Non-contact blood spectral imaging blood pressure monitoring system and method

Abstract

The application discloses a non-contact blood spectrum imaging blood pressure monitoring system and method, relates to the technical field of blood pressure monitoring, captures the reflection spectrum fluctuation of a key detection area of a face, carries out signal processing on the reflection spectrum fluctuation, extracts blood flow information, extracts hemoglobin optical response characteristics and blood flow dynamics conduction characteristics, constructs a multi-modal feature fusion network, forms a comprehensive feature representation, establishes a blood pressure mapping model based on the comprehensive feature representation, wherein the blood pressure mapping model comprises a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model, a basic blood pressure value is calculated through the static blood pressure estimation sub-model, a blood pressure fluctuation value is calculated through the dynamic blood pressure change prediction sub-model, the basic blood pressure value is combined with the blood pressure fluctuation value, and a final blood pressure monitoring result is generated; and non-contact, continuous and accurate blood pressure monitoring is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood pressure monitoring, in particular to a non-contact blood spectrum imaging blood pressure monitoring system and method. BACKGROUND

[0002] Blood pressure monitoring is an important indicator for assessing cardiovascular health. Traditional blood pressure monitoring methods mainly rely on cuff blood pressure monitors, which require applying pressure to the limbs to measure blood pressure values. This contact measurement method has several limitations: first, the traditional cuff blood pressure monitor causes discomfort to the patient during the measurement process, especially for patients who need long-term monitoring. Repeated inflation and pressure can cause skin compression and pain. Second, the traditional method can only obtain intermittent blood pressure data and cannot achieve continuous real-time monitoring, making it difficult to capture the instantaneous changes and diurnal fluctuations of blood pressure. Third, for special groups such as the elderly, infants, and skin-sensitive individuals, the applicability of traditional measurement methods is poor. Finally, in certain clinical scenarios such as surgery and intensive care, the traditional measurement method may interfere with medical operations or cannot be implemented.

[0003] In recent years, although some non-invasive blood pressure monitoring technologies have emerged, such as pulse wave transit time (PTT) and photoplethysmography (PPG), these methods still require direct contact between the sensor and the human body, and the measurement accuracy and stability are affected by various factors such as sensor position, contact pressure, and environmental temperature. In addition, existing technologies often only focus on a single physiological signal, ignoring the complex multi-factor mechanism of blood pressure formation, resulting in insufficient generalization ability and robustness of the prediction model.

[0004] Therefore, it is of great clinical value and broad application prospect to develop a non-contact blood pressure monitoring system that does not require direct contact with the human body, can achieve continuous monitoring, and has high accuracy and stability.

[0005] To this end, the present application proposes a non-contact blood spectrum imaging blood pressure monitoring system and method. SUMMARY

[0006] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a non-contact blood spectrum imaging blood pressure monitoring system and method, which realizes non-contact, continuous, and accurate blood pressure monitoring.

[0007] To achieve the above-mentioned purpose, a non-contact blood spectrum imaging blood pressure monitoring method is proposed, comprising the following steps:

[0008] Step one, capturing the reflection spectrum fluctuation of the key detection area of the human face through the camera;

[0009] Step two, signal processing of the reflection spectrum fluctuation to extract blood flow information;

[0010] Step three, based on the blood flow information, extracting hemoglobin optical response features and blood flow hemodynamic conduction features;

[0011] Step four, based on the hemoglobin optical response features and the blood flow hemodynamic conduction features, constructing a multi-modal feature fusion network to form a comprehensive feature representation;

[0012] Step five, based on the comprehensive feature representation, establishing a blood pressure mapping model, wherein the blood pressure mapping model includes a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model;

[0013] Step six, calculating the basic blood pressure value through the static blood pressure estimation sub-model, and calculating the blood pressure fluctuation value through the dynamic blood pressure change prediction sub-model;

[0014] Step seven, combining the basic blood pressure value and the blood pressure fluctuation value to generate the final blood pressure monitoring result.

[0015] The reflection spectrum fluctuation of the key detection area of the human face captured by the camera includes the following steps:

[0016] Step 11: arranging a camera to align the key detection area of the human face, the key detection area including the forehead, cheek and nose wing;

[0017] Step 12: when detecting that the human face enters the collection area, triggering the camera to continuously collect a human face image sequence at a preset frame rate;

[0018] Step 13: locating the key area of the human face image sequence, and determining the detection area to be detected by a facial feature point detection algorithm;

[0019] Step 14: using multispectral imaging technology to capture reflection spectrum information under different wavelengths in the detection area, and recording the fluctuation data of the reflection spectrum information as spectrum fluctuation data;

[0020] The signal processing of the reflection spectrum fluctuation to extract blood flow information includes the following steps:

[0021] Step 21: applying a blind source separation algorithm to separate a periodic signal related to heartbeats from the reflection spectrum fluctuation data;

[0022] Step 22: converting the periodic signal into time domain and frequency domain features through a time-frequency analysis method to form blood flow information data;

[0023] Further, the extraction method of the hemoglobin optical response features includes the following steps:

[0024] Based on the spectral absorption characteristics at different wavelengths, the hemoglobin concentration changes in the blood flow information are calculated using the Beer-Lambert law to extract the hemoglobin optical response characteristics; the hemoglobin optical response characteristics include the proportional changes of oxyhemoglobin and deoxyhemoglobin, the amplitude of hemoglobin concentration fluctuation and the phase difference;

[0025] The extraction method of the hemodynamic conduction characteristics is:

[0026] By analyzing the blood flow pulsatile waveform in the blood flow information, the hemodynamic conduction characteristics are extracted using a fluid mechanics model, including pulse wave propagation speed, pulse waveform features, vascular elasticity index and blood flow resistance estimate;

[0027] Based on the hemoglobin optical response characteristics and the hemodynamic conduction characteristics, a multi-modal feature fusion network is constructed to form a comprehensive feature representation, including the following steps:

[0028] Step 41: Design an optical feature processing subnetwork to process the hemoglobin optical response characteristics using a convolutional neural network structure to extract the representation of blood component changes;

[0029] Step 42: Design a hemodynamic feature processing subnetwork to process the hemodynamic conduction characteristics using a long short-term memory network structure to extract blood vessel state and blood flow characteristic representation;

[0030] Step 43: Construct a multi-modal feature fusion subnetwork using an attention mechanism and a cross-modal transformer structure to adaptively fuse the output features of the optical feature processing subnetwork and the hemodynamic feature processing subnetwork to generate the comprehensive feature representation;

[0031] Based on the comprehensive feature representation, a blood pressure mapping model is established, including the following steps:

[0032] Step 51: Construct a blood pressure sample dataset containing the comprehensive feature representation of the subject under different physiological states and the corresponding standard blood pressure measurement value, which is used as historical blood pressure data;

[0033] Step 52: Based on the blood pressure sample dataset, a support vector regression algorithm is used to design the static blood pressure estimation submodel for obtaining the basic blood pressure value; the static blood pressure estimation submodel uses a deep neural network structure to map the comprehensive feature representation to the basic values of systolic and diastolic blood pressure;

[0034] The static blood pressure estimation sub-model adopts a multi-layer perception structure and includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is the same as the dimension of the comprehensive feature representation, and is configured to receive the comprehensive feature representation. The hidden layer adopts a full connection structure, each hidden layer is equipped with a normalization layer and a ReLU activation function, and is configured to extract deep features in the comprehensive feature representation. The output layer includes two neurons, which correspond to the prediction results of the systolic pressure base value and the diastolic pressure base value respectively.

[0035] In the training process of the static blood pressure estimation sub-model, the training set in the blood pressure sample data set is used for model parameter learning. The training target is to minimize the mean square error between the predicted blood pressure value and the standard blood pressure value. The optimization algorithm adopts an Adam optimizer.

[0036] Step 53: Based on the time series data in the blood pressure sample data set, a recurrent neural network structure is designed for the dynamic blood pressure change prediction sub-model, which is used to predict the fluctuation trend of blood pressure in a short period. The time series data is a sequence of comprehensive feature representations sorted in time order;

[0037] The dynamic blood pressure change prediction sub-model adopts a bidirectional long short-term memory network structure and includes an input layer, a recurrent layer, and an output layer. The input layer receives the time series of the comprehensive feature representation in a preset time window; the recurrent layer is composed of bidirectional LSTM units; and the output layer maps the hidden state of the LSTM unit to the blood pressure change trend prediction value through a full connection network.

[0038] In the training process of the dynamic blood pressure change prediction sub-model, the input is the time series of the comprehensive feature representation in a continuous time window, and the output target is the blood pressure change sequence in the corresponding time window. The loss function of the dynamic blood pressure change prediction sub-model adopts a mean square error loss function between the predicted blood pressure change and the real blood pressure change, and simultaneously weights and fuses the point prediction error and the trend prediction error. The point prediction error measures the difference between the predicted value and the real value at each time point, and the trend prediction error measures the consistency of the change trend between the predicted sequence and the real sequence.

[0039] Step 54: The dynamic blood pressure change prediction sub-model combines the time series of the comprehensive feature representation and the historical blood pressure data to predict the change trend and amplitude of the blood pressure through a gated recurrent unit network.

[0040] The calculation of the base blood pressure value by the static blood pressure estimation sub-model and the calculation of the blood pressure fluctuation value by the dynamic blood pressure change prediction sub-model include the following steps:

[0041] Step 61: Input the latest collected comprehensive feature representation into the static blood pressure estimation sub-model to calculate the base systolic pressure and diastolic pressure values at the current time through forward propagation;

[0042] Step 62: Based on the historical blood pressure data and the current comprehensive feature representation, a blood pressure fluctuation trend in the short term is calculated through the recursive structure of the dynamic blood pressure change prediction sub-model;

[0043] Step 63: According to the blood pressure fluctuation trend, the blood pressure fluctuation value, including the systolic blood pressure fluctuation value and the diastolic blood pressure fluctuation value, is calculated using a time series analysis method;

[0044] The way to generate the final blood pressure monitoring result is:

[0045] The base blood pressure value and the blood pressure fluctuation value are combined through an adaptive weighting algorithm to generate the final systolic and diastolic blood pressure as the final blood pressure monitoring result.

[0046] A non-contact blood spectrum imaging blood pressure monitoring system is proposed, which includes an image acquisition module, a signal processing module, a feature extraction module, a feature fusion module, a blood pressure mapping module, and a result output module; wherein each module is connected through electrical means;

[0047] The image acquisition module captures the reflected spectral fluctuation of the key detection area of the human face through the camera and sends the reflected spectral fluctuation data to the signal processing module;

[0048] The signal processing module receives the reflected spectral fluctuation data, processes the reflected spectral fluctuation data, extracts blood flow information, and sends the blood flow information to the feature extraction module;

[0049] The feature extraction module receives the blood flow information, extracts hemoglobin optical response features and blood flow hemodynamics conduction features based on the blood flow information, and sends the hemoglobin optical response features and the blood flow hemodynamics conduction features to the feature fusion module;

[0050] The feature fusion module receives the hemoglobin optical response features and the blood flow hemodynamics conduction features, constructs a multi-modal feature fusion network, forms a comprehensive feature representation, and sends the comprehensive feature representation to the blood pressure mapping module;

[0051] The blood pressure mapping module receives the comprehensive feature representation, establishes a blood pressure mapping model based on the comprehensive feature representation, wherein the blood pressure mapping model includes a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model; the blood pressure mapping module calculates the base blood pressure value through the static blood pressure estimation sub-model and calculates the blood pressure fluctuation value through the dynamic blood pressure change prediction sub-model, and sends the base blood pressure value and the blood pressure fluctuation value to the result output module;

[0052] A result output module receives the basic blood pressure value and the blood pressure fluctuation value, combines the basic blood pressure value and the blood pressure fluctuation value, and generates a final blood pressure monitoring result.

[0053] Compared with the prior art, the present application has the following advantages:

[0054] The present application first captures the reflection spectrum fluctuation of the key area of the human face through the camera, extracts the clear blood flow pulsation signal through a specific signal processing algorithm, further analyzes the optical absorption characteristics of hemoglobin at different wavelengths based on the extracted blood flow signal, obtains the optical response characteristics of hemoglobin, at the same time, analyzes the propagation characteristics of the blood flow pulsation signal, extracts the hemodynamic conduction characteristics, then inputs these multi-modal features into a carefully designed feature fusion network, automatically learns the complex relationship between the features through a deep learning algorithm, forms a comprehensive feature representation, based on this comprehensive feature representation, a dual blood pressure mapping model is constructed, the static blood pressure estimation sub-model is responsible for calculating the basic blood pressure value related to the individual basic physiological state, and the dynamic blood pressure change prediction sub-model focuses on capturing the short-term blood pressure fluctuation caused by factors such as emotional changes and body position changes, finally, the basic blood pressure value and the blood pressure fluctuation value are combined to generate an accurate blood pressure monitoring result; solve the technical problems of traditional blood pressure monitoring methods, such as discomfort, intermittent measurement, poor applicability for special groups, etc., realize non-contact, continuous and accurate blood pressure monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A flowchart of a non-contact blood spectrum imaging blood pressure monitoring method in embodiment 1 of the present application;

[0056] Figure 2 A model structure diagram of a multi-modal feature fusion network in embodiment 1 of the present application;

[0057] Figure 3 A module connection relationship diagram of a non-contact blood spectrum imaging blood pressure monitoring system in embodiment 2 of the present application. DETAILED DESCRIPTION

[0058] The technical solutions of the present application will be described below in conjunction with the embodiments, obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0059] Embodiment 1

[0060] As shown in Figure 1 A non-contact blood spectrum imaging blood pressure monitoring method, comprising the following steps:

[0061] Step one, capture the reflection spectrum fluctuation of the key detection area of the human face through the camera;

[0062] Step two, signal processing of the reflection spectrum fluctuation, extracting blood flow information;

[0063] Step three, based on the blood flow information, extracting hemoglobin optical response characteristics and blood flow dynamics conduction characteristics;

[0064] Step four, based on the hemoglobin optical response characteristics and the blood flow dynamics conduction characteristics, constructing a multi-modal feature fusion network to form a comprehensive feature representation;

[0065] Step five, based on the comprehensive feature representation, establishing a blood pressure mapping model, wherein the blood pressure mapping model includes a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model;

[0066] Step six, calculating the basic blood pressure value through the static blood pressure estimation sub-model, and calculating the blood pressure fluctuation value through the dynamic blood pressure change prediction sub-model;

[0067] Step seven, combining the basic blood pressure value with the blood pressure fluctuation value to generate the final blood pressure monitoring result.

[0068] In the embodiments of the present application, the reflection spectrum fluctuation of the key detection area of the human face captured by the camera includes the following steps:

[0069] Step 11: arranging a high-definition camera to align the key detection area of the human face, including the forehead, cheek, and nasal ala, etc. rich in capillary blood vessels;

[0070] Specifically, the high-definition camera is composed of multiple types of imaging devices, including a visible light camera, a near-infrared camera, and a multi-spectral camera.

[0071] Among them, the visible light camera is arranged at the front position of the monitoring system, which is used to capture the overall contour and surface features of the human face, and to monitor the color change, texture features and surface reflection characteristics of the facial skin in real time; the overall contour of the human face refers to the complete facial structure including eyebrows, eyes, nose, mouth and chin;

[0072] Among them, the near-infrared camera is installed adjacent to the visible light camera, which is used to penetrate the epidermis to obtain the blood vessel distribution and blood flow change information of the subcutaneous tissue, and to monitor the pulsation mode, blood flow velocity change and blood vessel volume change of the subcutaneous blood vessels;

[0073] Among them, the multi-spectral camera is equipped with an adjustable spectral filter array, which can capture the skin reflection characteristics at different wavelengths at the same time, and monitor the spectral absorption characteristics of hemoglobin under different oxygenation states.

[0074] In the specific implementation of the present application, the high-definition camera adopts a ring-shaped light source layout, with several light source units evenly distributed around the camera, forming uniform facial lighting conditions. The light source units include light sources of different wavelengths, covering the spectral range from visible light to near-infrared, to enhance the imaging contrast of different depth skin tissues. The optical system of the high-definition camera adopts an autofocus mechanism, which can adjust the focal length in real time according to the distance change of the measured object, ensuring the clarity of the facial image.

[0075] The installation position of the high-definition camera needs to meet the following conditions: maintaining an appropriate distance from the measured object to ensure that the complete facial area can be captured; the installation angle presents a proper angle with the front of the face to avoid strong reflection interference caused by direct light source from the front; the height matches the height of the face in the conventional sitting posture, reducing the need for posture adjustment.

[0076] Step 12: When detecting that the face enters the collection area, trigger the high-definition camera to continuously collect the face image sequence at a preset frame rate;

[0077] Specifically, the process of detecting the face entering the collection area adopts a multi-level detection mechanism, including a motion detection stage, a face preliminary screening stage, and a face confirmation stage.

[0078] In the motion detection stage, the background difference method is used to continuously monitor the moving targets in the collection area, and when persistent motion is detected, the face preliminary screening stage is triggered; in the face preliminary screening stage, a face detection algorithm is used to quickly scan the motion area to determine whether there is a face feature.

[0079] When it is confirmed that the face enters the collection area, the high-definition camera is automatically triggered to start working, and the high-definition camera continuously and uninterruptedly collects face images at a preset frame rate.

[0080] The preset frame rate is set according to the characteristics of different types of high-definition cameras and monitoring requirements: the visible light camera uses a higher frame rate to capture subtle facial color changes; the near-infrared camera uses a medium frame rate to balance image quality and information quantity; the multispectral camera sets an appropriate frame rate according to the spectral switching speed to ensure that sufficient image samples can be obtained at each wavelength.

[0081] Step 13: Face key area positioning is performed on the face image sequence, and a face feature point detection algorithm is used to determine the detection area that needs to be detected;

[0082] Specifically, the face key area positioning process is divided into two stages: face feature point detection and blood flow rich area determination.

[0083] In the facial feature point detection stage, a multi-level feature point detection method is adopted. First, a face alignment algorithm is used to standardize each frame in the sequence of face images, eliminating the influence of posture changes and scale differences. Then, a deep learning model is applied to extract key feature points of the face, including eye corners, nose tip, and mouth corners. Finally, a face coordinate system is established based on these fixed feature points, providing a reference framework for subsequent region positioning.

[0084] In the blood flow rich area determination stage, anatomical knowledge and image feature analysis are combined to determine suitable detection areas. First, based on the facial anatomy, candidate regions with dense capillary distribution such as forehead, cheek, and nasal ala are marked. Then, by analyzing the time-domain variation characteristics of the candidate regions in the image sequence, the blood flow signal intensity and signal-to-noise ratio of each region are evaluated. Finally, several regions that meet the pre-set intensity range and signal-to-noise ratio range in terms of signal intensity and signal-to-noise ratio are selected as the final detection regions.

[0085] The detection regions that need to be detected generally have the following characteristics: smooth skin surface, reducing the interference of texture changes on light reflection; rich capillary network, ensuring the strength of blood flow signals; thin subcutaneous fat layer, reducing the attenuation of light signals during propagation; less affected by facial expression changes, reducing interference caused by non-blood flow factors.

[0086] Step 14: Using multispectral imaging technology, capture the reflected spectral information of the detection region under different wavelengths, and record the fluctuation data of the reflected spectral information as spectral fluctuation data.

[0087] Specifically, the multispectral imaging technology captures the reflection characteristics of skin tissue to different wavelengths of light by imaging the face under different spectral bands.

[0088] The multispectral imaging system consists of a light source unit, a wavelength selection unit, and an image acquisition unit. The light source unit provides a wide spectrum light source covering visible light to near-infrared range, ensuring sufficient illumination intensity at each wavelength. The wavelength selection unit uses a programmable spectral filter array that can quickly switch different wavelength channels according to a pre-set sequence. The image acquisition unit is responsible for synchronously acquiring reflected images under each wavelength channel.

[0089] During the multispectral imaging process, different spectral channels are cyclically switched according to a pre-set wavelength sequence. Each cycle contains several discrete wavelength points that cover the characteristic absorption peaks and valley values of oxyhemoglobin and deoxyhemoglobin. At each wavelength point, several frames of images are acquired and averaged in real time to reduce the influence of random noise.

[0090] The record of the reflection spectrum fluctuation adopts a multi-dimensional data structure, containing the following information: a time dimension, recording the accurate time stamp of each acquisition; a space dimension, recording the accurate position and size of each detection area; a wavelength dimension, recording the center wavelength and bandwidth of each spectral channel; and an intensity dimension, recording the reflection intensity value at each time point, each area, and each wavelength.

[0091] It can be understood that the present application is applied to the field of blood pressure monitoring, and in the field, the present application is applied only when the user actively proposes a blood pressure monitoring requirement, and thus the face image and spectral information of the user are obtained after the user agrees and actively cooperates;

[0092] The signal processing of the reflection spectrum fluctuation to extract blood flow information includes the following steps:

[0093] Step 21: applying a blind source separation algorithm to separate a periodic signal related to heartbeats from the reflection spectrum fluctuation data;

[0094] Specifically, the blind source separation algorithm adopts an independent component analysis method, and the blind source separation algorithm is based on the signal statistical independence principle to decompose the mixed reflection spectrum fluctuation data into a plurality of independent source signals.

[0095] In the actual operation of the embodiment, first, the reflection spectrum fluctuation data is preprocessed, and the reflection spectrum data collected at different wavelengths is organized into a multi-dimensional matrix form. Then, the multi-dimensional matrix is decomposed by an independent component analysis method, and the independent component analysis method realizes the separation of the source signals by maximizing the non-Gaussianity of the source signals. In the decomposition process, the Newton iteration method is used to optimize the objective function, so as to improve the convergence speed and stability of the independent component analysis method.

[0096] Among the plurality of independent source signals obtained by the independent component analysis method, the periodic signal related to heartbeats is identified by spectrum analysis and pattern matching method. The specific periodic signal identification method is as follows: first, the power spectral density of each source signal is calculated, and the component with obvious periodicity (i.e., periodicity and repetition) is identified; then, correlation analysis is performed with the pre-collected standard heartbeat template, and the component with the highest correlation coefficient is selected as the periodic signal related to heartbeats; finally, the adaptive threshold method is used to filter out possible false peaks, so as to improve the purity of the periodic signal.

[0097] Step 22: converting the periodic signal into time domain and frequency domain features by a time-frequency analysis method to form blood flow information data;

[0098] Specifically, the time-frequency analysis method adopts a continuous wavelet transform technique, which can provide time domain and frequency domain information of a signal simultaneously and is suitable for analyzing non-stationary signals such as blood flow pulsation signals.

[0099] In actual operation, first, a Mexican hat wavelet is selected as a mother wavelet function, the mother wavelet function has good time-frequency positioning characteristics and is suitable for capturing transient changes in a blood flow signal. Then, by adjusting a wavelet scale parameter, a periodic signal is subjected to multi-scale decomposition to obtain wavelet coefficients of several frequency components.

[0100] Based on the obtained wavelet coefficients, time domain features and frequency domain features are extracted. The time domain features include: a pulse wave peak position determined by a local maximum value detection algorithm; a pulse wave valley position determined by a local minimum value detection algorithm; a pulse wave rising time obtained by calculating a time interval from a neighboring valley value to a peak value; a pulse wave falling time obtained by calculating a time interval from a peak value to a next valley value; and a double-peak feature extracted by detecting a secondary peak value after a main peak. The frequency domain features include: a main frequency component obtained by wavelet energy spectrum analysis; a frequency spectrum energy distribution obtained by calculating an energy proportion of different frequency bands; and a harmonic structure feature extracted by analyzing an energy proportion relationship between a main frequency and a harmonic.

[0101] Finally, the extracted time domain features and frequency domain features are combined to form a feature vector, and the blood flow information is constructed. In the process of forming the feature vector, feature normalization processing is adopted to convert features of different dimensions into a unified numerical range, so as to facilitate subsequent processing.

[0102] Further, the extraction method of the hemoglobin optical response feature is as follows:

[0103] Based on the spectral absorption characteristics at different wavelengths, the hemoglobin concentration change in the blood flow information is calculated by using the Beer-Lambert law to extract the hemoglobin optical response feature; the hemoglobin optical response feature includes a proportional change of oxyhemoglobin and deoxyhemoglobin, a hemoglobin concentration fluctuation amplitude, and a phase difference;

[0104] The hemoglobin optical response feature is due to the fact that when blood flows in a blood vessel, the volume of the blood vessel changes periodically with the heart beat, a pulse wave generated by the heart is conducted to a superficial facial artery, the expansion of the superficial facial artery pushes the blood in the capillary to the skin surface, the change in the number of hemoglobin changes the absorption and reflection of the environment light by the face, and then causes a weak fluctuation of the skin surface reflection spectrum;

[0105] Specifically, the extraction process of the hemoglobin optical response characteristics first decomposes the blood flow information by wavelength through a multi-wavelength spectral analysis technique. The multi-wavelength spectral analysis technique uses a spectral separation algorithm to classify the reflected spectral fluctuation data according to different wavelength ranges, forming spectral signals in several wavelength intervals.

[0106] Among them, the multi-wavelength spectral separation algorithm is based on the absorption characteristics of hemoglobin at different wavelengths, especially the absorption differences of oxyhemoglobin and deoxyhemoglobin in the green light band and the infrared band. For example, in the green light band, the oxyhemoglobin and the deoxyhemoglobin have different absorption peaks; while in the infrared band, the absorption rate difference of the oxyhemoglobin and the deoxyhemoglobin is more significant, which constitutes the basis for extracting the hemoglobin optical response characteristics.

[0107] Among them, the application process of the Beer-Lambert law includes establishing a relationship model between optical density and hemoglobin concentration. This relationship model eliminates the influence of skin tissue scattering and other pigment interference by calculating the optical density ratio at different wavelengths, thereby extracting the concentration change information of the hemoglobin.

[0108] In the specific implementation process of the present application, first, the spectral absorption matrix of the oxyhemoglobin and the deoxyhemoglobin is constructed, which contains the molar absorption coefficients of the two kinds of hemoglobin at different wavelengths. Then, by solving the Beer-Lambert law, the relative concentrations of the oxyhemoglobin and the deoxyhemoglobin are calculated. In the calculation process, an iterative optimization algorithm is used to continuously adjust the parameters until the fitting error reaches the minimum, thereby obtaining the most accurate concentration estimate; through the above calculation, the hemoglobin optical response characteristics are extracted, including the proportion change of the oxyhemoglobin and the deoxyhemoglobin, the hemoglobin concentration fluctuation amplitude and the phase difference.

[0109] Among them, the proportion change of the oxyhemoglobin and the deoxyhemoglobin reflects the oxygenation state of blood and can be used to evaluate the oxygen supply of microcirculation; the hemoglobin concentration fluctuation amplitude characterizes the blood volume change caused by heartbeats and is closely related to blood vessel elasticity and blood pressure; the phase difference reflects the time difference of hemoglobin optical response at different wavelengths, containing blood vessel wall characteristics and hemodynamic information.

[0110] And the extraction method of the hemodynamic conduction characteristics is:

[0111] By analyzing the blood flow pulsatile waveform in the blood flow information, the hemodynamic conduction characteristics are extracted using a fluid mechanics model, including pulse wave propagation speed, pulse waveform features, blood vessel elasticity index and blood flow resistance estimate value.

[0112] wherein the hemodynamic conduction feature is a lateral pressure of blood on the vessel wall due to the flow of blood, i.e. blood pressure, which has a strong correlation with hemodynamic parameters, and the increase of blood pressure leads to the increase of vessel wall pressure, the change of blood flow velocity and the increase of pulse wave propagation velocity;

[0113] Specifically, the extracting the hemodynamic conduction feature by using the fluid dynamics model comprises the following steps:

[0114] Step 31: performing time domain analysis on the blood flow pulsatile waveform in the blood flow information to extract the pulse wave propagation velocity;

[0115] Specifically, the time domain analysis adopts a pulse wave propagation time measurement method and is performed by identifying feature points in the blood flow pulsatile waveform. First, key feature points such as wave peak, wave trough, rising branch starting point and falling branch ending point are marked in the blood flow pulsatile waveform; then, the pulse wave arrival time difference between adjacent key feature points is calculated, and the pulse wave propagation velocity is obtained in combination with the anatomical blood vessel path distance.

[0116] wherein the identification of the key feature points adopts a differential analysis method, and the inflection points and extreme points of the waveform are located by calculating the first derivative and the second derivative of the blood flow pulsatile waveform. For the identification of the wave peak and the wave trough, a peak detection algorithm is adopted, and a self-adaptive threshold method is used to exclude false peak value interference; for the identification of the rising branch starting point, a tangent method is used to determine the position where the waveform starts to obviously rise; for the identification of the falling branch ending point, the position where the waveform tends to be flat is determined by analyzing the curvature change of the waveform.

[0117] Step 32: performing morphological analysis on the blood flow pulsatile waveform to extract the pulse waveform morphological features;

[0118] Specifically, the morphological analysis is realized by a waveform parameterization method. The blood flow pulsatile waveform is decomposed into several key morphological parameters, including waveform width, waveform height, waveform area, waveform slope, waveform symmetry and double peak feature, etc.

[0119] In the extraction process of the key morphological parameters, first, the blood flow pulsatile waveform is normalized to eliminate the influence of amplitude difference; then, the time width features of the normalized blood flow pulsatile waveform are calculated, including full wave width, half height width and time interval between feature points; then, the shape features of the blood flow pulsatile waveform are analyzed, including the slope ratio of the rising branch and the falling branch, the roundness of the wave peak and the double peak ratio; finally, the frequency domain features of the blood flow pulsatile waveform are extracted, including the energy distribution of the main frequency component and the harmonic component.

[0120] The doublet feature analysis adopts a wavelet decomposition method to separate the main wave and the doublet wave components, calculate the amplitude ratio and time delay of the two components, and reflect the changes in peripheral vascular resistance and compliance. The waveform symmetry analysis quantifies the waveform symmetry by calculating the ratio of the waveform rise time to the fall time and the waveform centroid position offset.

[0121] Step 33: Based on the pulse wave propagation speed and the pulse waveform features, a vascular elasticity index is calculated by combining a fluid mechanics model.

[0122] Specifically, the fluid mechanics model adopts an improved Windkessel model, which simplifies the vascular system as a pipeline network with elastic and resistance characteristics. The Windkessel model contains several parameters representing the elasticity of large arteries, the resistance of small arteries, and the resistance of microcirculation.

[0123] During the calculation of the vascular elasticity index, first, based on the pulse wave propagation speed, the initial value of the vascular wall elastic modulus is estimated using the Moens-Korteweg equation. Then, by combining the time features and shape features in the pulse waveform features, the parameters of the Windkessel model are adjusted through a parameter optimization algorithm to make the output theoretical waveform match the actual measured blood flow pulsatile waveform to the greatest extent. Finally, the index representing vascular elasticity is extracted from the optimized model parameters to form the vascular elasticity index.

[0124] The parameter optimization algorithm adopts an iterative method to minimize the error function between the theoretical waveform and the actual waveform. The error function considers the amplitude difference, phase difference, and shape difference of the waveforms to ensure that the optimization result accurately reflects the physical characteristics of the vascular system.

[0125] Step 34: Based on the pulse waveform features and the vascular elasticity index, a blood flow resistance estimate is estimated by combining hemodynamic principles.

[0126] Specifically, the estimation of the blood flow resistance estimate is based on a modified Poiseuille's law, which describes the relationship between blood flow, vessel radius, and blood pressure gradient. Using the waveform area and waveform slope information in the pulse waveform features, combined with the vascular elasticity index, a blood flow resistance estimation model is constructed.

[0127] During the blood flow resistance estimation process, first, the indicators reflecting blood flow velocity and blood flow volume are extracted from the pulse waveform features, including the waveform rise slope, waveform area, and waveform decay rate. Then, combined with the vascular elasticity index, the dynamic changes in the vessel cross-sectional area are calculated through hemodynamic equations. Finally, based on the blood flow velocity, vessel cross-sectional area, and blood pressure gradient estimate, the fluid mechanics principle is applied to calculate the blood flow resistance estimate.

[0128] The blood flow equation considers the non-Newtonian fluid characteristics of blood and the elastic characteristics of blood vessels, and the applicability of the traditional Poiseuille law in the microcirculation system is corrected by introducing a correction coefficient. The blood pressure gradient estimation value is obtained by analyzing the wave steepness and wave attenuation characteristics in the pulse waveform.

[0129] The multi-modal feature fusion network is constructed based on the hemoglobin optical response characteristics and the blood flow conduction characteristics, and a comprehensive feature representation is formed, including the following steps:

[0130] Step 41: design an optical feature processing subnetwork to process the hemoglobin optical response characteristics using a convolutional neural network structure to extract the representation of blood component changes;

[0131] Specifically, the optical feature processing subnetwork is composed of several convolutional layers, pooling layers and fully connected layers. In the network structure design, first, the hemoglobin optical response characteristics are organized into a two-dimensional feature map, including the proportion change of oxyhemoglobin and deoxyhemoglobin, the amplitude of hemoglobin concentration fluctuation and the phase difference, etc. Then, several convolutional layers are used to process the feature map, and each convolutional layer uses several different size convolutional kernels to extract local feature patterns. After convolution operation, the pooling layer is used to reduce the feature dimension and retain the key information. Finally, the extracted features are mapped to a fixed-dimensional feature vector through the fully connected layer, which is used as the representation of blood component changes.

[0132] In the specific implementation of the present application, the optical feature processing subnetwork can also use a residual connection structure to directly pass the shallow features to the deep network through a jump connection method, which can alleviate the gradient vanishing problem and retain the original feature information. At the same time, batch normalization layers are inserted between the convolutional layers to standardize the features and improve the stability of network training. The activation function of the optical feature processing subnetwork uses the ReLU function, which effectively solves the gradient vanishing problem while keeping the calculation simple.

[0133] Step 42: design a blood flow dynamics feature processing subnetwork to process the blood flow conduction characteristics using a long short-term memory network structure to extract the blood vessel state and blood flow characteristic representation;

[0134] Specifically, the hemodynamic feature processing subnetwork is composed of an input layer, a plurality of long short-term memory units, and an output layer. In the network structure design, first, the hemodynamic conduction features, including the pulse wave propagation speed, the pulse waveform features, the vascular elasticity index, and the blood flow resistance estimate, etc., are organized into a time series sequence. Then, the sequence is gradually processed through a plurality of long short-term memory units, each of which contains three gating mechanisms of input gate, forget gate, and output gate, which control the input of new information, the forgetting of historical information, and the output of the current state, respectively.

[0135] In the specific implementation of the present application, the hemodynamic feature processing subnetwork adopts a bidirectional long short-term memory network structure, which processes the time series features from both forward and reverse directions to capture more comprehensive time series dependencies. Each long short-term memory unit in the network contains a plurality of neurons inside, which selectively remember and forget information through gating mechanisms. To enhance the expression ability of the network, a dropout layer is added between the long short-term memory layers, which randomly disconnects a plurality of neuron connections to prevent network overfitting. Finally, the output of the long short-term memory network is mapped to a fixed-dimensional feature vector through a fully connected layer, which serves as a representation of the vascular state and blood flow characteristics.

[0136] Step 43: Construct a multi-modal feature fusion subnetwork, which adopts an attention mechanism and a cross-modal transformer structure to adaptively fuse the output features of the optical feature processing subnetwork and the hemodynamic feature processing subnetwork, generating the comprehensive feature representation;

[0137] Specifically, the multi-modal feature fusion subnetwork is composed of an attention module, a cross-modal Transformer module, and a feature fusion module. In the network structure design, first, the features output by each subnetwork are weighted by the attention module according to their importance. The attention module calculates the attention weight of each feature dimension, highlighting the key features and suppressing the secondary features. Then, the cross-modal Transformer module is used to realize the interaction and information transmission between different modal features. The cross-modal Transformer module contains a plurality of multi-head self-attention layers and a feedforward neural network layer, which captures the correlation between different modal features through a self-attention mechanism.

[0138] In the implementation of the present application, the multi-modal feature fusion sub-network first projects the output features of the optical feature processing sub-network and the output features of the hemodynamic feature processing sub-network into the same dimensional feature space respectively to obtain corresponding sub-network feature vectors. Then, through a multi-head attention mechanism, the sub-network feature vectors are taken as query vectors to calculate the attention scores between the sub-network feature vectors and other sub-network feature vectors, and a context-aware perception feature vector is generated. Next, through residual connection and layer normalization processing, the expression ability and training stability of the context-aware feature vector are enhanced, and a multi-modal feature vector is obtained. Finally, through a feature fusion module, the multi-modal features of each sub-network are connected and mapped to a unified feature space to generate the comprehensive feature representation. The feature fusion module adopts a gating mechanism to dynamically adjust the contribution weight of the sub-network feature vector output by each sub-network in the fusion process according to the reliability and correlation of different sub-networks.

[0139] The model structure diagram of the multi-modal feature fusion network in the present embodiment is shown in Figure 2

[0140] The step of establishing a blood pressure mapping model based on the comprehensive feature representation includes the following steps:

[0141] Step 51: Construct a blood pressure sample data set containing the comprehensive feature representation of the subject under different physiological states and the corresponding standard blood pressure measurement value, which is used as historical blood pressure data.

[0142] Specifically, the construction process of the blood pressure sample data set includes three stages of subject recruitment, data acquisition and data annotation. In the subject recruitment stage, subjects of different ages, different genders and different health conditions are selected to ensure the diversity and representativeness of the samples. In the data acquisition stage, the comprehensive feature representation and the standard blood pressure value are simultaneously collected for each subject under different physiological states such as resting state, post-exercise state and emotional change state. Among them, the standard blood pressure value is obtained by measuring with a medical-grade electronic sphygmomanometer, including systolic pressure value and diastolic pressure value. In the data annotation stage, each group of comprehensive feature representation is paired with the corresponding standard blood pressure value to form an annotated sample. All annotated samples are organized in chronological order to constitute the blood pressure sample data set.

[0143] The blood pressure sample data set is divided into a training set, a validation set and a test set. Among them, the training set is used to train the static blood pressure estimation sub-model and the dynamic blood pressure change prediction sub-model; the validation set is used for model parameter optimization and model selection; the test set is used to evaluate the model performance and generalization ability.

[0144] ​Step 52: based on the blood pressure sample dataset, a support vector regression algorithm is used to design the static blood pressure estimation sub-model for estimating the basic blood pressure value; the static blood pressure estimation sub-model uses a deep neural network structure to map the comprehensive feature representation to the basic values of systolic pressure and diastolic pressure;

[0145] Specifically, the static blood pressure estimation sub-model uses a multi-layer perceptron structure, including an input layer, several hidden layers and an output layer. The number of neurons in the input layer is the same as the dimension of the comprehensive feature representation, which is used to receive the comprehensive feature representation; the hidden layer uses a fully connected structure, each hidden layer is equipped with a batch normalization layer and a ReLU activation function, which is used to extract deep features in the comprehensive feature representation; the output layer contains two neurons, which correspond to the prediction results of the basic values of systolic pressure and diastolic pressure, respectively.

[0146] In the training process of the static blood pressure estimation sub-model, the training set in the blood pressure sample dataset is used for model parameter learning. The training target is to minimize the mean square error between the predicted blood pressure value and the standard blood pressure value. The optimization algorithm uses the Adam optimizer, and the learning rate uses a dynamic adjustment strategy, with a large initial value that gradually decreases as the training round increases. To prevent overfitting, L2 regularization and early stopping mechanism are used, and training is stopped when the performance on the validation set no longer improves.

[0147] Step 53: based on the time series data in the blood pressure sample dataset, a recurrent neural network structure is used to design the dynamic blood pressure change prediction sub-model for predicting the fluctuation trend of blood pressure in the short term; the time series data is a sequence of comprehensive feature representations sorted by time order;

[0148] Specifically, the dynamic blood pressure change prediction sub-model uses a bidirectional long short-term memory network structure, including an input layer, several recurrent layers and an output layer. The input layer receives the time series of the comprehensive feature representation within a pre-set time window; the recurrent layer is composed of bidirectional LSTM units, which can consider past and future time series information simultaneously, and capture the time dependence of blood pressure changes; the output layer maps the hidden state of the LSTM unit to the blood pressure change trend prediction value through a fully connected network.

[0149] The training of the dynamic blood pressure change prediction sub-model uses a sequence-to-sequence learning paradigm. In the training process, the input is the time series of the comprehensive feature representation within a continuous time window, and the output target is the blood pressure change sequence within the corresponding time window. The loss function of the dynamic blood pressure change prediction sub-model uses the mean square error loss function between the predicted blood pressure change and the true blood pressure change, while weighting the point prediction error and the trend prediction error, where the point prediction error measures the difference between the predicted value and the true value at each time point, and the trend prediction error measures the consistency of the change trend between the predicted sequence and the true sequence.

[0150] Step 54: The dynamic blood pressure change prediction sub-model combines the time series of the comprehensive feature representation and historical blood pressure data to predict the trend and amplitude of blood pressure change through a gated recurrent unit network;

[0151] Specifically, the gated recurrent unit network in the dynamic blood pressure change prediction sub-model is composed of an input gate, a forget gate, and an output gate. The input gate controls the degree of new information entering the memory unit, the forget gate controls the degree of historical information retention, and the output gate controls the degree of memory unit information output. The gating mechanism enables the model to adaptively determine how much historical information to retain and how much new information to receive, making it particularly suitable for processing time series data such as blood pressure changes that have long-term dependencies and short-term fluctuations.

[0152] During the prediction process, the gated recurrent unit network first receives the comprehensive feature representation and historical blood pressure data at the current time as input. Historical blood pressure data includes systolic and diastolic pressure records at past time points. The network filters and integrates the input information through the gating mechanism, updates the internal state, and then outputs the prediction result of blood pressure change. The prediction result includes the direction of blood pressure change (up, down, or stable) and the change amplitude for systolic and diastolic pressures.

[0153] The output of the gated recurrent unit network is adjusted by the post-processing module to ensure that the prediction result is physiologically reasonable. The post-processing module contains constraint rules that limit the predicted blood pressure change amplitude within a reasonable range and ensure that the systolic pressure is always greater than the diastolic pressure, maintaining the physiological significance of the prediction result.

[0154] The calculation of the basic blood pressure value through the static blood pressure estimation sub-model and the calculation of the blood pressure fluctuation value through the dynamic blood pressure change prediction sub-model include the following steps:

[0155] Step 61: Input the latest comprehensive feature representation into the static blood pressure estimation sub-model to calculate the basic systolic and diastolic pressure values at the current time through forward propagation, as the basic blood pressure value;

[0156] Specifically, the process of forward propagation calculates the output of each layer in order from the input layer to the output layer of the static blood pressure estimation sub-model. First, input the comprehensive feature representation into the input layer of the static blood pressure estimation sub-model; then, calculate the intermediate feature representation layer by layer through the weight matrix and activation function of the hidden layer of the static blood pressure estimation sub-model; finally, generate the basic systolic and diastolic pressure values through the linear transformation of the output layer of the static blood pressure estimation sub-model.

[0157] In the implementation of the present application, the forward propagation is calculated in batch mode, the comprehensive feature representations of consecutive time points are organized into a feature batch, and the calculation of the basic systolic and diastolic pressure values is completed at one time, thereby improving the calculation efficiency. The intermediate calculation results in the forward propagation process are cached for subsequent model updating and adjustment.

[0158] Step 62: Based on the historical blood pressure data and the current comprehensive feature representation, a blood pressure fluctuation trend in the short term is calculated through the recursive structure of the dynamic blood pressure change prediction sub-model;

[0159] Specifically, in the blood pressure fluctuation trend calculation process, first, an input sequence is constructed, which includes the comprehensive feature representations of historical time points; then, through the gating mechanism of the long short-term memory network, each time step in the input sequence is processed, and the internal state of the network is updated; finally, based on the updated network state, the blood pressure change trend at future time points is predicted.

[0160] Step 63: According to the blood pressure fluctuation trend, the blood pressure fluctuation values, including the systolic pressure fluctuation value and the diastolic pressure fluctuation value, are calculated using a time series analysis method;

[0161] Specifically, the time series analysis method uses a trend decomposition technique to decompose the blood pressure fluctuation trend into a trend term, a periodic term, and a residual term.

[0162] In the blood pressure fluctuation value calculation process, first, the moving average filter is applied to the blood pressure fluctuation trend to extract the long-term change trend; then, the Fourier analysis method is used to identify the periodic change pattern to extract the periodic term; finally, the difference between the original trend and the sum of the trend term and the periodic term is calculated to obtain the residual term. The systolic pressure fluctuation value and the diastolic pressure fluctuation value are obtained by weighted combination of the corresponding trend term, periodic term, and residual term.

[0163] Among them, the trend term reflects the long-term change direction of blood pressure, which is related to the sustained change of physiological state; the periodic term captures the regular fluctuation of blood pressure, which is related to the respiratory cycle and autonomic nervous system activity; the residual term represents short-term random fluctuations, which is related to instantaneous external stimuli and internal regulation.

[0164] In the implementation of the present application, the weighted combination adopts an adaptive weight strategy, which dynamically adjusts the weight coefficients according to the stability and prediction value of each term under different physiological states. For example, in the resting state, the weights of the trend term and the periodic term are higher; in the active state, the weight of the residual term is increased accordingly to capture more short-term fluctuation information.

[0165] The way to generate the final blood pressure monitoring result is:

[0166] combining the base blood pressure value and the blood pressure fluctuation value through an adaptive weighting algorithm to generate final systolic and diastolic pressures as the final blood pressure monitoring result;

[0167] Specifically, the adaptive weighting algorithm first constructs a multi-level weight distribution structure and calculates the combination weight coefficients of systolic and diastolic pressures respectively, wherein the systolic pressure combination weight coefficient mainly considers the influence of cardiac contractility and aortic elasticity, and the diastolic pressure combination weight coefficient mainly considers the influence of peripheral vascular resistance and microcirculation state; then, the base blood pressure value and the blood pressure fluctuation value are weighted and summed according to the combination weight coefficients to generate preliminary estimated values of systolic and diastolic pressures respectively; next, the preliminary estimated values are subjected to physiological constraint processing, for example, including reasonableness test and time series smoothing, to ensure that the blood pressure monitoring result conforms to the physiological law; finally, the processed systolic and diastolic pressure values are organized into a standard format blood pressure data structure containing measurement time stamp, systolic pressure value, diastolic pressure value, mean arterial pressure value and pulse pressure difference value, etc. as the output of the final blood pressure monitoring result.

[0168] Embodiment 2

[0169] As shown in Figure 3 A non-contact blood spectrum imaging blood pressure monitoring system, comprising an image acquisition module, a signal processing module, a feature extraction module, a feature fusion module, a blood pressure mapping module and a result output module; wherein each module is connected through electrical means;

[0170] The image acquisition module captures the reflected spectral fluctuation of the key detection area of the human face through a high-resolution camera and sends the reflected spectral fluctuation data to the signal processing module;

[0171] The signal processing module receives the reflected spectral fluctuation data, processes the reflected spectral fluctuation data, extracts blood flow information, and sends the blood flow information to the feature extraction module;

[0172] The feature extraction module receives the blood flow information, extracts hemoglobin optical response features and blood flow hemodynamics conduction features based on the blood flow information, and sends the hemoglobin optical response features and the blood flow hemodynamics conduction features to the feature fusion module;

[0173] The feature fusion module receives the hemoglobin optical response features and the blood flow hemodynamics conduction features, constructs a multi-modal feature fusion network, forms a comprehensive feature representation, and sends the comprehensive feature representation to the blood pressure mapping module;

[0174] a blood pressure mapping module receiving the comprehensive feature representation, establishing a blood pressure mapping model based on the comprehensive feature representation, wherein the blood pressure mapping model comprises a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model; the blood pressure mapping module calculates a basic blood pressure value through the static blood pressure estimation sub-model, and calculates a blood pressure fluctuation value through the dynamic blood pressure change prediction sub-model, and sends the basic blood pressure value and the blood pressure fluctuation value to a result output module;

[0175] a result output module receiving the basic blood pressure value and the blood pressure fluctuation value, combining the basic blood pressure value and the blood pressure fluctuation value to generate a final blood pressure monitoring result.

[0176] In addition, the parts of the above technical solutions in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive repetition.

[0177] The specific embodiments described above further illustrate the purposes, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0178] The above preset parameters or preset thresholds are set by a person skilled in the art according to actual conditions or obtained through a large amount of data simulation.

[0179] The above embodiments are only used to illustrate the technical method of the present application and are not intended to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A non-contact blood spectral imaging blood pressure monitoring method, characterized in that: The following steps are involved: Step 1: Use a camera to capture the reflection spectrum fluctuations of the key detection area of ​​the face; Step 2: performing signal processing on the reflected spectrum fluctuation to extract blood flow information; Step 3: extracting hemoglobin optical response characteristics and hemodynamic conduction characteristics based on the blood flow information; Step 4: constructing a multimodal feature fusion network based on the hemoglobin optical response characteristics and the hemodynamic conduction characteristics to form a comprehensive feature representation; Step 5: establishing a blood pressure mapping model based on the comprehensive feature representation, wherein the blood pressure mapping model includes a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model; Step 6: Calculate the basal blood pressure value using the static blood pressure estimation sub-model, and calculate the blood pressure fluctuation value using the dynamic blood pressure change prediction sub-model; Step 7: Combine the basal blood pressure value with the blood pressure fluctuation value to generate a final blood pressure monitoring result.

2. The non-contact blood spectral imaging blood pressure monitoring method according to claim 1, characterized in that: The method of capturing the reflection spectrum fluctuation of the key detection area of ​​the face by the camera includes the following steps: Step 11: Arrange the camera and aim it at the key detection areas of the face, including the forehead, cheeks, and nose wings; Step 12: When a face is detected entering the collection area, the camera is triggered to continuously collect a sequence of face images at a preset frame rate; Step 13: Positioning key facial regions of the facial image sequence, and determining detection regions to be detected using a facial feature point detection algorithm; Step 14: Use multispectral imaging technology to capture the reflection spectrum information at different wavelengths in the detection area, and record the fluctuation data of the reflection spectrum information as spectrum fluctuation data.

3. The non-contact blood spectral imaging blood pressure monitoring method according to claim 2, characterized in that: The signal processing of the reflection spectrum fluctuation to extract blood flow information comprises the following steps: Step 21: Apply a blind source separation algorithm to separate the periodic signal related to the heart beat from the reflectance spectrum fluctuation data; Step 22: Convert the periodic signal into time domain and frequency domain features through a time-frequency analysis method to form blood flow information data.

4. The non-contact blood spectral imaging blood pressure monitoring method according to claim 3, characterized in that: The method for extracting the optical response characteristics of hemoglobin comprises the following steps: Based on the spectral absorption characteristics at different wavelengths, the Beer-Lambert law is used to calculate the hemoglobin concentration change in the blood flow information and extract the hemoglobin optical response characteristics; the hemoglobin optical response characteristics include the change in the ratio of oxyhemoglobin and deoxyhemoglobin, the hemoglobin concentration fluctuation amplitude and the phase difference.

5. The non-contact blood spectral imaging blood pressure monitoring method according to claim 4, characterized in that: The hemodynamic conduction characteristics are extracted as follows: By analyzing the blood flow pulsation waveform in the blood flow information, the hemodynamic conduction characteristics are extracted using a fluid mechanics model. The hemodynamic conduction characteristics include pulse wave propagation velocity, pulse wave morphology characteristics, vascular elasticity index and blood flow resistance estimation value.

6. The non-contact blood spectral imaging blood pressure monitoring method according to claim 5, characterized in that: The step of constructing a multimodal feature fusion network based on the hemoglobin optical response feature and the hemodynamic conduction feature to form a comprehensive feature representation includes the following steps: Step 41: Design an optical feature processing subnetwork, use a convolutional neural network structure to process the optical response characteristics of the hemoglobin, and extract the representation of the changes in blood components; Step 42: Design a hemodynamic feature processing subnetwork, use a long short-term memory network structure to process the hemodynamic conduction features, and extract vascular status and blood flow characteristic representations; Step 43: Construct a multimodal feature fusion subnetwork, adopt an attention mechanism and a cross-modal transformer structure, adaptively fuse the output features of the optical feature processing subnetwork and the hemodynamic feature processing subnetwork, and generate the comprehensive feature representation.

7. The non-contact blood spectral imaging blood pressure monitoring method according to claim 6, characterized in that: The establishing of the blood pressure mapping model based on the comprehensive feature representation comprises the following steps: Step 51: constructing a blood pressure sample data set, which includes the comprehensive feature representation and corresponding standard blood pressure measurement values ​​of the subject in different physiological states, and the standard blood pressure measurement values ​​are used as historical blood pressure data; Step 52: Based on the blood pressure sample dataset, a support vector regression algorithm is used to design the static blood pressure estimation sub-model for obtaining the basal blood pressure value; the static blood pressure estimation sub-model uses a deep neural network structure to map the comprehensive feature representation into the basal values ​​of systolic and diastolic blood pressure; Step 53: Based on the time series data in the blood pressure sample dataset, a recurrent neural network structure is used to design the dynamic blood pressure change prediction sub-model to predict the fluctuation trend of blood pressure in the short term; the time series data is a sequence represented by comprehensive features arranged in chronological order; Step 54: The dynamic blood pressure change prediction sub-model combines the time series represented by the comprehensive features and the historical blood pressure data to predict the change trend and amplitude of the blood pressure through a gated recurrent unit network.

8. The non-contact blood spectral imaging blood pressure monitoring method according to claim 7, characterized in that: The static blood pressure estimation submodel adopts a multi-layer perceptron structure, comprising an input layer, a hidden layer, and an output layer; the number of neurons in the input layer is the same as the dimension of the comprehensive feature representation, and is used to receive the comprehensive feature representation; the hidden layer adopts a fully connected structure, and each hidden layer is equipped with a normalization layer and a ReLU activation function to extract deep features from the comprehensive feature representation; The output layer contains two neurons, corresponding to the prediction results of the systolic blood pressure base value and the diastolic blood pressure base value respectively; During the training process of the static blood pressure estimation sub-model, the training set in the blood pressure sample data set is used to learn the model parameters; The training goal is to minimize the mean square error between the predicted blood pressure value and the standard blood pressure value; The optimization algorithm uses the Adam optimizer.

9. The non-contact blood spectral imaging blood pressure monitoring method according to claim 8, characterized in that: The dynamic blood pressure change prediction sub-model adopts a bidirectional long short-term memory network structure, including an input layer, a recurrent layer, and an output layer; the input layer receives the time series represented by the comprehensive features within a preset time window; the recurrent layer is composed of bidirectional LSTM units; and the output layer maps the hidden state of the LSTM unit to a predicted value of the blood pressure change trend through a fully connected network. During the training process of the dynamic blood pressure change prediction sub-model, the input is the time series represented by the comprehensive features within the continuous time window, and the output target is the blood pressure change sequence within the corresponding time window; the loss function of the dynamic blood pressure change prediction sub-model adopts the mean square error loss function between the predicted blood pressure change and the actual blood pressure change, and simultaneously weightedly integrates the point prediction error and the trend prediction error, where the point prediction error measures the difference between the predicted value and the actual value at each time point, and the trend prediction error measures the consistency of the change trend between the predicted sequence and the actual sequence.

10. The non-contact blood spectral imaging blood pressure monitoring method according to claim 9, characterized in that: Calculating the basal blood pressure value by the static blood pressure estimation submodel and calculating the blood pressure fluctuation value by the dynamic blood pressure change prediction submodel comprises the following steps: Step 61: Input the latest collected comprehensive feature representation into the static blood pressure estimation sub-model, and calculate the current basal systolic and diastolic blood pressure values ​​through forward propagation; Step 62: Based on the historical blood pressure data and the current comprehensive feature representation, the short-term blood pressure fluctuation trend is calculated through the recursive structure of the dynamic blood pressure change prediction sub-model; Step 63: Calculate the blood pressure fluctuation value, including the systolic pressure fluctuation value and the diastolic pressure fluctuation value, based on the blood pressure fluctuation trend using a time series analysis method.

11. The non-contact blood spectral imaging blood pressure monitoring method according to claim 10, characterized in that: The method for generating the final blood pressure monitoring result is: The basal blood pressure value and the blood pressure fluctuation value are combined through an adaptive weighting algorithm to generate final systolic pressure and diastolic pressure as the final blood pressure monitoring result.

12. A non-contact blood spectral imaging blood pressure monitoring system, used to implement the non-contact blood spectral imaging blood pressure monitoring method according to any one of claims 1 to 11, characterized in that: It includes an image acquisition module, a signal processing module, a feature extraction module, a feature fusion module, a blood pressure mapping module, and a result output module; wherein each module is electrically connected; The image acquisition module captures the reflection spectrum fluctuations of the key detection area of ​​the face through the camera and sends the reflection spectrum fluctuation data to the signal processing module; a signal processing module, receiving the reflection spectrum fluctuation data, performing signal processing on the reflection spectrum fluctuation data, extracting blood flow information, and sending the blood flow information to a feature extraction module; a feature extraction module that receives the blood flow information, extracts a hemoglobin optical response feature and a hemodynamic conduction feature based on the blood flow information, and sends the hemoglobin optical response feature and the hemodynamic conduction feature to a feature fusion module; a feature fusion module that receives the hemoglobin optical response feature and the hemodynamic conduction feature, constructs a multimodal feature fusion network, forms a comprehensive feature representation, and sends the comprehensive feature representation to the blood pressure mapping module; a blood pressure mapping module that receives the comprehensive feature representation and establishes a blood pressure mapping model based on the comprehensive feature representation, wherein the blood pressure mapping model includes a static blood pressure estimation sub-model and a dynamic blood pressure change prediction sub-model; the blood pressure mapping module calculates a basal blood pressure value using the static blood pressure estimation sub-model and calculates a blood pressure fluctuation value using the dynamic blood pressure change prediction sub-model, and sends the basal blood pressure value and the blood pressure fluctuation value to a result output module; The result output module receives the basal blood pressure value and the blood pressure fluctuation value, combines the basal blood pressure value with the blood pressure fluctuation value, and generates a final blood pressure monitoring result.

Images