Training methods for blood pressure prediction models and methods for predicting blood pressure data

Abstract

This invention provides a training method for a blood pressure prediction model and a method for predicting blood pressure data. The training method for the blood pressure prediction model includes: acquiring data from a wearable device to obtain a raw data sequence; preprocessing the raw data sequence to obtain a standard data sequence; extracting features from the standard pressure signal sequence and the standard blood pressure sequence based on peak and trough values ​​to obtain a pressure feature sequence, a blood pressure feature sequence, and a pulse wave propagation time sequence; extracting features from the standard acceleration sequence and the standard angular velocity sequence based on the rate of change and frequency to obtain a motion intensity sequence; fusing the pressure feature sequence, the blood pressure feature sequence, the pulse wave propagation time sequence, and the motion intensity sequence to form a multidimensional feature sequence; and training the blood pressure prediction pre-training model based on the multidimensional feature sequence to obtain the blood pressure prediction model. The blood pressure prediction model obtained by this invention has higher accuracy in predicting blood pressure.

Description

Technical Field

[0001] This invention relates to the field of blood pressure measurement, and more specifically to a training method for a blood pressure prediction model for a wearable device and a method for predicting blood pressure data. Background Technology

[0002] Blood pressure, as a crucial indicator of cardiovascular health, is essential for the accurate monitoring and prevention of cardiovascular diseases such as hypertension. Traditional blood pressure measurement methods primarily rely on cuff-type devices, measuring systolic and diastolic pressure through intermittent inflation and deflation. These methods suffer from drawbacks such as inconvenience, discontinuous measurement, and limitations on user activity. With the development of wearable technology, continuous, real-time blood pressure monitoring using wearable devices has become a research hotspot. However, existing wearable blood pressure monitoring methods still face numerous challenges in terms of data acquisition accuracy, real-time processing capabilities, and prediction accuracy.

[0003] In existing technologies, single sensors often fail to fully reflect the complex physiological mechanisms of blood pressure changes. While multi-sensor fusion data processing methods have improved upon these methods, further optimization is still needed in areas such as data preprocessing, feature extraction, and model prediction. Furthermore, traditional machine learning algorithms often struggle to fully capture deep-seated features in high-dimensional, time-series data, leading to insufficient accuracy and stability in blood pressure prediction.

[0004] Therefore, there is an urgent need for a new data processing method that can effectively integrate multi-sensor data, improve the accuracy and real-time performance of blood pressure prediction, and have good generalization ability to meet the needs of different user groups. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a method for training a blood pressure prediction model and a method for predicting blood pressure data.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A method for training a blood pressure prediction model includes the following steps:

[0008] Data from a wearable device is acquired to obtain a raw data sequence, wherein the data from the wearable device includes at least pressure signals, blood pressure signals, acceleration, and angular velocity;

[0009] The original data sequence is preprocessed to obtain a standard data sequence;

[0010] Based on the peak and trough values, the features of the standard pressure signal sequence and the standard blood pressure sequence are extracted to obtain the pressure feature sequence, the blood pressure feature sequence, and the pulse wave propagation time sequence.

[0011] Based on the rate of change and frequency, features of the standard acceleration sequence and standard angular velocity sequence are extracted to obtain the motion intensity sequence;

[0012] By integrating pressure feature sequences, blood pressure feature sequences, pulse wave propagation time sequences, and exercise intensity sequences, a multidimensional feature sequence is formed.

[0013] The blood pressure prediction model is obtained by training a pre-trained model based on multidimensional feature sequences.

[0014] As one possible implementation, the pressure characteristic sequence includes at least a pressure peak sequence, a pressure trough sequence, a pressure rise time sequence, and a pressure fall time sequence. Obtaining the pressure characteristic sequence includes the following steps:

[0015] The peak and trough values ​​of each heartbeat cycle are extracted from the standard pressure signal sequence to obtain the pressure peak-trough value sequence, pressure peak value sequence, and pressure trough value sequence;

[0016] The pressure rise time series is obtained by subtracting the timestamp of the target peak value from the timestamp of the valley value preceding the target peak value in the pressure peak-valley value sequence.

[0017] The pressure drop time series is obtained by subtracting the timestamp of the target valley value from the timestamp of the peak value preceding the target valley value in the pressure peak-valley value sequence.

[0018] As one possible implementation, the blood pressure characteristic sequence includes at least a blood voltage peak sequence, a blood voltage rise rate sequence, and a blood voltage fall rate sequence. Obtaining the blood pressure characteristic sequence includes the following steps:

[0019] The peak and trough values ​​of blood voltage in each blood volume change cycle are extracted from the standard blood pressure sequence to obtain the blood voltage peak-trough value sequence and the blood voltage peak value sequence;

[0020] Calculate the rate at which the blood voltage rises from a trough to the next peak in the blood voltage peak-trough sequence to obtain the blood voltage rise rate sequence;

[0021] The rate at which blood voltage decreases from the peak value to the next trough value in the blood voltage peak-trough sequence is calculated to obtain the blood voltage decrease rate sequence.

[0022] As one possible implementation method, obtaining the pulse wave propagation time series includes the following steps:

[0023] Peak values ​​for each heartbeat cycle are extracted from the standard pressure signal sequence to obtain the pressure peak sequence;

[0024] Peak values ​​were extracted from the standard blood pressure sequence for each cycle of blood volume change to obtain the blood voltage peak sequence;

[0025] The pulse wave propagation time series is obtained by performing a time-stamp-based difference operation on the pressure peak sequence and the blood voltage peak sequence.

[0026] As one possible implementation, the motion intensity sequence includes at least a motion intensity level sequence, an acceleration rate of change sequence, and a motion frequency sequence. Obtaining the motion intensity sequence includes the following steps:

[0027] The standard acceleration sequence and the standard angular velocity sequence are summed based on rate to obtain a comprehensive rate sequence;

[0028] Based on the comprehensive rate sequence, the first preset rate threshold, and the second preset rate threshold, a motion intensity level sequence is obtained;

[0029] Calculate the rate of change of acceleration in the standard acceleration sequence to obtain the acceleration rate of change sequence;

[0030] Based on Fourier transform, the frequency of the standard acceleration sequence is extracted within each preset period to obtain the acceleration frequency feature sequence.

[0031] Extract the maximum frequency, i.e. the motion frequency, from each preset period in the acceleration frequency feature sequence to obtain the motion frequency sequence.

[0032] As one possible implementation, the blood pressure prediction pre-training model includes a convolutional recursive fusion module, an attention weighting module, and a prediction module;

[0033] The convolutional recursive fusion module performs convolution and bidirectional recursive operations on the input vector to obtain convolutional feature sequences and bidirectional feature sequences; it then fuses the convolutional feature sequences and bidirectional feature sequences to obtain a comprehensive feature sequence.

[0034] The attention weighting module processes the comprehensive feature sequence using an attention mechanism to obtain an attention weight sequence, and then performs weighted processing on the comprehensive feature sequence and the attention weight sequence to obtain a weighted feature sequence.

[0035] The prediction module maps the weighted feature sequence through a fully connected layer to obtain the predicted blood pressure data.

[0036] The predicted blood pressure data is represented as follows:

[0037]

[0038] in, This represents the predicted systolic blood pressure value. This represents the predicted diastolic blood pressure value. Indicates matrix transpose. Indicates the weights of the fully connected layer. Represents a weighted feature sequence. This indicates the bias of the fully connected layer.

[0039] A method for predicting blood pressure data includes the following steps:

[0040] Acquire target data from wearable devices;

[0041] The target data is processed based on the blood pressure prediction model to obtain the predicted blood pressure data value;

[0042] The blood pressure prediction model is a model obtained by using the training method of any one of the blood pressure prediction models described above.

[0043] A training system for a blood pressure prediction model, used to implement the training method for the blood pressure prediction model described in any one of the above-mentioned methods, the training system comprising:

[0044] The data acquisition module is used to acquire data from the wearable device to obtain a raw data sequence, wherein the data from the wearable device includes at least pressure signals, blood pressure signals, acceleration, and angular velocity;

[0045] The data processing module is used to preprocess the original data sequence to obtain a standard data sequence;

[0046] The first feature extraction module is used to extract features from the standard pressure signal sequence and the standard blood pressure sequence based on the peak and trough values, respectively, to obtain the pressure feature sequence, the blood pressure feature sequence, and the pulse wave propagation time sequence.

[0047] The second feature extraction module is used to extract features from the standard acceleration sequence and the standard angular velocity sequence based on the rate of change and frequency to obtain the motion intensity sequence.

[0048] The feature fusion module is used to fuse pressure feature sequences, blood pressure feature sequences, pulse wave propagation time sequences, and exercise intensity sequences to form multidimensional feature sequences.

[0049] The model training module is used to train the pre-trained blood pressure prediction model based on multi-dimensional feature sequences to obtain the blood pressure prediction model.

[0050] A blood pressure data prediction system, the prediction system comprising:

[0051] The acquisition module is used to acquire target data from wearable devices;

[0052] The prediction module is used to infer the target data based on the blood pressure prediction model to obtain the predicted blood pressure value; wherein the blood pressure prediction model is a model obtained by training the blood pressure prediction model using any of the above-described methods.

[0053] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the training method for the blood pressure prediction model described in any one of the preceding claims.

[0054] A training device for a blood pressure prediction model includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the training method for the blood pressure prediction model described above.

[0055] This invention, by employing the above technical solutions, achieves significant technical effects: It provides a training method for a blood pressure prediction model and a method for predicting blood pressure data. Through the collaborative processing and deep fusion of multi-sensor data, it effectively overcomes the limitations of a single sensor data source, significantly improving the comprehensiveness and reliability of data acquisition. Through innovative model training strategies and data processing flows, the system can adapt to user groups with different physiological characteristics, maintaining stable prediction performance in diverse application scenarios. The various technical features of this invention are mutually supportive and closely integrated, improving the accuracy and real-time performance of the blood pressure prediction model. Simultaneously, the blood pressure prediction model possesses good generalization ability, adapting to the needs of different user groups, and exhibits significant technical advantages and application value. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic flowchart of an embodiment of the training method for the blood pressure prediction model of the present invention;

[0058] Figure 2 This is an overall schematic diagram of an embodiment of the training system for the blood pressure prediction model of the present invention;

[0059] Figure 3 This is a flowchart illustrating an embodiment of the blood pressure data prediction method of the present invention;

[0060] Figure 4 This is an overall schematic diagram of an embodiment of the blood pressure data prediction system of the present invention. Detailed Implementation

[0061] The present invention will be further described in detail below with reference to embodiments. These embodiments are illustrative of the invention and the invention is not limited thereto. Unless otherwise specified, the features in the following embodiments can be combined with each other.

[0062] Example 1:

[0063] A training method for a blood pressure prediction model, such as Figure 1 As shown, it includes the following steps:

[0064] S110: Acquire data from the wearable device to obtain a raw data sequence, wherein the data from the wearable device includes at least pressure signals, blood pressure signals, acceleration, and angular velocity;

[0065] S120: Preprocess the original data sequence to obtain a standard data sequence;

[0066] S130: Based on the peak and trough values, extract the features of the standard pressure signal sequence and the standard blood pressure sequence to obtain the pressure feature sequence, blood pressure feature sequence and pulse wave propagation time sequence;

[0067] S140: Based on the rate of change and frequency, extract the features of the standard acceleration sequence and the standard angular velocity sequence to obtain the motion intensity sequence;

[0068] S150: Integrates pressure feature sequence, blood pressure feature sequence, pulse wave propagation time sequence and exercise intensity sequence to form a multidimensional feature sequence;

[0069] S160: The blood pressure prediction pre-training model is trained based on multi-dimensional feature sequences to obtain the blood pressure prediction model.

[0070] In S110, data from the wearable device is acquired to obtain a raw data sequence. The wearable device data includes at least pressure signals, blood pressure signals, acceleration, and angular velocity. This data is collected through built-in sensors, including at least a pressure sensor, a blood volume sensor, an acceleration sensor, and an angular velocity sensor. Specifically:

[0071] (1) Pressure sensor

[0072] Pressure sensors are used to acquire data related to pressure signals from wearable devices. Examples include MEMS (Micro-Electro-Mechanical Systems) pressure sensors and PVDF thin-film pressure sensors. High-sensitivity MEMS pressure sensor wafers are preferred as the primary pressure detection element. Pressure sensors are typically located on the wrist, with the bottom of the watch directly contacting the skin to collect the pressure applied by the user in real time.

[0073] The pressure collected by the thin film is transmitted to the MEMS pressure sensor die through a narrow fluid channel using internal silicone oil (which has good viscosity and stability). The silicone oil conduction path is a cylindrical pipe. Under steady-state flow conditions, the flow rate in the silicone oil... With pressure gradient The relationship between them is: Among them, the flow rate in silicone oil The unit is cubic meters per second. Pressure gradient. This represents the pressure difference across the silicone oil conduction path, i.e., the pressure applied by the thin film. Pressure received by the sensor ,Right now . This indicates the radius of the silicone oil conduction path, in meters. This indicates the dynamic viscosity of silicone oil, measured in Pascals per second. This indicates the length of the silicone oil conduction path, in meters. Therefore, the pressure received by the sensor is obtained. for: Because pressure will attenuate to some extent during transmission, it is necessary to ensure near Then it needs to satisfy: This requires As small as possible, Make it as large as possible.

[0074] The relationship between the output voltage of the pressure sensor and the input pressure is as follows: ,in, This represents the voltage signal output by the sensor, measured in millivolts. This indicates the sensitivity of the sensor, measured in mV / Pa (millivolts per pad). This indicates the pressure received by the sensor, which is the actual pressure applied to the sensor, and the unit is Pa.

[0075] Pressure transmitted to the membrane Force and the force applied to the film by the user The relationship between them is: ,in, The effective contact area of ​​the membrane is expressed in square meters. The force applied by the user to the membrane is expressed in Newtons (N). The pressure transmitted to the membrane is expressed in Pa.

[0076] The pressure collected by the thin film is transmitted to the MEMS sensor through silicone oil. The sensor converts the pressure into a voltage signal, and the voltage signal output by the pressure sensor... Represented as: The voltage signal output by the pressure sensor is measured in mV. The force applied by the user to the film Related.

[0077] It should be noted that, as is known to those skilled in the art, although pressure data is collected by sensors, in the context of blood pressure measurement, pressure intensity is a form of pressure. Therefore, pressure and pressure intensity are actually only a conceptual conversion, not a numerical conversion. Based on the conventions of those skilled in the art, this application still uses the term "pressure" in the text, but the symbol is represented as... (pressure).

[0078] (2) Blood volume sensor

[0079] Blood volume sensors are used to measure changes in blood volume; examples include photoplethysmography (PPG) sensors and bioimpedance impedance sensors (BIA). In this embodiment, a PPG sensor is used to measure changes in blood volume, and its output is a PPG signal voltage denoted as […]. ,but ,in, This represents the voltage signal output by the PPG sensor, measured in mV. The sensitivity of the PPG sensor is expressed in units of 1000 kJ / m². Change in unit light intensity. This indicates the change in light intensity caused by changes in blood volume.

[0080] (3) Accelerometer

[0081] There are many types of accelerometers, used to measure a user's acceleration. Accelerometers output an acceleration vector. ,but ,in, These represent the measurements taken by the accelerometer sensor. The acceleration value in the direction, in units of . The sensitivity of the accelerometer is expressed in units of: ). Indicates the output voltage of the accelerometer, then , These represent the measurements taken by the accelerometer sensor. The output voltage value in the direction, in mV. This indicates transpose.

[0082] (4) Angular velocity sensor

[0083] An angular velocity sensor is used to measure angular velocity. Angular velocity sensors are mainly divided into different types of gyroscopes. The output of an angular velocity sensor is the angular velocity vector measured by the gyroscope. ,but ,in, These represent the measurements taken by the gyroscope. The angular velocity value in the direction, in rad / s. The sensitivity of a gyroscope is expressed in units of 1000 kilometres per second. . This represents the output voltage of the gyroscope. , These represent the measurements taken by the gyroscope. The output voltage value in the direction, in mV.

[0084] In S120, the original data sequence is preprocessed to obtain a standard data sequence, including the following steps:

[0085] S121: Bandpass filtering is performed on the original pressure signal sequence and the original blood voltage signal sequence respectively to obtain the bandpass pressure signal sequence and the bandpass blood voltage signal sequence. Details are as follows:

[0086] (1) Identify the frequency of the original pressure signal sequence to obtain the pressure signal frequency sequence.

[0087] (2) Based on the pressure frequency range formed by the first preset pressure frequency threshold and the second preset pressure frequency threshold, if the frequency in the pressure signal frequency sequence is within the pressure frequency range, the corresponding transfer value is 1, thereby obtaining the pressure transfer value sequence, which can be expressed as: ,in, Represents the first value in the pressure transmission value sequence. One passed value, Represents the first frequency in the pressure signal frequency sequence One frequency, These represent the first preset pressure frequency threshold and the second preset pressure frequency threshold, respectively.

[0088] (3) Perform an inverse Fourier transform on the transmission values ​​with a value of 1 in the pressure transmission value sequence to obtain the pressure time-domain response value sequence. The pressure time-domain response value sequence can then be expressed as: ,in, Represents the first value in the time-domain response sequence of pressure. Each time-domain response value, Represents the imaginary unit. Indicates frequency, Indicates time.

[0089] (4) Convolve the pressure time-domain response value sequence and the original pressure signal sequence to achieve bandpass filtering, and obtain the bandpass pressure signal sequence; the bandpass pressure signal sequence is represented as follows: ,in, This represents a bandpass pressure signal sequence. This represents the original pressure signal sequence. This represents the time-domain response value sequence of pressure. This indicates convolution processing.

[0090] (5) Bandpass filtering is performed on the original blood voltage signal sequence to obtain the bandpass blood voltage signal sequence.

[0091] The original pressure signal sequence and the original blood voltage signal sequence are bandpass filtered to obtain the bandpass pressure signal sequence and the bandpass blood voltage signal sequence.

[0092] Bandpass filtering can remove power supply interference (e.g., 50Hz) and high-frequency noise from the original pressure signal sequence and the original blood voltage signal sequence, while retaining low-frequency components related to blood pressure (e.g., 0.5Hz to 5Hz).

[0093] S122: Perform low-pass filtering on the original acceleration sequence and the original angular velocity sequence respectively to obtain the low-pass acceleration sequence and the low-pass angular velocity sequence, as follows:

[0094] (1) Identify the frequency of the original acceleration sequence to obtain the acceleration frequency sequence.

[0095] (2) If the frequency in the acceleration frequency sequence is lower than or equal to the preset acceleration frequency threshold, the corresponding transfer value is 1, thus obtaining the acceleration transfer value sequence. The acceleration transfer value sequence can then be expressed as: ,in, Represents the first value in the sequence of acceleration transfer values. One passed value, Represents the first in the acceleration frequency sequence One frequency, This indicates the preset acceleration frequency threshold.

[0096] (3) Perform an inverse Fourier transform on the transfer values ​​with a value of 1 in the acceleration transfer value sequence to obtain the acceleration time-domain response value sequence. Then the acceleration time-domain response value sequence can be expressed as: ,in, Represents the first value in the time-domain response sequence of acceleration. Each time-domain response value, Represents the imaginary unit. Indicates frequency, Indicates time.

[0097] (4) Convolve the acceleration time-domain response sequence and the original acceleration sequence to achieve low-pass filtering, thereby obtaining a low-pass acceleration sequence; the low-pass acceleration sequence is expressed as: ,in, Represents a low-pass acceleration sequence. Represents the original acceleration sequence. Represents the time-domain response value sequence of acceleration. This indicates convolution processing.

[0098] Low-pass filtering can remove high-frequency motion noise from the original acceleration and angular velocity sequences while preserving the user's main motion state information (such as stationary, walking, or running).

[0099] Filtering data sequences aims to remove noise and irrelevant frequency components, improving data quality. This is a crucial step in ensuring that raw data acquired from sensors can be effectively utilized by pre-trained blood pressure prediction models. Appropriate bandpass and low-pass filters can be selected based on the specific application environment and sensor characteristics to optimize data sequence quality. Digital filters (such as FIR or IIR filters) can be used to implement data sequence filtering, reducing computational complexity. Butterworth filters can be used to design bandpass and / or low-pass filters, offering advantages such as smooth frequency response and minimal ripple.

[0100] In this invention, bandpass filtering and low-pass filtering are applied to different types of raw data to improve data quality and the accuracy of the blood pressure prediction model. Bandpass filtering is mainly used to process pressure and blood pressure signals, removing high-frequency noise and low-frequency interference from the raw signals while retaining key low-frequency components of blood pressure changes, thus effectively enhancing data validity and prediction accuracy. Low-pass filtering is applied to acceleration and angular velocity signals to remove high-frequency motion noise and retain low-frequency information related to the user's motion state. This step helps to accurately capture changes in motion intensity, improves the model's ability to recognize motion states, and thus enhances the stability and reliability of the blood pressure prediction model. The combined application of bandpass and low-pass filtering significantly optimizes the data preprocessing workflow and improves the overall performance of the blood pressure prediction model.

[0101] S123: Data Alignment (Synchronization) Processing. If the sampling frequencies of the bandpass pressure signal sequence, bandpass blood voltage signal sequence, low-pass acceleration sequence, and low-pass angular velocity sequence are inconsistent, data alignment processing based on a unified timestamp is performed on these sequences. Details are as follows:

[0102] (1) Obtain the minimum and maximum timestamps in the bandpass pressure signal sequence, bandpass blood voltage signal sequence, low-pass acceleration sequence and low-pass angular velocity sequence, and then combine them with a preset time interval to obtain a unified timestamp sequence.

[0103] (2) Based on a unified timestamp sequence, the bandpass pressure signal sequence, bandpass blood voltage signal sequence, low-pass acceleration sequence and low-pass angular velocity sequence are aligned. The alignment process includes one or more of interpolation, forward padding, backward padding and resampling.

[0104] Data synchronization ensures that data from different sensors are aligned in time, facilitating subsequent feature extraction and model training.

[0105] S124: Standardization Processing. The bandpass pressure signal sequence, bandpass blood voltage signal sequence, low-pass acceleration sequence, and low-pass angular velocity sequence are standardized respectively to obtain standard pressure signal sequences, standard blood voltage signal sequences, standard acceleration sequences, and standard angular velocity sequences. This includes the following steps:

[0106] (1) Based on the bandpass pressure signal sequence, the mean and standard deviation of the pressure signal are obtained.

[0107] (2) Based on the bandpass pressure signal sequence, the mean of the pressure signal, and the standard deviation of the pressure signal, the standard pressure signal sequence is obtained, which can be expressed as: ,in, This represents the first [number] in the standard pressure signal sequence. A pressure signal, This represents the first [number] signal in the bandpass pressure signal sequence. A pressure signal, This represents the average pressure signal. This represents the standard deviation of the pressure signal. Represents a constant. This represents the learnable scaling parameter. This represents the learnable translation parameters.

[0108] (3) Standardize the bandpass blood voltage signal sequence, lowpass acceleration sequence and lowpass angular velocity sequence respectively to obtain standard blood voltage signal sequence, standard acceleration sequence and standard angular velocity sequence.

[0109] Standardization converts data from different sensors and features into the same units, eliminating dimensional differences between different features, thereby improving the efficiency and stability of blood pressure prediction pre-training model training.

[0110] In S130, features of the standard pressure signal sequence and the standard blood pressure sequence are extracted based on peak and trough values, respectively, to obtain the pressure feature sequence, the blood pressure feature sequence, and the pulse wave propagation time sequence, including the following steps:

[0111] S131: Obtain the pressure characteristic sequence, which includes at least a pressure peak sequence, a pressure trough sequence, a pressure rise time sequence, and a pressure fall time sequence, specifically including:

[0112] (1) Extract the peak (systolic pressure) and trough (diastolic pressure) of each heartbeat cycle from the standard pressure signal sequence to obtain the pressure peak-trough sequence, pressure peak sequence and pressure trough sequence. The pressure peak-trough sequence is a sequence containing the peak and trough of each heartbeat cycle.

[0113] (2) The difference between the timestamp of the target peak value and the timestamp of the valley value before the target peak value in the pressure peak-valley value sequence is processed to obtain the pressure rise time sequence. The pressure rise time represents the time required for the pressure to rise from the valley value to the peak value.

[0114] (3) The timestamp of the target valley value and the timestamp of the peak value before the target valley value in the pressure peak valley value sequence are subtracted to obtain the pressure drop time sequence. The pressure drop time represents the time required for the pressure to drop from the peak value to the valley value of the next cycle.

[0115] The time series of pressure rise can be represented as: The time series of pressure drop can be represented as: ,in, Indicating the first time series of pressure rise A period of rise, This represents the first value in the pressure peak-to-trough sequence. The timestamps corresponding to each pressure peak This represents the first value in the pressure peak-to-trough sequence. The timestamp corresponding to the previous pressure trough value before each pressure peak. Indicating the time series of pressure decreases, the first... A descent time, Indicates the first value in the pressure peak-valley sequence The timestamps corresponding to each pressure trough value Indicates the first value in the pressure peak-valley sequence The timestamp corresponding to the previous pressure peak value of each pressure trough.

[0116] S132: Obtain the blood pressure characteristic sequence, which includes at least a blood voltage peak sequence, a blood voltage rise rate sequence, and a blood voltage fall rate sequence, as follows:

[0117] (1) Extract the peak and trough values ​​of each blood volume change cycle from the standard blood pressure sequence to obtain the blood voltage peak and trough value sequence and the blood voltage peak value sequence.

[0118] (2) Calculate the rate at which the blood voltage rises from the valley to the next peak in the blood voltage peak-valley sequence to obtain the blood voltage rise rate sequence.

[0119] (3) Calculate the rate at which the blood voltage drops from the peak value to the next valley value in the blood voltage peak-valley sequence to obtain the blood voltage drop rate sequence.

[0120] The blood voltage rise rate sequence can be represented as: The blood voltage decrease rate sequence can be represented as: ,in, The sequence representing the rate of rise of blood voltage is the first... The rate of increase in blood voltage, This represents the first peak in the blood voltage peak-to-trough sequence. A blood voltage trough value, This represents the first peak in the blood voltage peak-to-trough sequence. The next peak blood voltage value after the trough of blood voltage This indicates the timestamp corresponding to the next peak blood voltage value in the blood voltage peak-valley sequence. This represents the first peak in the blood voltage peak-to-trough sequence. The timestamps corresponding to the lowest blood voltage values This represents the first peak in the blood voltage peak-to-trough sequence. A peak blood voltage, This represents the first peak in the blood voltage peak-to-trough sequence. The next blood voltage trough after the peak blood voltage value. This represents the first peak in the blood voltage peak-to-trough sequence. The timestamps corresponding to the peak blood voltage values. This indicates the timestamp corresponding to the next blood voltage valley in the blood voltage peak-valley sequence.

[0121] Blood voltage peak value is used to characterize the peak position of the blood voltage waveform. The blood voltage rise rate and blood voltage fall rate are characteristics related to the shape of the blood voltage waveform, reflecting the dynamic characteristics of blood volume changes.

[0122] S133: Obtain the pulse wave propagation time sequence, as follows:

[0123] (1) Extract the peak value of each heartbeat cycle from the standard pressure signal sequence to obtain the pressure peak sequence.

[0124] (2) Extract the peak value of each blood volume change cycle from the standard blood pressure sequence to obtain the blood voltage peak value sequence.

[0125] (3) Perform time-stamp-based subtraction on the blood voltage peak sequence and pressure peak sequence to obtain the pulse wave propagation time sequence, which can then be expressed as: ,in, This represents the pulse wave propagation time series. A period of time for transmission, Indicates the first peak in the pressure peak sequence The timestamps corresponding to each pressure peak Indicates the first peak in the blood voltage sequence The timestamp corresponding to each peak blood voltage.

[0126] The heartbeat is the driving force behind changes in blood volume, which in turn are the result of the heartbeat's propagation through the vascular system. Pulse wave propagation time represents the time it takes for the pulse wave to travel along the conduction path (from heartbeat to changes in blood volume), reflecting the elasticity and hemodynamic characteristics of the arteries. Since the heartbeat occurs earlier than the blood volume change, the pulse wave propagation time is positive. If the pulse wave propagation time is negative due to the timing of the data collection, a standard blood pressure sequence for one blood volume change cycle is discarded. Under physiological conditions, the duration of the heartbeat cycle and the blood volume change cycle are equal because they are both driven by the heartbeat and represent different manifestations of the same physiological process. Therefore, the pulse wave propagation time can be considered as being derived from the heartbeat cycle.

[0127] In S140, features of the standard acceleration sequence and standard angular velocity sequence are extracted based on the rate of change and frequency to obtain the motion intensity sequence. The motion intensity sequence includes at least a motion intensity level sequence, an acceleration rate of change sequence, and a motion frequency sequence, and includes the following steps:

[0128] S141: Perform rate-based summation on the standard acceleration sequence and the standard angular velocity sequence to obtain the composite rate sequence, which can then be expressed as: ,in, Represents the first in the integrated motion sequence The rate of a timestamp These represent the first and second accelerations in the standard acceleration sequence, respectively. The acceleration of each timestamp The acceleration value in the axial direction, These represent the 6th angular velocity in the standard angular velocity sequence. The angular velocity of each timestamp is Angular velocity value in the axial direction.

[0129] S142: Based on the comprehensive rate sequence, the first preset rate threshold, and the second preset rate threshold, a motion intensity level sequence is obtained, which can then be expressed as: ,in, This represents the first in the sequence of exercise intensity levels. The exercise intensity level value of each timestamp This indicates the first preset rate threshold. This indicates the second preset rate threshold. An exercise intensity level of 0 indicates the user is stationary, 1 indicates the user is walking, and 2 indicates the user is running.

[0130] S143: Calculate the rate of change of acceleration in the standard acceleration sequence to obtain the acceleration rate of change sequence, which is then expressed as: ,in, Represents the first in the sequence of acceleration rate of change The rate of change of acceleration at each timestamp In the standard acceleration sequence, the first... The acceleration of a timestamp In the standard acceleration sequence, the first... The acceleration of a timestamp In the standard acceleration sequence, the first... The timestamp corresponding to the acceleration of each timestamp. In the standard acceleration sequence, the first... The timestamp corresponding to the acceleration of each timestamp.

[0131] S144: Based on Fourier transform, extract the frequency of the standard acceleration sequence within each preset period to obtain the acceleration frequency feature sequence.

[0132] S145: Extract the maximum frequency, i.e., the motion frequency, from each preset period in the acceleration frequency feature sequence, thus obtaining the motion frequency sequence. The motion frequency can then be expressed as: ,in, Indicates the frequency of motion. Indicates Fourier transform, This represents the range of acceleration frequencies in the acceleration frequency characteristic sequence. This represents the acceleration frequency corresponding to the maximum value of the function. The Fourier transform is used to perform frequency domain analysis on acceleration data to extract the main motion frequency components; the Fast Fourier Transform (FFT) is preferably used.

[0133] The preset period is an integer multiple of the heartbeat period ≥ 2. The timestamp corresponding to the heartbeat period can be obtained from the standard pressure signal sequence. The timestamp corresponding to the preset period can be extracted accordingly.

[0134] The rate of change of acceleration and motion frequency are used to capture the dynamic changes in the user's motion state, thereby further refining the impact of exercise on blood pressure.

[0135] The order of steps S130 and S140 is not limited.

[0136] In S150, pressure feature sequences, blood pressure feature sequences, pulse wave propagation time sequences, and exercise intensity sequences are fused to form a multidimensional feature sequence.

[0137] Each value in each sequence of the pressure characteristic series (pressure peak sequence, pressure trough sequence, pressure rise time series, and pressure fall time series) is derived based on each heartbeat cycle, and the length of each sequence in the pressure characteristic series is the same as the total number of heartbeat cycles. Similarly, each value in each sequence of the blood pressure characteristic series (blood voltage peak sequence, blood voltage rise rate sequence, and blood voltage fall rate sequence) is derived based on each blood volume change cycle, and the length of each sequence in the blood pressure characteristic series is the same as the total number of blood volume change cycles. As mentioned earlier, each value in the pulse wave propagation time series can be considered as derived based on each heartbeat cycle. Since the duration of the heartbeat cycle and the blood volume change cycle is equal, the lengths of the pressure characteristic series, blood pressure characteristic series, and pulse wave propagation time series are generally the same, occasionally differing by one value. In such cases, the first or last value of the longer sequence is discarded.

[0138] Each value in the exercise intensity level sequence and acceleration rate of change sequence in the exercise intensity sequence is obtained based on each timestamp. The sequence length is significantly greater than the total number of heartbeat cycles. The processing is as follows: (1) Extract the timestamp of each heartbeat cycle in the standard pressure signal sequence; (2) Calculate the average acceleration rate of change in each heartbeat cycle based on the timestamp of each heartbeat cycle to obtain the average acceleration rate of change sequence; (3) Obtain the exercise intensity level that appears most frequently in each heartbeat cycle based on the timestamp of each heartbeat cycle to obtain the main exercise intensity level sequence. In this way, the processed exercise intensity level sequence and acceleration rate of change sequence are consistent with the sequence length of the pressure feature sequence / blood pressure feature sequence / pulse wave propagation time sequence.

[0139] The motion frequency sequence is obtained based on a preset period, which is an integer multiple of the heart rate cycle (≥2). Therefore, the length of each sequence in the pressure feature sequence is also an integer multiple of the length of the motion frequency sequence. Thus, it is only necessary to extend the motion frequency within the corresponding preset period in the motion frequency sequence to the corresponding heart rate cycle to obtain the extended motion frequency sequence. The length of the extended motion frequency sequence is consistent with the length of the other (processed) sequences.

[0140] Since all the processed sequences have the same length, the fusion process simply matches the values ​​of each sequence in order.

[0141] In S160, the blood pressure prediction pre-training model is trained based on multidimensional feature sequences to obtain the blood pressure prediction model, including the following steps:

[0142] S161: Construct a pre-trained blood pressure prediction model. This model can directly utilize existing AI models or be built independently using AI technology. Building a custom pre-trained model allows for a better balance between performance, efficiency, and scenario requirements. For example, the model architecture can be customized based on specific sensor data from wearable devices (such as MEMS pressure sensors, PPG, etc.) for targeted optimization, thereby improving prediction accuracy. In this embodiment, a self-built pre-trained model is used. The pre-trained blood pressure prediction model includes a convolutional recursive fusion module, an attention weighting module, and a prediction module, as detailed below:

[0143] (1) The convolutional recursive fusion module performs convolution and bidirectional recursive operations on the input vector to obtain convolutional feature sequences and bidirectional feature sequences; it then fuses the convolutional feature sequences and bidirectional feature sequences to obtain a comprehensive feature sequence. In this embodiment, the convolutional operation is implemented using a convolutional neural network (CNN), and the bidirectional recursive operation is implemented using a bidirectional recursive neural network (Bi-RNN).

[0144] CNNs process the waveform features of PPG signals to extract local time-frequency features. Bi-RNNs can process the temporal features of pressure signals and motion data to capture long-term dependencies.

[0145] The convolutional feature sequence can then be represented as: A bidirectional feature sequence can be represented as: The comprehensive feature sequence can be represented as: ,in, Represents the convolutional feature sequence. Represents the input vector. This represents the activation function. Indicates the convolution kernel weights. Indicates the bias term. Represents a bidirectional feature sequence. This indicates a bidirectional recursive operation. Represents a comprehensive feature sequence. This represents the convolution operation. This indicates a fusion operation.

[0146] (2) The attention weighting module processes the comprehensive feature sequence using an attention mechanism to weight the importance of different features, resulting in an attention weight sequence. The comprehensive feature sequence and the attention weight sequence are then weighted to obtain a weighted feature sequence. The attention weight sequence can then be expressed as: The weighted feature sequence can be represented as: ,in, Represents the first [number] in the attention weight sequence Each attention weight, Indicates the first Each importance score , All of these represent learnable parameters. Indicates matrix transpose. Represents the first in the comprehensive feature sequence One characteristic, This represents the dimension of the integrated feature sequence. This represents a weighted feature sequence.

[0147] (3) The prediction module maps the weighted feature sequence through a fully connected layer to obtain the predicted blood pressure data. The predicted blood pressure data can then be expressed as: ,in, This represents the predicted value of blood pressure data. This represents the predicted systolic blood pressure value. This represents the predicted diastolic blood pressure value. Indicates the weights of the fully connected layer. This indicates the bias of the fully connected layer.

[0148] S162: Based on a preset partitioning ratio, the multidimensional feature sequence is partitioned to obtain training, testing, and validation datasets. User datasets containing different ages, genders, health conditions, and lifestyles can be constructed to improve the model's generalization ability. Furthermore, data augmentation methods (such as noise addition and time warping) can be applied to expand the training dataset, enhancing the model's robustness and resistance to interference.

[0149] S163: Input the training dataset into the pre-trained blood pressure prediction model, obtain the model's predicted output through the forward propagation algorithm, and combine it with the actual values ​​to obtain the loss function value. The loss function value is the sum of the mean squared error of diastolic blood pressure, the mean squared error of systolic blood pressure, and the regularization value. This prevents overfitting and improves the model's generalization ability. Furthermore, sharing hidden layers can improve the model's overall understanding of blood pressure changes.

[0150] The loss function value can then be expressed as: The mean square error of systolic blood pressure can be expressed as: The mean square error of diastolic blood pressure can be expressed as: The regularization value can be expressed as: ,in, This represents the value of the loss function. This represents the mean square error of systolic blood pressure. These represent the mean square error of diastolic blood pressure, respectively. These represent the weighting coefficients for the mean square error of systolic blood pressure and the mean square error of diastolic blood pressure, respectively. This indicates the number of samples in the training dataset. These represent the predicted systolic and diastolic blood pressure values ​​from the pre-trained blood pressure prediction model, respectively. These represent the actual values ​​of systolic and diastolic blood pressure, respectively. Represents the regularization value. Indicates the regularization strength. Indicates the regularization weight. This indicates a regularization bias.

[0151] S164: Calculate the gradient of the parameters of the pre-trained blood pressure prediction model based on the loss function value using the backpropagation algorithm.

[0152] S165: Based on the learning rate over the training cycle, the parameters of the pre-trained blood pressure prediction model are updated. Specifically, the Adam optimizer is used for efficient gradient descent to quickly converge to the optimal solution, and a learning rate decay strategy is employed to gradually reduce the learning rate, stabilize the training process, and improve model performance. The learning rate can then be expressed as: ,in, Indicates the first The learning rate for each training cycle. This represents the initial learning rate. This represents the learning rate decay coefficient. Indicates the training period.

[0153] S166: Systematically adjust hyperparameters (such as learning rate, batch size, and number of network layers) using grid search or Bayesian optimization methods to find the optimal model configuration. Evaluate the generalization ability of the pre-trained blood pressure prediction model using a validation dataset.

[0154] S167: Iterate the training until the iteration termination condition is met to obtain the blood pressure prediction model, wherein the iteration termination condition includes the convergence of the loss function value or reaching the preset number of training iterations.

[0155] Example 2:

[0156] A training system for a blood pressure prediction model, used to implement the training method for the blood pressure prediction model in any of the above embodiments, such as... Figure 2 As shown, the training system includes:

[0157] The data acquisition module 110 is used to acquire data from the wearable device to obtain a raw data sequence, wherein the data from the wearable device includes at least pressure signals, blood pressure signals, acceleration, and angular velocity;

[0158] Data processing module 120 is used to preprocess the original data sequence to obtain a standard data sequence;

[0159] The first feature extraction module 130 is used to extract features of the standard pressure signal sequence and the standard blood pressure sequence based on the peak and trough values, respectively, to obtain the pressure feature sequence, the blood pressure feature sequence, and the pulse wave propagation time sequence.

[0160] The second feature extraction module 140 is used to extract features from the standard acceleration sequence and the standard angular velocity sequence based on the rate of change and frequency to obtain the motion intensity sequence.

[0161] The feature fusion module 150 is used to fuse pressure feature sequences, blood pressure feature sequences, pulse wave propagation time sequences, and exercise intensity sequences to form multidimensional feature sequences.

[0162] The model training module 160 is used to train the pre-trained blood pressure prediction model based on multi-dimensional feature sequences to obtain the blood pressure prediction model.

[0163] Example 3:

[0164] A method for predicting blood pressure data, such as Figure 3 As shown, it includes the following steps:

[0165] S210: Acquire target data from wearable devices;

[0166] S220: Process the target data based on the blood pressure prediction model to obtain the predicted blood pressure data value;

[0167] The blood pressure prediction model is a model obtained by training the blood pressure prediction model using any of the above embodiments.

[0168] Example 4:

[0169] A blood pressure data prediction system, such as Figure 4 As shown, the prediction system includes:

[0170] Acquisition module 210 is used to acquire target data of wearable devices;

[0171] Prediction module 220 is used to infer the target data based on the blood pressure prediction model to obtain the predicted blood pressure data value;

[0172] The blood pressure prediction model is a model obtained by training the blood pressure prediction model using any of the above embodiments.

[0173] Various changes and modifications made without departing from the spirit and scope of this invention, and all equivalent technical solutions, also fall within the scope of this invention.

[0174] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0175] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0176] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0177] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0178] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0179] It should be noted that:

[0180] The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0181] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.

Claims

1. A training method for a blood pressure prediction model, characterized in that, The method comprises the following steps: obtaining data of a wearable device to obtain an original data sequence, wherein the data of the wearable device at least includes a pressure signal, a blood pressure signal, acceleration and angular velocity; preprocessing the original data sequence to obtain a standard data sequence; extracting features of the standard pressure signal sequence and the standard blood pressure sequence based on peaks and valleys to obtain a pressure feature sequence, a blood pressure feature sequence and a pulse wave propagation time sequence; extracting features of the standard acceleration sequence and the standard angular velocity sequence based on a change rate and a frequency to obtain a motion intensity sequence; fusing the pressure feature sequence, the blood pressure feature sequence, the pulse wave propagation time sequence and the motion intensity sequence to form a multi-dimensional feature sequence; training a blood pressure prediction pre-training model based on the multi-dimensional feature sequence to obtain a blood pressure prediction model; wherein the blood pressure prediction pre-training model comprises a convolutional recurrent fusion module, an attention weighting module and a prediction module; the convolutional recurrent fusion module performs convolution operation and bidirectional recurrent operation on the input vector respectively to obtain a convolution feature sequence and a bidirectional feature sequence; and fuses the convolution feature sequence and the bidirectional feature sequence to obtain a comprehensive feature sequence; the attention weighting module performs attention mechanism processing on the comprehensive feature sequence to obtain an attention weight sequence, and performs weighting processing on the comprehensive feature sequence and the attention weight sequence to obtain a weighted feature sequence; the prediction module maps the weighted feature sequence through a fully connected layer to obtain a blood pressure data prediction value. 2.The method of claim 1, wherein, The pressure feature sequence at least includes a pressure peak value sequence, a pressure valley value sequence, a pressure rising time sequence and a pressure falling time sequence, and the pressure feature sequence is obtained by the following steps: extracting peaks and valleys of each heartbeat cycle of the standard pressure signal sequence to obtain a pressure peak-valley value sequence, a pressure peak value sequence and a pressure valley value sequence; performing difference processing on the timestamp of a target peak value and the timestamp of a previous valley value before the target peak value in the pressure peak-valley value sequence to obtain a pressure rising time sequence; performing difference processing on the timestamp of a target valley value and the timestamp of a previous peak value before the target valley value in the pressure peak-valley value sequence to obtain a pressure falling time sequence. 3.The method of claim 1, wherein, The blood pressure feature sequence at least includes a blood voltage peak value sequence, a blood voltage rising rate sequence and a blood voltage falling rate sequence, and the blood pressure feature sequence is obtained by the following steps: extracting peaks and valleys of each blood volume change cycle of the standard blood pressure sequence to obtain a blood voltage peak-valley value sequence and a blood voltage peak value sequence; calculating the rate at which the blood voltage rises from a valley value to a subsequent peak value in the blood voltage peak-valley value sequence to obtain a blood voltage rising rate sequence; calculating the rate at which the blood voltage falls from a peak value to a subsequent valley value in the blood voltage peak-valley value sequence to obtain a blood voltage falling rate sequence. 4.The method of claim 1, wherein, The pulse wave propagation time sequence is obtained by the following steps: extracting peaks of each heartbeat cycle of the standard pressure signal sequence to obtain a pressure peak value sequence; extracting peaks of each blood volume change cycle of the standard blood pressure sequence to obtain a blood voltage peak value sequence; The pressure peak sequence and the blood voltage peak sequence are timestamp-based subtracted to obtain a pulse wave propagation time sequence. 5.The method of claim 1, wherein, The motion intensity sequence includes at least a motion intensity level sequence, an acceleration change rate sequence and a motion frequency sequence, and the motion intensity sequence is obtained by the following steps: The standard acceleration sequence and the standard angular velocity sequence are rate-based summed to obtain a comprehensive rate sequence; The motion intensity level sequence is obtained based on the comprehensive rate sequence, a first preset rate threshold and a second preset rate threshold; The change rate of acceleration in the standard acceleration sequence is calculated to obtain an acceleration change rate sequence; The acceleration frequency feature sequence is obtained by extracting the frequency in each preset period of the standard acceleration sequence based on Fourier transform; The motion frequency sequence is obtained by extracting the maximum frequency in each preset period in the acceleration frequency feature sequence as the motion frequency. 6.The method of claim 1, wherein, The blood pressure data prediction value is represented as follows: wherein, represents the systolic blood pressure prediction value, represents the diastolic blood pressure prediction value, represents the matrix transpose, represents the fully connected layer weight, represents the weighted feature sequence, represents the fully connected layer bias.

7. A blood pressure data prediction method, characterized by, The following steps are included: Obtaining target data of a wearable device; Processing the target data based on a blood pressure prediction model to obtain a blood pressure data prediction value; The blood pressure prediction model is a model obtained by the training method of the blood pressure prediction model according to any one of claims 1 to 6.

8. A training system for a blood pressure prediction model, characterized in that, The training system for implementing the training method of the blood pressure prediction model according to any one of claims 1 to 6 includes: A data acquisition module for acquiring data of a wearable device to obtain an original data sequence, wherein the data of the wearable device includes at least a pressure signal, a blood pressure signal, acceleration and angular velocity; A data processing module for pre-processing the original data sequence to obtain a standard data sequence; A first feature extraction module for extracting features of the standard pressure signal sequence and the standard blood pressure sequence based on peaks and valleys, respectively, to obtain a pressure feature sequence, a blood pressure feature sequence and a pulse wave propagation time sequence; A second feature extraction module for extracting features of the standard acceleration sequence and the standard angular velocity sequence based on change rates and frequencies to obtain a motion intensity sequence; A feature fusion module for fusing the pressure feature sequence, the blood pressure feature sequence, the pulse wave propagation time sequence and the motion intensity sequence to form a multi-dimensional feature sequence; A model training module for training a blood pressure prediction pre-training model based on the multi-dimensional feature sequence to obtain a blood pressure prediction model.

9. A blood pressure data prediction system, characterized by, The prediction system includes: An acquisition module for acquiring target data of a wearable device; A prediction module for inferring the target data based on a blood pressure prediction model to obtain a blood pressure data prediction value; wherein the blood pressure prediction model is a model obtained by the training method of the blood pressure prediction model according to any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the training method of the blood pressure prediction model according to any one of claims 1 to 6. 11.A device for training a blood pressure prediction model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the device is configured to perform the method according to any one of claims 1-10. The processor executes the computer program to implement the training method of the blood pressure prediction model according to any one of claims 1 to 6.

Images