocardiogram (ECG) that reflects ventricular depolarization (contraction caused by electrical excitation of myocardial cells). The R wave peak marks the beginning of ventricular contraction (blood is ejected from the heart into the arteries). The PPG main wave peak refers to the peak value of vascular volume expansion caused by cardiac ejection in the photoelectric volumetric pulse wave (PPG) signal. The time interval from the R wave peak to the PPG main wave peak is called the pulse wave transmission time (PAT, Pulse Arrival Time).

[0039] In a specific embodiment, the synchronized signal needs to be transmitted to the main control unit of the system. SPI (Serial Peripheral Interface) or I2C (Inter-Integrated Circuit) bus can be used to achieve high-speed communication between modules inside the device; wireless transmission can also be used to transmit through low-power Bluetooth (BLE, Bluetooth Low Energy) to meet the needs of wearable devices.

[0040] Specifically, the physiological information collected refers to static physiological parameters such as gender, age, height and weight input by the user into the system. A user-specific database is established to associate and store multimodal signals with physiological parameters to provide training data for subsequent personalized regression models.

[0041] S2. Denoising and multimodal feature extraction Through multi-scale signal analysis and feature engineering, noise suppression, time alignment and accurate extraction of physiological features of ECG and PPG signals are achieved. The specific process is as follows: (1) Denoising and baseline correction based on wavelet transform The ECG signal and the PPG signal are subjected to discrete wavelet transform (DWT), and the db4 wavelet basis (Daubechies 4th order wavelet) is selected to decompose the ECG signal and the PPG signal into 5 layers, and the approximate coefficient (low frequency) and detail coefficient (high frequency) of each layer are obtained, and then the noise variance estimation and soft threshold processing operations are performed, and finally the final denoised signal is obtained by stepwise signal reconstruction. In this embodiment, represents the raw signal of the ECG signal, represents the denoised signal of the ECG signal; represents the raw signal of the PPG signal, It represents the denoised signal of the PPG signal. The specific process is expressed as follows: First, discrete wavelet decomposition is performed, including Layer decomposition, is the layer number, =1, 2, 3, 4, 5; thus, the discrete wavelet decomposition process is expressed as:

[0042]

[0043] in, is the approximate coefficient, is the detail coefficient; is the low-pass filter coefficient of the db4 wavelet basis, is the high-pass filter coefficient of the db4 wavelet basis; It is the time index or space index of the signal, indicating the position of the signal in the time domain or space domain; is the filter length. In this embodiment =8.

[0044] Secondly, the noise variance is estimated and the median absolute deviation (MAD) is used to estimate the noise level. , the process of estimating the noise level is expressed as:

[0045] Approximation coefficient , the process of estimating the noise level is expressed as:

[0046] in, represents the noise variance of each layer detail function, represents the noise variance of the approximate function of each layer, and both are used for noise level estimation.

[0047] Then, the SUREShrink (Stein's UnbiasedRiskEstimateShrinkage) risk function is calculated for any set of coefficients, such as the detail coefficients in this embodiment. The set and approximate coefficients of The risk function is defined as follows:

[0048] in, is the threshold to be optimized, indicating the optimal threshold that can minimize the SURE function; is the number of coefficients, is the noise variance, Represents an index, Indicates coefficients; is an indicative function. In this embodiment, the indicative function In meeting the conditions When the condition is not met, the value is 1. , the value is 0.

[0049] After that, the optimal threshold is selected and the optimal threshold of each layer is determined by minimizing the SURE function. The process is expressed as:

[0050]

[0051] in, Indicates The optimal threshold of the layer detail coefficient, Indicates Optimal threshold for layer approximation coefficients.

[0052] Based on the optimal threshold, the detail coefficient and approximate coefficients Apply a soft threshold function to perform soft thresholding, where is a symbol function; the process of soft threshold processing is expressed as:

[0053]

[0054] Finally, the denoised signal is reconstructed by inverse discrete wavelet transform (IDWT), and the reconstruction is performed layer by layer from the 5th layer to the 1st layer, and finally the denoised signal of the 1st layer is obtained. The signal reconstruction process is expressed as:

[0055] Finally, the denoised signal is obtained, that is, the ECG signal and PPG signal with high-frequency noise and baseline drift suppressed.

[0056] In a specific embodiment, in order to solve the time offset caused by physiological delay or noise between the R peak of the ECG signal and the starting point of the main peak of the PPG signal, the R peak of the ECG signal and the starting point of the main peak of the PPG signal are aligned by DTW to optimize the PAT calculation accuracy. The dynamic time warping (DTW) algorithm is a timing alignment algorithm that can nonlinearly align two signals on the time axis to more accurately match the corresponding feature points. It has good robustness to local time deformation and can effectively handle time deviations caused by measurement noise or physiological changes. DTW is expressed as:

[0057] in, is the ECG signal R peak sequence, is the PPG signal starting point sequence, e , f All are index subscripts.

[0058] (2) Multimodal feature extraction Feature extraction is performed on ECG signals, including obtaining the R wave position through R wave peak detection, and calculating the average RR interval, average heart rate, and heart rate variability.

[0059] The RR interval is the time interval between two consecutive R wave peaks, reflecting the rhythmicity of the heartbeat. The mean RR interval (MeanRRInterval) is the arithmetic mean of multiple RR intervals. The mean heart rate (MeanHeartRate) is the number of heartbeats per unit time. Heart rate variability (HeartRateVariability, HRV) is the physiological fluctuation degree of consecutive heartbeat intervals (RR intervals).

[0060] First, the filtered ECG signal Take the derivative and get the derivative , to emphasize the slope changes, thus highlighting the characteristics of the QRS complex; Derivative Perform a square operation to make all values ​​positive and emphasize larger values ​​to get the resulting value , the process is expressed as follows:

[0061] The moving window integral smoothing signal is used to further highlight the width information of the QRS complex. The integral formula is expressed as:

[0062] Represents the total number of summations, that is, the number of data samples of the derivative.

[0063] After the above treatment, the obtained The curve will have a distinct peak. The curve can be used to represent the QRS complex. By setting a dynamic threshold (adaptive threshold update method based on signal-to-noise ratio), The peak value in the R wave is detected to determine the R wave position of the ECG signal.

[0064] Specifically, setting the dynamic threshold includes setting the peak threshold for detecting the peak of the R wave. and the noise threshold used to distinguish noise from valid signals . For peak threshold and the noise threshold Perform initialization processing and set the initial values ​​of the peak threshold and noise threshold to:

[0065]

[0066] Then, the ECG signal is traversed to detect peaks that meet the peak condition and the noise condition.

[0067] The peak condition is: the amplitude of the current sampling point is the local maximum and the amplitude of the current sampling point is greater than , the peak value that meets the peak condition is the effective peak value.

[0068] The amplitude of the current sampling point is a local maximum, that is, the amplitude of the current sampling point is greater than the amplitude of the previous sampling point and greater than the amplitude of the next sampling point.

[0069] The noise condition is: the amplitude of the current sampling point is less than , then it is considered as noise.

[0070] In a specific embodiment, the peak threshold is set based on the detected peak value and noise. and noise threshold Update dynamically.

[0071] When a valid peak is detected, the is a certain ratio of the current peak amplitude (in this embodiment, it can be 0.7 times); the formula for updating the peak threshold is expressed as:

[0072] in, is the smoothing coefficient, In this embodiment, the value may be 0.1.

[0073] If a noise point is detected, update is a certain ratio of the current noise amplitude (in this embodiment, it can be 0.8 times). The formula for updating the noise threshold is expressed as:

[0074] in, is the smoothing coefficient, In this embodiment, the value may be 0.1.

[0075] Finally, for each detected peak, check whether it meets the characteristics of the R wave (such as R wave amplitude, QRS complex width, etc. In this embodiment, the R wave amplitude can be 0.5 mV to 2.5 mV, and the QRS complex width can be 0.08 seconds to 0.12 seconds); according to the physiological characteristics of the R wave (in this embodiment, the average RR interval can range from 0.6 seconds to 1.2 seconds), remove the pseudo peaks that do not meet the conditions, and finally obtain the R wave position of the ECG signal.

[0076] Specifically, the average RR interval is obtained The expression is:

[0077] in, is the time interval between two adjacent R waves (in seconds); The unit of average heart rate is beats per minute (bpm). Get the average heart rate The expression is:

[0078] More specifically, SDNN (Standard Deviation of NN intervals), as one of the most commonly used time domain indicators in HRV (heart rate variability) analysis, can be used to measure the overall variability of the average RR interval. The expression for obtaining SDNN is:

[0079] In a specific embodiment, feature extraction is performed on the PPG signal, including obtaining an average pulse rate (PR), a pulse wave transmission time (PAT), a rising interval, and a falling interval.

[0080] The slopeup interval refers to the time period from the pulse wave starting point (corresponding to the ECGR peak) to the main wave peak (the point of maximum vascular volume) in a single PPG cycle. The slopedown interval refers to the time period from the main wave peak to the trough (or the terminal stable segment) in a single PPG cycle.

[0081] The number of pulse beats per minute measured by the PPG signal, the average pulse rate The calculation formula is:

[0082] in, is the total number of PPG cycles (the time interval between two adjacent PPG pulse peaks), in seconds; PPG cycles.

[0083] Pulse wave transmission time (PAT) is the time difference between the cardiac electrical activity (ECGR peak) and the peripheral pulse wave (PPG starting point), measured in milliseconds (ms).

[0084] Specifically, the pulse wave transit time is calculated as:

[0085] in, is the time point of the R wave in the ECG signal, It is the time point at which the corresponding pulse wave arrives in the PPG signal.

[0086] More specifically, the rising interval slope The calculation process is:

[0087] in, is the amplitude difference from the lowest point to the highest point in a single PPG cycle, For the time difference.

[0088] Falling interval slope The calculation process is:

[0089] in, is the amplitude difference from the highest point to the PPG end point in a single PPG cycle, For the time difference.

[0090] In a specific embodiment, the user's physiological parameters are normalized to obtain gender. (Male = 1, Female = 0), Age (actual age in whole numbers), body mass index .

[0091] S3. Nonlinear dynamic system modeling and personalized regression A weighted kernel function regression model integrating multimodal features was constructed, and the blood pressure prediction value was generated by combining random forest feature selection with the dynamic switching mechanism of piecewise linear model.

[0092] (1) Multimodal feature fusion In order to fully capture the multi-dimensional information of the user's physiological characteristics, this embodiment designs a multimodal feature fusion strategy. The features extracted from the ECG signal and the PPG signal are organically combined with the user's individual physiological parameters to form an input vector Specifically, the input vector is composed of:

[0093] This feature fusion strategy helps to make full use of multimodal data and enhance the adaptability of blood pressure prediction models to individual differences.

[0094] (2) Nonlinear dynamic regression model In order to capture the complex nonlinear relationship between blood pressure and multimodal features, weighted kernel function regression (WKFR) is used in this embodiment for modeling. The model is expressed as:

[0095] in, is the blood pressure value predicted by the model; is a Gaussian kernel function used to capture the input feature vector With the sample feature vector Nonlinear similarity between is the weight coefficient of the kernel function, whose value is optimized by the machine learning algorithm during the training process to maximize the regression performance of the model; The coupling relationship between body mass index (BMI) and pulse wave arrival time (PAT) was introduced to dynamically adjust the impact of BMI on blood pressure prediction.

[0096] Specifically, the Gaussian kernel function Defined as:

[0097] in, is the kernel width parameter, which is used to control the range of nonlinear mapping.

[0098] (3) Lightweight model design The importance score of each feature is calculated through the Random Forest algorithm, and the top five features (Top-5 features) that contribute most to blood pressure prediction are selected. This feature selection method is based on the feature splitting frequency and information gain in the random forest, ensuring that the selected features can fully express the core information of the data, while removing redundant features and reducing computational costs. The top-5 features that are finally retained can significantly reduce the input dimension while avoiding performance degradation, laying the foundation for model compression.

[0099] Based on feature selection, for the established model, the original weighted kernel function regression (WKFR) model is transformed into a piecewise linear model, which significantly reduces the computational complexity. Specifically, the model structure is dynamically switched according to the activity state (stationary or moving), which simplifies the prediction process. Specifically, the mathematical expression of the piecewise linear model is:

[0100] in, is the linear weight vector in the static state, is the linear weight vector in motion state; is the bias term in the static state; It is the bias item in the motion state; MAI (MovementActivityIndicator) is the activity state indicator, indicating whether the current state is motion. MAI=0 represents the static state, and MAI=1 represents the motion state.

[0101] S4. Incremental learning and dynamic calibration In order to adapt to the long-term physiological changes of individuals and the influence of dynamic environment, a model optimization strategy based on incremental learning and dynamic calibration is designed in this embodiment, including online parameter update, adaptive dual-mode calibration mechanism and long-term physiological adaptability adjustment to ensure the continuous optimization of model prediction performance.

[0102] (1) Online parameter update During the model operation, the recursive least squares (RLS) algorithm is used to dynamically update the model weights to achieve real-time incremental learning. The RLS algorithm can efficiently optimize the model parameters step by step through new input data, and its expression is as follows:

[0103]

[0104] in, is the model weight at the current moment. is the current weight covariance matrix. is the current input feature vector. This online update mechanism adjusts the model weights through continuous iterations to adapt to the user's dynamic physiological state changes, thereby improving the accuracy of real-time predictions.

[0105] (2) Adaptive dual-mode calibration In this embodiment, in order to improve the adaptability of the individual blood pressure prediction model to physiological dynamic changes, a dual-mode calibration mechanism integrating the time dimension and error feedback is designed, including strategies for initial calibration and periodic calibration.

[0106] Specifically, the initial calibration is based on more than 2 standard blood pressure measurements to establish the initial personalized parameters of the model. That is, 3 standard cuff blood pressure measurements are completed continuously in a resting state, and the average is taken as the reference value; the feature vectors at the corresponding time are collected during the blood pressure measurement interval, and the model weights are initialized by the least squares method. In this embodiment, the initial calibration is based on 3 standard blood pressure measurements, the interval between standard cuff blood pressure measurements is 2 minutes / time, and the resting state is maintained for 10 minutes.

[0107] Periodic calibration uses periodic triggering mechanism and event triggering mechanism to achieve dynamic optimization of parameters.

[0108] Specifically, the periodic trigger is a forced start when the preset running time has accumulated since the last calibration, and the model parameters are personalized based on the latest data; in this embodiment, the preset time is 72 hours (3 days).

[0109] The event trigger is when the absolute value of the prediction residual exceeds the clinically permitted threshold for three consecutive times within the preset prediction times, the system is started in advance, and the model parameters are personalized based on the latest data. In this embodiment, the preset prediction times are 10 times.

[0110] (3) Long-term physiological adaptation: In order to cope with long-term physiological changes, all incremental data are summarized every 7 days, and calibration values ​​are obtained through a blood pressure cuff monitor; every 7 days or when the user's physiological state changes significantly (such as weight fluctuation > 5%), the global fine-tuning of the model is triggered to optimize the global parameters. The weights are updated using the following formula:

[0111] in, is the learning rate, which is used to control the amplitude of weight adjustment; is the regularization coefficient, used to suppress overfitting; is the input feature matrix of the calibration phase; is the actual output value in the calibration phase; Indicates transpose.

[0112] By dynamically adjusting the model parameters, the predicted blood pressure values ​​can be continuously optimized, ensuring the long-term accuracy of the measurement results under changes in individual physiological status and environmental interference, and ultimately obtaining personalized, high-precision real-time blood pressure measurements.

[0113] Therefore, this step uses multiple technical means such as real-time parameter update, intelligent trigger calibration, and physiological adaptive optimization to ultimately obtain high-precision, individualized, and interference-resistant blood pressure measurement values, avoiding the measurement inaccuracy problems caused by traditional methods due to rigid models and lack of calibration.

[0114] In a specific embodiment, a system using the above-mentioned blood pressure measurement method includes a multimodal signal acquisition module, a denoising and multimodal feature extraction module, a modeling and personalized regression, and an incremental learning and dynamic calibration module.

[0115] Specifically, the multimodal signal acquisition module is used to synchronously acquire ECG signals and PPG signals, pre-process the ECG signals and PPG signals, and align the timestamps of the ECG signals and the PPG signals using a synchronous trigger mechanism; The denoising and multimodal feature extraction module is used to denoise ECG signals and PPG signals based on discrete wavelet transform, align the starting point of the ECG signal R peak and the PPG signal main peak through the dynamic time warping algorithm, and extract features from ECG signals and PPG signals; The modeling and personalized regression module is used to build a weighted kernel function regression model that integrates multimodal features, combining random forest feature selection with the dynamic switching mechanism of the piecewise linear model to generate blood pressure prediction values; The incremental learning and dynamic calibration module is used to update the model weights based on the recursive least squares method, adopt a hybrid mechanism of event triggering and period triggering for dual-mode calibration, and obtain the final blood pressure value through global fine-tuning.

[0116] Among them, the multimodal signal acquisition module includes a main control unit, which is responsible for receiving and processing synchronized signals, performing calculations or logical operations, and transmitting data via Bluetooth.

[0117] The blood pressure measurement method of the present invention significantly improves the accuracy of blood pressure prediction and enhances the physiological relevance of blood pressure estimation through multi-signal fusion and personalized modeling; secondly, by combining individual difference information, the method is applicable to subjects of different ages, genders and weights, improving individual adaptability and the universality of the prediction model; and through lightweight model design, the computational complexity is reduced, so that it can realize real-time blood pressure monitoring on mobile devices to meet daily health management needs; in addition, the optimized signal processing algorithm effectively reduces noise interference, improves the stability and reliability of measurement, and enhances the practicality of the method; finally, through the incremental learning mechanism, historical data are used to dynamically update the model parameters to adapt to long-term physiological changes, further improving the applicability and accuracy of the model.

[0118] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited to this, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all fall within the scope of protection of this patent. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A blood pressure measurement method comprising the following steps: Signal acquisition and processing: synchronously acquire ECG and PPG signals, pre-process them, and use a synchronous trigger mechanism to align the timestamps of ECG and PPG signals; Denoising and multimodal feature extraction: Denoise the ECG and PPG signals based on discrete wavelet transform, align the starting point of the ECG signal R peak with the PPG signal main peak through the dynamic time warping algorithm, and extract features from the ECG and PPG signals; Modeling and personalized regression: Construct a weighted kernel function regression model that integrates multimodal features, combine random forest feature selection with the dynamic switching mechanism of the piecewise linear model, and generate blood pressure prediction values; Incremental learning and dynamic calibration: The model weights are updated based on the recursive least squares method, and a hybrid mechanism of event triggering and period triggering is used for dual-mode calibration. The final blood pressure value is obtained through global fine-tuning.

2. The blood pressure measurement method according to claim 1, characterized in that: The synchronization trigger mechanism includes: controlling the ADC sampling clock of ECG and PPG signals through a field programmable gate array to achieve hardware-level synchronization; The PPG signal acquisition is started with the R peak of the ECG signal as the trigger signal, and a unified timestamp and synchronization flag are added to each frame of the signal.

3. The blood pressure measurement method according to claim 1, characterized in that: The ECG signal preprocessing process includes: using a preamplifier to amplify the signal; using a bandpass filter to eliminate baseline drift and high-frequency noise, and combining a notch filter to suppress power frequency interference; and converting the collected ECG signal from an analog signal to a digital signal through an analog-to-digital converter.

4. The blood pressure measurement method according to claim 1, characterized in that: The discrete wavelet transform process includes: selecting the db4 wavelet basis, decomposing the ECG signal and PPG signal into 5 layers respectively, calculating the median absolute deviation of the approximate coefficient and detail coefficient of each layer respectively, and estimating the noise variance; The optimal threshold is determined based on the SUREShrink risk function, the soft threshold function is used for coefficient processing, and the denoised signal is reconstructed by inverse discrete wavelet transform.

5. The blood pressure measurement method according to claim 1, characterized in that: Feature extraction of ECG signals includes: obtaining the R wave position through R wave peak detection, calculating the average RR interval, average heart rate and heart rate variability; The process of obtaining the R wave position includes: taking the derivative of the filtered ECG signal, squaring the derivative, and then smoothing the signal using a moving window integral to obtain a curve representing the QRS complex; Set the dynamic threshold, detect the peak value in the curve, and determine the R wave position of the ECG signal.

6. The blood pressure measurement method according to claim 5, characterized in that: Get the average RR interval The expression is: ； in, is the time interval between two adjacent R waves, in seconds; Get average heart rate The expression is: 。 7. The blood pressure measurement method according to claim 1, characterized in that: Feature extraction of PPG signals includes obtaining the average pulse rate and pulse wave transmission time; Get average pulse rate The process is expressed as: ； is the total number of PPG cycles, in seconds; is the serial number of the PPG cycle; The process of obtaining the pulse wave transmission time is expressed as: ； in, is the time point of the R wave in the ECG signal, It is the time point at which the corresponding pulse wave arrives in the PPG signal.

8. The blood pressure measurement method according to claim 1 or 7, characterized in that: Feature extraction of the PPG signal also includes obtaining a rising interval and a falling interval; Get the slope of the rising interval The process is expressed as: ； is the amplitude difference from the lowest point to the highest point in a single PPG cycle, is the time difference; Get the slope of the descending interval The process is expressed as: ； is the amplitude difference from the highest point to the PPG end point in a single PPG cycle, For the time difference.

9. The blood pressure measurement method according to claim 1, characterized in that: The dynamic switching mechanism of the piecewise linear model includes: when the activity state indicator MAI=0, the linear weight vector in the static state and the bias term in the static state are used; When MAI=1, switch to the linear weight vector in motion state and the bias term in motion state.

10. The blood pressure measurement method according to claim 1, characterized in that: The dual-mode calibration mechanism includes initial calibration and periodic calibration; the initial calibration initializes the model weights based on the mean of more than 2 standard blood pressure measurements in a resting state; Periodic calibration triggers the update of model parameters when the preset cumulative running time is reached.

Images