Multi-modal synchronous monitoring method and system based on traditional chinese medicine pulse condition and microcirculation

Abstract

The application discloses a multi-modal synchronous monitoring method and system based on traditional Chinese medicine pulse conditions and microcirculation, and relates to the technical field of intelligent diagnosis of traditional Chinese medicine. A radial styloid process point is recognized through a wrist depth image, a pulse position coordinate set is generated in combination with a preset space mapping relationship, a pulse pressure signal and a microcirculation blood flow image are synchronously collected, frequency domain transformation is performed on the pulse pressure signal and pathological risk probability L is calculated, a comprehensive node feature vector is constructed to form a graph structure, graph data is input into a pre-trained graph convolution network, a pathological mode recognition result is output by matching a pathological subgraph template library, dynamic correlation analysis of traditional Chinese medicine pulse conditions and microcirculation is realized, and the technical bottleneck of strong subjectivity, multi-modal asynchrony and difficult pathological quantification of traditional pulse diagnosis is broken through.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traditional Chinese medicine intelligent diagnosis, in particular to a multi-modal synchronous monitoring method and system based on traditional Chinese medicine pulse and microcirculation. BACKGROUND

[0002] Traditional Chinese medicine pulse diagnosis relies on the doctor's tactile perception of radial artery pulsation characteristics (such as force, rhythm), and has two major defects of strong subjective experience dependence and lack of quantitative standard. Although electronic pulse diagnosis equipment can collect pulse pressure signals, the sensor attachment relies on manual positioning (error often > 5mm), resulting in distortion of the spatial mapping of "cun, guan, chi" pulse sites, and the inability to accurately reflect the gradient changes of the blood state of different pulse sites (such as "cun floating chi sinking" virtual and real differentiation).

[0003] When laser speckle imaging technology is introduced to obtain microcirculation blood flow parameters, there is a hardware acquisition time sequence deviation (typical value 10-50ms) between the pulse pressure signal (sampling rate > 1kHz) and the microcirculation image (frame rate 30fps). More importantly, the involuntary wrist tremor of the human body (amplitude 50-200μm) will further cause a physiological response time delay, resulting in a time misalignment between the pulse wave peak and the capillary perfusion peak (actual measurement error 12.3±2.1ms), completely blocking the dynamic correlation verification of the traditional Chinese medicine "pulse blood homology" theory.

[0004] The time and space asynchrony of multi-modal signals causes the dynamic correlation of pulse and microcirculation to be destroyed, making the pathological quantitative analysis guided by the traditional Chinese medicine holistic view lose a scientific basis. In view of the above problems, the present application urgently needs to develop an intelligent analysis method that can integrate the time and space characteristics of pulse and the dynamic changes of microcirculation, and eliminate the asynchronous error of signals. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a multi-modal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation.

[0006] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0007] In the first aspect, the present application discloses a multi-modal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation, comprising the following steps:

[0008] Obtaining pulse pressure time sequence signal data and microcirculation blood flow image sequence data of the radial styloid process point position of the user;

[0009] Calculating the pathological risk probability L by pathological risk calculation on the pulse pressure time sequence signal data and judging:

[0010] When the pathological risk probability L is greater than or equal to a preset first threshold value, the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data are time domain aligned, and the specific steps are as follows:

[0011] acquire motion compensation data of the user;

[0012] calculate a displacement amount of a dynamic time warping path based on the pulse pressure time series signal data, the microcirculation blood flow image sequence data and the motion compensation data through a path cost function;

[0013] trigger a path correction function to generate a displacement compensation amount when the displacement amount is greater than a predetermined displacement threshold;

[0014] obtain a comprehensive node feature vector through time domain alignment processing based on the displacement amount of the dynamic time warping path and the displacement compensation amount;

[0015] calculate an edge weight value between adjacent pulse position nodes based on a correlation measure of signals between nodes of the comprehensive node feature vector, and construct graph structure data;

[0016] input the graph structure data into a pre-trained graph convolutional neural network model, match a preset pathological sub-atlas template library, and output a pathological pattern recognition result.

[0017] In a second aspect, the present application discloses a multi-modal synchronous monitoring system based on traditional Chinese medicine pulse and microcirculation, which uses the multi-modal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation described above, comprising:

[0018] The acquisition module is configured to acquire pulse pressure time series signal data and microcirculation blood flow image sequence data of a radial styloid process point position of a user.

[0019] The risk judgment module is configured to calculate a pathological risk probability L based on the pulse pressure time series signal data and make a judgment.

[0020] The signal synchronization module is configured to perform time domain alignment on the pulse pressure time series signal data and the microcirculation blood flow image sequence data when the pathological risk probability L is greater than or equal to a preset first threshold, and the specific steps are as follows:

[0021] acquire motion compensation data of the user;

[0022] calculate a displacement amount of a dynamic time warping path based on the pulse pressure time series signal data, the microcirculation blood flow image sequence data and the motion compensation data through a path cost function;

[0023] trigger a path correction function to generate a displacement compensation amount when the displacement amount is greater than a predetermined displacement threshold;

[0024] obtain a comprehensive node feature vector through time domain alignment processing based on the displacement amount of the dynamic time warping path and the displacement compensation amount;

[0025] a graph modeling module, configured to calculate edge weight values between adjacent pulse position nodes based on a correlation measure of inter-node signals of the integrated node feature vectors, and construct graph structure data based on the edge weight values;

[0026] a pathology recognition module, configured to input the graph structure data into a pre-trained graph convolutional neural network model, and match the graph structure data with a pre-set pathology sub-atlas template library, and output a pathology pattern recognition result.

[0027] Compared with the prior art, the present application has the following beneficial effects:

[0028] 1. The present application introduces the synergistic effect of motion compensation data and path cost function, and activates the time domain alignment mechanism when the pathological risk probability is above the threshold value. By capturing the acceleration changes caused by limb micro-movement in real time, and triggering the compensation function based on the dynamic time warping path displacement, the physiological response time delay between the pulse pressure signal and the microcirculation blood flow image sequence is effectively eliminated. This mechanism breaks through the signal phase misalignment caused by the hardware acquisition time sequence deviation and the unconscious tremor of the human body in traditional devices, and makes the pulse wave peak strictly correspond to the capillary perfusion peak, providing a verifiable technical basis for the Chinese medicine theory of "pulse blood homology".

[0029] 2. The present application constructs a topological network containing pulse position nodes and mutual information edge weights based on the integrated node feature vectors generated by time domain alignment. By quantifying the pulse pressure waveform entropy, blood flow texture and other multi-dimensional features as node attributes, and calculating the edge weights based on the correlation of adjacent pulse signals, a dynamic interaction atlas conforming to the Chinese medicine theory of "three parts and nine signs" is formed. This structure not only completely retains the spatial gradient characteristics of the cun, guan and chi pulse positions, but also more accurately quantifies the conduction correlation of the blood state between pulse positions, and for the first time realizes the mathematical model mapping of the Chinese medicine holistic diagnosis view at the data level.

[0030] 3. The present application analyzes the topological network through the message passing mechanism of the pre-trained graph convolutional neural network, and identifies the coupled pathology pattern of pulse feature and microcirculation parameter by combining the attention matching mechanism of the pathology sub-atlas template library. This process breaks through the limitations of traditional single-dimensional analysis, and forms a collaborative diagnosis basis for cross-modal features such as "pulse channel stagnation" and "capillary granular flow", significantly improving the recognition ability of complex pathological states. The final output of the pathology pattern recognition result provides a scientific basis with spatio-temporal correlation and pathological specificity for the objective diagnosis of Chinese medicine syndromes. BRIEF DESCRIPTION OF DRAWINGS

[0031] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the present application. In the drawings, the same reference numerals are used to refer to the same parts. Among them:

[0032] Figure 1 is a step flowchart of the present application;

[0033] Figure 2 The generation step chart of the comprehensive node feature vector of the present application;

[0034] Figure 3 The system flow chart of the present application;

[0035] Figure 4 The system working schematic of the present application. DETAILED DESCRIPTION

[0036] It is easy to understand that, according to the technical solution of the present application, a person skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following detailed description and the accompanying drawings are only exemplary description of the technical solution of the present application, and should not be regarded as the whole or regarded as the limitation or restriction of the technical solution of the present application.

[0037] SUMMARY

[0038] In the prior art, traditional Chinese medicine pulse diagnosis is mostly dependent on the subjective tactile judgment of doctors, and microcirculation detection is independently carried out by a single optical mode, so it is difficult to realize the spatiotemporal synchronous correlation analysis of pulse and microcirculation. When the patient's limbs are displaced, the positioning of the pulse pressure sensor is inaccurate, resulting in feature drift. In the deep blood vessel detection scene, microcirculation imaging has information loss due to signal attenuation, and there is a millisecond-level time delay between the pulse pressure wave and the capillary blood flow wave, resulting in mismatch of pathological features. Especially in the case of complex pulse such as floating pulse, the existing system cannot simultaneously meet the needs of high time accuracy and multi-source signal spatial registration, which seriously restricts the accurate analysis of pathological mechanisms.

[0039] In order to solve the above problems, the inventors found that the radial styloid process point can be used as an anatomical constant identifier to construct a wrist spatial coordinate reference. Experiments have proved that there is a certain phase difference between the pulse pressure signal feature point and the microcirculation perfusion fluctuation, and the blood vessel depth and the signal attenuation have a nonlinear mapping relationship. Therefore, the core scheme of the "spatiotemporal synchronous framework" is proposed: a spatial coordinate system is established by depth image positioning, and multi-source acquisition is triggered synchronously based on position mapping; in order to solve the time delay problem, an improved DTW compensation algorithm is used to real-time correct the microcirculation acquisition time axis according to the radial artery pulse pressure waveform.

[0040] Specifically, first, the wrist point cloud data is acquired by a 3D depth sensor, the radial styloid process point is automatically identified, and a three-dimensional space coordinate system is established. At the preset 32 space grid points, the following is performed synchronously: 1) millisecond-level pulse pressure time sequence signals are collected based on a MEMS pressure sensor array; 2) a laser speckle contrast imaging technology is used to capture microcirculation blood flow dynamic image sequences. When the preliminary pathological risk assessment probability exceeds a threshold value, the system starts a dynamic time domain alignment mechanism: through an acceleration compensation algorithm, limb displacement interference is eliminated, an improved DTW algorithm is used to calibrate the time delay of multiple source signals, and a time-space synchronization verification parameter δ is generated. The aligned signals are input into a graph convolution network, the pulse image time domain features and microcirculation space features are fused, and after matching a pathological sub-atlas template library, a diagnosis result and a confidence evaluation are output.

[0041] Compared with the prior art, the traditional Chinese medicine detection equipment can only acquire a single dimension of physiological signals, and lacks a multi-modal signal space registration mechanism. The scheme solves the multi-source sensor space registration problem through space coordinate mapping; at the same time, the improved DTW algorithm is used to realize microsecond-level time domain synchronization, and a pulse-microcirculation fusion graph neural network model is constructed. The method establishes a new paradigm for multi-modal dynamic monitoring of objective Chinese pulse diagnosis, and is particularly suitable for early warning scenarios of cardiovascular and cerebrovascular diseases.

[0042] After introducing the basic concept of the application, the embodiments of the application will be specifically introduced below with reference to the drawings.

[0043] Embodiment one:

[0044] As shown in Figure 1 The multi-modal synchronous monitoring method based on Chinese pulse and microcirculation includes the following steps:

[0045] Obtain the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data of the user's radial styloid process point position;

[0046] Calculate the pathological risk probability L from the pulse pressure time sequence signal data, and judge:

[0047] When the pathological risk probability L is greater than or equal to a preset first threshold value, the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data are time domain aligned, and the specific steps are as follows:

[0048] Obtain the motion compensation data of the user;

[0049] Based on the pulse pressure time sequence signal data, the microcirculation blood flow image sequence data and the motion compensation data, calculate the displacement amount of the dynamic time warping path through a path cost function;

[0050] When the displacement amount is greater than a predetermined displacement threshold value, trigger a path correction function to generate a displacement compensation amount;

[0051] The displacement amount and the displacement compensation amount based on the dynamic time warping path are used to obtain a comprehensive node feature vector through time domain alignment processing.

[0052] Based on the correlation measurement of the signals between the nodes based on the comprehensive node feature vector, the edge weight values between the adjacent pulse nodes are calculated, and a graph structure data is constructed.

[0053] The graph structure data is input into a pre-trained graph convolutional neural network model, matched with a pre-set pathological sub-atlas template library, and a pathological pattern recognition result is output.

[0054] The dynamic time domain alignment trigger mechanism is activated when the pathological risk probability L is greater than or equal to a pre-set first threshold value, which is 0.7 in this embodiment. The threshold value is dynamically adjusted according to the user's basic physiological parameters (the threshold value is set by optimizing the clinical ROC curve), such as lowering the threshold value to 0.6 when the age is greater than 60 years old. The acceleration sensor data is used to correct the limb displacement error, and the microcirculation image time sequence is reconstructed by cubic spline interpolation. The calculation formula of the pathological risk probability L is as follows:

[0055]

[0056] , which represents the proportion of the target frequency band energy (0.5-5Hz);

[0057] is the comprehensive score of other physiological factors (such as age, heart rate, metabolic indicators, etc.), with a value range of [0, 1].

[0058] The space-time synchronization verification parameter δ is the time deviation quantization value of the pulse pressure signal and the microcirculation image sequence, which is used to determine the time alignment accuracy of the multi-source data by calculating the waveform similarity using the improved DTW (dynamic time warping) algorithm. Specifically, first, physiological feature points are extracted based on the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data, including the wave peak, wave trough, inflection point in the pulse pressure time sequence signal data, and the extreme value point of blood flow rate change in the microcirculation blood flow image sequence data frame, to ensure one-to-one correspondence in physiological meaning. On this basis, the improved DTW compensation algorithm is used to pair the two modal feature point time sequences, quantify the time deviation, and calculate the space-time synchronization verification parameter δ. The calculation formula of the space-time synchronization verification parameter δ is as follows:

[0059]

[0060] where N is the total number of paired feature points, is the time stamp of the i-th pulse pressure signal feature point, ​The time stamp corresponding to the microcirculation feature point is the space-time synchronization verification parameter δ, which is the average time deviation value finally obtained, with the unit of milliseconds (ms).

[0061] In the matching process of the improved DTW algorithm in this embodiment, an acceleration compensation factor and a window constraint mechanism are introduced to correct the influence of limb micro-movement on time domain registration, thereby improving the accuracy and stability of multi-modal signal alignment.

[0062] In order to ensure the reliability and physiological reasonableness of data synchronization, the system presets a synchronization tolerance threshold to determine the effectiveness of the collected data. The setting of the synchronization tolerance threshold takes into account both the hardware sampling accuracy error and the physiological response delay fluctuation. The hardware error mainly comes from the difference in sampling frequency and synchronization deviation between the piezoelectric sensor and the laser speckle imaging system, and the comprehensive error is about ±10 ms. The physiological response fluctuation mainly considers the natural delay of the pulse wave propagation to the microcirculation blood flow, and the fluctuation range is generally not more than ±5 ms. Based on the above factors, a safety margin of 50% is introduced in the synchronization tolerance setting for weighted adjustment, and the preset synchronization tolerance is finally determined to be ±15 ms. When the calculated space-time synchronization verification parameter δ is greater than or equal to the preset synchronization tolerance threshold, the system will automatically trigger the signal reacquisition instruction to ensure the effectiveness and consistency of the input data for subsequent feature extraction and graph model training.

[0063] Through the introduction of the above-mentioned space-time synchronization verification mechanism, the time domain registration accuracy of the pulse pressure signal and the microcirculation image sequence under dynamic body state is effectively improved, providing high-quality multi-modal input data support for the subsequent node feature construction, edge weight calculation and diagnostic model of graph convolutional neural network.

[0064] Among them, the comprehensive node feature vector refers to the mathematical expression of fusing the time domain pulse parameters and the spatial microcirculation features, which is specifically constructed by extracting the wavelet packet coefficients (frequency range 0.5-20 Hz) of the pulse pressure signal and the LSB (laser speckle contrast) values of the microcirculation image as the input unit of the graph neural network.

[0065] Among them, the edge weight value refers to the physiological correlation strength measurement between adjacent pulse position nodes, which specifically calculates the pulse signal correlation by using mutual information entropy, and generates a 0-1 standardized weight matrix combined with the vascular topological distance.

[0066] Among them, the pathological subgraph template library refers to the pre-trained cardiovascular disease feature graph structure set, which is specifically constructed based on the GCN (graph convolutional network) feature space clustering of clinically diagnosed cases, and contains 128-dimensional feature templates of 8 pathological modes such as coronary heart disease and hypertension.

[0067] The confidence evaluation refers to a reliability verification mechanism of the pathological recognition result, and is specifically calculated by weighted calculation of a Softmax probability distribution and a template matching degree, and a preset confidence threshold is set to 0.85.

[0068] The confidence calculation formula is as follows:

[0069]

[0070] wherein, The model output probability represents a confidence probability of the neural network model for pathological recognition, and the value range is [0, 1]. The atlas similarity represents a matching degree of an actual feature map and a pathological template, and the value range is [0, 1]. The output confidence value is

[0071] The model output probability weight is 60% (a dominant factor), and the atlas similarity weight is 40% (auxiliary verification), and the weights are dynamically configured according to the pathological subatlas type, for example, the coronary heart disease model weight is increased to 70%.

[0072] Through comprehensive confidence calculation, weighted fusion of model prediction and objective atlas verification is realized, and the final output value is in the [0, 1] interval, and the greater the value, the higher the reliability of the diagnosis result. When the preset confidence threshold 0.85 is reached, the pathological pattern recognition result is output as the diagnosis result data.

[0073] The graph structure data refers to a dynamic topological graph composed of pulse position nodes, and the node attribute dimensions include pulse time domain features and microcirculation space features, and the edge weight reflects the physiological connectivity of adjacent pulse position blood vessel transmission paths.

[0074] The core innovation of the present application is to construct a closed-loop diagnosis system based on time-space synchronous verification and multi-modal feature fusion, to solve the data distortion problem caused by limb displacement by dynamically triggering a time domain compensation mechanism, to realize cross-modal correlation analysis of pulse-microcirculation by graph structure modeling, and to break through the limitations of traditional single-dimensional detection of equipment.

[0075] As Figure 2As shown, the working process and principle of the present application are as follows: first, the wrist depth image data is acquired and the radial styloid point position is identified, and a spatial coordinate set is generated based on a preset spatial mapping relationship; the pulse pressure time sequence signal and the microcirculation blood flow image sequence data of the coordinate point are synchronously collected; the pulse pressure signal is subjected to fast Fourier transform to generate a frequency domain feature, and a pathological risk probability L is calculated; when L is greater than or equal to a preset first threshold value (0.7), an improved DTW compensation algorithm is started for time domain alignment, and a time-space synchronization verification parameter δ is generated; if δ exceeds a tolerance value (±15 ms), the data is re-collected, otherwise the pulse condition feature parameters and the microcirculation blood flow feature parameters are extracted to construct a comprehensive node feature vector; the edge weight is calculated based on the mutual information entropy between nodes to form a dynamic graph structure data; the pre-trained graph convolutional neural network is input to match the pathological sub-graph template library; the output result is subjected to confidence evaluation (threshold value 0.85), and if the standard is met, the diagnosis result data is output. Through the closed-loop process, the pathological recognition accuracy under complex conditions is significantly improved.

[0076] The present application further proposes that the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data of the user's radial styloid point position are acquired, and the specific steps of acquiring the user's radial styloid point position include:

[0077] The feature point data is extracted by wrist contour scanning to collect depth image data;

[0078] The feature point data is processed based on a preset radial styloid positioning algorithm to identify and output the radial styloid point position data;

[0079] The positioning error of the radial styloid point position data is not greater than a predetermined error threshold value, and the predetermined error threshold value is 0.8 mm.

[0080] Wrist contour scanning refers to collecting tissue surface geometric information by a three-axis synchronous depth imaging system, specifically using a structure light projection and a binocular infrared imaging cooperative working mode to generate a millimeter-level point cloud (accuracy ±0.2 mm) containing skin surface deformation characteristics. The radial styloid positioning algorithm refers to an intelligent recognition engine that integrates deep learning and anatomical rules, specifically using an improved PointNet++ architecture to integrate prior knowledge of hand bone spatial distribution, and training the network by an anatomical topological constraint loss function to locate key bone landmarks.

[0081] The positioning error control mechanism refers to a closed-loop feedback system for real-time verification of positioning accuracy, specifically using a multi-point stereo calibration verification technology to generate an error distribution heat map, and automatically starting a sub-pixel level position compensation algorithm when the local area error exceeds 0.5 mm.

[0082] Specifically, three sets of depth sensors are used to synchronously project coded structured light at 120° intervals, completing a 270° wrap-around scan of the wrist in 8 milliseconds. The point cloud data obtained from the user's wrist is input into a FPN model trained on 12,000 clinical data. The first network extracts macro-anatomical contour features (such as the ulnar and radial shafts), and the deep network focuses on the geometric mutation features of the styloid process region through an anatomical constraint module. The positioning engine outputs real-time three-dimensional coordinates, while the error compensation system performs stereo calibration on 12 reference points in the scanning area. When a specific quadrant error exceeds the threshold, a secondary positioning process is triggered. The final output coordinate data not only includes spatial positions but also carries confidence and a predetermined error threshold (0.8 mm), fully meeting the accuracy requirements of subsequent spatial mapping.

[0083] Compared with the prior art, the traditional wrist positioning scheme relies on single-view two-dimensional image processing, which is easily affected by soft tissue obstruction, resulting in positioning deviation often exceeding 2 mm; the existing recognition method based on grayscale features has significant error when the skin color difference is large. The present scheme achieves:

[0084] 1) Multi-angle three-dimensional scanning eliminates the problem of visual dead angles, and the integrity of the point cloud data reaches 99.3%;

[0085] 2) The deep learning network fused with anatomical constraints significantly improves the robustness of bone landmark recognition, maintaining an average accuracy of 0.85 mm in the test group with BMI>30;

[0086] 3) The closed-loop error compensation mechanism reduces the positioning failure rate.

[0087] Through the above technical solutions, the present application successfully solves the problem of reference point positioning in wrist dynamic monitoring. Multi-sensor cooperative scanning ensures data integrity from the source, intelligent positioning model realizes accurate analysis of anatomical structures, and real-time error compensation ensures system stability. Clinical verification shows that in dynamic scenarios such as clenched fists and rotations, the positioning accuracy of the radial styloid point is always maintained at 0.68-0.79 mm (n=153), which is an improvement over traditional optical positioning schemes, providing a reliable spatial coordinate system reference for subsequent pulse detection.

[0088] The present application further proposes that after obtaining the position of the user's radial styloid point, it further includes generating and associating a spatial coordinate set based on a preset spatial mapping relationship and the position of the radial styloid point, and the specific steps include:

[0089] Based on the anatomical mapping relationship and the position data of the radial styloid point, a spatial coordinate transformation function is applied to calculate and generate millimeter-level coordinate point data sets corresponding to the inch part, the joint part, and the cubit part, forming a spatial coordinate set data;

[0090] Based on the spatial coordinate set data and the preset soft tissue pressure data, the position is calibrated.

[0091] wherein the spatial coordinate transformation function refers to a mathematical model describing the mapping relationship from the radial styloid point to the pulse diagnosis area, specifically using a rigid body transformation matrix in the Lie group SE(3) space, and solving the rotation and translation parameters (rotation accuracy of 0.05 radian, translation resolution of 0.1 mm) through a pre-trained orthogonal projection network. Wherein the soft tissue pressure data refers to the dynamic distribution parameter representing the biomechanical characteristics of the contact surface, specifically collecting the tissue deformation characteristics (sampling frequency of 200 Hz) based on the pressure sensor array, and inversely calculating the elastic modulus distribution map through the Hertz contact mechanics model. Wherein the position calibration mechanism refers to a real-time feedback system for compensating physiological tissue deformation, specifically using a deformation gradient field and pressure field coupling algorithm to dynamically correct the coordinate depth parameter according to the tissue deformation characteristics (maximum compensation amplitude of ±1.2 mm).

[0092] Specifically, after the system obtains the three-dimensional coordinates of the radial styloid point (accuracy of 0.8 mm), the spatial mapping engine is started immediately: in the first stage, a depth projection network trained based on 2000 CT data is called to generate 32 basic coordinate nodes within a 6 cm³ space around the styloid point; in the second stage, a thin plate spline interpolation algorithm (TPS) is used to map the nodes into the three coordinate sets of the cun part (proximal end), the guan part (directly above the styloid), and the chi part (distal end); in the final stage, a pressure-sensitive array monitors the contact pressure distribution in real time, and when a local pressure gradient exceeding 5 kPa / cm is detected, an elastic modulus correction model is automatically triggered to adjust the coordinate height value.

[0093] Through the above technical solution, the present application effectively solves the positioning misalignment problem caused by anatomical variation and tissue deformation in traditional Chinese medicine pulse diagnosis equipment. The dynamic conversion of anatomical reference lays the foundation for sub-millimeter positioning, the biomechanical feedback system intelligently corrects the contact deformation, and the real-time calibration mechanism ensures the stability of dynamic detection. Clinical verification shows that, under different body positions (sitting, lying, and side-lying), the standard deviation of the cun-guan-chi coordinate is only 0.21 mm (n=210), which is an improvement in accuracy compared to traditional optical projection schemes, and establishes a precise spatial reference framework for subsequent multi-modal data synchronous acquisition.

[0094] The present application further proposes specific steps for obtaining pulse pressure time series signal data and microcirculation blood flow image sequence data and associating spatial coordinate set data, including:

[0095] Controlling the fitting position of the user's wrist based on the spatial coordinate set data.

[0096] Under the condition that the acquisition clock deviation of the pulse pressure time series signal data and the microcirculation blood flow image sequence data meets the condition that the acquisition clock deviation is less than a predetermined time deviation threshold, the pulse pressure time series signal data and the microcirculation blood flow image sequence data are collected in parallel according to the first sampling rate and the second sampling rate, respectively.

[0097] Through the cooperative control mechanism of space-time reference coupling, the physical perception synchronization and spatial accurate positioning integrated collection of pulse condition and microcirculation dual-mode data are realized. The sub-millimeter level spatial coordinates (error <0.8mm) established by the radial styloid process point are taken as the hardware control reference. Under the control of the pressure adaptive contact interface, the dual-channel millisecond level parallel collection is realized through the master-slave type hardware trigger architecture. Finally, the constructed space-time correlation data set meets the standards of spatial positioning accuracy 0.3mm and time synchronization error <2ms, and provides the space-time consistent original input for subsequent pathological analysis.

[0098] Among them, the fitting position control refers to dynamically regulating and controlling the pressure and angle of the detection probe according to the spatial coordinates, and specifically adopts a three-axis micro-distance servo mechanism to cooperate with a pressure feedback closed loop (response frequency 200Hz) to real-time correct the normal contact angle (angle deviation <0.8°) of the probe and the skin surface. Among them, the collection clock deviation control refers to a hardware mechanism that ensures the time alignment of the pulse pressure signal (sampling rate 1kHz) and the microcirculation image (acquisition rate 30fps), and specifically adopts a master clock frequency trigger signal generated by an FPGA to eliminate the cumulative error (clock jitter <10μs) of the crystal oscillator through a double-channel phase-locked loop.

[0099] Specifically, after the user's wrist completes the spatial coordinate positioning: the contact surface control system generates a three-dimensional motion trajectory based on the spatial coordinate set, the three-axis micro-motion platform drives the sensor array to fit the skin surface with constant contact pressure (1.5N±0.2N), while real-time detects the soft tissue deformation characteristics and feeds back the compensation depth coordinates; the hardware trigger circuit generates a reference trigger pulse based on the master clock, the pulse pressure sensor starts 1000Hz high-frequency sampling with the pulse front edge, and at the same time sends a hardware trigger signal to the microcirculation camera through an optocoupler isolation circuit; the microcirculation system starts image acquisition after receiving the hardware trigger with a delay of 15μs (preset compensation optical exposure delay), and each frame of image automatically embeds a time base marker, which shares the same clock source with the pulse pressure signal.

[0100] Through the above technical solutions, the technical bottleneck of spatial positioning drift and multi-source data mismatch in traditional Chinese medicine detection equipment is systematically solved. The spatial coordinates are converted into mechanical control instructions through the servo mechanism, which guarantees the accurate fitting of the sensor from the physical layer; the hardware trigger architecture eliminates the millisecond level delay of traditional software start; the pre-compensation mechanism avoids the inherent time difference of the optical system. Clinical operation verification shows that this scheme can maintain spatial positioning accuracy of 0.35mm (SD±0.06mm) under the condition of slight wrist movement, and the time alignment error of dual-mode data is 1.7ms, which establishes an unreliable reference system for pulse-microcirculation space-time correlation analysis.

[0101] The application further proposes that, based on the pulse pressure time sequence signal data, the microcirculation blood flow image sequence data and the motion compensation data, the displacement amount of the dynamic time warping path is calculated through a path cost function, and the path cost function calculation formula is:

[0102]

[0103] wherein i, j represent time sequence indexes, represents a pulse pressure signal value, represents a microcirculation characteristic value, represents an inter-modal distance, represents a Z-axis acceleration, and a vertical motion weight factor λ = 0.35;

[0104] The time domain drift amount is calculated through time domain alignment processing, and when it is detected that the time domain drift amount in the time domain alignment processing process is greater than a predetermined drift threshold, resampling correction processing is started on the pulse pressure time sequence signal data or the microcirculation blood flow image sequence data through an adaptive window function.

[0105] wherein the motion compensation data refers to a quantitative parameter representing three-dimensional displacement of a limb, and is specifically collected (sampling rate 500 Hz) through a MEMS acceleration sensor worn on a detection wristband, and is output in real time to contain a translation acceleration component separated from a gravity vector. The path cost function refers to a comprehensive distance measurement model fusing signal similarity and motion interference, and a vertical motion weight factor (λ = 0.35) of Z-axis acceleration is introduced on the basis of a standard DTW Euclidean distance, and secondary path optimization is triggered through a displacement amount threshold.

[0106] To determine the optimal weight factor λ, the gradient descent method is used to perform iterative optimization in the interval λ ∈ [0.1, 0.5] in this study. The objective function is defined as the maximum signal mutual information:

[0107]

[0108] wherein and respectively represent the pulse condition characteristics and the microcirculation indicators of the nth sample, and N is the total number of samples.

[0109] The optimization result is shown in Table 1. When λ = 0.35, the signal mutual information reaches a peak value of 0.89, and the microvibration suppression rate (5-15 Hz frequency band) is the highest (91.7%). The gradient descent process converges to the global optimum (loss value < 0.001) after 15 iterations.

[0110] Table 1: Weight factor optimization experiment result

[0111] Lambda value Signal mutual information Microtremor suppression rate (5-15 Hz) 0.25 0.73±0.05 68.2%±3.1% 0.35 0.89±0.03 91.7%±1.8% 0.45 0.81±0.04 83.4%±2.6%

[0112] Note: Data are expressed as mean ± SD.

[0113] Conclusion: The experiment shows that λ = 0.35 is the optimal parameter configuration, at this time the system realizes 91.7% microtremor suppression rate while maintaining high signal mutual information (0.89), which is significantly better than other parameter combinations (p < 0.01).

[0114] Among them, the adaptive window function refers to a dynamic adjustment of signal sampling time domain correction mechanism, specifically using a variable width Hanning window function, when detecting that the time domain drift exceeds the physiological rhythm period (the default threshold is 15% of the heart cycle), start local resampling correction, based on the pulse pressure signal automatically calculate real-time heart cycle.

[0115] Specifically, the time domain alignment engine works in three stages:

[0116] 1) The preprocessing stage fuses the pulse pressure signal (1000Hz), microcirculation image feature sequence (30fps) and motion compensation data, and calculates the signal distance matrix at each time point;

[0117] 2) The dynamic regularization stage calculates the optimal bending path by improving the path cost function, and activates the acceleration compensation algorithm when the path displacement exceeds the preset threshold (dynamically adjusted according to the heart cycle), to generate a displacement compensation vector to correct the regularization path;

[0118] 3) The post-processing stage calculates the time domain drift (standard deviation of adjacent peak spacing), triggers the window function to perform local interpolation resampling on the pulse pressure signal, and eliminates the cumulative phase error. The whole process compresses the time deviation to the microsecond level while maintaining the integrity of the signal physiological characteristics.

[0119] Through the above technical solutions, the application systematically solves the problem of time domain mismatch of multi-modal signals in dynamic monitoring. Acceleration sensing data provides physical motion reference for time regularization, weight factor accurately quantifies the intensity of vertical interference, and window function mechanism maintains the physiological feature fidelity of the resampling process. Clinical verification shows that the processing flow makes the time domain alignment error stable within 1% of the heart cycle, effectively overcoming the significant phase deviation produced by traditional methods when the body position changes.

[0120] The application further proposes that when the displacement is greater than a predetermined displacement threshold, a displacement compensation amount is generated by triggering a path correction function, and the path correction function calculation formula is:

[0121]

[0122] Among them, is the displacement compensation amount at time t, representing the velocity offset accumulated by acceleration; is the vertical motion weight factor, ; and Z-axis acceleration; Signal start time; t is the current time;

[0123] If >0.5m / s, pause displacement compensation and trigger motion interference alarm;

[0124] To 5-15Hz band-pass filter to eliminate non-physiological high-frequency noise.

[0125] The application realizes path correction closed-loop compensation, breaks through the limitation of traditional DTW sensitive to limb micro-motion, and provides mathematical guarantee for time domain alignment accuracy δ≤1ms.

[0126] The application further proposes that based on the displacement amount and displacement compensation amount of the dynamic time warping path, a comprehensive node feature vector is obtained through time domain alignment processing, and the comprehensive node feature vector specifically includes:

[0127] Pulse condition feature parameters extracted from pulse pressure time sequence signal data, including pulse pressure waveform entropy value Data and rhythm variation coefficient data;

[0128] Microcirculation blood flow feature parameters extracted from microcirculation blood flow image sequence data, including blood flow average flow rate Data, blood flow perfusion unit PU data, and LBP-TOP blood flow texture feature value data;

[0129] The pulse pressure waveform entropy value Data and rhythm variation coefficient data form a 32-dimensional pulse pressure time sequence feature vector;

[0130] The blood flow average flow rate Data, blood flow perfusion unit PU data, and LBP-TOP blood flow texture feature value data form a 64-dimensional microcirculation blood flow feature vector;

[0131] Merge the 32-dimensional pulse pressure time sequence feature vector and the 64-dimensional microcirculation blood flow feature vector, and generate comprehensive node feature vector data based on the displacement amount and displacement compensation amount of the dynamic time warping path.

[0132] Wherein, the hardware level clock synchronization circuit (nanosecond level clock source based on PTP protocol) triggers the synchronous acquisition of the pulse pressure sensor and the microcirculation imaging device, and eliminates the device start-up delay difference; independent DMA channels are allocated for the dual-mode data stream, and the delay fluctuation caused by bus competition is avoided through parallel transmission mechanism. The global timestamp marking module attaches a unified time reference (accuracy ±1ms) to each data packet, ensuring that the pulse pressure peak time strictly corresponds to the capillary perfusion peak, and the physiological synchronicity of pulsation and blood flow in the traditional Chinese medicine theory of "pulsation and blood homology".

[0133] Two types of key pathological indicators are extracted from the aligned pulse pressure signals. The pulse pressure waveform entropy quantifies the chaotic degree of pulsation energy distribution (formula: ), reflecting the degree of vascular elasticity degeneration); the rhythm variation coefficient calculates the ratio of the standard deviation to the mean of adjacent pulse periods ( ), representing the autonomic nervous regulation function; the above parameters constitute a 32-dimensional pulse pressure time sequence feature vector, which captures the dynamic evolution law of "string pulse to sliding pulse" through sliding window analysis.

[0134] From the spatiotemporal aligned microcirculation blood flow image sequence, the microcirculation blood flow feature parameters are extracted: the average flow velocity and the perfusion unit (PU) quantify the capillary level blood flow intensity; the spatiotemporal texture feature: the LBP-TOP operator is used to extract the texture pattern of blood flow direction and velocity distribution, and to identify "stagnation-excitation" abnormalities; the above parameters are fused to generate a 64-dimensional microcirculation feature vector, realizing the holographic description from macro-perfusion to micro-flow state.

[0135] The 32-dimensional pulse pressure vector and the 64-dimensional microcirculation vector are spliced into a 96-dimensional comprehensive node feature vector, each dimension corresponding to a specific pathological sensitive indicator. This vector serves as the embedded representation of the graph node, providing atomized input for the subsequent construction of the pulse position topological network.

[0136] Through the above technical solutions, the cross-modal time alignment error is ≤1ms, and the signal phase shift caused by wrist micro-vibration is completely eliminated. The time difference between the pulse pressure wave peak and the microcirculation perfusion peak is controlled within the physiological synchronization range, providing a distortionless data basis for feature extraction. The pulse pressure waveform entropy sensitivity to the early vascular stiffness changes of arteriosclerosis is improved; the microcirculation LBP-TOP feature can identify capillary cluster abnormalities with a diameter <100μm, breaking through the traditional imaging resolution limit. The two are complementarily enhanced through vector splicing, improving the detection rate of complex pathological patterns such as "string pulse with foot blood flow stagnation". Through the three-order optimization of hardware synchronization guarantee, feature depth extraction, and vector cross-modal fusion, the system constructs a comprehensive node feature vector with spatiotemporal consistency and pathological specificity. This vector serves as the basic unit of the pulse position topological network, driving the graph convolution model to realize the precise analysis of the dynamic correlation between "pulse condition" and "microcirculation" in traditional Chinese medicine, and providing a quantifiable scientific basis for early diagnosis of arteriosclerosis.

[0137] The present application further proposes that based on the correlation measure of the signals between nodes of the comprehensive node feature vector, the edge weight value between adjacent pulse position nodes is calculated, and the specific steps of constructing the graph structure data include:

[0138] For each pair of adjacent pulse position comprehensive node feature vector data, the mutual information value between the comprehensive node feature vector data is calculated;

[0139] The mutual information value is normalized standardization calculation is performed;

[0140] wherein, represents an edge weight value, and Vi represents a pulse pressure time sequence feature vector, represents a microcirculation blood flow feature vector, represents an entropy, and represents uncertainty of a comprehensive node feature vector;

[0141] The mutual information value obtained by calculation is taken as weight data of the edge weight value;

[0142] Based on all nodes and corresponding comprehensive node feature vector data and all adjacent node pairs and corresponding edge weight value weight data, graph structure data is constructed.

[0143] For adjacent pulse position nodes (such as the inch part and the joint part) after time domain alignment, the system calculates the standardization mutual information value of the 96-dimensional feature vector (including 32-dimensional pulse pressure time sequence features and 64-dimensional microcirculation blood flow features) thereof:

[0144]

[0145] wherein, represents pulse pressure time sequence features (pulse pressure waveform entropy value , rhythm variation coefficient CVRR), represents microcirculation blood flow features (average flow rate , perfusion unit PU, LBP-TOP texture), represents, and measures statistical dependence of two pulse position features, is an information entropy function. The formula eliminates the influence of feature dimension difference on correlation calculation, so that the edge weight objectively reflects the physiological correlation strength between pulse positions.

[0146] The mutual information value is taken as an edge weight connecting adjacent pulse position nodes, and a weighted undirected graph is constructed:

[0147] Node set V: high-dimensional feature vectors corresponding to inch, joint, and chi pulse positions;

[0148] Edge set E: anatomical connections between adjacent pulse positions (inch-joint, joint-chi);

[0149] Weight set W: weight matrix composed of standardization mutual information values ;

[0150] This structure completely reproduces the traditional Chinese medicine “pulse network view”, and upgrades the traditional discrete pulse position to a dynamic interactive system. After the generated graph data G is input into a pre-trained graph convolutional network (GCN), a message passing mechanism is performed along high-weight edges (edges with large mutual information values): >0.6), the attention layer strengthens the diagnostic weight of the abnormal correlation path such as "inch part string pulse- foot part stasis", and the output layer matches the pathological sub-atlas template library to identify the early arteriosclerosis "pulse path astringency-microcirculation disorder" coupling mode.

[0151] Mutual information edge weight The classical theory of "the inch pulse being the end and beginning of the five internal organs and six bowels" is quantified. Clinical verification shows that the mutual information value of the Guan-Foot part of healthy people is =0.82±0.05, and that of early arteriosclerosis patients is reduced to , objectively revealing the pathological correlation of "weak foot pulse indicating kidney qi deficiency".

[0152] When the inch part appears high pulse pressure waveform entropy value ( , a sign of vascular rigidity) and the mutual information value of Guan-Foot part is , the system automatically triggers the "pulse condition-microcirculation decoupling" warning, which improves the early detection rate of diabetic microangiopathy.

[0153] High weight edge ( ) guides the GCN to focus on the key pathological path. When the foot microcirculation PU value is abnormal, the GCN strengthens the attention to the pulse pressure coefficient of variation of the inch part through the Guan-Foot edge weight , avoiding misjudgment caused by isolated features.

[0154] The pulse position topological network constructed by the mutual information edge weight converts the TCM holistic view into a computable dynamic graph model. The model and the hardware-level synchronization mechanism and high-dimensional feature vector depth synergy make the physiological interaction relationship of the "inch, Guan and foot" three pulse positions first quantified, providing a scientific basis for early warning of arteriosclerosis and other diseases based on group pathological atlas.

[0155] The application further proposes that the specific steps of inputting the graph structure data into the pre-trained graph convolutional neural network model and outputting the pathological pattern recognition result by matching the pre-set pathological sub-atlas template library include:

[0156] The graph convolutional neural network model updates the node state data in the graph structure data based on the message passing mechanism;

[0157] During the updating process, attention weight calculation is performed based on the current comprehensive node feature vector data and the data in the pathological sub-atlas template library;

[0158] After the graph convolutional neural network model is iteratively updated based on the attention weight, the matching operation with the template data in the pathological sub-atlas template library is performed in combination with the calculated attention weight and the final node state data, and the pathological pattern recognition result data is output.

[0159] The three-layer graph convolution network is used to process the structure and fuse the dual-channel attention mechanism.

[0160]

[0161] wherein Q = node feature, K = pathological sub-atlas template, represents the dimension of the pathological sub-atlas template vector.

[0162] By matching the pre-stored 12 types of pathological sub-atlas (such as “slippery vein-high perfusion” and “tense vein-particle flow stop”), the pathological pattern recognition result is output.

[0163] The graph convolution operation adopts a message passing mechanism.

[0164]

[0165] wherein, In general, in the formula related to neural networks, h is often used to represent the hidden state. The superscripts ( +1) and ( ) represent the hidden state of the neural network at the +1 layer and the layer, respectively. is the hidden state of the i-th node in the +1 layer is the hidden state of the j-th node in the layer;

[0166] represents the connection weight from the j-th node of the layer to the i-th node of the layer. It determines the importance degree of information transmission from the previous layer to the current layer;

[0167] is a ReLU activation function, which is used to introduce a non-linear factor to the neuron, so that the neural network can approximate any non-linear function;

[0168] represents the bias parameter of the layer. The bias parameter is used to adjust the activation threshold of the neuron, so that the neuron can have a certain output even without input;

[0169] represents the degree of node i and node j, which is used to measure the connection number and other information of the node. Here they are used for normalization and other operations on the weight to adjust the intensity of information transmission.

[0170] The graph convolution iteratively updates the node state through a three-layer network, and the final output layer combines the pathological subgraph attention weight to generate a diagnosis result.

[0171] The graph convolutional neural network model updates the node state data in the graph structure data based on a message passing mechanism. In this process, the model uses the message passing mechanism to pass information between the nodes of the graph, updates the node state data according to the connection relationship and related weight between the nodes, so that it can reflect the internal characteristics and relationship changes of the graph structure data.

[0172] During the update process, attention weight calculation is performed based on the current comprehensive node feature vector data and the data in the pathological subgraph template library. Here, the relevance of the current comprehensive node feature vector and each type of template data in the pathological subgraph template library is calculated to determine the attention weight of different templates for the current node, so as to highlight important features and associations.

[0173] The graph convolutional neural network model iteratively updates based on the attention weight, combines the calculated attention weight with the final node state data, and performs matching operations with the template data in the pathological subgraph template library to output pathological pattern recognition result data. In this process, the model continuously adjusts and optimizes the node state data according to the calculated attention weight, and matches and compares with each type of template in the pathological subgraph template library, and finally outputs the corresponding pathological pattern recognition result.

[0174] Specifically, a three-layer graph convolutional network is used to process the structure, and a double-channel attention mechanism is used to more accurately calculate the attention weight. The graph convolution operation uses a message passing mechanism to iteratively update the new node state through a three-layer network, and the final output layer combines the pathological subgraph attention weight to generate a diagnosis result. Through the overall process and the interaction of each part, the graph convolutional neural network model can effectively utilize the graph structure data and the pathological subgraph template library and other information, accurately perform pathological pattern recognition and output reliable results, and provide strong support for related medical diagnosis.

[0175] The present application further proposes that the pre-training process of the graph convolutional neural network model is as follows:

[0176] Inject a triple contrast enhancement mechanism, which includes:

[0177] Add a preset amplitude of Gaussian noise data to the training sample data belonging to the same class to construct a positive sample pair , );

[0178] Randomly select training sample data of different classes to construct a negative sample );

[0179] based on positive sample pairs ( , ) and negative samples ( ), a triplet contrast loss function is used to optimize the parameters of the graph convolutional neural network model, and the triplet contrast loss function is set with a boundary distance threshold, which is used to force the graph convolutional neural network model to expand the feature distance between different pathological patterns.

[0180] The triplet contrast loss function is defined as:

[0181]

[0182] wherein the boundary distance threshold a = 0.4.

[0183] The present application has been clinically verified in three institutions including Liaoning University of Traditional Chinese Medicine Affiliated Hospital, and the data set construction aims to cover the statistical significance of pulse- microcirculation correlation analysis and verify the pathological subgraph recognition ability. The total number of samples is 350, including 210 males and 140 females. The age range of the subjects is 25 to 75 years old, with an average age of 48.6 ± 11.3 years (standard deviation), which covers the high-risk age group of arteriosclerosis, ensuring sufficient characterization of related pathological mechanisms. The disease types are divided into three groups: early arteriosclerosis group (98 cases), diabetic microcirculation disorder group (85 cases) and healthy control group (167 cases), to systematically evaluate the recognition ability of pathological subgraph. The data collection period is from March 2022 to June 2023, and each sample is collected by three repeated sampling and averaging to eliminate temporal variation and ensure data stability.

[0184] Table 2: Parameter distribution of the data set and its technical relevance

[0185] Parameter Data distribution Technical relevance Total sample amount 350 cases (210 male, 140 female) Statistical significance of the correlation analysis between pulse and microcirculation Age distribution 25-75 years old (mean 48.6±11.3 years old) Including the age group with high incidence of arteriosclerosis Disease type ① Early arteriosclerosis (98 cases) ② Diabetic microcirculation disorder (85 cases) ③ Healthy control group (167 cases) Verification of pathological subgraph recognition ability

[0186] The present application further proposes to inject a triplet contrast enhancement mechanism in the pre-training process of the graph convolutional neural network model, so as to optimize the model parameters and improve its discrimination ability between different pathological patterns.

[0187] Among them, the specific operation of the triplet contrast enhancement mechanism is as follows: first, add pre-set amplitude Gaussian noise data to the training sample data belonging to the same class, so as to construct positive sample pairs ( , ), which can increase the diversity of samples of the same class, so that the model can better learn the common features of samples of the same class while adapting to a certain degree of noise interference. Second, randomly select training sample data of different classes to construct negative samples ( ), by introducing negative samples, the model can clearly distinguish the sample features of different categories. Then, based on the positive sample pairs ( , ) and negative samples (negative samples ), and the model parameters of the graph convolutional neural network are optimized by using a triplet contrast loss function. In this process, the triplet contrast loss function is set with a boundary distance threshold, and the value of the boundary distance threshold is set to 0.4. The boundary distance threshold is used to force the graph convolutional neural network model to expand the feature distance between different pathological patterns, so that the model can more clearly distinguish the sample features of different pathological patterns. Through the calculation and optimization of the triplet contrast loss function, the model can continuously adjust its parameters to minimize the distance between samples of the same class and expand the distance between samples of different classes.

[0188] Through the above technical solutions, the application effectively solves the problem that the graph convolutional neural network model has insufficient ability to distinguish between different pathological patterns. The way of constructing positive sample pairs in the triplet contrast enhancement mechanism increases the richness of samples of the same class, and the operation of constructing negative samples enables the model to better distinguish different classes. The process of optimizing the model parameters based on the specific boundary distance threshold of the triplet contrast loss function enables the model to expand the feature distance between different pathological patterns, and finally enables the model to significantly improve the accuracy in distinguishing different pathological patterns, and more accurately identify various pathological patterns.

[0189] The following is a specific embodiment of the above multi-modal synchronous monitoring method based on TCM pulse and microcirculation:

[0190] The specific implementation process of a patient (58 years old, with a history of hypertension for 7 years, BMI 26.3) receiving monitoring in the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine is as follows:

[0191] The system performs a 270° wrap-around scan on the wrist by using three groups of depth sensors, and generates point cloud data with millimeter-level precision (±0.2mm) within 8 milliseconds. The data is input into the FPN model trained by 12,000 clinical data, and the anatomical constraint module is used to identify the radial styloid point, and output the three-dimensional coordinates (32.1mm, 15.4mm, 8.3mm). After 12 reference point stereo calibration verification, the maximum positioning error is 0.5mm (lower than the threshold of 0.8mm).

[0192] Based on the radial styloid point coordinates, the SE(3) rigid transformation matrix (rotation accuracy 0.05 radian, translation resolution 0.1mm) is used to generate the coordinates of the cun part (30.2mm, 14.8mm, 6.5mm), the guan part (32.1mm, 15.4mm, 8.3mm), and the chi part (34.0mm, 16.1mm, 9.7mm). The pressure gradient detected by the pressure-sensitive array is 3.8 kPa / cm (lower than the calibration threshold of 5kPa / cm), and the depth calibration is not started.

[0193] The pulse pressure sensor (1000 Hz sampling rate) collects a 10-second time series (10,000 sampling points), and the microcirculation camera (30 fps) captures 300 frames of blood flow images (resolution 1024x768). The actual clock skew is 8.2 μs (less than the 10 μs threshold).

[0194] The pulse pressure signal is analyzed by fast Fourier transform, and the energy ratio in the 1-5 Hz frequency band is only 48% (normal range 55%-70%). The calculated pathological risk probability L=0.72 (exceeding the threshold value 0.7) triggers the time domain alignment process.

[0195] The time domain alignment engine fuses the Z-axis micro-vibration data detected by the MEMS accelerometer (amplitude 80 μm, frequency 8 Hz), and calculates the path cost:

[0196]

[0197] The detected displacement D=7 steps (exceeding the threshold value =5 steps). The path correction function is activated: The calculated displacement compensation Δ(t)=0.18 m / s (corresponding to 80 μm micro-vibration).

[0198] Time domain index correction Three times of spline interpolation are performed on the microcirculation sequence to generate alignment feature values .

[0199] The system then extracts the feature parameters of the time domain alignment data: pulse pressure waveform entropy (vessel elasticity degeneration marker, normal <3.8), rhythm variation coefficient CVRR=18% (autonomic nervous function abnormality, normal <15%), average flow velocity of blood flow =1.2 mm / s (capillary perfusion insufficient, normal >1.5 mm / s), LBP-TOP blood flow texture feature value 0.63 (blood flow stasis, normal <0.50). These parameters constitute a 32-dimensional pulse pressure time sequence feature vector and a 64-dimensional microcirculation blood flow feature vector, which are combined into a 96-dimensional comprehensive node feature vector.

[0200] Based on the comprehensive node feature vector, the system calculates the standardized mutual information value of adjacent pulse positions:

[0201] The measured Cun-Guan department weight is 0.41, and the Guan-Chi department weight is 0.38 (normal value >0.6), reflecting that the physiological correlation between pulse positions is significantly weakened. The constructed graph structure data is input into the pre-trained graph convolution network, and the node state is updated through a three-layer message passing mechanism:

[0202] Combined with attention weight calculation ) The pathological sub-pattern template library is matched, and the highest matching degree is the "early arteriosclerosis" template (similarity 0.89). After verifying the triple comparison loss (positive sample distance 0.31, negative sample distance 1.02> boundary threshold 0.71), the final confidence evaluation value is:

[0203]

[0204] The output diagnostic report is "early arteriosclerosis (confidence 90.8%)", and the core evidence includes the increase of the entropy value of the pulse pressure, the decrease of the mutual information of the Guan-Shi part, and the blood flow stasis of the Shi part. The subsequent coronary CTA examination verifies that the left anterior descending branch is narrowed by 35%, which is consistent with the system prediction. Clinical data show that the positioning accuracy of the system in 153 dynamic tests is 0.68-0.79 mm, and the time domain alignment error is stable within 1% of the cardiac cycle, which is improved by 62% compared with the traditional optical scheme.

[0205] Embodiment Two

[0206] As shown in Figure 3 and Figure 4 , a multi-modal synchronous monitoring system based on traditional Chinese medicine pulse and microcirculation uses the multi-modal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation described above, comprising:

[0207] A collection module is used to acquire pulse pressure time series signal data and microcirculation blood flow image sequence data of the user's radial styloid process point position;

[0208] A risk judgment module is used to calculate the pathological risk probability L by calculating the pathological risk of the pulse pressure time series signal data and to judge;

[0209] A signal synchronization module is used to perform time domain alignment on the pulse pressure time series signal data and the microcirculation blood flow image sequence data when the pathological risk probability L is greater than or equal to a preset first threshold, and the specific steps are as follows:

[0210] Obtain the motion compensation data of the user;

[0211] Based on the pulse pressure time series signal data, the microcirculation blood flow image sequence data, and the motion compensation data, the displacement amount of the dynamic time warping path is calculated through the path cost function;

[0212] When the displacement amount is greater than a predetermined displacement threshold, trigger the path correction function to generate a displacement compensation amount;

[0213] Based on the displacement amount of the dynamic time warping path and the displacement compensation amount, the comprehensive node feature vector is obtained through time domain alignment processing;

[0214] The atlas modeling module is configured to calculate edge weight values between adjacent pulse position nodes based on a correlation measurement of inter-node signals of the comprehensive node feature vector, and construct graph structure data based on the edge weight values.

[0215] The pathology recognition module is configured to input the graph structure data into a pre-trained graph convolutional neural network model, and match the graph structure data with a pre-set pathology sub-atlas template library, and output a pathology pattern recognition result.

[0216] The system takes the acquisition module as a data inlet, synchronously captures the pulse pressure time sequence signal and the microcirculation blood flow image sequence at the radial styloid process position, and ensures the spatial consistency of the original signal. The risk judgment module performs real-time pathological risk probability calculation on the pulse pressure signal, and activates a high-precision alignment process when the probability value exceeds a pre-set threshold value. At this time, the signal synchronization module intervenes, fuses motion compensation data and a dynamic time warping algorithm, quantifies the displacement deviation of the multi-modal signals through a path cost function, and triggers a path correction function to generate a physical compensation amount when the displacement exceeds a limit, thereby completely eliminating the physiological phase misalignment caused by wrist micro-vibration.

[0217] The signal processed in the time domain is input into the atlas modeling module. The atlas modeling module fuses cross-modal features such as pulse pressure waveform entropy and blood flow texture into a comprehensive node vector, dynamically calculates edge weights based on the mutual information entropy of adjacent pulse signals, and constructs a pulse condition topological network in line with the "three parts and nine indications" theory of traditional Chinese medicine. Finally, the pathology recognition module drives the graph convolutional neural network to analyze the network structure, fuses the node state and the attention weight of the pathology sub-atlas template library through a message passing mechanism, and outputs a pathology pattern recognition result with spatiotemporal correlation.

[0218] Compared with the fragmented analysis of pulse conditions and microcirculation in the traditional scheme, the system realizes dynamic correlation verification of "pulse condition pulsation-microcirculation perfusion" at the hardware level, and completes pathological network modeling under the guidance of the holistic view of traditional Chinese medicine at the algorithm level, thereby providing a cross-modal collaborative diagnosis paradigm for early warning of cardiovascular and cerebrovascular diseases. The atlas modeling module calculates edge weight values between adjacent pulse position nodes based on a correlation measurement of inter-node signals of the comprehensive node feature vector, and constructs graph structure data based on the edge weight values, thereby providing a suitable data structure form for subsequent pathology recognition. The pathology recognition module inputs the graph structure data into a pre-trained graph convolutional neural network model, matches the graph structure data with a pre-set pathology sub-atlas template library, and outputs a pathology pattern recognition result.

[0219] Through the above technical solutions, the present application solves the problem of inaccurate multi-modal data acquisition position of traditional Chinese medicine pulse conditions and microcirculation, realizes multi-source data acquisition based on accurate position, overcomes the defect of large time domain synchronization error of multi-modal data, achieves high-precision time domain alignment, improves the accuracy of pathology recognition, can more effectively integrate multi-modal data for pathology analysis, and ensures the reliability of the output diagnosis result, thereby providing strong support for medical diagnosis through confidence evaluation.

[0220] The technical scope of the present application is not limited to the above-described embodiments, and various modifications and changes can be made to the above-described embodiments without departing from the technical idea of the present application, and these modifications and changes should be included in the scope of the present application.

Claims

1. A multi-modal synchronous monitoring method based on traditional Chinese medicine pulse condition and microcirculation, characterized in that: The method comprises the following steps: obtaining pulse pressure time sequence signal data and microcirculation blood flow image sequence data of a user's radial styloid process point position; calculating a pathological risk probability L from the pathological risk of the pulse pressure time sequence signal data and determining: when the pathological risk probability L is greater than or equal to a preset first threshold, aligning the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data in the time domain, and the specific steps are as follows: obtaining motion compensation data of the user; calculating the displacement of the dynamic time warping path based on the pulse pressure time sequence signal data, the microcirculation blood flow image sequence data, and the motion compensation data; the path cost function calculation formula is: ； wherein i, j represent time series indexes, represents the pulse pressure signal value, represents the microcirculation characteristic value, represents the inter-modal distance, represents the Z-axis acceleration, the vertical motion weight factor λ = 0.35; calculating the time domain drift through time domain alignment processing, and when it is detected that the time domain drift in the time domain alignment processing process is greater than a predetermined drift threshold, starting the adaptive window function to resample and correct the pulse pressure time sequence signal data or the microcirculation blood flow image sequence data; when the displacement is greater than a predetermined displacement threshold, triggering a path correction function to generate a displacement compensation amount; the path correction function calculation formula is: ； wherein, is the displacement compensation amount at time t, representing the velocity offset accumulated by the acceleration; is the vertical motion weight factor, ; is the Z-axis acceleration; is the signal start time; t is the current time; If > 0.5 m / s, pause displacement compensation and trigger motion disturbance alarm; To 5-15Hz band-pass filter to eliminate non-physiological high-frequency noise; obtaining a comprehensive node feature vector through time domain alignment processing based on the displacement of the dynamic time warping path and the displacement compensation amount; wherein the comprehensive node feature vector specifically comprises: Pulse condition characteristic parameters extracted from the pulse pressure time series signal data, the pulse condition characteristic parameters including pulse pressure waveform entropy values data and rhythm variation coefficient data; microcirculation blood flow feature parameters extracted from the microcirculation blood flow image sequence data, the microcirculation blood flow feature parameters including blood flow average flow velocity data, blood flow perfusion unit PU data, and LBP-TOP blood flow texture feature value data; said pulse pressure waveform entropy value data and said rhythm variability coefficient data form a 32-dimensional pulse pressure time series feature vector; averaging the blood flow velocity The data, the blood flow perfusion unit PU data and the LBP-TOP blood flow texture feature value data constitute a 64-dimensional microcirculation blood flow feature vector. merging the 32-dimensional pulse pressure time sequence feature vector and the 64-dimensional microcirculation blood flow feature vector, and generating the comprehensive node feature vector data based on the displacement of the dynamic time warping path and the displacement compensation amount; based on the correlation measurement of the nodes of the comprehensive node feature vector, calculating the edge weight value between adjacent pulse position nodes, and constructing graph structure data; The constructing the graph structure data comprises: calculating mutual information values between the integrated node feature vector data of each pair of adjacent pulses ; The mutual information value is calculated by normalization; wherein, denotes an edge weight value, V i denotes the pulse pressure time series feature vector, denotes the microcirculatory blood flow feature vector, denotes an entropy, characterizing the uncertainty of the integrated node feature vector; using the calculated mutual information value as the weight data of the edge weight value; based on all nodes and corresponding comprehensive node feature vector data and all adjacent node pairs and corresponding edge weight value weight data, constructing the graph structure data; inputting the graph structure data into a pre-trained graph convolutional neural network model, matching with a preset pathological sub-atlas template library, and outputting a pathological pattern recognition result.

2. The multi-modal synchronous monitoring method based on pulse conditions and microcirculation according to claim 1, characterized in that: Obtaining pulse pressure time sequence signal data and microcirculation blood flow image sequence data of a user's radial styloid process point position, the specific steps of obtaining the user's radial styloid process point position include: extracting feature point data by wrist contour scanning and collecting depth image data; processing the feature point data based on a preset radial styloid process positioning algorithm, identifying and outputting the radial styloid process point position data; wherein the positioning error of the radial styloid process point position data is not greater than a predetermined error threshold, and the predetermined error threshold is 0.8mm.

3. The multi-modal synchronous monitoring method based on pulse conditions and microcirculation according to claim 1, characterized in that: After obtaining the radial styloid process point position of the user, it further includes generating and associating a spatial coordinate set based on a preset spatial mapping relationship and the radial styloid process point position, and the specific steps include: based on the anatomical mapping relationship and the radial styloid process point position data, applying a spatial coordinate transformation function to calculate and generate millimeter-level coordinate point data sets corresponding to the inch part, the joint part, and the foot part, forming spatial coordinate set data; Perform position calibration based on the spatial coordinate set data and preset soft tissue pressure data; Perform fitting position control of the user's wrist based on the spatial coordinate set data; When the acquisition clock deviation of the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data is less than a predetermined time deviation threshold, trigger the acquisition of the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data according to the first sampling rate and the second sampling rate respectively.

4. The multi-modal synchronous monitoring method based on pulse conditions and microcirculation according to claim 1, characterized in that: The specific steps of inputting the graph structure data into the pre-trained graph convolutional neural network model and outputting the pathological pattern recognition result by matching the preset pathological sub-atlas template library include: The graph convolutional neural network model updates the node state data in the graph structure data based on a message passing mechanism; During the updating process, attention weight calculation is performed based on the current comprehensive node feature vector data and the data in the pathological sub-atlas template library; After the graph convolutional neural network model is iteratively updated based on the attention weight, the matching operation with the template data in the pathological sub-atlas template library is performed by combining the calculated attention weight and the final node state data, and the pathological pattern recognition result data is output.

5. The multi-modal synchronous monitoring method based on pulse conditions and microcirculation according to claim 1, characterized in that: The pre-training process of the graph convolutional neural network model is as follows: Inject a triple contrast enhancement mechanism, which includes: Add the preset amplitude of Gaussian noise data to the training sample data belonging to the same class to construct a positive sample pair , ); Randomly select different types of training sample data to construct negative samples ); based on the positive sample pair ( , ) and the negative sample pair ( ), a triplet contrast loss function is used to optimize the parameters of the graph convolutional neural network model, and the triplet contrast loss function is provided with a boundary distance threshold value α, and the boundary distance threshold value α is used to force the graph convolutional neural network model to expand the feature distance between different pathological patterns. The triple contrast loss function is defined as: ； Wherein, the boundary distance threshold α = 0.

4.

6. A multi-modal synchronous monitoring system based on pulse conditions of traditional Chinese medicine and microcirculation, characterized in that: Using the multi-modal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation as claimed in any one of claims 1-5, comprising: The acquisition module is used to acquire the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data of the user's radial styloid process point position; The risk judgment module is used to calculate the pathological risk probability L by performing pathological risk calculation on the pulse pressure time sequence signal data and to judge; The signal synchronization module is used to perform time domain alignment on the pulse pressure time sequence signal data and the microcirculation blood flow image sequence data when the pathological risk probability L is greater than or equal to a preset first threshold, and the specific steps are as follows: Obtain the motion compensation data of the user; Based on the pulse pressure time sequence signal data, the microcirculation blood flow image sequence data, and the motion compensation data, calculate the displacement of the dynamic time warping path through a path cost function; the path cost function calculation formula is: ； where i, j represent time series indexes, represents the pulse pressure signal value, represents the microcirculation characteristic value, represents the inter-modal distance, represents the Z-axis acceleration, the vertical motion weight factor λ = 0.35; Calculate the time domain drift amount through time domain alignment processing. When it is detected that the time domain drift amount is greater than a predetermined drift threshold during the time domain alignment processing, start the adaptive window function to perform resampling correction processing on the pulse pressure time sequence signal data or the microcirculation blood flow image sequence data; When the displacement is greater than a predetermined displacement threshold, trigger the path correction function to generate a displacement compensation amount; the path correction function calculation formula is: ； wherein, is the displacement compensation amount at time t, representing the velocity offset accumulated by the acceleration; is the vertical motion weight factor, ; is the Z-axis acceleration; is the signal start time; t is the current time; If > 0.5 m / s, pause displacement compensation and trigger motion disturbance alarm; For 5-15Hz band-pass filtering to eliminate non-physiological high-frequency noise; Based on the displacement of the dynamic time warping path and the displacement compensation amount, obtain the comprehensive node feature vector through time domain alignment processing; The comprehensive node feature vector specifically includes: Pulse condition characteristic parameters extracted from the pulse pressure time series signal data, the pulse condition characteristic parameters including pulse pressure waveform entropy values data and rhythm variation coefficient data; microcirculation blood flow feature parameters extracted from the microcirculation blood flow image sequence data, the microcirculation blood flow feature parameters including blood flow average flow velocity data, blood perfusion unit (PU) data, and LBP-TOP blood flow texture feature value data; said pulse pressure waveform entropy value data and said rhythm variability coefficient data form a 32-dimensional pulse pressure time series feature vector; averaging the blood flow velocity The data, the blood flow perfusion unit PU data and the LBP-TOP blood flow texture feature value data constitute a 64-dimensional microcirculation blood flow feature vector. merge the 32-dimensional pulse pressure timing feature vector and the 64-dimensional microcirculation blood flow feature vector, and generate the comprehensive node feature vector data based on the displacement amount of the dynamic time warping path and the displacement compensation amount; a graph modeling module configured to calculate edge weight values between adjacent pulse position nodes based on a correlation measure of inter-node signals of the comprehensive node feature vectors, and construct graph structure data based on the edge weight values; The constructing the graph structure data comprises: calculating mutual information values between the integrated node feature vector data of each pair of adjacent pulses ; The mutual information value is calculated by normalization calculation; wherein, denotes an edge weight value, V i denotes the pulse pressure time series feature vector, denotes the microcirculatory blood flow feature vector, denotes an entropy, characterizing the uncertainty of the integrated node feature vector; use the calculated mutual information values as weight data of the edge weight values; construct the graph structure data based on all nodes and corresponding comprehensive node feature vector data and all adjacent node pairs and corresponding edge weight value weight data; a pathology recognition module configured to input the graph structure data into a pre-trained graph convolutional neural network model, match the graph structure data with a pre-set pathology sub-atlas template library, and output a pathology pattern recognition result.

Images