Blood pressure measurement method, apparatus and system

Abstract

The application relates to the technical field of physiological signal processing, and discloses a blood pressure measurement method, device and system, the method comprising the following steps: acquiring a digital audio signal and a pressure value sequence which are synchronously collected in a blood pressure measurement process of a subject; decomposing the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition; determining a target Korotkoff sound signal according to an intrinsic mode function with a center frequency located in a target frequency band range of 15 Hz-80 Hz; determining a variation rate parameter of a spectral energy centroid according to the spectral energy centroid of each Korotkoff sound pulse in the target Korotkoff sound signal; acquiring a heart rate parameter of the subject; and taking the target Korotkoff sound signal, the pressure value sequence, the variation rate parameter and the heart rate parameter as input data of a blood pressure prediction model to determine a blood pressure value of the subject. The application can effectively separate noise and Korotkoff sound, guarantee time-frequency high fidelity of the extracted target Korotkoff sound signal, and improve the accuracy of diastolic pressure feature extraction.

Description

Technical Field

[0001] This application relates to the field of physiological signal processing technology, specifically to blood pressure measurement methods, devices, and systems. Background Technology

[0002] Currently, non-invasive blood pressure (NIBP) measurement mainly employs oscillometric or conventional auscultatory methods. The oscillometric method calculates blood pressure based on the pressure oscillation envelope during the inflation and deflation of the cuff, combined with an empirical proportionality coefficient. However, this empirical proportionality coefficient is usually obtained based on population statistics. When subjects have variations in vascular compliance, arteriosclerosis, or arrhythmia, the characteristic morphology of the oscillation wave will change, easily leading to reduced accuracy of individual measurement results.

[0003] To mitigate the limitations of single-modal measurement, related technologies have proposed blood pressure measurement schemes that combine multimodal signals such as electrocardiograms (ECG) and photo-plethysmography (PPG). These schemes comprehensively evaluate blood pressure by extracting features such as the time difference between multiple physiological signals. However, multimodal measurement schemes have the following drawbacks: First, they require the integration of multiple types of sensors and independent analog front-end circuits into the device, increasing the physical size, hardware cost, and power consumption, thus limiting their application in miniaturized wearable devices. Second, due to the inherent physical delays and sampling rate differences among sensor components, phase deviations can easily occur during the acquisition of multi-source signals, increasing the difficulty of achieving high-precision synchronization and alignment of signals in the time domain.

[0004] To reduce the number of hardware components, blood pressure measurement methods based on single-modal Korotkoff sounds combined with neural networks have been gradually proposed. However, this approach typically extracts Korotkoff sound signals using bandpass filtering, making it difficult to accurately separate noise from Korotkoff sounds, resulting in errors in the final measurement results. Summary of the Invention

[0005] In view of this, this application provides a blood pressure measurement method, device and system to solve the problem of large errors in blood pressure measurement results.

[0006] In a first aspect, this application provides a method for measuring blood pressure, the method comprising: Acquire digital audio signals and pressure value sequences simultaneously collected during blood pressure measurement of subjects; The digital audio signal is decomposed into multiple intrinsic mode functions based on variational mode decomposition. The intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz are extracted. The extracted intrinsic mode functions are linearly superimposed to determine the target Korotkoff tone signal. The rate of change parameter of the spectral energy centroid is determined based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal. Obtain the heart rate parameters of the subject; The target Korotkoff sound signal, the pressure value sequence, the rate of change parameter, and the heart rate parameter are used as input data for the blood pressure prediction model, and the subject's blood pressure value is determined based on the output of the blood pressure prediction model.

[0007] Secondly, this application provides a blood pressure measuring device, the device comprising: The data acquisition module is used to acquire digital audio signals and pressure value sequences that are synchronously collected during the blood pressure measurement of the subject. The reconstruction module is used to decompose the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition, extract the intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz, and linearly superimpose the extracted intrinsic mode functions to determine the target Korotkoff tone signal. The processing module is used to determine the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal. Heart rate acquisition module, used to acquire the heart rate parameters of the subject; The prediction module is used to take the target Korotkoff sound signal, the pressure value sequence, the rate of change parameter and the heart rate parameter as input data of the blood pressure prediction model, and determine the subject's blood pressure value based on the output of the blood pressure prediction model.

[0008] Thirdly, this application provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the blood pressure measurement method of the first aspect or any corresponding embodiment described above.

[0009] Fourthly, this application provides a computer-readable storage medium storing computer instructions for causing a computer to perform the blood pressure measurement method of the first aspect or any corresponding embodiment described above.

[0010] Fifthly, this application provides a computer program product, including computer instructions for causing a computer to execute the blood pressure measurement method of the first aspect or any corresponding embodiment described above.

[0011] Sixthly, this application provides a blood pressure measurement system, the system comprising: an audio acquisition module, a pressure acquisition module, and a controller; The audio acquisition module is used to acquire the acoustic signal of the subject's blood pressure, extract the corresponding digital audio signal, and send the digital audio signal to the controller; The pressure acquisition module is used to synchronously acquire the pressure values ​​of the subject, form a pressure value sequence, and send the pressure value sequence to the controller; The controller is used to perform the blood pressure measurement method of the first aspect described above or any corresponding embodiment thereof.

[0012] The blood pressure measurement method provided in this application uses a variational mode decomposition algorithm to adaptively decompose the original digital audio signal, achieving effective separation between noise and Korotkoff sounds, ensuring high time-frequency fidelity of the extracted target Korotkoff sound signal; furthermore, it identifies the spectral energy centroid of each Korotkoff sound pulse in the target Korotkoff sound signal and determines the corresponding rate of change parameter, thereby extracting the feature reference point where the Korotkoff sound energy changes drastically during the diastolic blood pressure transition phase. Combined with heart rate parameters, this enables the blood pressure prediction model to accurately extract relevant features from the acoustic signal, reducing noise interference in feature optimization, improving the accuracy of diastolic blood pressure feature extraction, and predicting accurate blood pressure values. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of a blood pressure measurement system according to an embodiment of this application; Figure 2 This is a schematic flowchart of a first blood pressure measurement method according to an embodiment of this application; Figure 3 This is a schematic diagram of a second blood pressure measurement method according to an embodiment of this application; Figure 4 This is a schematic diagram of the original digital audio signal according to an embodiment of this application; Figure 5 This is a schematic diagram of a target Korotkoff tone signal according to an embodiment of this application; Figure 6 This is a schematic diagram of the spectral energy centroid according to an embodiment of this application; Figure 7 This is a detailed schematic diagram of a spectral energy centroid sequence curve according to an embodiment of this application; Figure 8This is a schematic diagram of the Shannon energy envelope waveform of the target Korotkoff tone signal according to an embodiment of this application; Figure 9 This is a schematic diagram of the architecture of a blood pressure prediction model according to an embodiment of this application; Figure 10 This is a structural block diagram of a blood pressure measuring device according to an embodiment of this application; Figure 11 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0017] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] The Korotkoff sound method is a non-invasive method that indirectly measures blood pressure by detecting changes in the sound of arterial blood flow. The Korotkoff sounds are a series of sounds heard when measuring blood pressure, and it was discovered by the Russian physician Kolotkov.

[0019] Korotkoff sounds can be divided into five phases. The first sound (systolic pressure) is crisp and gradually intensifies, marking phase one. As the pressure within the cuff continues to decrease, the crisp sound softens, becoming more like a heart murmur, marking phase two. As the pressure decreases further, the sound returns to an intensified sound similar to phase one, marking phase three. When the pressure drops to the point where the sound suddenly weakens and deepens (tone change), this is phase four (Korotkoff phase four). When the pressure decreases further until the sound disappears (disappearance sound), this is phase five (Korotkoff phase five).

[0020] Specifically, the first phase: the acoustic characteristic point of systolic blood pressure (initial appearance). When the cuff pressure decreases to the critical point equivalent to the cardiac systolic blood pressure, local blood flow breaks through the compression surface to form an initial jet, triggering transient deformation of the blood vessel wall. Its acoustic signal characteristics are: the first appearance of a high-frequency acoustic pulse with a significant envelope peak in the time domain.

[0021] Phase Two: The Characteristic Phase of Blood Flow Turbulence. As the cuff pressure further decreases, the effective cross-sectional area of ​​the blood vessel increases, and the fluid Reynolds number changes. Its acoustic signal characteristics are as follows: a mid-to-low frequency broadband continuous signal is generated by superposition between high-frequency pulse signals, presenting a complex spectral structure of characteristic pulses and broadband background noise.

[0022] Phase 3: Peak amplitude characteristic period. A specific equilibrium is reached between vascular wall compliance and dynamic blood flow resistance. Its acoustic signal characteristics are: attenuation of the background broadband signal, and the time-domain amplitude and frequency-domain energy density of the high-frequency acoustic pulse reaching local maxima throughout the entire pressure measurement cycle.

[0023] Phase 4: The transitional phase of diastolic blood pressure. Vascular deformation tends to return to a natural, uncompressed state, and fluid dynamics transition from jet-dominated to stable laminar flow. Its acoustic signal characteristics are: a sharp attenuation of high-frequency components, a significant shift of the spectral energy centroid to the low-frequency band, and a nonlinear decreasing trend in the total signal envelope energy.

[0024] Phase 5: Acoustic Characteristic Point of Diastolic Pressure (Disappearance Phase). Vascular compression is completely relieved, blood flow fully returns to laminar flow, and hemodynamics no longer excites vibrations in the vessel wall sufficient to generate measurable sound waves. Its acoustic signal characteristics are: the characteristic acoustic pulse signal completely disappears, and the total energy decays and approaches the baseline threshold of the ambient background noise.

[0025] To reduce hardware components, blood pressure measurement can be achieved by acquiring and identifying single-modal Korotkoff sounds and combining them with neural networks (such as blood pressure prediction networks). However, in practical applications, the above methods still have the following technical problems: The accuracy of Korotkoff sound feature extraction is insufficient. Korotkoff sounds are non-stationary acoustic signals, often accompanied by low-frequency muscle tremor artifacts and broadband mechanical aerodynamic noise during measurement. Existing signal preprocessing methods mostly employ fixed-bandwidth bandpass filtering or simple time-domain envelope analysis, failing to effectively separate the Korotkoff sound signal components. Particularly during the fourth (chromatic) to fifth (disappearance) phases of the Korotkoff sound corresponding to diastolic blood pressure, the signal acoustic energy significantly attenuates and the spectral structure changes, making it difficult for existing feature extraction methods to achieve effective separation at low signal-to-noise ratios, easily leading to errors in diastolic blood pressure feature extraction.

[0026] As one optional application scenario in the embodiments of this application, such as Figure 1As shown, the blood pressure measurement system provided in this embodiment includes: a blood pressure measurement hardware body and a controller; wherein, the blood pressure measurement hardware body specifically includes an audio acquisition module, an airway component and a pressure acquisition module.

[0027] The audio acquisition module is used to acquire acoustic signals related to blood pressure during blood pressure measurement of a subject, such as digital audio signals from which Korotkoff sounds can be extracted. Specifically, for example... Figure 1 As shown, the audio acquisition module may specifically include a pickup sensor, a low-noise amplifier, and an analog-to-digital converter.

[0028] The microphone is positioned inside the cuff and configured to acquire acoustic signals from the subject's artery (e.g., the brachial artery). The signal output of the microphone is sequentially connected to a low-noise amplifier (LNA) and an analog-to-digital converter (ADC). The LNA amplifies the acoustic signal acquired by the microphone, while the ADC converts the analog electrical signal (i.e., the amplified acoustic signal) from the LNA into a digital audio signal, which is then output. The sampling frequency of the ADC is configurable to be no less than 2 kHz.

[0029] The gas path components are conventional gas path components used to achieve blood pressure measurement, such as... Figure 1 As shown, it includes a miniature air pump, an airbag, and a three-way solenoid valve. The three ports of the three-way solenoid valve are respectively connected to the miniature air pump, the airbag, and the external atmosphere. The controller can be electrically connected to the three-way solenoid valve and configured to output a pulse-width modulation (PWM) signal to control the opening and closing state and duty cycle of the three-way solenoid valve, thereby controlling the inflation, pressure holding, and linear deflation of the airbag at a preset rate (such as 2-3 mmHg / s). The working principle of the air circuit components is not described in detail in this embodiment.

[0030] The pressure acquisition module includes a piezoresistive pressure sensor, whose air path is connected to the airbag and its electrical connection is connected to the controller. It is used to output the absolute pressure value of the airbag in real time and form a corresponding pressure value sequence.

[0031] This controller is used to process the acquired digital audio signals to achieve blood pressure measurement, such as... Figure 1As shown, it can be specifically divided into: a data caching unit, a signal processing and feature extraction unit, a prediction model unit, and an adaptive calibration unit. The data caching unit is used to cache digital audio signals; the signal processing and feature extraction unit is used to perform signal processing based on VMD and extract the required features (such as spectral energy centroid change rate, heart rate, etc.); the prediction model unit is used to deploy the blood pressure measurement model and predict blood pressure values; the adaptive calibration unit is used to fine-tune some weight parameters of the blood pressure prediction model based on accurate blood pressure reference values. Each unit will be described in detail later.

[0032] The blood pressure measurement method provided in this embodiment uses a variational mode decomposition algorithm to adaptively decompose the original digital audio signal, achieving effective separation between noise and Korotkoff sounds, and ensuring high time-frequency fidelity of the extracted target Korotkoff sound signal. Furthermore, it identifies the centroid of the spectral energy of each Korotkoff sound pulse in the target Korotkoff sound signal and determines the corresponding rate of change parameter, thereby extracting the feature reference point where the Korotkoff sound energy changes drastically during the diastolic blood pressure transition phase. Combined with heart rate parameters, this allows the blood pressure prediction model to accurately extract relevant features from the acoustic signal, reducing the interference of noise on feature optimization, improving the accuracy of diastolic blood pressure feature extraction, and predicting accurate blood pressure values.

[0033] According to an embodiment of this application, a blood pressure measurement method embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0034] This embodiment provides a blood pressure measurement method, which can be used in a blood pressure measurement system, specifically applied to components with processing functions within the blood pressure measurement system, such as... Figure 1 The controller shown. Figure 2 This is a flowchart of a blood pressure measurement method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps.

[0035] Step S201: Acquire the digital audio signal and pressure value sequence that are synchronously collected during the blood pressure measurement of the subject.

[0036] In this embodiment, after receiving the measurement start signal, the controller controls the micro air pump to operate to inflate the airbag; obtains the pressure value inside the airbag, and controls the micro air pump to stop operating when the pressure value reaches the preset pressure threshold; then controls the three-way solenoid valve to open, so that the airbag deflates at a preset deflation rate.

[0037] During the deflation process, a digital audio signal is acquired through an audio acquisition module and stored in a local data buffer unit; at the same time, a pressure acquisition module synchronously acquires real-time absolute pressure values ​​to form a pressure value sequence.

[0038] Among them, the digital audio signal and the corresponding real-time pressure value can be dynamically slid and intercepted by a set time window and sliding step size to construct a time-series mapping relationship between the digital audio signal and the real-time pressure value containing at least one complete cardiac cycle.

[0039] In this embodiment, the data buffer unit can be a circular buffer allocated in the local microcontroller's static random access memory (SRAM). The local microcontroller continuously writes the digital audio signal output from the analog-to-digital converter into the circular buffer via a direct memory access (DMA) channel to avoid data overflow caused by asynchronous sampling. When signal processing is required, the digital audio signal can be read from this data buffer unit.

[0040] Step S202: Based on variational mode decomposition, the digital audio signal is decomposed into multiple intrinsic mode functions, and the intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz are extracted. The extracted intrinsic mode functions are linearly superimposed to determine the target Korotkoff tone signal.

[0041] In this embodiment, the digital audio signal in the data buffer unit is called, and the variational mode decomposition (VMD) algorithm is used to decompose the digital audio signal into multiple intrinsic mode functions (IMFs).

[0042] Specifically, for the original acquired digital audio signal Variational mode decomposition yields K eigenmode functions, i.e. , Indicates the first One intrinsic mode function.

[0043] Among them, the eigenmode functions It is an amplitude-frequency modulation (AM-FM) signal, which can be represented as , Represents the instantaneous amplitude at time t; Represents the instantaneous phase at time t; each eigenmode function Both have spectral energy highly concentrated at a single center frequency. .

[0044] Based on the characteristics of the Korotkoff sound signal during blood pressure measurement, the required frequency band is preset, and the center frequency is set accordingly. The intrinsic mode functions falling within the target frequency band are used as the target intrinsic mode functions. By linearly superimposing each target intrinsic mode function, the Korotkoff tone signal, which is used for subsequent processing, can be reconstructed.

[0045] In practical blood flow acoustics, signals with a center frequency of 10Hz-15Hz are usually transitional noise caused by cuff friction against the skin or extremely low-frequency peristalsis of the blood vessel wall, while signals with a center frequency of 80Hz-100Hz are usually high-frequency pneumatic aftershock noise signals from air pump valves. Therefore, the target frequency band can be 15Hz-80Hz.

[0046] Specifically, high-frequency intrinsic mode functions with center frequencies between 40Hz and 80Hz and mid-frequency intrinsic mode functions with center frequencies between 15Hz and 40Hz can be extracted. The two intrinsic mode functions are then linearly superimposed to reconstruct the target Korotkoff tone signal. This can filter out the remaining intrinsic mode function components whose center frequencies do not fall within the continuous frequency range of 15Hz-80Hz, i.e., filter out low-frequency motion artifact modes with center frequencies <15Hz and high-frequency aerodynamic noise modes with center frequencies >80Hz.

[0047] The relevant scheme uses bandpass filtering to extract Korotkoff sound signals, such as static physical truncation based on Butterworth filter. This filtering method will treat all signals within the required frequency range (such as the target frequency band range mentioned above) as valid signals. If the subject's muscle tremors (which are a kind of noise) happen to be within this frequency range, it will cause the noise and Korotkoff sounds to be mixed together, making it impossible to separate the two.

[0048] Variational mode decomposition (VMD) is a data-driven adaptive decomposition that does not pre-define fixed filtering boundaries. Based on VMD, it decomposes digital audio signals and can adaptively find multiple intrinsic mode functions (IMFs) in the signal. This can effectively separate noise (such as muscle tremors) and Korotkoff sounds into two independent IMFs, even if they partially overlap in the frequency domain, thus achieving effective separation between noise and Korotkoff sounds.

[0049] Step S203: Determine the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal.

[0050] In this embodiment, the target Korotkoff sound signal is a continuous signal over a period of time, which includes multiple cardiac cycles. The Korotkoff sound pulses corresponding to each cardiac cycle can be extracted from the target Korotkoff sound signal, thereby determining the spectral energy centroid of each Korotkoff sound pulse.

[0051] Each contraction of the heart pumps out a stream of blood. When this blood stream impacts the brachial artery, which is compressed by the cuff, it generates a transient sound wave vibration. On the time axis, this appears as a concentrated energy wave packet, forming a Korotkoff sound pulse; that is, one cardiac cycle corresponds to one Korotkoff sound pulse.

[0052] For any Korotkoff pulse, signal processing can be performed to determine the spectral energy centroid. The spectral energy centroid refers to the centroid of the spectral energy distribution of the Korotkoff pulse, which can represent an average frequency position or frequency energy. Furthermore, based on the spectral energy centroid of each Korotkoff pulse, the variation of the spectral energy centroid can be determined, thereby obtaining the corresponding rate of change parameter.

[0053] In this embodiment, during the fourth and fifth stages corresponding to diastolic pressure, the acoustic energy of the Korotkoff sound decays sharply. Therefore, the critical point of the steep slope of the Korotkoff sound energy collapse can be determined based on the change in the centroid of the spectral energy. For example, the point where the rate of change of the centroid of the spectral energy is negative and minimal (maximum absolute value) can be used as a characteristic reference point for the diastolic pressure transition stage. The parameters related to the rate of change at this point can be used as the required rate of change parameters later. For example, the rate of change parameter may include the magnitude of the rate of change at the characteristic reference point, the timestamp, and the corresponding pressure value (determined based on the pressure value sequence).

[0054] Step S204: Obtain the subject's heart rate parameters.

[0055] In this embodiment, to ensure the accuracy of subsequent blood pressure measurements, it is also necessary to determine the subject's heart rate parameters. These parameters can be acquired using a separate hardware component (such as a heart rate sensor), or they can be determined based on extracted target Korotkoff sound signals; this embodiment does not limit the specific method used.

[0056] Specifically, the heart rate parameter can include the heart rate value and heart rate changes.

[0057] In step S205, the target Korotkoff sound signal, pressure value sequence, rate of change parameter, and heart rate parameter are used as input data for the blood pressure prediction model, and the subject's blood pressure value is determined based on the output of the blood pressure prediction model.

[0058] After the above processing, the target Korotkoff sound signal, pressure value sequence, rate of change parameter and heart rate parameter can be integrated and used as input data for the blood pressure prediction model. This allows the blood pressure prediction model to infer based on the input data and predict the subject's blood pressure value, such as the subject's systolic blood pressure and diastolic blood pressure.

[0059] The blood pressure prediction model is a pre-trained neural network model that can be trained offline using a training sample set. The training sample set preferably includes subject data from multiple age groups, different blood pressure distribution ranges, and different vascular compliance characteristics. Furthermore, the training sample includes at least: a target Korotkoff sound signal, real-time cuff pressure values ​​(i.e., a pressure value sequence) synchronized with the target Korotkoff sound signal, a reference systolic blood pressure label, and a reference diastolic blood pressure label. This embodiment does not detail the model training process.

[0060] The blood pressure measurement method provided in this embodiment uses a variational mode decomposition algorithm to adaptively decompose the original digital audio signal, achieving effective separation between noise and Korotkoff sounds, and ensuring high time-frequency fidelity of the extracted target Korotkoff sound signal. Furthermore, it identifies the centroid of the spectral energy of each Korotkoff sound pulse in the target Korotkoff sound signal and determines the corresponding rate of change parameter, thereby extracting the feature reference point where the Korotkoff sound energy changes drastically during the diastolic blood pressure transition phase. Combined with heart rate parameters, this allows the blood pressure prediction model to accurately extract relevant features from the acoustic signal, reducing the interference of noise on feature optimization, improving the accuracy of diastolic blood pressure feature extraction, and predicting accurate blood pressure values.

[0061] This embodiment provides a blood pressure measurement method, which can be used in a blood pressure measurement system, specifically applied to components with processing functions within the blood pressure measurement system, such as... Figure 1 The controller shown. Figure 3 This is a flowchart of a blood pressure measurement method according to an embodiment of this application, such as... Figure 3 As shown, the process includes the following steps.

[0062] Step S301: Acquire the digital audio signal and pressure value sequence that are synchronously collected during the blood pressure measurement of the subject.

[0063] Please see details Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0064] Step S302: Based on variational mode decomposition, the digital audio signal is decomposed into multiple intrinsic mode functions, and the intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz are extracted. The extracted intrinsic mode functions are linearly superimposed to determine the target Korotkoff tone signal.

[0065] Please see details Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0066] In some optional implementations, variational mode decomposition has certain decomposition hyperparameters, which can be preset fixed values ​​or finely adjusted in real time based on the current digital audio signal. Specifically, before step S302, the method further includes steps a1 to a2.

[0067] Step a1: Perform a Fast Fourier Transform on the digital audio signal to obtain the spectral energy distribution characteristics of the digital audio signal; the spectral energy distribution characteristics include the number of main energy peaks and the bandwidth of the main peaks.

[0068] Step a2: Determine the decomposition hyperparameters of variational mode decomposition based on the number of energy peaks and the bandwidth of the peaks; the decomposition hyperparameters include the number of intrinsic mode functions and the quadratic penalty factor.

[0069] Furthermore, the aforementioned step S302, "decomposing the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition," specifically includes: performing variational mode decomposition on the digital audio signal according to the decomposition hyperparameters to decompose the digital audio signal into multiple intrinsic mode functions.

[0070] In this embodiment, the decomposition hyperparameters of the variational mode decomposition (VMD) algorithm are adaptively determined based on the number of energy peaks and the bandwidth of the peaks in the spectral energy distribution characteristics of the digital audio signal. The decomposition hyperparameters include the number of intrinsic mode functions K and the quadratic penalty factor α.

[0071] Specifically, a Fast Fourier Transform (FFT) is performed on the original digital audio signal to statistically analyze the spectral energy distribution within the 0Hz-150Hz range. This allows for observation of the spectrum within this range, determining the number and bandwidth of the main energy peaks. If three distinct energy peaks appear in the frequency domain (i.e., the number of main energy peaks is three, such as extremely low-frequency friction sounds, core Korotkoff sounds, or high-frequency air pumping sounds), the value of K can be adaptively set, for example, to 4 or 5. This avoids situations where K is too small, leading to inability to decompose due to signal mode aliasing, or K is too large, causing a signal to be fragmented, i.e., over-decomposition of modes.

[0072] Experiments have shown that when K is less than 4, it is difficult to effectively separate high-frequency Korotkoff pulses from low-frequency motion artifacts; when K is greater than 7, mode over-splitting is likely to occur, resulting in energy dispersion in the target frequency band. Therefore, the value range of K is configured to be 4 to 7. A reference value of K can be preset, for example, K=5, and K can be dynamically adjusted based on this reference value.

[0073] The second-order penalty factor α determines the tightness of the filtering bandwidth. By evaluating the bandwidth of the main energy peak of the FFT, α can be adaptively adjusted. If the main peak is wide, α is reduced to allow a wider bandwidth to pass; if the main peak is sharp, α is increased for strict filtering.

[0074] Specifically, the ratio of the integral value of the spectral power in the 0-10Hz frequency band to the integral value of the spectral power in the 0-150Hz full frequency band is calculated. When the ratio exceeds a first preset threshold, it is determined that there is a strong motion artifact in the current signal, and the system proceeds according to a preset step size (e.g., ...). =500 The penalty factor α is increased incrementally until the change in the center frequency of the intrinsic mode function between two adjacent iterations is less than the set convergence threshold. For example, the first preset threshold is configured to be 25%-40%, and the value range of α is configured to be 1500-3000.

[0075] When the full width at half maximum (FWHM) of the mid-frequency band belonging to Korotkoff tones (e.g., the 40Hz-80Hz band) exceeds a second preset threshold, it is determined that the current Korotkoff audio spectrum exhibits band broadening. The secondary penalty factor α (a parameter controlling the filter width) or the number of decompositions K can be dynamically adjusted (e.g., increasing the number of decompositions K to reduce the risk of modal aliasing between different physiological components). For example, the second preset threshold is 15Hz.

[0076] During operation, the half-peak width (i.e., the frequency difference between the peak frequency and the half-power point) of the current main peak can be dynamically calculated, and then the half-peak width is compared with the second preset threshold.

[0077] For example, if the system calculates that the half-peak width is greater than the second preset threshold, it indicates that the energy peak is wide, the signal composition is complex, or that the real Korotkoff tones and mechanical aerodynamic noise are severely mixed together, resulting in blurred boundaries. In this case, the corresponding mechanism is triggered: the secondary penalty factor α is automatically reduced (for example, set to 1500) to widen the filter threshold and prevent the useful signal edges from being forcibly cut off.

[0078] If the full width at half maximum (FWHM) is less than or equal to the second preset threshold, it indicates that the energy peak is narrow and sharp. In this case, the signal is relatively pure, the characteristics are more obvious, and there is almost no noise interference. The mechanism at this time is to automatically increase the secondary penalty factor α (for example, set it to 2500) to filter out the extra small background noise on both sides.

[0079] Furthermore, if the half-peak width is within the first excess range (e.g., 15Hz < half-peak width ≤ 25Hz), in addition to reducing the second penalty factor α to the first low-order value (e.g., α = 1500) to broaden the variational basis width, the quantity K can also be adaptively increased, for example, to 6.

[0080] If the half-peak width is within the second excess range (e.g., half-peak width > 25Hz), it is determined that the current state is a broadband fluid flow state caused by extreme noise interference or severe arteriosclerosis. In this case, the quantity K is adaptively increased to the upper limit value of the boundary, such as 7, to maximize the frequency domain resolution and achieve forced stripping of deeply overlapping components.

[0081] After determining the decomposition hyperparameters, step S302 can be executed, and when using VMD to decompose the signal, the decomposition is performed according to the decomposition hyperparameters determined above.

[0082] Figure 4 A schematic diagram of the original digital audio signal is shown, including baseline drift and airflow noise. The target Korotkoff tone signal, reconstructed after VMD processing of this digital audio signal, frequency band filtering, and modal superposition, can be obtained as follows: Figure 5 As shown, it effectively eliminates low-frequency artifacts and high-frequency noise.

[0083] In this embodiment, the decomposition hyperparameters of the variational mode decomposition algorithm are adaptively determined based on the number of energy peaks and the bandwidth of the peaks in the spectral energy distribution characteristics. This allows for adaptive decomposition of the original digital audio signal, avoiding problems such as mode aliasing or over-decomposition caused by fixed parameters.

[0084] Optionally, the Korotkow tone signal obtained by linearly superimposing the target intrinsic mode functions whose center frequency falls within the target frequency band can be used as the target Korotkow tone signal. Alternatively, the validity of the directly obtained Korotkow tone signal (i.e., the Korotkow tone signal to be determined below) can be verified, and only if it is valid will it be used as the target Korotkow tone signal for subsequent use. Specifically, the above step S302, "extracting the intrinsic mode functions whose center frequency is located within the target frequency band of 15Hz-80Hz, linearly superimposing the extracted intrinsic mode functions, and determining the target Korotkow tone signal", specifically includes steps b1 to b3.

[0085] Step b1: Linearly superimpose the intrinsic mode functions of the target whose center frequency falls within the target frequency band of 15Hz-80Hz to reconstruct the Korotkoff tone signal to be determined.

[0086] Step b2: Extract the total acoustic energy and envelope periodicity of the Korotkoff tone signal to be determined.

[0087] Step b3: When the total acoustic energy is greater than the preset energy threshold and the heart rate corresponding to the envelope periodic characteristics is within the preset heart rate range, the undetermined Korotkoff sound signal is taken as the target Korotkoff sound signal.

[0088] For ease of description, the Korotkow tone signal directly obtained from the linear superposition of the intrinsic mode functions of each target is referred to as the "Korotkow tone signal to be determined". For this Korotkow tone signal to be determined, a preliminary validity check is first performed.

[0089] Specifically, the total acoustic energy of the Korotkoff sound signal within a preset time window is calculated, and the envelope periodicity features of the signal are initially extracted. If the total acoustic energy does not reach the preset background noise energy threshold, or the initial heart rate calculated from the envelope periodicity features does not fall within the preset human physiological heart rate range (e.g., 30 beats / min to 200 beats / min), it is determined that the currently acquired segment does not contain a valid Korotkoff sound pulse. The background noise energy threshold is specifically the average acoustic energy value obtained by acquiring audio signals within a preset time window (e.g., the first 2 seconds) in a resting state without inflating the airbag. The system directly generates an abnormal interrupt command and clears the current data cache, suspending subsequent feature extraction and blood pressure prediction operations to prevent invalid noise from entering the subsequent neural network prediction stage.

[0090] Step S303: Determine the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal.

[0091] Please see details Figure 2 Step S203 of the illustrated embodiment will not be described again here.

[0092] In some optional implementations, the above step S303, "determining the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal", includes steps c1 to c6.

[0093] Step c1: Extract the Korotkoff sound pulses corresponding to each cardiac cycle from the target Korotkoff sound signal.

[0094] Step c2: For any Korotkoff pulse, determine the energy value of the Korotkoff pulse at each frequency point.

[0095] Step c3: Determine the spectral energy centroid of the Korotkoff pulse based on the energy values ​​corresponding to each frequency point of the Korotkoff pulse.

[0096] Step c4: Based on the spectral energy centroid sequence generated from the spectral energy centroids of each Korotkoff pulse, a sliding window method is used to linearly fit the spectral energy centroid sequence to obtain the local slope.

[0097] Step c5: When multiple consecutive local slopes are less than the preset low-frequency collapse threshold, extract the critical point with the largest absolute value of slope change.

[0098] Step c6: The local slope corresponding to the critical point and the cuff pressure value matching the timestamp corresponding to the critical point are used as the rate of change parameter.

[0099] In this embodiment, the continuous target Korotkoff tone signal oscillates at high frequencies around the zero axis (e.g., Figure 5(As shown). First, the Shannon energy envelope of the target Korotkoff sound signal is calculated, which is equivalent to extracting a low-frequency smooth envelope curve from the high-frequency oscillation signal. Then, a dynamic threshold is used to find the energy peak on the smoothed curve. The timestamp corresponding to each peak corresponds to the position of a Korotkoff sound pulse. Taking this peak timestamp as the center, a fixed time length is taken to the left and right (e.g., -50ms to +150ms). The segment of the target Korotkoff sound signal within this time period is a complete Korotkoff sound pulse.

[0100] After the pulse positioning and segmentation output described above, each independent Korotkoff pulse can be obtained, which is a one-dimensional time-amplitude sequence, such as a signal segment with a length of 200ms.

[0101] For any Korotkoff pulse, the energy value of the Korotkoff pulse at each frequency point is determined based on the short-time Fourier transform.

[0102] Specifically, a specific smoothing window function (such as a Hamming window) is selected, and the Korotkoff pulse (such as a single pulse sequence of 200ms) is truncated into several overlapping short signal frames according to the set frame length (such as 25ms) and frame shift step length (such as 10ms).

[0103] Perform a Fast Fourier Transform (FFT) on each short-time signal frame to calculate the energy of the Korotkoff pulse at each frequency point and obtain the corresponding energy value.

[0104] The processing results of all Korotkoff sound pulses are sorted according to the time corresponding to the cardiac cycle (e.g., the timestamp of the Korotkoff sound pulse), thereby upgrading the original one-dimensional target Korotkoff sound signal (a time-domain signal) into a two-dimensional time-frequency distribution matrix. For example, each row of this time-frequency distribution matrix represents the energy value corresponding to each frequency point of a Korotkoff sound pulse, and each column represents the energy value of each Korotkoff sound pulse at the same frequency point.

[0105] For any Korotkoff pulse, the spectral energy centroid of the Korotkoff pulse can be calculated based on the time-frequency distribution matrix. The calculation formula is as follows: Where C represents the spectral energy centroid, The first Korotkoff pulse after short-time Fourier transform The actual physical frequency value corresponding to each frequency point The actual physical frequency value The corresponding energy value of the Korotkoff pulse.

[0106] By connecting multiple consecutive spectral energy centroids in a time series manner to generate a spectral energy centroid sequence, the spectral energy centroid change rate characteristics during the diastolic pressure transition phase can be extracted.

[0107] Specifically, let the sequence of spectral energy centroids corresponding to n consecutive Korotkoff tone pulses be: By using a sliding window approach to linearly fit the spectral energy centroid sequence, a local slope can be obtained, for example, a slope of: .

[0108] in, For the first The spectral energy centroid of a Korotkoff tone pulse. For the first Each Korotkoff pulse corresponds to the peak timestamp of the local maxima of the Shannon energy envelope. The peak-to-peak value is the time difference between the peak values ​​of two adjacent Korotkoff tone pulses.

[0109] To facilitate the input of targeted rate of change features into the blood pressure prediction model and avoid irrelevant interference, the rate of change parameter of the spectral energy centroid during the diastolic blood pressure transition phase can be extracted.

[0110] Specifically, a low-frequency collapse threshold is preset. The threshold As a negative scalar, it represents the critical rate at which the spectral centroid shifts to lower frequencies during the diastolic pressure transition phase. This negative value can be understood as representing the physical decay direction of the spectral energy centroid (a frequency value) over time. This threshold... It can be obtained based on historical training sample statistics, and the preferred configuration is 1.5 to 3 times the average rate of change of the centroid of spectral energy per unit time.

[0111] If multiple consecutive local slopes satisfy: Then the spectral energy centroid of the current Korotkoff pulse is shifting not only towards lower frequencies ( (Negative), and the severity of its drop to lower frequencies (absolute slope) has exceeded the system's preset physiological interference threshold. The absolute value of the pressure (the absolute value of the pressure) presents a steep, cliff-like collapse. This cliff-like collapse is the most accurate physical characteristic reference point for capturing diastolic pressure (when blood vessels have fully recovered laminar flow), so at this point, it can be determined that the current spectral energy centroid has entered the diastolic pressure transition phase.

[0112] Furthermore, the critical point with the largest absolute value of slope change is extracted. The cuff pressure value corresponding to the time of the maximum critical point can be used as a reference feature point for diastolic blood pressure and is also used as part of the rate of change parameter. The rate of change parameter specifically includes the local slope corresponding to the critical point and the cuff pressure value that matches the timestamp corresponding to the critical point.

[0113] Figure 6 The graph shows the spectral energy centroid sequence curve (horizontal axis is time, vertical axis is the frequency value corresponding to the spectral energy centroid). ) and the rate of change curve of the spectral energy centroid (horizontal axis is time, vertical axis is rate of change) Based on this rate of change curve, the critical point where the slope abruptly changes when the rate of change is less than the threshold T can be determined, i.e. Figure 6 The point corresponding to the left boundary of the gray area is the critical point of the slope change, which can be mapped as a reference feature point for diastolic blood pressure. For example, based on the timestamp corresponding to the critical point of the slope change, the pressure value sequence can be queried to obtain the absolute cuff pressure value at this time, which is the feature reference point for diastolic blood pressure.

[0114] Throughout the entire cuff deflation process, for the vast majority of the time, the change in the spectral energy centroid between two adjacent heartbeats is minimal, and its slope is very small. It oscillates slightly around 0. Within a specific timeframe (a brief window typically containing only 2 to 3 consecutive heartbeats, lasting approximately 2 to 3 seconds), the slope... A sudden negative step change will occur, falling below the pre-set threshold T. At this point, it can be considered that the region has entered a low-frequency collapse phase, which corresponds to the time axis... The time period.

[0115] Figure 7 A detailed schematic diagram of the spectral energy centroid sequence curve is shown, such as... Figure 7 As shown, it includes the normal stable period, the pre-diastolic blood pressure phase, the voice change transition period, the low-frequency collapse threshold, and the disappearance period, with a slope of The diastolic pressure reference characteristic point is determined at that time, and the segment between it and the subsequent low-frequency collapse critical point ( Figure 7 The gray area in the image represents the low-frequency collapse zone.

[0116] In this embodiment, because the high-frequency energy of Korotkoff sounds disappears abruptly during the diastolic pressure transition period of blood pressure measurement (when vascular compression is about to be completely relieved), causing the centroid trajectory line to drop sharply, this embodiment uses a sliding window to calculate the local geometric slope on the trajectory line. When the slope change is detected to cross the preset low-frequency collapse threshold T, the limiting slope value is extracted. This limiting slope value serves as a key scalar characterizing the diastolic pressure feature and can be directly input into the blood pressure prediction model for the final blood pressure calculation.

[0117] Step S304: Obtain the subject's heart rate parameters.

[0118] Please see details Figure 2 Step S204 of the illustrated embodiment will not be described again here.

[0119] In some alternative implementations, step S304, "acquiring the subject's heart rate parameters," may include steps d1 to d3.

[0120] Step d1: Extract the Shannon energy envelope of the target Korotkoff sound signal.

[0121] Step d2 involves finding peaks on the Shannon energy envelope, determining the difference between the timestamps of two adjacent energy peaks, and generating a peak-to-peak sequence.

[0122] Step d3: Extract heart rate parameters based on the peak-to-peak sequence; heart rate parameters include heart rate feature values ​​and heart rate variability feature values.

[0123] In this embodiment, the heart rate parameters of the subject can be directly determined based on the collected target Korotkoff sound signal, without the need for additional hardware such as heart rate sensors.

[0124] Specifically, the Shannon energy envelope of the target Korotkoff tone signal can be extracted; the Shannon energy envelope can also be extracted from the spectral energy centroid of the Korotkoff tone pulse determined above. It can be understood that this step d1 only needs to be performed once.

[0125] Peaks can be found on the Shannon energy envelope using a dynamic threshold. The timestamps of two adjacent energy peaks can be extracted and the difference calculated. This difference represents the time difference between two adjacent Korotkoff beats. This time difference is subsequently referred to as the peak-to-peak period, and multiple peak-to-peak periods form a peak-to-peak sequence. The dynamic threshold is not an absolute value; it is calculated in real-time based on the average energy of the past few heartbeats and the current background noise level, generating the next judgment threshold accordingly. When encountering extremely loud Korotkoff sounds, the threshold automatically increases to filter out loud noise; when encountering weak Korotkoff sounds, the threshold automatically decreases to sensitively capture weak pulses.

[0126] For example, the dynamic threshold is configured to be updated in real time based on a sliding time window. The threshold update logic is as follows: extract the average Shannon energy peak value of the preceding N (e.g., N=3) confirmed Korotkoff sound pulses. And the average background noise energy within the current sliding time window. ; Mean peak Compared with the mean energy of the background noise The judgment threshold at the current time is calculated by linearly weighting the results according to preset weighting coefficients. The calculation formula can be: Among them, the weighting coefficient The value range is configured to be 0.3 to 0.5, and the weighting coefficient is... The value range is configured to be 1.2 to 1.8.

[0127] Based on the peak-to-peak sequence, heart rate feature values ​​and heart rate variability feature values ​​can be extracted and output, both of which can be used as heart rate parameters required subsequently.

[0128] Specifically, the peak-to-peak interval represents the time difference between two consecutive heartbeats (reflected in Korotkoff sounds as two energy peaks). For example, the interval between the first and second beats is 0.80 seconds, between the second and third beats is 0.82 seconds, between the third and fourth beats is 0.79 seconds, and so on. This can form an array of the form [0.80, 0.82, 0.79...], which is the peak-to-peak interval sequence. Heart rate The corresponding calculation model is: Specifically, the instantaneous heart rate or the average heart rate can be used as the characteristic value of the heart rate.

[0129] Furthermore, time-domain statistical analysis is performed on the fluctuation dispersion of the peak-to-peak sequence, such as calculating the root mean square (RMSSD) of the difference between adjacent peaks or the overall standard deviation (SDNN) of the sequence, thereby extracting and outputting heart rate variability feature values. These heart rate variability feature values ​​are configured as physiological parameters characterizing the subject's autonomic nervous system tension and vascular compliance status.

[0130] Figure 8 The Shannon energy envelope waveform of the target Korotkoff tone signal is shown (the horizontal axis is time, and the vertical axis is signal energy), with the energy peak timestamps determined by dynamic peak finding marked on it. Figure 8 The identified energy peaks are represented by black dots, totaling 13. The time difference between these 13 energy peaks is the peak-to-peak period corresponding to the adjacent peaks, forming a peak-to-peak period sequence containing 12 peak-to-peak periods. By calculating the average of the 12 peak-to-peak periods, the average heart rate within these tens of seconds can be obtained. By calculating the variance or fluctuation degree (dispersion) of the 12 peak-to-peak periods, heart rate variability can be extracted.

[0131] In this embodiment, heart rate and heart rate variability features are derived by extracting peak-to-peak sequences. This expands the feature dimensions without requiring additional modal sensors or analog front-end hardware, reducing hardware costs and avoiding the signal phase deviation and synchronization problems common in multi-sensor systems. Furthermore, the target Korotkoff sound signal is obtained based on variational mode decomposition. After linear superposition and reconstruction of the IMF, phase fidelity can be achieved without destroying the energy envelope shape of the original Korotkoff sound. This ensures the accuracy of subsequent slope and heart rate calculations and avoids the problem of phase delay and waveform distortion (phase shift) caused by frequency cutoff in traditional bandpass filters, which leads to inaccurate energy peak timestamp positioning.

[0132] Step S305: The target Korotkoff sound signal, pressure value sequence rate of change parameter, and heart rate parameter are used as input data for the blood pressure prediction model, and the subject's blood pressure value is determined based on the output of the blood pressure prediction model.

[0133] Please see details Figure 2 Step S205 of the illustrated embodiment will not be described again here.

[0134] In some alternative implementations, the blood pressure prediction model includes: a first input layer, a second input layer, a feature extraction network, a channel attention layer, a feature fusion layer, and an output layer; specifically, it may be a lightweight one-dimensional convolutional neural network (1D-CNN) embedded in a local controller. Figure 9 A schematic diagram of one architecture of this blood pressure prediction model is shown.

[0135] The above step S305, "using the target Korotkoff sound signal, pressure value sequence, rate of change parameter and heart rate parameter as input data for the blood pressure prediction model, and determining the subject's blood pressure value based on the output results of the blood pressure prediction model," may specifically include the following steps e1 to e6.

[0136] Step e1: Construct a first feature vector based on the first input layer, which includes the target Korotkoff sound signal and the pressure value sequence.

[0137] Step e2: Construct a second feature vector based on the second input layer, including the rate of change parameter and the heart rate parameter.

[0138] Step e3: Temporal feature extraction is performed on the first feature vector based on the feature extraction network to obtain a temporal feature map. The feature extraction network includes at least one one-dimensional depthwise separable convolutional layer, which is configured to perform spatial convolution and pointwise convolution operations on the first feature vector sequentially.

[0139] Step e4: Obtain the attention weight coefficients of the temporal feature map based on the channel attention layer, and perform weighted calculation on the temporal feature map according to the attention weight coefficients to output the target feature vector.

[0140] Step e5: Based on the feature fusion layer, the target feature vector and the second feature vector are concatenated and fused to generate a joint feature vector.

[0141] Step e6: Based on the output layer, perform regression calculation on the joint feature vector to predict the subject's blood pressure value.

[0142] In this embodiment, the blood pressure prediction model employs a dual-input approach, comprising a first input layer and a second input layer. The first input layer receives the time sequence of the target Korotkoff sound signal and the synchronized absolute pressure value sequence, forming a first feature vector. The second input layer forms a second feature vector based on a rate of change parameter (e.g., geometric slope) and heart rate parameters (e.g., heart rate eigenvalues ​​and heart rate variability eigenvalues).

[0143] The first input layer receives dynamic time-series data, including a global time-series segment of the target Korotkoff sound signal and a synchronized absolute pressure sequence. The second input layer receives static scalar features, including a limit collapse slope scalar (i.e., the global maximum negative slope extremum extracted from the low-frequency collapse segment), as well as heart rate feature values ​​and heart rate variability feature values. These static scalar features enable the prediction model to obtain global physiological boundary constraints characterizing diastolic blood pressure mutation points and subject cardiovascular compliance when performing one-dimensional convolutional feature extraction.

[0144] The second feature vector receives static scalar features, so its dimension is fixed. For example, the second feature vector is a one-dimensional array of length 3 [slope, heart rate, heart rate variability]. Throughout the entire cuff deflation process, the spectral centroid of most heartbeat pulses (early and mid-stages) is very stable, with a slope approximately 0. The only factor that decisively contributes to predicting diastolic blood pressure is the extreme slope of the cliff-like collapse when crossing the threshold T. Therefore, only the slope feature corresponding to that moment is needed.

[0145] The feature extraction network employs at least one one-dimensional depthwise separable convolutional layer to perform temporal implicit feature extraction on the first feature vector. Specifically, the feature extraction network sequentially performs spatial convolution and pointwise convolution operations on the first feature vector. To adapt to the computational and storage resources of the local controller, the kernel size of the convolutional layer can be configured to 3 or 5, the number of channels can be configured to 16 to 64, and a non-linear activation function is used to map the computation results to obtain a feature map containing audio and pressure temporal implicit features, i.e., a temporal feature map.

[0146] The channel attention layer is used to perform global average pooling on the extracted temporal feature map, compressing the temporal feature map into a one-dimensional vector. The one-dimensional vector is input into a multilayer perceptron network containing two fully connected layers to generate attention weight coefficients corresponding to each channel. The attention weight coefficients are multiplied and weighted by channel with the temporal feature map to output the target feature vector.

[0147] The feature fusion layer is used to concatenate and fuse the target feature vector with the second feature vector to generate a joint feature vector. Specifically, the feature fusion layer can concatenate and fuse the target feature vector with the second feature vector along the feature dimension to generate a high-dimensional joint feature vector, thereby achieving a dimensionality reduction combination of local temporal features and global topological / physiological features.

[0148] The output layer contains a fully connected network that receives the weighted target feature vector and performs regression prediction based on the fully connected layer to predict the subject's blood pressure value. Specifically, the predicted systolic blood pressure and diastolic blood pressure values ​​can be output in parallel by the dual-end output nodes.

[0149] Traditional prediction models typically concatenate all input data into a long vector at the input end. This causes the one-dimensional convolution kernel to be unable to extract effective temporal local correlations on static scalars (such as heart rate and centroid collapse slope) during sliding calculations, resulting in dimensional redundancy and noise interference in mathematical matrix operations.

[0150] In this embodiment, a dual-path independent topology is adopted at the input end. The first path (dynamic temporal path) is dedicated to inputting a one-dimensional high-frequency audio temporal sequence and synchronization pressure sequence, which is then fed into a convolutional network to extract microscopic spatiotemporal features. The second path (static feature path) completely strips away macroscopic physiological / topological scalars (heart rate HR, heart rate variability HRV, geometric slope, etc.), bypassing the convolutional layers until the feature fusion layer performs late-stage fusion. This model architecture retains the temporal local feature extraction capability of one-dimensional convolution while avoiding the gradient vanishing problem caused by heterogeneous data aliasing.

[0151] Furthermore, standard one-dimensional convolutional neural networks have a large number of parameters and high computational cost. Because the convolutional kernels require fully connected implicit merging across the entire channel, their computational complexity increases quadratically, making them unsuitable for integration into local controllers with extremely limited physical resources (such as SRAM / Flash). This embodiment completely abandons standard convolution and adopts one-dimensional depthwise separable convolution, splitting the computation into two concatenated steps: single-channel convolution followed by individual convolutions and 1×1 point-to-channel combination. While maintaining the same feature extraction depth, this reduces the model's parameter count and floating-point computation to the level of 1 / N of standard convolution (where N is the number of channels).

[0152] Step S306: Obtain the subject's blood pressure reference value.

[0153] Step S307: With the weight parameters of the blood pressure prediction model locked except for the output layer, the weight parameters of the output layer are fine-tuned according to the blood pressure reference value to update the output layer of the blood pressure prediction model.

[0154] In this embodiment, the blood pressure prediction model can also be adaptively calibrated.

[0155] When the blood pressure value predicted by the model consistently deviates from the clinical gold standard, for example, due to the subject's specific physical condition (such as extreme obesity or severe vascular calcification) leading to inaccurate blood pressure prediction, other measurement methods such as a mercury sphygmomanometer can be used to measure the subject's accurate blood pressure value, i.e., the blood pressure reference value. This blood pressure reference value can specifically include the systolic blood pressure reference value. Compared with diastolic blood pressure reference value The data is then input into the device for subsequent adaptive adjustments to correct the blood pressure measurement model.

[0156] The controller is preloaded with global base model weights based on pre-trained samples and configured with a local unfreezing strategy. This local unfreezing strategy is to only unfreeze and update the weight parameters of the fully connected layers in the output layer, while locking the weight parameters of other parts, such as the feature extraction network and the channel attention layer.

[0157] After receiving the blood pressure reference value, the global network weights corresponding to parts such as the feature extraction network and channel attention layer in the blood pressure prediction model are frozen based on the local unfreezing strategy. The residual between the predicted blood pressure value and the blood pressure reference value is calculated, and a lightweight gradient descent algorithm is used to perform local fine-tuning operations only on the bias parameters and mapping weights in the output layer. Through the edge-side local update mechanism, personalized targeted calibration of the blood pressure prediction model for the individual physiological characteristics of the subject is achieved while avoiding controller computing power and memory overflow.

[0158] The fully connected layer network is configured with a static initial weight matrix. Its input dimension corresponds to the dimension of the joint feature vector output by the feature fusion layer, and its output dimension is configured to be two-dimensional, mapping the systolic blood pressure prediction node and the diastolic blood pressure prediction node respectively.

[0159] Specifically, the above fine-tuning process can employ a mean squared error loss function that includes an L2 regularization term. Specifically, it is possible to: ; in, , The blood pressure values ​​predicted in step S305 are the predicted systolic blood pressure and the predicted diastolic blood pressure, respectively. , These are the reference values ​​for systolic blood pressure and diastolic blood pressure, respectively. The weight parameters of the fully connected layer network are for output layer. This is the regularization coefficient.

[0160] The controller uses the gradient descent algorithm to adjust the weight parameters of the fully connected layer network. Iterative calculations are performed. Once the loss function value converges to a preset condition, the controller calls the Advanced Encryption Standard (AES) algorithm to update the weight parameters. Encrypt it and overwrite it into non-volatile memory.

[0161] In this embodiment, if the loss function only includes the first half of the mean squared error term, the gradient descent algorithm will drive the weight parameters to force the fitting error of local samples to be reduced to 0. The numerical inflation of weights can lead to severe overfitting. While the model may perfectly fit the current few calibrated samples, the network mapping loses its smoothness due to the excessively large absolute values ​​of the weights. In subsequent actual measurements, even the slightest fluctuations (such as noise) in the input feature signal can be amplified by the massive weight multiplication, resulting in drastic distortions or large jumps in the predicted blood pressure values.

[0162] By introducing into the loss function This allows for the addition of a decay constraint to the original error optimization objective. During backpropagation gradient calculation, this term continuously penalizes large numerical weights, forcing the algorithm to make trade-offs during optimization: finding a set of weight parameters with the smallest possible values ​​and a relatively uniform distribution. This approach aims to achieve the optimal blood pressure prediction error. It helps to curb the surge in weights caused by fine-tuning with small samples, ensuring that the fine-tuned blood pressure prediction model remains robust against environmental noise and maintains the generalization ability acquired during the pre-training phase.

[0163] The blood pressure measurement method provided in this embodiment employs a VMD algorithm based on specific frequency band selection. This algorithm can separate the transient intrinsic components of Korotkoff sounds from broadband motion artifacts and aerodynamic noise in the extremely aliased original digital audio signal, ensuring high time-frequency fidelity of the blood flow acoustic signal. Building upon this, spectral energy centroid topology tracking is introduced. During the diastolic blood pressure assessment stage, by dynamically identifying the slope critical point where the high-frequency Korotkoff sound energy experiences a cliff-like collapse, the interference of residual low-frequency muscle noise on feature optimization can be reduced, improving the accuracy of diastolic blood pressure feature extraction. During model updates, the weights of the basic network are locked, and gradient optimization is performed only on the terminal fully connected layers. This mechanism effectively controls the computational load of the fine-tuning process, enabling the device to complete model adaptation to different individual vascular compliance differences without cloud computing power. Furthermore, local encryption methods enhance data security.

[0164] This embodiment also provides a blood pressure measurement system, which includes an audio acquisition module, a pressure acquisition module, and a controller.

[0165] The audio acquisition module is used to acquire the acoustic signal of the subject's blood pressure, extract the corresponding digital audio signal, and send the digital audio signal to the controller; The pressure acquisition module is used to synchronously acquire the pressure values ​​of the subjects, form a pressure value sequence, and send the pressure value sequence to the controller; The controller is used to execute the blood pressure measurement method provided in the above embodiments.

[0166] For details on the working principle of this blood pressure measurement system, please refer to [link / reference needed]. Figure 1 The relevant descriptions shown will not be repeated here.

[0167] This embodiment also provides a blood pressure measuring device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0168] This embodiment provides a blood pressure measuring device, such as... Figure 10 As shown, the device includes: Data acquisition module 1001 is used to acquire digital audio signals and pressure value sequences that are synchronously collected during the blood pressure measurement of the subject; The reconstruction module 1002 is used to decompose the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition, extract the intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz, and linearly superimpose the extracted intrinsic mode functions to determine the target Korotkoff tone signal. Processing module 1003 is used to determine the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal; Heart rate acquisition module 1004 is used to acquire the heart rate parameters of the subject; The prediction module 1005 is used to take the target Korotkoff sound signal, the pressure value sequence, the rate of change parameter and the heart rate parameter as input data of the blood pressure prediction model, and determine the subject's blood pressure value based on the output result of the blood pressure prediction model.

[0169] In some optional embodiments, the processing module 1003 is further configured to: The digital audio signal is subjected to a Fast Fourier Transform to obtain the spectral energy distribution characteristics of the digital audio signal; the spectral energy distribution characteristics include the number of main energy peaks and the bandwidth of the main peaks; The decomposition hyperparameters of variational mode decomposition are determined based on the number of energy peaks and the bandwidth of the peaks; the decomposition hyperparameters include the number of intrinsic mode functions and the quadratic penalty factor. The step of decomposing the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition includes: The digital audio signal is subjected to variational mode decomposition according to the decomposition hyperparameters, and the digital audio signal is decomposed into multiple intrinsic mode functions.

[0170] In some optional implementations, the step of extracting intrinsic mode functions with center frequencies within the target frequency band of 15Hz-80Hz, and linearly superimposing the extracted intrinsic mode functions to determine the target Korotkoff tone signal includes: Linear superposition of the intrinsic mode functions of the target whose center frequency falls within the target frequency band of 15Hz-80Hz is used to reconstruct the undetermined Korotkoff tone signal. Extract the total acoustic energy and envelope periodicity of the Korotkoff tone signal to be determined; If the total acoustic energy is greater than a preset energy threshold and the heart rate corresponding to the envelope periodic feature is within a preset heart rate range, the undetermined Korotkoff sound signal is taken as the target Korotkoff sound signal.

[0171] In some optional implementations, determining the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal includes: Extract the Korotkoff sound pulses corresponding to each cardiac cycle from the target Korotkoff sound signal; For any Korotkoff pulse, determine the energy value of the Korotkoff pulse at each frequency point; The spectral energy centroid of the Korotkoff pulse is determined based on the energy values ​​corresponding to each frequency point of the Korotkoff pulse. Based on the spectral energy centroid sequence generated from the spectral energy centroids of each Korotkoff pulse, a sliding window method is used to linearly fit the spectral energy centroid sequence to obtain the local slope. When multiple consecutive local slopes are less than a preset low-frequency collapse threshold, the critical point with the largest absolute value of slope change is extracted. The local slope corresponding to the critical point and the cuff pressure value matching the timestamp corresponding to the critical point are used as the rate of change parameter.

[0172] In some optional implementations, obtaining the subject's heart rate parameters includes: Extract the Shannon energy envelope of the target Korotkoff sound signal; Peak finding is performed on the Shannon energy envelope to determine the difference between the timestamps of two adjacent energy peaks, generating a peak-to-peak sequence; Heart rate parameters are extracted based on the peak-to-peak sequence; the heart rate parameters include heart rate feature values ​​and heart rate variability feature values.

[0173] In some optional implementations, the blood pressure prediction model includes: a first input layer, a second input layer, a feature extraction network, a channel attention layer, a feature fusion layer, and an output layer; The step of using the target Korotkoff sound signal, the pressure value sequence, the rate of change parameter, and the heart rate parameter as input data for the blood pressure prediction model, and determining the subject's blood pressure value based on the output of the blood pressure prediction model, includes: A first feature vector is constructed based on the first input layer, comprising the target Korotkoff sound signal and the pressure value sequence; A second feature vector including the rate of change parameter and the heart rate parameter is constructed based on the second input layer; Temporal feature extraction is performed on the first feature vector based on the feature extraction network to obtain a temporal feature map; wherein, the feature extraction network includes at least one one-dimensional depthwise separable convolutional layer, and the one-dimensional depthwise separable convolutional layer is configured to perform spatial convolution and pointwise convolution operations on the first feature vector in sequence; The attention weight coefficients of the temporal feature map are obtained based on the channel attention layer, and the temporal feature map is weighted according to the attention weight coefficients to output the target feature vector. Based on the feature fusion layer, the target feature vector and the second feature vector are concatenated and fused to generate a joint feature vector; The joint feature vector is regressed based on the output layer to predict the subject's blood pressure value.

[0174] In some optional implementations, the data acquisition module 1001 is further configured to: acquire the blood pressure reference value of the subject; The processing module 1003 is further configured to: fine-tune the weight parameters of the output layer according to the blood pressure reference value, while locking other weight parameters in the blood pressure prediction model except for the output layer, so as to update the output layer of the blood pressure prediction model.

[0175] In some alternative implementations, the loss function when fine-tuning the weight parameters of the output layer for: ; in, , For example, the predicted values ​​are systolic blood pressure and diastolic blood pressure, respectively. , These are the reference values ​​for systolic blood pressure and diastolic blood pressure, respectively. The weight parameters of the output layer, This is the regularization coefficient.

[0176] The blood pressure measuring device provided in this disclosure can execute the blood pressure measuring method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0177] Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0178] The following is a detailed reference. Figure 11 The diagram illustrates a structural schematic suitable for implementing the electronic device described in the embodiments of this application. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 1101, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1102 or a program loaded from memory 1108 into random access memory (RAM) 1103. The RAM 1103 also stores various programs and data required for the operation of the electronic device. The processor 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0179] Typically, the following devices can be connected to I / O interface 1105: input devices 1106 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 1107 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 1108 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1109. Communication device 1109 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 11 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0180] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 1109, or installed from memory 1108, or installed from ROM 1102. When the computer program is executed by processor 1101, it performs the functions defined in the blood pressure measurement method of embodiments of this application.

[0181] Figure 11The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0182] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the blood pressure measurement method shown in the above embodiments is implemented.

[0183] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0184] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for measuring blood pressure, characterized in that, The method includes: Acquire digital audio signals and pressure value sequences that are simultaneously collected during blood pressure measurement of subjects; The digital audio signal is decomposed into multiple intrinsic mode functions based on variational mode decomposition. The intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz are extracted. The extracted intrinsic mode functions are linearly superimposed to determine the target Korotkoff tone signal. The rate of change parameter of the spectral energy centroid is determined based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal. Obtain the heart rate parameters of the subject; The target Korotkoff sound signal, the pressure value sequence, the rate of change parameter, and the heart rate parameter are used as input data for the blood pressure prediction model, and the blood pressure value of the subject is determined based on the output of the blood pressure prediction model. The step of determining the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal includes: Extract the Korotkoff sound pulses corresponding to each cardiac cycle from the target Korotkoff sound signal; For any Korotkoff pulse, determine the energy value of the Korotkoff pulse at each frequency point; The spectral energy centroid of the Korotkoff pulse is determined based on the energy values ​​corresponding to each frequency point of the Korotkoff pulse. Based on the spectral energy centroid sequence generated from the spectral energy centroids of each Korotkoff pulse, a sliding window method is used to linearly fit the spectral energy centroid sequence to obtain the local slope. When multiple consecutive local slopes are less than a preset low-frequency collapse threshold, the critical point with the largest absolute value of slope change is extracted. The local slope corresponding to the critical point and the cuff pressure value matching the timestamp corresponding to the critical point are used as the rate of change parameter. The acquisition of the subject's heart rate parameters includes: Extract the Shannon energy envelope of the target Korotkoff sound signal; Peak finding is performed on the Shannon energy envelope to determine the difference between the timestamps of two adjacent energy peaks, generating a peak-to-peak sequence; Heart rate parameters are extracted based on the peak-to-peak sequence; the heart rate parameters include heart rate feature values ​​and heart rate variability feature values. The blood pressure prediction model includes: a first input layer, a second input layer, a feature extraction network, a channel attention layer, a feature fusion layer, and an output layer; The step of using the target Korotkoff sound signal, the pressure value sequence, the rate of change parameter, and the heart rate parameter as input data for the blood pressure prediction model, and determining the subject's blood pressure value based on the output of the blood pressure prediction model, includes: A first feature vector is constructed based on the first input layer, comprising the target Korotkoff sound signal and the pressure value sequence; A second feature vector including the rate of change parameter and the heart rate parameter is constructed based on the second input layer; Temporal feature extraction is performed on the first feature vector based on the feature extraction network to obtain a temporal feature map; wherein, the feature extraction network includes at least one one-dimensional depthwise separable convolutional layer, and the one-dimensional depthwise separable convolutional layer is configured to perform spatial convolution and pointwise convolution operations on the first feature vector in sequence; The attention weight coefficients of the temporal feature map are obtained based on the channel attention layer, and the temporal feature map is weighted according to the attention weight coefficients to output the target feature vector. Based on the feature fusion layer, the target feature vector and the second feature vector are concatenated and fused to generate a joint feature vector; The joint feature vector is regressed based on the output layer to predict the subject's blood pressure value.

2. The method according to claim 1, characterized in that, The method further includes: The digital audio signal is subjected to a Fast Fourier Transform to obtain the spectral energy distribution characteristics of the digital audio signal; the spectral energy distribution characteristics include the number of main energy peaks and the bandwidth of the main peaks; The decomposition hyperparameters of variational mode decomposition are determined based on the number of energy peaks and the bandwidth of the peaks; the decomposition hyperparameters include the number of intrinsic mode functions and the quadratic penalty factor. The step of decomposing the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition includes: The digital audio signal is subjected to variational mode decomposition according to the decomposition hyperparameters, and the digital audio signal is decomposed into multiple intrinsic mode functions.

3. The method according to claim 1, characterized in that, The process of extracting intrinsic mode functions (IMFs) with center frequencies within the target frequency band of 15Hz-80Hz, and then linearly superimposing the extracted IMFs to determine the target Korotkoff tone signal includes: Linear superposition of the intrinsic mode functions of the target whose center frequency falls within the target frequency band of 15Hz-80Hz is used to reconstruct the undetermined Korotkoff tone signal. Extract the total acoustic energy and envelope periodicity of the Korotkoff tone signal to be determined; If the total acoustic energy is greater than a preset energy threshold and the heart rate corresponding to the envelope periodic feature is within a preset heart rate range, the undetermined Korotkoff sound signal is taken as the target Korotkoff sound signal.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the reference blood pressure value of the subject; With all weight parameters in the blood pressure prediction model except for the output layer locked, the weight parameters of the output layer are fine-tuned according to the blood pressure reference value to update the output layer of the blood pressure prediction model.

5. The method according to claim 4, characterized in that, Loss function when fine-tuning the weight parameters of the output layer for: ； in, , These are the predicted systolic blood pressure and the predicted diastolic blood pressure, respectively. , These are the reference values ​​for systolic blood pressure and diastolic blood pressure, respectively. The weight parameters of the output layer, is the regularization coefficient.

6. A blood pressure measuring device, characterized in that, The device includes: The data acquisition module is used to acquire digital audio signals and pressure value sequences that are simultaneously collected during the blood pressure measurement of the subject. The reconstruction module is used to decompose the digital audio signal into multiple intrinsic mode functions based on variational mode decomposition, extract the intrinsic mode functions with center frequencies in the target frequency band of 15Hz-80Hz, and linearly superimpose the extracted intrinsic mode functions to determine the target Korotkoff tone signal. The processing module is used to determine the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal. Heart rate acquisition module, used to acquire the heart rate parameters of the subject; The prediction module is used to take the target Korotkoff sound signal, the pressure value sequence, the rate of change parameter and the heart rate parameter as input data of the blood pressure prediction model, and determine the blood pressure value of the subject based on the output result of the blood pressure prediction model; The step of determining the rate of change parameter of the spectral energy centroid based on the spectral energy centroid of each Korotkoff pulse in the target Korotkoff signal includes: Extract the Korotkoff sound pulses corresponding to each cardiac cycle from the target Korotkoff sound signal; For any Korotkoff pulse, determine the energy value of the Korotkoff pulse at each frequency point; The spectral energy centroid of the Korotkoff pulse is determined based on the energy values ​​corresponding to each frequency point of the Korotkoff pulse. Based on the spectral energy centroid sequence generated from the spectral energy centroids of each Korotkoff pulse, a sliding window method is used to linearly fit the spectral energy centroid sequence to obtain the local slope. When multiple consecutive local slopes are less than a preset low-frequency collapse threshold, the critical point with the largest absolute value of slope change is extracted. The local slope corresponding to the critical point and the cuff pressure value matching the timestamp corresponding to the critical point are used as the rate of change parameter. The acquisition of the subject's heart rate parameters includes: Extract the Shannon energy envelope of the target Korotkoff sound signal; Peak finding is performed on the Shannon energy envelope to determine the difference between the timestamps of two adjacent energy peaks, generating a peak-to-peak sequence; Heart rate parameters are extracted based on the peak-to-peak sequence; the heart rate parameters include heart rate feature values ​​and heart rate variability feature values. The blood pressure prediction model includes: a first input layer, a second input layer, a feature extraction network, a channel attention layer, a feature fusion layer, and an output layer; The step of using the target Korotkoff sound signal, the pressure value sequence, the rate of change parameter, and the heart rate parameter as input data for the blood pressure prediction model, and determining the subject's blood pressure value based on the output of the blood pressure prediction model, includes: A first feature vector is constructed based on the first input layer, comprising the target Korotkoff sound signal and the pressure value sequence; A second feature vector including the rate of change parameter and the heart rate parameter is constructed based on the second input layer; Temporal feature extraction is performed on the first feature vector based on the feature extraction network to obtain a temporal feature map; wherein, the feature extraction network includes at least one one-dimensional depthwise separable convolutional layer, and the one-dimensional depthwise separable convolutional layer is configured to perform spatial convolution and pointwise convolution operations on the first feature vector in sequence; The attention weight coefficients of the temporal feature map are obtained based on the channel attention layer, and the temporal feature map is weighted according to the attention weight coefficients to output the target feature vector. Based on the feature fusion layer, the target feature vector and the second feature vector are concatenated and fused to generate a joint feature vector; The joint feature vector is regressed based on the output layer to predict the subject's blood pressure value.

7. A blood pressure measurement system, characterized in that, The system includes: an audio acquisition module, a pressure acquisition module, and a controller; The audio acquisition module is used to acquire the acoustic signal of the subject's blood pressure, extract the corresponding digital audio signal, and send the digital audio signal to the controller; The pressure acquisition module is used to synchronously acquire the pressure values ​​of the subject, form a pressure value sequence, and send the pressure value sequence to the controller; The controller is used to perform the blood pressure measurement method according to any one of claims 1 to 5.

Images