Multi-mode synchronous monitoring method and system based on traditional Chinese medicine pulse condition and microcirculation

Abstract

The invention discloses a multi-mode synchronous monitoring method and system based on traditional Chinese medicine pulse conditions and microcirculation, and relates to the technical field of traditional Chinese medicine intelligent diagnosis. A pulse position coordinate set is generated by combining a preset space mapping relation; a pulse pressure signal and a microcirculation blood flow image are synchronously acquired; frequency domain transformation is performed on the pulse pressure signal and a pathological risk probability L is calculated; a comprehensive node feature vector is constructed to form a graph structure; a pathology mode recognition result is output by matching a pathology sub-atlas template library, dynamic correlation analysis of traditional Chinese medicine pulse conditions and microcirculation is achieved, and the technical bottlenecks that traditional pulse diagnosis is high in subjectivity, multi-mode asynchronous and difficult in pathology quantification are broken through.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnosis of traditional Chinese medicine, and in particular to a multimodal synchronous monitoring method and system based on traditional Chinese medicine pulse and microcirculation. Background Art

[0002] Traditional Chinese Medicine (TCM) pulse diagnosis relies on the physician's tactile perception of radial artery pulse characteristics (such as strength and rhythm), resulting in a strong reliance on subjective experience and a lack of quantitative standards. While electronic pulse diagnosis devices can collect pulse pressure signals, sensor attachment relies on manual positioning (with errors often exceeding 5mm), resulting in distorted spatial mapping of the "Cun, Guan, and Chi" pulse positions. This makes it impossible to accurately reflect the gradient changes in Qi and blood status at different pulse positions (e.g., the "Cun floating, Chi sinking" pattern in distinguishing deficiency and excess).

[0003] When laser speckle imaging technology is used to acquire microcirculatory blood flow parameters, there is a hardware acquisition timing deviation (typically 10-50ms) between the pulse pressure signal (sampling rate >1kHz) and the microcirculatory image (frame rate 30fps). More critically, unconscious wrist tremors (amplitude 50-200μm) can further induce physiological response delays, leading to a temporal misalignment between the pulse peak and the peak of capillary perfusion (measured error 12.3±2.1ms), completely blocking the dynamic correlation verification of the Traditional Chinese Medicine theory of "pulse and blood share a common origin."

[0004] The spatiotemporal asynchrony of multimodal signals disrupts the dynamic correlation between pulse and microcirculation, depriving the scientific basis for quantitative pathological analysis guided by the holistic perspective of Traditional Chinese Medicine. To address this issue, there is an urgent need to develop an intelligent analysis method that can integrate the spatiotemporal characteristics of pulse and the dynamic changes of microcirculation while eliminating signal asynchrony errors. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a multimodal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation.

[0006] In order to achieve the above object, the technical solution of the present invention is as follows:

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

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

[0009] Perform pathological risk calculation on the pulse pressure time series signal data to obtain the pathological risk probability L and determine:

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

[0011] Acquiring motion compensation data of the user;

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

[0013] When the displacement amount is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount;

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

[0015] Based on the correlation measurement of the inter-node signals of the comprehensive node feature vector, the edge weight values between adjacent pulse position nodes are calculated, and graph structure data is constructed;

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

[0017] In a second aspect, the present invention discloses a multimodal synchronous monitoring system based on TCM pulse and microcirculation, which uses the above-mentioned multimodal synchronous monitoring method based on TCM pulse and microcirculation, including:

[0018] An acquisition module, used to obtain pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position of the user;

[0019] a risk judgment module, configured to calculate the pathological risk of the pulse pressure time series signal data to obtain the pathological risk probability L and make a judgment;

[0020] A 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 value. The specific steps are as follows;

[0021] Acquiring motion compensation data of the user;

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

[0023] When the displacement amount is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount;

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

[0025] A graph modeling module is used to calculate the edge weight values between adjacent pulse nodes based on the correlation measurement of the node signals between the comprehensive node feature vectors, and to construct graph structure data based on this;

[0026] The pathology recognition module is used to input the graph structure data into a pre-trained graph convolutional neural network model, match it with a preset pathology sub-atlas template library, and output pathology pattern recognition results.

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

[0028] 1. This invention utilizes motion compensation data and a path cost function to activate a time-domain alignment mechanism when the pathological risk probability exceeds a threshold. By capturing the acceleration changes generated by limb micro-movements in real time and triggering a compensation function based on the dynamic time warping path displacement, it effectively eliminates the physiological response delay between the pulse pressure signal and the microcirculatory blood flow image sequence. This mechanism overcomes the signal phase misalignment caused by hardware acquisition timing deviations and unconscious human tremors in traditional devices, ensuring that pulse peaks strictly correspond to peak capillary perfusion, providing a verifiable technical foundation for the Traditional Chinese Medicine theory of "pulse and blood share a common origin."

[0029] 2. Based on the comprehensive node feature vectors generated by time-domain alignment, this invention constructs a topological network consisting of pulse position nodes and mutual information edge weights. By quantifying multidimensional features such as pulse pressure waveform entropy and blood flow texture as node attributes and calculating edge weights based on the correlation between adjacent pulse position signals, a dynamic interactive graph is formed that conforms to the Traditional Chinese Medicine (TCM) theory of "Three Parts and Nine Signs." This structure not only fully preserves the spatial gradient characteristics of the Cun, Guan, and Chi pulse positions, but also more accurately quantifies the conduction correlation between Qi and blood states between pulse positions, achieving for the first time a mathematical model mapping of TCM's holistic syndrome differentiation at the data level.

[0030] 3. This invention uses a pre-trained graph convolutional neural network to analyze topological networks through a message-passing mechanism, combined with an attention-matching mechanism within a pathology sub-graph template library, to identify coupled pathological patterns of pulse characteristics and microcirculatory parameters. This process transcends the limitations of traditional single-dimensional analysis, enabling cross-modal features such as "pulse stagnation" and "capillary granular flow" to form a collaborative diagnostic basis, significantly improving the ability to identify complex pathological conditions. The resulting pathological pattern recognition results provide a scientific basis for objective diagnosis of TCM syndromes, combining temporal and spatial correlation with pathological specificity. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0032] Figure 1 is a flow chart of the steps of the present invention;

[0033] Figure 2 A diagram showing the steps for generating a comprehensive node feature vector according to the present invention;

[0034] Figure 3 is a system flow chart of the present invention;

[0035] Figure 4 Schematic diagram of the system operation of the present invention. DETAILED DESCRIPTION

[0036] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present invention and should not be regarded as the entire invention or as a limitation or restriction of the technical solution of the present invention.

[0037] Application Overview:

[0038] In existing technologies, traditional Chinese medicine pulse diagnosis mostly relies on the physician's subjective tactile judgment, while microcirculation detection is performed independently using a single optical modality, making it difficult to achieve temporal and spatial synchronous correlation analysis of pulse and microcirculation. In traditional methods, when the patient's limbs move, the pulse pressure sensor is mispositioned, resulting in feature drift. In deep blood vessel detection scenarios, microcirculation imaging loses information due to signal attenuation, and there is a millisecond-level delay between the pulse pressure wave and the capillary blood flow wave, resulting in mismatched pathological feature matching. Especially in the case of complex pulse conditions such as floating and slippery pulses, existing systems are unable to simultaneously meet the requirements of high temporal accuracy and multi-source signal spatial registration, which seriously restricts the accurate analysis of pathological mechanisms.

[0039] To address these issues, the inventors discovered that the radial styloid process serves as an anatomically constant landmark, establishing a spatial coordinate reference for the wrist. Experiments confirmed that there is a deterministic phase difference between pulse pressure signal signatures and microcirculatory perfusion fluctuations, while vascular depth and signal attenuation exhibit a nonlinear mapping relationship. Consequently, they proposed a core solution, the "Spatiotemporal Synchronization Framework," which utilizes depth image positioning to establish a spatial coordinate system and synchronizes multi-source acquisition based on position mapping. To address latency issues, an improved DTW compensation algorithm is employed to calibrate the microcirculatory acquisition timeline in real time based on the radial artery pulse pressure waveform.

[0040] Specifically, a 3D depth sensor first acquires wrist point cloud data, automatically identifies the radial styloid process, and establishes a three-dimensional spatial coordinate system. Simultaneously, at 32 preset spatial grid points, the following operations are performed: 1) millisecond-level pulse pressure timing signals are acquired using a MEMS pressure sensor array; and 2) dynamic microcirculatory blood flow image sequences are captured using laser speckle contrast imaging. When the initial pathology risk assessment probability exceeds a threshold, the system initiates a dynamic time-domain alignment mechanism: an acceleration compensation algorithm is used to eliminate limb displacement interference, an improved DTW algorithm is used to calibrate multi-source signal delays, and a spatiotemporal synchronization verification parameter δ is generated. The aligned signals are then fed into a graph convolutional network, which fuses the pulse temporal characteristics with the microcirculatory spatial characteristics, matches the pathology sub-atlas template library, and outputs a diagnostic result and confidence assessment.

[0041] Compared with existing technologies, traditional Traditional Chinese Medicine (TCM) testing equipment can only capture single-dimensional physiological signals and lacks a mechanism for spatiotemporal registration of multimodal signals. This solution addresses the spatial registration challenge of multi-source sensors through spatial coordinate mapping. It also utilizes an improved DTW algorithm to achieve microsecond-level time-domain synchronization and constructs a graph neural network model for pulse-microcirculation fusion. This method establishes a new paradigm for multimodal dynamic monitoring in the objectification of TCM pulse diagnosis, particularly suitable for early warning scenarios for cardiovascular and cerebrovascular diseases.

[0042] After introducing the basic concept of the present invention, embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Example 1:

[0044] like Figure 1 As shown, the multimodal synchronous monitoring method based on traditional Chinese medicine pulse and microcirculation includes the following steps:

[0045] Acquire pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position of the user;

[0046] Calculate the pathological risk of the pulse pressure time series signal data to obtain the pathological risk probability L and judge:

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

[0048] Obtaining motion compensation data of the user;

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

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

[0051] Based on the displacement and displacement compensation of the dynamic time warping path, the comprehensive node feature vector is obtained through time domain alignment processing;

[0052] Based on the correlation measurement of the node signals between the integrated node feature vectors, the edge weight values between adjacent pulse nodes are calculated and the graph structure data is constructed;

[0053] The graph structure data is input into the pre-trained graph convolutional neural network model, matched with the preset pathology sub-atlas template library, and the pathology pattern recognition results are output.

[0054] The dynamic time domain alignment trigger mechanism activates the compensation algorithm when the pathological risk probability L ≥ a preset first threshold. In this embodiment, the preset first threshold is set to 0.7, and the threshold is dynamically adjusted based on the user's basic physiological parameters (this threshold is optimized and set through a clinical ROC curve). For example, for users aged > 60, the threshold is lowered to 0.6. Specifically, limb displacement errors are corrected using acceleration sensor data, and the microcirculation image time series is reconstructed through cubic spline interpolation. The calculation formula for the pathological risk probability L is:

[0055]

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

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

[0058] Among them, the spatiotemporal synchronization verification parameter δ refers to the quantitative value of the time deviation between the pulse pressure signal and the microcirculation image sequence. Specifically, the improved DTW (dynamic time warping) algorithm is used to calculate the waveform similarity, which is used to determine the time alignment accuracy of multi-source data. Specifically, physiological feature points are first extracted based on the pulse pressure time series signal data and the microcirculation blood flow image sequence data. The feature points include peaks, troughs, and inflection points in the pulse pressure time series signal data, and extreme points of blood flow rate changes in the image frames of the microcirculation blood flow image sequence data, to ensure the one-to-one correspondence of the two modal data in physiological meaning. On this basis, the improved DTW compensation algorithm is used to pair the time series of the two modal feature points, quantify the time deviation, and calculate the spatiotemporal synchronization verification parameter δ. The calculation formula of the spatiotemporal synchronization verification parameter δ is as follows:

[0059]

[0060] Among them, N is the total number of paired feature points, is the timestamp of the i-th pulse pressure signal feature point, is the timestamp of the corresponding microcirculatory feature point, and the spatiotemporal synchronization verification parameter δ is the final average time deviation value, in milliseconds (ms).

[0061] In this embodiment, the improved DTW algorithm introduces an acceleration compensation factor and a window constraint mechanism in the matching process to correct the impact of limb micro-movement on time domain registration and improve the accuracy and stability of multimodal signal alignment.

[0062] In order to ensure the reliability and physiological rationality of data synchronization, the system presets a synchronization tolerance threshold to determine the validity of the collected data. The setting of the synchronization tolerance threshold comprehensively considers two factors: hardware sampling accuracy error and physiological response delay fluctuation. Among them, the hardware error mainly comes from the sampling frequency difference and synchronization deviation between the piezoelectric sensor and the laser speckle imaging system, and the comprehensive error is about ±10ms; the physiological response fluctuation mainly considers the natural delay of the pulse wave propagating to the microcirculation blood flow, and the fluctuation range generally does not exceed ±5ms. Based on the above factors, this embodiment introduces a 50% safety margin weighted adjustment in the synchronization tolerance setting, and finally determines the preset synchronization tolerance to be ±15ms. When the calculated spatiotemporal synchronization verification parameter δ is greater than or equal to the preset synchronization tolerance threshold, the system will automatically trigger the signal re-acquisition instruction to ensure the validity and consistency of the subsequent feature extraction and graph model training input data.

[0063] By introducing the above-mentioned spatiotemporal synchronization verification mechanism, the time-domain registration accuracy of pulse pressure signals and microcirculation image sequences in dynamic postures is effectively improved, providing high-quality multimodal input data support for subsequent node feature construction, edge weight calculation, and graph convolutional neural network diagnostic models.

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

[0065] Among them, the edge weight value refers to the measure of the physiological correlation strength between adjacent pulse nodes. Specifically, the mutual information entropy is used to calculate the correlation of pulse signals, and the 0-1 standardized weight matrix is generated in combination with the vascular topological distance.

[0066] Among them, the pathological sub-graph template library refers to a collection of pre-trained cardiovascular disease feature graph structures, which is specifically constructed based on the GCN (graph convolutional network) feature space clustering of clinically confirmed cases, and includes 128-dimensional feature templates for 8 types of pathological patterns such as coronary heart disease and hypertension.

[0067] Among them, confidence assessment refers to the reliability verification mechanism of pathology recognition results, which is specifically calculated through weighted calculation of Softmax probability distribution and template matching degree, and the preset confidence threshold is set to 0.85;

[0068] The confidence calculation formula is as follows:

[0069]

[0070] in, Represents the model output probability, with a value range of [0,1], indicating the confidence probability of the neural network model for pathology recognition; Indicates the atlas similarity, with a value range of [0,1], which quantifies the degree of matching between the actual feature map and the pathological template; is the output confidence value;

[0071] Among them, the model output probability weight is 60% (dominant factor), the atlas similarity weight is 40% (auxiliary verification), and the weight is dynamically configured according to the pathological sub-atlas type. For example, the weight of the coronary heart disease model is increased to 70%.

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

[0073] Among them, graph structure data refers to a dynamic topological graph composed of pulse nodes. The node attribute dimensions include: pulse time domain characteristics and microcirculation spatial characteristics. The edge weight reflects the physiological connectivity of the vascular conduction pathways of adjacent pulse positions.

[0074] The core innovation of this application lies in constructing a closed-loop diagnostic system based on spatiotemporal synchronous verification and multimodal feature fusion, solving the data distortion problem caused by limb displacement through dynamically triggered time domain compensation mechanism, and using graph structure modeling to realize cross-modal correlation analysis of pulse-microcirculation, breaking through the limitations of single-dimensional detection of traditional equipment.

[0075] like Figure 2As shown, the working process and principle of this application are as follows: first, wrist depth image data is acquired and the location of the radial styloid process is identified. A spatial coordinate set is generated based on a preset spatial mapping relationship; the pulse pressure time series signal and microcirculation blood flow image sequence data of the coordinate point are synchronously acquired; the pulse pressure signal is fast Fourier transformed to generate frequency domain features and the pathological risk probability L is calculated; when L ≥ the preset first threshold (0.7), the improved DTW compensation algorithm is activated for time domain alignment to generate a spatiotemporal synchronization verification parameter δ; if δ exceeds the tolerance value (±15ms), the data is re-acquired; otherwise, the pulse characteristic parameters and microcirculation blood flow characteristic parameters are extracted to construct a comprehensive node feature vector; the edge weights are calculated based on the mutual information entropy between nodes to form 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 confidence evaluated (threshold 0.85), and the diagnostic result data is output if it meets the standard. Through this closed-loop process, the accuracy of pathology recognition in complex body postures is significantly improved.

[0076] The present application further proposes to obtain pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position of the user. The specific steps of obtaining the radial styloid process position of the user include:

[0077] Extract feature point data and collect depth image data by scanning the wrist contour;

[0078] Processing feature point data based on a preset radial styloid process positioning algorithm, identifying and outputting radial styloid process point position data;

[0079] The positioning error of the radial styloid process position data is no greater than a predetermined error threshold, which is 0.8 mm.

[0080] Wrist contour scanning uses a three-axis synchronized depth imaging system to capture tissue surface geometry. It utilizes structured light projection and binocular infrared imaging to generate a millimeter-level point cloud (with an accuracy of ±0.2mm) that captures skin surface deformation characteristics. The radial styloid localization algorithm is an intelligent recognition engine that integrates deep learning and anatomical rules. It utilizes an improved PointNet++ architecture to integrate prior knowledge of the spatial distribution of hand bones and uses an anatomical topology constraint loss function to train a network to locate key bony landmarks.

[0081] Among them, the positioning error control mechanism refers to a closed-loop feedback system that verifies positioning accuracy in real time. It specifically uses multi-point stereo calibration verification technology to generate an error distribution heat map, and automatically starts the sub-pixel position compensation algorithm when the local area error exceeds 0.5mm.

[0082] Specifically, three sets of depth sensors are used to synchronously project coded structured light at 120° intervals, completing a 270° circumferential scan of the wrist within 8 milliseconds. The point cloud data of the user's wrist is input into the FPN model trained with 12,000 clinical data. The first-layer network extracts macroscopic anatomical contour features (such as the axis of the radius and ulna), and the deep network focuses on the geometric mutation features of the styloid process through the anatomical constraint module. While the positioning engine outputs three-dimensional coordinates in real time, the error compensation system performs stereo calibration on the 12 reference points in the scanning area. When it is detected that the error in a specific quadrant exceeds the threshold, the secondary positioning process is immediately triggered. The coordinate data finally output not only contains the spatial position, but also carries a predetermined error threshold for confidence (the predetermined error threshold is 0.8mm), which fully meets the accuracy requirements of the subsequent spatial mapping link.

[0083] Compared with existing technologies, traditional wrist positioning solutions rely on single-view two-dimensional image processing and are easily affected by soft tissue occlusion, resulting in positioning errors often exceeding 2mm; existing recognition methods based on grayscale features have significant errors when skin color varies greatly. This solution achieves:

[0084] 1) Multi-angle 3D scanning eliminates visual blind spots, and the point cloud data integrity reaches 99.3%;

[0085] 2) A deep learning network incorporating anatomical constraints significantly improved the robustness of bony landmark recognition, maintaining an average accuracy of 0.85 mm in the test group with a BMI > 30.

[0086] 3) Closed-loop error compensation mechanism reduces positioning failure rate.

[0087] Through the above technical solution, this application successfully solves the problem of benchmark positioning in dynamic wrist monitoring. Multi-sensor collaborative scanning ensures data integrity from the source, the intelligent positioning model enables accurate analysis of anatomical structures, and real-time error compensation ensures system stability. Clinical verification has shown that in dynamic scenarios such as fist clenching and rotation, the positioning accuracy of the radial styloid process remains at 0.68-0.79mm (n=153), which is higher than the accuracy of traditional optical positioning solutions, establishing a reliable spatial coordinate system benchmark for subsequent pulse detection.

[0088] The present application further proposes that after obtaining the user's radial styloid process point position, it also 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:

[0089] Based on the anatomical mapping relationship and the radial styloid process position data, the spatial coordinate transformation function is applied to calculate and generate the millimeter-level coordinate point data set corresponding to the Cun, Guan, and Chi parts to form the spatial coordinate set data;

[0090] Position calibration is performed based on spatial coordinate set data and preset soft tissue pressure data.

[0091] The spatial coordinate transformation function refers to a mathematical model that describes the mapping relationship between the radial styloid process and the pulse diagnosis area. Specifically, it uses a rigid body transformation matrix in the Lie group SE(3) space to solve the rotation and translation parameters through a pre-trained orthogonal projection network (the rotation transformation accuracy reaches 0.05 radians, and the translation resolution is 0.1 mm). Soft tissue pressure data refers to the dynamic distribution parameters that characterize the biomechanical properties of the contact surface. Specifically, it collects tissue deformation characteristics based on a pressure-sensitive sensor array (sampling frequency 200 Hz) and inverts the elastic modulus distribution map through the Hertz contact mechanics model. The position calibration mechanism refers to a real-time feedback system that compensates for physiological tissue deformation. Specifically, it uses a deformation gradient field and pressure field coupling algorithm to dynamically correct the coordinate depth parameters according to the tissue deformation characteristics (maximum compensation amplitude ±1.2 mm).

[0092] Specifically, once the system obtains the three-dimensional coordinates of the radial styloid process (with an accuracy of 0.8mm), the spatial mapping engine starts immediately: in the first stage, a deep projection network trained based on 2,000 CT data sets is called to generate 32 basic coordinate nodes in a 6cm³ space around the styloid process; in the second stage, the thin plate spline interpolation algorithm (TPS) is used to map the nodes into coordinate sets of three zones: Cunbu (proximal end), Guanbu (directly above the styloid process), and Chibu (distal end); in the final stage, the pressure-sensitive array monitors the contact pressure distribution in real time. When the local pressure gradient is detected to exceed 5kPa / cm, the elastic modulus correction model is automatically triggered to adjust the coordinate height value.

[0093] Through the above-mentioned technical solution, this application effectively solves the problem of positioning inaccuracy caused by anatomical variation and tissue deformation in Traditional Chinese Medicine pulse diagnosis equipment. The dynamic transformation of anatomical references lays the foundation for submillimeter positioning, the biomechanical feedback system intelligently corrects contact deformation, and a real-time calibration mechanism ensures dynamic detection stability. Clinical verification has shown that the standard deviation of the Cun, Guan, and Chi coordinates in different body positions (sitting / lying / sideways) is only 0.21mm (n=210), which is higher in accuracy than traditional optical projection solutions and establishes a precise spatial reference framework for subsequent synchronous multimodal data acquisition.

[0094] The present application further proposes that the specific steps of obtaining pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position and associating the spatial coordinate set include:

[0095] Controlling the fit position of the user's wrist based on 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 is less than a predetermined time deviation threshold, parallel triggering is performed to respectively acquire the pulse pressure time series signal data and the microcirculation blood flow image sequence data according to the first sampling rate and the second sampling rate.

[0097] Through a collaborative control mechanism coupled with spatiotemporal references, the system achieves integrated acquisition of pulse and microcirculatory dual-modal data, synchronized with physical perception and precise spatial positioning. Using submillimeter spatial coordinates (error <0.8mm) established at the radial styloid process as the hardware control benchmark, a master-slave hardware trigger architecture enables dual-channel millisecond-level parallel acquisition under pressure-adaptive contact interface control. The resulting spatiotemporal correlation dataset simultaneously meets the standards of 0.3mm spatial positioning accuracy and <2ms temporal synchronization error, providing spatiotemporally consistent raw input for subsequent pathological analysis.

[0098] Fit position control dynamically adjusts the pressure and angle of the detection probe based on spatial coordinates. This involves using a three-axis macro servo mechanism coupled with a closed-loop pressure feedback loop (200Hz response frequency) to correct the normal contact angle between the probe and the skin surface in real time (angle deviation <0.8°). Acquisition clock deviation control is a hardware mechanism that ensures time alignment between the pulse pressure signal (sampling rate 1kHz) and the microcirculation image (acquisition rate 30fps). This utilizes a master clock frequency-divided trigger signal generated by an FPGA, and a dual-channel phase-locked loop to eliminate accumulated crystal oscillator errors (clock jitter <10μs).

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

[0100] Through the above-mentioned technical solution, this application systematically addresses the technical bottlenecks of spatial positioning drift and multi-source data mismatch in Traditional Chinese Medicine (TCM) testing equipment. Spatial coordinates are converted into mechanical control instructions via a servo mechanism, physically ensuring precise sensor alignment. A hardware trigger architecture eliminates the millisecond-level delay associated with traditional software startup, and a pre-compensation mechanism circumvents the inherent time lag of optical systems. Clinical validation has demonstrated that this solution can maintain a spatial positioning accuracy of 0.35mm (SD ± 0.06mm) and a dual-modal data time alignment error of 1.7ms under conditions of slight wrist movement, establishing an unreliable benchmark system for pulse-microcirculation spatiotemporal correlation analysis.

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

[0102]

[0103] Among them, i,j represent the time series index, Indicates the pulse pressure signal value, represents the microcirculation characteristic value, represents the inter-modal distance, represents the Z-axis acceleration, and the vertical motion weight factor λ=0.35;

[0104] The time domain drift is calculated through time domain alignment processing. When it is detected that the time domain drift is greater than a predetermined drift threshold during the time domain alignment processing, the adaptive window function is started to resample and correct the pulse pressure time series signal data or the microcirculation blood flow image series data.

[0105] Motion compensation data refers to quantitative parameters representing the three-dimensional displacement of the limbs. This data is collected by a MEMS accelerometer (sampling rate 500Hz) worn on a detection wristband, which outputs real-time translational acceleration components separated from the gravity vector. The path cost function is a comprehensive distance metric model that integrates signal similarity and motion interference. It incorporates a vertical motion weighting factor (λ = 0.35) for Z-axis acceleration based on the standard DTW Euclidean distance, and triggers secondary path optimization based on a displacement threshold.

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

[0107]

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

[0109] The optimization results are shown in Table 1. When λ = 0.35, the signal mutual information reaches a peak of 0.89, while the micro-tremor 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 experimental results

[0111] λ value Signal mutual information Microtremor suppression rate (5-15Hz) 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 ± standard deviation.

[0113] Conclusion: Experiments show that λ=0.35 is the optimal parameter configuration. At this time, the system achieves a 91.7% microtremor suppression rate while maintaining a high signal mutual information (0.89), which is significantly better than other parameter combinations (p<0.01). Among them, the adaptive window function refers to a time domain correction mechanism that dynamically adjusts signal sampling. It specifically adopts a variable-width Hanning window function. When it is detected that the time domain drift exceeds the physiological rhythm cycle (the default threshold is 15% of the cardiac cycle), local resampling correction is initiated, and the real-time cardiac cycle is automatically calculated based on the pulse pressure signal.

[0114] Specifically, the time domain alignment engine works in three stages:

[0115] 1) In the preprocessing stage, the pulse pressure signal (1000 Hz), microcirculatory image feature sequence (30 fps), and motion compensation data are fused to calculate the signal distance matrix at each time point;

[0116] 2) In the dynamic regularization phase, the optimal curved path is calculated by improving the path cost function. When the path displacement exceeds a preset threshold (dynamically adjusted according to the cardiac cycle), the acceleration compensation algorithm is activated to generate a displacement compensation vector to correct the regularized path.

[0117] 3) In the post-processing stage, the time domain drift (standard deviation of the distance between adjacent peaks) is calculated and a window function is triggered to perform local interpolation and resampling of the pulse pressure signal to eliminate the accumulated phase error. The entire process compresses the time deviation to the microsecond level while maintaining the integrity of the signal's physiological characteristics.

[0118] Through the above technical solution, this application systematically solves the problem of time-domain mismatch in multimodal signals during dynamic monitoring. Accelerometer data provides a physical motion reference for time regularization, weighting factors accurately quantify the vertical interference intensity, and a window function mechanism maintains the fidelity of physiological characteristics during the resampling process. Clinical validation has shown that this processing flow stabilizes the time-domain alignment error to within 1% of the cardiac cycle, effectively overcoming the significant phase deviation caused by traditional methods when body position changes.

[0119] The present application further proposes that when the displacement is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount. The calculation formula of the path correction function is:

[0120]

[0121] in, is the displacement compensation at time t, representing the velocity offset accumulated by acceleration; is the vertical motion weight factor, ; is the Z-axis acceleration; is the signal starting time; t is the current time;

[0122] like >0.5m / s, displacement compensation will be suspended and motion interference alarm will be triggered;

[0123] right 5-15 Hz band-pass filtering was performed to eliminate non-physiological high-frequency noise.

[0124] This application implements closed-loop compensation for path correction, breaking through the limitation of traditional DTW that is sensitive to micro-movements of limbs and providing mathematical guarantee for time domain alignment accuracy δ≤1ms.

[0125] This application further proposes to obtain a comprehensive node feature vector based on the displacement and displacement compensation of the dynamic time warping path through time domain alignment processing. The comprehensive node feature vector specifically includes:

[0126] Pulse characteristic parameters extracted from pulse pressure time series signal data, including pulse pressure waveform entropy value Data and rhythm coefficient of variation data;

[0127] Microcirculation blood flow characteristic parameters extracted from microcirculation blood flow image sequence data, including the average blood flow velocity Data, blood perfusion unit PU data and LBP-TOP blood flow texture feature value data;

[0128] Pulse pressure waveform entropy The data and rhythm variation coefficient data constitute a 32-dimensional pulse pressure time series feature vector;

[0129] Average blood flow velocity The data, blood perfusion unit PU data and LBP-TOP blood flow texture feature value data constitute a 64-dimensional microcirculation blood flow feature vector;

[0130] The 32-dimensional pulse pressure time series feature vector and the 64-dimensional microcirculation blood flow feature vector are merged, and the comprehensive node feature vector data is generated based on the displacement and displacement compensation of the dynamic time warping path.

[0131] A hardware-level clock synchronization circuit (a nanosecond-level clock source based on the PTP protocol) triggers the pulse pressure sensor and microcirculation imaging device to synchronize data acquisition, eliminating device startup delay differences. Independent DMA channels are allocated for the dual-modal data streams, and a parallel transmission mechanism mitigates latency fluctuations caused by bus contention. A global timestamp module adds a unified time reference (accuracy of ±1ms) to each data packet, ensuring that the pulse pressure peak coincides with the peak of capillary perfusion, confirming the physiological synchronization between pulsation and blood flow, as embodied in Traditional Chinese Medicine (TCM) theory of "pulse and blood share a common origin."

[0132] Two key pathological indicators are extracted from the aligned pulse pressure signal. The entropy of the pulse pressure waveform quantifies the disorder of the pulse energy distribution (formula: ), reflecting the degree of vascular elasticity degradation); the rhythm variation coefficient is calculated as the ratio of the standard deviation of adjacent pulse cycles to the mean ( ), characterizing the autonomic nervous system regulatory function; the above parameters constitute a 32-dimensional pulse pressure time series feature vector, and the dynamic evolution laws such as "string pulse to slippery pulse" are captured through sliding window analysis.

[0133] Extracting microcirculatory blood flow characteristic parameters from spatiotemporally aligned microcirculatory blood flow image sequences: mean flow velocity The capillary-level blood flow intensity is quantified using perfusion units (PUs); spatiotemporal texture features: the LBP-TOP operator is used to extract the texture patterns of blood flow direction and velocity distribution, and to identify "stasis-agitation" anomalies; the above parameters are fused to generate a 64-dimensional microcirculation feature vector, realizing a holographic characterization from macro perfusion to micro flow state.

[0134] The 32-dimensional pulse pressure vector and the 64-dimensional microcirculation vector are concatenated into a 96-dimensional integrated node feature vector, with each dimension corresponding to a specific pathological sensitivity indicator. This vector serves as an embedded representation of the graph node, providing atomic input for the subsequent construction of the pulse location topology network.

[0135] Through this technical solution, cross-modal time alignment error is maintained at ≤1ms, completely eliminating signal phase shifts caused by wrist tremors. The time difference between the pulse pressure peak and the microcirculatory perfusion peak is kept within the physiological synchronization range, providing a distortion-free data foundation for feature extraction. The pulse pressure waveform entropy increases sensitivity to changes in vascular stiffness in the early stages of arteriosclerosis; the microcirculatory LBP-TOP feature can identify abnormal capillary clusters with a diameter of less than 100μm, surpassing the resolution limit of traditional imaging. Vector concatenation of these two complements and enhances the detection rate of complex pathological patterns such as "stringy pulse with stagnation of blood flow in the radial region." Through a three-step optimization approach combining hardware synchronization, deep feature extraction, and cross-modal vector fusion, the system constructs a comprehensive node feature vector that is both temporally and spatially consistent and pathologically specific. This vector, serving as the fundamental unit of the pulse position topology network, drives a graph convolutional model to accurately analyze the dynamic relationship between pulse condition and microcirculation in Traditional Chinese Medicine, providing a quantifiable scientific basis for the early diagnosis of arteriosclerosis.

[0136] This application further proposes that the specific steps of calculating the edge weight values between adjacent pulse nodes based on the correlation measurement of the inter-node signals of the integrated node feature vector and constructing the graph structure data include:

[0137] For each pair of adjacent pulse position comprehensive node feature vector data, calculate the mutual information value between the comprehensive node feature vector data ;

[0138] Mutual information value Perform standardized calculations;

[0139] in, represents the edge weight value, Vi represents the pulse pressure time series feature vector, represents the microcirculatory blood flow feature vector, represents entropy, which characterizes the uncertainty of the comprehensive node feature vector;

[0140] The calculated mutual information value is used as the weight data of the edge weight value;

[0141] Graph structure data is constructed based on all nodes and their corresponding comprehensive node feature vector data and all adjacent node pairs and their corresponding edge weight values.

[0142] For adjacent pulse nodes (such as Cunbu and Guanbu) after time domain alignment, the system calculates the normalized mutual information value of their 96-dimensional feature vector (including 32-dimensional pulse pressure time series features and 64-dimensional microcirculation blood flow features):

[0143]

[0144] in, Characterize the pulse pressure time series characteristics (pulse pressure waveform entropy value , rhythm variation coefficient CVRR), Characterize microcirculatory blood flow characteristics (average flow velocity , perfusion unit PU, LBP-TOP texture), It represents the statistical dependence of the two pulse position features. is the information entropy function. This formula eliminates the influence of feature dimension differences on correlation calculation, making the edge weight Objectively reflect the strength of physiological correlation between pulse positions.

[0145] The mutual information value As the edge weight connecting adjacent pulse nodes, construct a weighted undirected graph :

[0146] Node set V: high-dimensional feature vectors corresponding to Cun, Guan, and Chi pulse positions;

[0147] Edge set E: anatomical connections between adjacent pulse positions (Cun-Guan, Guan-Chi);

[0148] Weight set W: Normalized mutual information value The weight matrix formed;

[0149] This structure fully reproduces the “Pulse Network View” of Traditional Chinese Medicine, elevating the traditional discrete pulse position to a dynamic interactive system. After the generated graph data G is input into the pre-trained graph convolutional network (GCN), the message passing mechanism is transmitted along the high-weight edge ( >0.6) aggregates adjacent node features, the attention layer strengthens the diagnostic weight of abnormal associated paths such as "cun pulse-chi stasis", and the output layer matches the pathological sub-atlas template library to identify the "pulse stagnation-microcirculation disorder" coupling pattern in the early stage of arteriosclerosis.

[0150] Mutual information edge weight The classic theory of "Cunkou pulse is the beginning and end of the five internal organs and six bowels" was quantified. Clinical verification shows that the mutual information value of Guan-Chi in healthy people =0.82±0.05, while patients with early arteriosclerosis decreased to , objectively revealing the pathological association of "weak radial pulse indicates kidney qi deficiency".

[0151] When the high pulse pressure waveform entropy value appears in the Cun area ( , a marker of vascular stiffness) and the mutual information value of the joints and wrists When the pulse is low, the system automatically triggers the "pulse-microcirculation decoupling" warning. This mechanism improves the early detection rate of diabetic microvascular lesions.

[0152] High-weight edges ( ) guides GCN to focus on key pathological pathways. When the PU value of the ulnar microcirculation is abnormal, GCN uses the Guan-ulnar edge weight Strengthen attention to the coefficient of variation of pulse pressure at the Cun point to avoid misjudgment caused by isolated features.

[0153] A pulse position topology network constructed through mutual information edge weights transforms Traditional Chinese Medicine's holistic view into a computable dynamic graph model. This model, in conjunction with hardware-level synchronization mechanisms and high-dimensional feature vectors, enables the first mathematical quantification of the physiological interactions between the three pulse positions (Cun, Guan, and Chi), providing a scientific basis for early warning of conditions such as arteriosclerosis based on population pathology maps.

[0154] This application further proposes that the specific steps of inputting graph structure data into a pre-trained graph convolutional neural network model and outputting pathology pattern recognition results by matching a preset pathology sub-atlas template library include:

[0155] The graph convolutional neural network model updates the node status data in the graph structure data based on the message passing mechanism;

[0156] During the update process, attention weights are calculated based on the current integrated node feature vector data and the data in the pathological sub-graph template library;

[0157] After the graph convolutional neural network model is iteratively updated based on the attention weights, it combines the calculated attention weights with the final node state data to perform a matching operation with the template data in the pathology sub-atlas template library and outputs the pathology pattern recognition result data.

[0158] A three-layer graph convolutional network is used to process the structure, integrating a dual-channel attention mechanism:

[0159]

[0160] Where Q = node feature, K = pathological sub-graph template, Represents the dimension of the pathology sub-atlas template vector;

[0161] By matching the 12 pre-stored pathological sub-atlases (such as "slippery pulse-high perfusion" and "wiry pulse-granular flow stasis"), the pathological pattern recognition results are output.

[0162] Graph convolution operations use a message passing mechanism:

[0163]

[0164] in, Usually in neural network related formulas, h is often used to represent the hidden state. +1) and ( ) represent the first +1st level and The hidden state of the layer. It is Hidden state of the i-th node in layer +1 It is The hidden state of the jth node in the layer;

[0165] Indicates that from From the jth node of the layer to the The connection weight of the i-th node in the layer. It determines the importance of the information from the previous layer when it is passed to the current layer;

[0166] The ReLU activation function introduces nonlinear factors into neurons, allowing the neural network to approximate any nonlinear function.

[0167] Indicates the The bias parameter of the layer is used to adjust the activation threshold of the neuron so that the neuron can have a certain output even when there is no input;

[0168] Indicates the degree of node i and node j, which is used to measure the number of connections between nodes. Here they are used to weight Perform normalization and other operations to adjust the intensity of information transmission.

[0169] Graph convolution iteratively updates the node status through a three-layer network, and the final output layer combines the attention weights of the pathological sub-graph to generate the diagnosis result.

[0170] The graph convolutional neural network model updates the node state data in the graph structure data based on a message passing mechanism. During this process, the model uses message passing to transmit information between nodes in the graph. Based on the connection relationships and related weights between nodes, the node state data is updated to reflect the inherent characteristics and relationship changes of the graph structure data.

[0171] During the update process, attention weights are calculated based on the current integrated node feature vector data and the data in the pathology sub-atlas template library. By calculating the correlation between the current integrated node feature vector and the various template data in the pathology sub-atlas template library, the attention weights of different templates for the current node are determined, thereby highlighting important features and associations.

[0172] The graph convolutional neural network model iteratively updates attention weights, combining the calculated attention weights with the final node state data to match the template data in the pathology sub-atlas template library, and outputs the pathology pattern recognition results. During this process, the model continuously adjusts and optimizes node state data based on the calculated attention weights, and compares and matches various templates in the pathology sub-atlas template library, ultimately outputting the corresponding pathology pattern recognition results.

[0173] Specifically, a three-layer graph convolutional network is used to process this structure, integrating a dual-channel attention mechanism to more accurately calculate attention weights. The graph convolution operation uses a message-passing mechanism to iteratively update new node states through the three-layer network. The final output layer combines the attention weights of the pathology sub-graph to generate a diagnosis. Through this overall process and the interactive collaboration of various components, the graph convolutional neural network model can effectively utilize information such as graph structure data and the pathology sub-graph template library to accurately recognize pathology patterns and output reliable results, providing strong support for related medical diagnoses.

[0174] This application further proposes that the pre-training process of the graph convolutional neural network model is as follows:

[0175] Inject triple contrast enhancement mechanism, the triple contrast enhancement mechanism includes:

[0176] Add Gaussian noise data of preset amplitude to the training sample data of the same type to construct positive sample pairs ( , );

[0177] Randomly select different types of training sample data to construct negative samples ( );

[0178] Based on the positive sample pair ( , ) and negative samples ( ), the triple contrast loss function is used to optimize the parameters of the graph convolutional neural network model. The triple 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;

[0179] The triplet contrast loss function is defined as:

[0180]

[0181] Among them, the boundary distance threshold α=0.4.

[0182] This application was clinically validated at three institutions, including the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine. The dataset was constructed to demonstrate statistical significance in pulse-microcirculation correlation analysis and to verify the pathological sub-atlas recognition capabilities. The total sample size was 350, including 210 males and 140 females. Subjects ranged in age from 25 to 75 years, with a mean age of 48.6 ± 11.3 years (standard deviation). This distribution encompasses the age group with a high incidence of arteriosclerosis and ensures a thorough characterization of the associated pathological mechanisms. Disease types were divided into three groups: early arteriosclerosis (98 cases), diabetic microcirculatory disorders (85 cases), and a healthy control group (167 cases) to systematically evaluate the recognition capabilities of the pathological sub-atlas. Data were collected from March 2022 to June 2023. Each sample was collected three times and the mean value was calculated to eliminate temporal variation and ensure data stability.

[0183] Table 2: Parameter distribution of datasets and their technical relevance

[0184] parameter Data distribution Technology relevance Total sample size 350 cases (210 males, 140 females) Covering the statistical significance of pulse-microcirculation correlation analysis Age distribution 25-75 years old (mean 48.6±11.3 years) Includes age groups with a high risk of arteriosclerosis Disease type ① Early arteriosclerosis (98 cases) ② Diabetic microcirculatory disorders (85 cases) ③ Healthy control group (167 cases) Verify the pathology sub-atlas recognition capability

[0185] This application further proposes to inject a triplet contrast enhancement mechanism into the pre-training process of the graph convolutional neural network model to optimize the model parameters and improve its ability to distinguish between different pathological patterns.

[0186] The specific operation of the triplet contrast enhancement mechanism is as follows: First, Gaussian noise data with a preset amplitude is added to the training sample data belonging to the same category to construct a positive sample pair ( , ), which can increase the diversity of similar samples, allowing the model to better learn the common characteristics of similar samples while being able to adapt to a certain degree of noise interference. Secondly, randomly select different types of training sample data 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 ( ), using a triplet contrastive loss function to optimize the graph convolutional neural network model parameters. This process uses a boundary distance threshold of 0.4 to force the graph convolutional neural network model to increase the feature distance between different pathological patterns, enabling the model to more clearly distinguish sample features from different pathological patterns. By calculating and optimizing the triplet contrastive loss function, the model continuously adjusts its parameters to minimize the distance between samples of the same type and increase the distance between samples of different types.

[0187] Through the above technical solutions, this application effectively solves the problem of insufficient ability of graph convolutional neural network models to distinguish between different pathological patterns. The method of constructing positive sample pairs in the triple contrast enhancement mechanism increases the richness of samples of the same type, and the operation of constructing negative samples enables the model to better distinguish different classes. The process of optimizing model parameters based on the triple contrast loss function with a specific boundary distance threshold enables the model to increase the feature distance between different pathological patterns. Ultimately, the model's accuracy in distinguishing different pathological patterns is significantly improved, and various pathological patterns can be more accurately identified.

[0188] The following is a specific embodiment of the multimodal synchronous monitoring method based on TCM pulse and microcirculation:

[0189] The specific implementation process of a patient (58 years old, 7-year history of hypertension, BMI 26.3) who was monitored at the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine is as follows:

[0190] The system uses three sets of depth sensors to perform a 270° scan of the wrist, generating point cloud data with millimeter-level accuracy (±0.2mm) within 8 milliseconds. This data is fed into an FPN model trained on 12,000 clinical data points. Combined with an anatomical constraint module, the system identifies the radial styloid process and outputs 3D coordinates (32.1mm, 15.4mm, and 8.3mm). Verified by stereo calibration using 12 reference points, the system achieves a maximum positioning error of 0.5mm (below the 0.8mm threshold).

[0191] Based on the coordinates of the radial styloid process, the SE(3) rigid body transformation matrix (rotational accuracy 0.05 radians, translation resolution 0.1 mm) was applied to generate the coordinate sets for the Cun (30.2 mm, 14.8 mm, 6.5 mm), Guan (32.1 mm, 15.4 mm, 8.3 mm), and Chi (34.0 mm, 16.1 mm, 9.7 mm) regions. The pressure-sensitive array detected a local pressure gradient of 3.8 kPa / cm (below the 5 kPa / cm calibration threshold), and depth calibration was not initiated.

[0192] The pulse pressure sensor (sampling rate of 1000 Hz) collected a 10-second time series (10,000 sampling points), and the microcirculation camera (30 fps) captured 300 frames of blood flow images (resolution 1024×768). The measured clock deviation was 8.2 μs (less than the 10 μs threshold).

[0193] After fast Fourier transform analysis of the pulse pressure signal, the energy in the 1-5Hz frequency band accounted for only 48% (normal range 55%-70%), and the calculated pathological risk probability L=0.72 (exceeding the threshold of 0.7), triggering the time domain alignment process.

[0194] The time-domain alignment engine integrates the Z-axis micro-tremor data (amplitude 80μm, frequency 8Hz) detected by the MEMS accelerometer and calculates the path cost:

[0195]

[0196] The displacement D=7 steps (exceeding the threshold) is detected = 5 steps). Path correction function activated: The calculated displacement compensation amount Δ(t) = 0.18 m / s (corresponding to 80 μm micro-tremor).

[0197] Time domain index correction Perform cubic spline interpolation on microcirculatory sequences to generate alignment eigenvalues .

[0198] The system then extracts the characteristic parameters of the time domain aligned data: pulse pressure waveform entropy (mark of vascular elasticity degradation, normal <3.8), rhythm variation coefficient CVRR = 18% (autonomic nervous system dysfunction, normal <15%), average blood flow velocity =1.2 mm / s (capillary hypoperfusion, normal >1.5 mm / s), and LBP-TOP blood flow texture eigenvalue 0.63 (blood stasis, normal <0.50). These parameters constitute a 32-dimensional pulse pressure time series feature vector and a 64-dimensional microcirculatory blood flow feature vector, which are combined into a 96-dimensional comprehensive node feature vector.

[0199] Based on the comprehensive node feature vector, the system calculates the normalized mutual information value of adjacent pulse positions:

[0200] The measured weights for the Cun-Guan region were 0.41, and the weights for the Guan-Chi region were 0.38 (normal values > 0.6), indicating a significant weakening of the physiological correlation between pulse positions. The constructed graph structure data was input into a pre-trained graph convolutional network, and the node states were updated through a three-layer message passing mechanism:

[0201] Combined with attention weight calculation ( ) matched the pathology sub-atlas template library, with the highest match being the "early stage arteriosclerosis" template (similarity 0.89). Verified by triplet contrast loss (positive sample distance 0.31, negative sample distance 1.02 > boundary threshold 0.71), the final confidence assessment value is:

[0202]

[0203] The output diagnosis was "early arteriosclerosis (90.8% confidence level)." Core evidence included elevated pulse pressure entropy at the Cun section, decreased mutual information between the Guan and Chi sections, and blood flow stasis at the Chi section. A subsequent coronary CT scan confirmed 35% stenosis of the left anterior descending artery, consistent with the system's prediction. Clinical data demonstrated that the system achieved a positioning accuracy of 0.68-0.79mm in 153 dynamic tests, with time-domain alignment error consistently within 1% of the cardiac cycle, a 62% improvement over traditional optical solutions.

[0204] Example 2:

[0205] like Figure 3 and Figure 4 As shown, a multimodal synchronous monitoring system based on TCM pulse and microcirculation uses the above-mentioned multimodal synchronous monitoring method based on TCM pulse and microcirculation, including:

[0206] An acquisition module, used to obtain pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position of the user;

[0207] The risk judgment module is used to calculate the pathological risk of the pulse pressure time series signal data to obtain the pathological risk probability L and make a judgment;

[0208] The signal synchronization module is used to align the pulse pressure time series signal data with the microcirculation blood flow image sequence data in the time domain when the pathological risk probability L is greater than or equal to a preset first threshold. The specific steps are as follows;

[0209] Obtaining motion compensation data of the user;

[0210] Based on pulse pressure time series signal data, microcirculation blood flow image sequence data and motion compensation data, the displacement of the dynamic time warping path is calculated through the path cost function;

[0211] When the displacement amount is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount;

[0212] Based on the displacement and displacement compensation of the dynamic time warping path, the comprehensive node feature vector is obtained through time domain alignment processing;

[0213] The graph modeling module is used to measure the correlation between node signals based on the comprehensive node feature vector, calculate the edge weights between adjacent pulse nodes, and construct graph structure data based on this;

[0214] The pathology recognition module is used to input graph structure data into the pre-trained graph convolutional neural network model, match it with the preset pathology sub-graph template library, and output the pathology pattern recognition results.

[0215] The system uses the acquisition module as its data entry point, synchronously capturing pulse pressure time series signals at the radial styloid process and microcirculatory blood flow image sequences to ensure spatial consistency of the original signals. The risk assessment module calculates the pathological risk probability of the pulse pressure signal in real time. When the probability exceeds a preset threshold, a high-precision alignment process is activated. At this point, the signal synchronization module intervenes, integrating motion compensation data with a dynamic time warping algorithm. This quantifies the displacement deviation of the multimodal signals using a path cost function. When the displacement exceeds the limit, a path correction function is triggered to generate physical compensation, completely eliminating the physiological phase misalignment caused by wrist tremors.

[0216] The time-aligned signal is fed into the graph modeling module, which fuses cross-modal features such as pulse pressure waveform entropy and blood flow texture into a comprehensive node vector. It dynamically calculates edge weights based on the mutual information entropy of adjacent pulse signals, constructing a pulse topology network consistent with the Traditional Chinese Medicine (TCM) theory of "Three Parts and Nine Signs." Finally, the pathology recognition module drives the graph convolutional neural network to analyze the network structure. Through a message-passing mechanism, it fuses node states with the attention weights of the pathology subgraph template library to output pathology pattern recognition results that are both temporally and spatially correlated.

[0217] Compared to traditional approaches that separate pulse and microcirculation analysis, this system implements dynamic correlation verification of "pulse pulsation-microcirculatory perfusion" at the hardware level and completes pathology network modeling guided by the holistic perspective of Traditional Chinese Medicine at the algorithm level, providing a cross-modal collaborative diagnosis paradigm for early warning of cardiovascular and cerebrovascular diseases. The graph modeling module calculates the edge weights between adjacent pulse nodes based on the correlation measure of the signals between nodes in the comprehensive node feature vector, and uses this to construct graph structure data, providing a suitable data structure for subsequent pathology identification. The pathology recognition module inputs the graph structure data into a pre-trained graph convolutional neural network model and matches it with a preset pathology sub-graph template library to output pathology pattern recognition results.

[0218] Through the above technical solution, this application solves the problem of inaccurate acquisition position of TCM pulse and microcirculation multimodal data, and realizes multi-source data acquisition based on precise position; overcomes the defect of large time domain synchronization error of multimodal data, and achieves high-precision time domain alignment; improves the accuracy of pathology recognition, and can more effectively integrate multimodal data for pathology analysis; ensures the reliability of the output diagnosis results, and provides strong support for medical diagnosis through confidence assessment.

[0219] The technical scope of the present invention is not limited to the contents of the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical concept of the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A multimodal synchronous monitoring method based on TCM pulse and microcirculation, characterized by: The following steps are involved: Acquire pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position of the user; Perform pathological risk calculation on the pulse pressure time series signal data to obtain the pathological risk probability L and determine: When the pathological risk probability L is greater than or equal to a preset first threshold, the pulse pressure time series signal data and the microcirculation blood flow image sequence data are time-domain aligned, and the specific steps are as follows; Acquiring motion compensation data of the user; Calculating the displacement of the dynamic time warping path using a path cost function based on the pulse pressure time series signal data, the microcirculation blood flow image sequence data, and the motion compensation data; When the displacement amount is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount; Based on the displacement of the dynamic time warping path and the displacement compensation amount, a comprehensive node feature vector is obtained through time domain alignment processing; Based on the correlation measurement of the inter-node signals of the comprehensive node feature vector, the edge weight values between adjacent pulse position nodes are calculated, and graph structure data is constructed; The graph structure data is input into a pre-trained graph convolutional neural network model, matched with a preset pathology sub-atlas template library, and the pathology pattern recognition result is output.

2. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 1, characterized in that: The pulse pressure time series signal data and microcirculation blood flow image sequence data of the radial styloid process position of the user are obtained. The specific steps of obtaining the radial styloid process position of the user include: Extract feature point data and collect depth image data through wrist contour scanning; Processing the feature point data based on a preset radial styloid process positioning algorithm, identifying and outputting the radial styloid process point position data; The positioning error of the radial styloid point position data is no greater than a predetermined error threshold, which is 0.8 mm.

3. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 1, characterized in that: After obtaining the radial styloid process position of the user, the method further includes generating and associating a spatial coordinate set based on a preset spatial mapping relationship and the radial styloid process position, the specific steps of which include: Based on the anatomical mapping relationship and the radial styloid process position data, a spatial coordinate transformation function is applied to calculate and generate a set of millimeter-level coordinate point data corresponding to the Cun, Guan, and Chi parts to form spatial coordinate set data; Performing position calibration based on the spatial coordinate set data and preset soft tissue pressure data; Controlling the fitting position of the user's wrist based on the spatial coordinate set data; Under the condition that the acquisition clock deviation of the pulse pressure time series signal data and the microcirculation blood flow image sequence data is less than a predetermined time deviation threshold, parallel triggering is performed to respectively acquire the pulse pressure time series signal data and the microcirculation blood flow image sequence data according to a first sampling rate and a second sampling rate.

4. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 3, characterized in that: Based on the pulse pressure time series signal data, the microcirculation blood flow image sequence data and the motion compensation data, the displacement of the dynamic time warping path is calculated by a path cost function. The calculation formula of the path cost function is: Among them, i,j represent the time series index, Indicates the pulse pressure signal value, represents the microcirculation characteristic value, represents the inter-modal distance, represents the Z-axis acceleration, and the vertical motion weight factor λ=0.35; The time domain drift is calculated by time domain alignment processing. When it is detected that the time domain drift is greater than a predetermined drift threshold during the time domain alignment processing, the adaptive window function is started to resample and correct the pulse pressure time series signal data or the microcirculation blood flow image series data.

5. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 4, characterized in that: When the displacement is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount. The calculation formula of the path correction function is: in, is the displacement compensation at time t, representing the velocity offset accumulated by acceleration; is the vertical motion weight factor, ; is the Z-axis acceleration; is the signal starting time; t is the current time; like >0.5m / s, displacement compensation will be suspended and motion interference alarm will be triggered; right 5-15 Hz band-pass filtering was performed to eliminate non-physiological high-frequency noise.

6. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 1, characterized in that: Based on the displacement of the dynamic time warping path and the displacement compensation, a comprehensive node feature vector is obtained through time domain alignment processing. The comprehensive node feature vector specifically includes: The pulse characteristic parameters extracted from the pulse pressure time series signal data include the pulse pressure waveform entropy value Data and rhythm coefficient of variation data; The microcirculation blood flow characteristic parameters extracted from the microcirculation blood flow image sequence data include the average blood flow velocity Data, blood perfusion unit PU data and LBP-TOP blood flow texture feature value data; The pulse pressure waveform entropy value The data and the rhythm variation coefficient data constitute a 32-dimensional pulse pressure time series feature vector; The mean blood flow velocity The data, the blood perfusion unit PU data and the LBP-TOP blood flow texture feature value data constitute a 64-dimensional microcirculation blood flow feature vector; The 32-dimensional pulse pressure time series feature vector and the 64-dimensional microcirculation blood flow feature vector are combined, and the comprehensive node feature vector data is generated based on the displacement amount of the dynamic time warping path and the displacement compensation amount.

7. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 1, characterized in that: The specific steps of measuring the correlation of signals between nodes based on the comprehensive node feature vector, calculating the edge weight values between adjacent pulse position nodes, and constructing graph structure data include: For each pair of adjacent pulse position comprehensive node feature vector data, calculate the mutual information value between the comprehensive node feature vector data ; The mutual information value is Perform standardized calculations; in, represents the edge weight value, Vi represents the pulse pressure time series feature vector, represents the microcirculation blood flow characteristic vector, represents entropy, which characterizes the uncertainty of the comprehensive node feature vector; The calculated mutual information value is used as the weight data of the edge weight value; The graph structure data is constructed based on all nodes and the corresponding comprehensive node feature vector data and all adjacent node pairs and the corresponding edge weight values.

8. The multimodal synchronous monitoring method based on TCM pulse and microcirculation according to claim 1, characterized in that: The specific steps of inputting the graph structure data into a pre-trained graph convolutional neural network model and outputting the pathology pattern recognition results by matching a preset pathology sub-atlas template library include: The graph convolutional neural network model updates the node status 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 integrated 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, it combines the calculated attention weight with the final node state data to perform a matching operation with the template data in the pathology sub-atlas template library and outputs the pathology pattern recognition result data.

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

4.

10. A multimodal synchronous monitoring system based on TCM pulse and microcirculation, characterized by: include: An acquisition module, used to obtain pulse pressure time series signal data and microcirculation blood flow image sequence data at the radial styloid process position of the user; a risk judgment module, configured to calculate the pathological risk of the pulse pressure time series signal data to obtain the pathological risk probability L and make a judgment; A 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 value. The specific steps are as follows; Acquiring motion compensation data of the user; Calculating the displacement of the dynamic time warping path using a path cost function based on the pulse pressure time series signal data, the microcirculation blood flow image sequence data, and the motion compensation data; When the displacement amount is greater than a predetermined displacement threshold, a path correction function is triggered to generate a displacement compensation amount; Based on the displacement of the dynamic time warping path and the displacement compensation amount, a comprehensive node feature vector is obtained through time domain alignment processing; A graph modeling module is used to calculate the edge weight values between adjacent pulse nodes based on the correlation measurement of the node signals between the comprehensive node feature vectors, and to construct graph structure data based on this; The pathology recognition module is used to input the graph structure data into a pre-trained graph convolutional neural network model, match it with a preset pathology sub-atlas template library, and output pathology pattern recognition results.

Images