Traditional Chinese medicine intelligent blood vessel health management model construction method

Abstract

The invention discloses a traditional Chinese medicine intelligent blood vessel health management model construction method, which comprises the steps of obtaining multi-modal data based on intelligent wearable equipment; encrypting and coding the multi-modal data based on the FHIR standard, and generating a multi-modal data structured report supporting cross-system sharing; analyzing the multi-modal data structured report to obtain a traditional Chinese medicine blood gas analysis result; the method comprises the following steps: on the basis of a PICO framework and a Living system review technology, constructing an evidence graph RAG supporting incremental learning; and inputting the traditional Chinese medicine blood gas analysis result into a pre-trained large language model, and obtaining a traditional Chinese medicine health management report in combination with an evidence graph RAG. According to the method, the traditional Chinese medicine four-diagnosis data, the nail fold microcirculation image and the modern physiological indexes are integrated, a structured report is constructed, and the problem that heterogeneous data are difficult to store together is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of blood vessel health management, and particularly relates to a traditional Chinese medicine intelligent blood vessel health management model construction method. BACKGROUND

[0002] The core concept of traditional Chinese medicine health management is "preventing disease", which implements individualized intervention strategies based on the four diagnostic methods and the principle of syndrome differentiation. Sun Peiyu mentioned in the article "Understanding and Application of 'Four Seasons Have Soil' Based on the Thought of Preventing Disease" that the spleen-wetting method based on the thought of "Four Seasons Have Soil" can play a key role in the three stages of "preventing disease before disease", "preventing disease after disease", and "preventing disease after recovery", providing important theoretical support for preventing disease. However, the traditional mode has significant limitations, such as strong subjectivity in the data collection process, insufficient data structuring, and difficulty in realizing interconnection and intercommunication between different institutions, leading to the formation of information silos. In addition, the depth of intelligent analysis also needs to be further improved. Song Shibo et al. pointed out in "Thinking on the Objectification of Traditional Chinese Medicine Four Diagnostics" that the current research on the objectification of traditional Chinese medicine four diagnostics still faces key challenges: the standard of collection equipment and methods is not perfect, which limits the comparability and reliability of data; the four diagnostic objectification equipment has deficiencies in clinical transformation, and the combination of equipment development and clinical actual needs is not close enough, failing to fully "listen to the voice of the clinic"; in the four diagnostic data fusion and modeling technology, the existing methods are difficult to effectively handle the high-dimensional complex relationship between multi-modal diagnostic data in traditional Chinese medicine multi-diagnosis, and the explanation of the causal relationship between four diagnostic data and syndrome differentiation results is not clear. In recent years, with the rapid development of modern technologies such as artificial intelligence and big data, traditional Chinese medicine health management is gradually realizing the transformation of digitization and intelligentization. For example, the traditional Chinese medicine four diagnostic integrated equipment mentioned by Tang Weichang in his book "Traditional Chinese Medicine Four Diagnostic Information Collection and Research" uses modern scientific and technological means such as sensor technology and data analysis mining to realize the digital collection of traditional diagnostic information such as pulse diagnosis and tongue diagnosis. However, traditional Chinese medicine health management still faces many challenges in the technical aspect. Specifically, it is reflected in the following aspects:

[0003] There are difficulties in multi-modal data integration and standardization. The four diagnostic data of traditional Chinese medicine "observation, listening, questioning and palpation" and modern Western physiological indicators belong to different data modalities, and there are technical barriers in fusion analysis. Although existing research has realized the objective collection of part of the data through image recognition, sensors and other means, it lacks a systematic integration framework. For example, Weng Heng mentioned a kind of traditional Chinese medicine auxiliary decision-making method and device combining clinical evidence graph and multi-modal information fusion, combining heterogeneous network embedding, graph neural network reasoning, deep learning interpretability technology, realizing the integration framework of traditional Chinese medicine four diagnostic multi-modal data collection, information extraction, correction filtering, information fusion decision-making under the traditional Chinese medicine evidence graph construction and representation learning, but the collected multi-modal data has not formed a unified, standardized structured report, so that the conclusion after fusion lacks a clear data support chain, also limits the backtracking analysis and large-scale clinical verification of the results.

[0004] The current analysis of traditional Chinese medicine AI mainly stays in data pattern recognition and experience replication, lacks systematic evaluation and hierarchical integration of evidence quality, and leads to decision bias relying on experience rather than evidence. This makes it difficult to respond to the requirements of modern medicine for repeatability of therapeutic effect and verifiability of safety, and hinders the dialogue between traditional Chinese medicine and modern medical evidence system. The root cause lies in the fact that a fusion mechanism considering the overall nature of traditional Chinese medicine theory and the rigor of evidence-based medicine has not been formed, and the effective conversion from data to reliable evidence cannot be realized, which reflects the problem of insufficient combination of intelligent analysis and evidence-based medicine.

[0005] Traditional Chinese medicine health management also has significant limitations in dynamic monitoring and intervention suggestion generation. Its data collection mode mainly stays at the static level, which is difficult to track the dynamic fluctuations of physiological indicators in real time, and cannot construct continuous data graph reflecting the trend of body changes, resulting in the lack of timeliness and integrity in the evaluation of health status. In terms of intervention suggestion generation, existing schemes mainly rely on empirical dialectical logic, lack the support of evidence-based medicine system, and it is difficult to clearly define the applicable boundaries and efficacy levels of intervention measures. Although related research has explored the methodological framework of traditional Chinese medicine evidence-based research, providing a theoretical basis for the standardization of intervention suggestions, it has not realized the conversion from evidence-based theory to intelligent recommendation, and cannot adjust the intervention scheme in real time according to individual dynamic health data, resulting in insufficient precision and adaptability of intervention, which is difficult to meet the dynamic and personalized needs of modern health management. SUMMARY

[0006] The present application provides a traditional Chinese medicine intelligent blood vessel health management model construction method to solve the technical problem that the traditional method cannot adjust the intervention scheme in real time according to individual dynamic health data, resulting in insufficient precision and adaptability of intervention, which is difficult to meet the dynamic and personalized needs of modern health management.

[0007] To solve the above technical problems, one technical solution adopted by the present application is: a traditional Chinese medicine intelligent blood vessel health management model construction method, comprising:

[0008] S1. Based on the intelligent wearable device, multi-modal data including traditional Chinese medicine four diagnostic data, image data and various physiological indicators are obtained;

[0009] S2. Based on the FHIR standard, the multi-modal data is encrypted and encoded to generate a multi-modal data structured report supporting cross-system sharing;

[0010] S3. Based on a plurality of learning models, the multi-modal data structured report is parsed to obtain traditional Chinese medicine blood gas analysis results; wherein the learning model includes an improved U-Net model.

[0011] S4. Based on the PICO framework and the Living systematic review technology, an evidence graph RAG supporting incremental learning is constructed;

[0012] S5. The traditional Chinese medicine blood gas analysis result is input into a pre-trained large language model, and combined with the evidence graph RAG to obtain a traditional Chinese medicine health management report.

[0013] Further, the method of step S1 comprises:

[0014] S11. Based on the camera component, facial diagnosis data including tongue image, face image and nail fold microcirculation image are obtained;

[0015] S12. Based on the audio recording component, the auscultation data are obtained;

[0016] S13. Based on the collection result of the traditional Chinese medicine constitution differentiation questionnaire, the interrogation data are obtained;

[0017] S14. Based on the pulse wave waveform collected by the pulse wave collection module, the palpation data are obtained;

[0018] S15. Based on the sensor monitor, physiological data including body temperature, blood oxygen, blood pressure, heart rate and arterial stiffness are obtained.

[0019] Further, the method of step S3 comprises:

[0020] S31. Based on the multi-modal data structured report, feature alignment and structured expression of the multi-modal data are performed to obtain an aligned feature matrix;

[0021] S32. Based on the aligned feature matrix, a gated attention fusion mechanism is constructed;

[0022] S33. Based on the gated attention fusion mechanism, traditional Chinese medicine blood gas classification is performed to obtain a traditional Chinese medicine blood gas classification result output.

[0023] Further, the method of step S32 comprises:

[0024] S321. Construct a global feature vector based on the aligned feature matrix;

[0025] S322. Encode the global feature vector to obtain an encoded feature;

[0026] S323. Obtain a gating mechanism based on the encoded feature and a gating neuron;

[0027] S324. Perform modulation output based on the gating mechanism.

[0028] Further, the blood and qi analysis result of traditional Chinese medicine comprises constitution type determination, blood stasis type, blood vessel aging degree classification, and semi-quantitative evaluation of zang-fu function.

[0029] Further, the method of step S4 comprises:

[0030] S41. Based on the PICO framework, combined with the blood and qi analysis result of traditional Chinese medicine, obtain the entities required to be contained in the original evidence graph RAG, and set the search terms;

[0031] S42. Based on the evidence-based medicine evidence level principle and PICO matching logic, construct a five-dimensional hierarchical scoring system containing research design scientificity, methodological rigor, PICO element adaptability, academic influence, and result practicality;

[0032] S43. Calculate the comprehensive score by weighted summation method, combine the total score threshold screening and single low score forced exclusion rule, and obtain the evidence scientific correlation target problem;

[0033] S44. Knowledge extraction is performed on the screened high-quality literature, and entity alignment, relationship and attribute fusion operations are performed to construct a unified initial evidence graph;

[0034] S45. Based on the initial evidence graph and the blood and qi analysis result of traditional Chinese medicine, design search terms that need to be updated;

[0035] S46. Based on the API literature retrieval interface, obtain the updated search terms, and then perform named entity recognition, relationship extraction, literature scoring analysis, and conflict verification to screen out the clinical evidence that needs to be updated;

[0036] S47. Update the updated clinical evidence into the initial evidence graph to obtain the evidence graph RAG.

[0037] Further, the traditional Chinese medicine health management report comprises the blood and qi analysis result, dietary therapy suggestion, seasonal conditioning suggestion, living conditioning suggestion, spiritual cultivation suggestion, exercise health preservation suggestion, and meridian point health preservation suggestion.

[0038] The application integrates traditional Chinese medicine four diagnostic data, nail capillary microcirculation images and modern physiological indicators to construct a structured report, solves the problem of heterogeneous data difficult to be saved together. The application carries out data standardization conversion and cross-system sharing based on FHIR, breaks the data barriers between medical institutions. The application combines machine learning and large language model to realize traditional Chinese medicine blood and gas analysis and provide user health management suggestions based on evidence-based medicine. The time trend curve of temperature, blood pressure and other indicators is generated through continuous data acquisition, and the health risk warning ability is improved. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of an embodiment of the traditional Chinese medicine intelligent blood vessel health management model construction method of the application;

[0040] Figure 2 is a flowchart of an embodiment of step S1 in Figure 1

[0041] Figure 3 is a FHIR structured report mode diagram in an embodiment of step S2 in Figure 1

[0042] Figure 4 is a flowchart of an embodiment of step S3 in Figure 1

[0043] Figure 5 is an improved gating attention fusion mechanism in an embodiment of step S32 in Figure 4

[0044] Figure 6 is a flowchart of an embodiment of step S4 in Figure 1 DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with specific embodiments.

[0046] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, therefore, the application is not limited to the specific embodiments disclosed in the following description.

[0047] Referring to Figure 1 , Figure 1 is a flowchart of an embodiment of the traditional Chinese medicine intelligent blood vessel health management model construction method of the application, which comprises:

[0048] S1. Based on the intelligent wearable device, multi-modal data including traditional Chinese medicine four diagnostic data, image data and various physiological indicators are acquired.​​​​​

[0049] Specifically, referring to Figures 2-3 , the method of step S1 comprises:

[0050] S11. Based on the camera assembly, obtain facial diagnosis data including tongue image, facial image and nail fold microcirculation image.

[0051] Specifically, the user enters the facial and tongue image capture interface through the camera assembly of the smart wearable device, and shoots the required facial and tongue images according to the prompt; then the user opens the nail fold microcirculation image capture module in the camera assembly through the touch screen, which is equipped with a macro lens, a high-definition image sensor and a directional light supplement module (containing 3 high-brightness miniature white LEDs) for shooting high-definition images.

[0052] S12. Based on the recording assembly, obtain the auscultation data.

[0053] Specifically, the user selects to enter the auscultation module through the touch screen in the smart wearable device, reads out a pre-set paragraph according to the prompt, completes the sound recording collection, and obtains the auscultation data.

[0054] S13. Based on the collection result of the traditional Chinese medicine constitution differentiation questionnaire, obtain the inquiry data.

[0055] Specifically, the user selects to enter the inquiry module through the touch screen in the smart wearable device, answers the pre-set traditional Chinese medicine constitution differentiation questionnaire according to the prompt, and completes the questionnaire result collection.

[0056] S14. Based on the pulse wave waveform collected by the pulse wave collection module, obtain the pulse diagnosis data.

[0057] Specifically, the smart wearable device collects the pulse wave waveform data of the user in real time, which adopts an integrated high-precision PPG (photoelectric volume pulse wave) sensor module, including 3 LEDs (1 660nm red LED, 1 940nm infrared LED, 1 520nm green LED), 1 high-sensitivity silicon photodiode and signal amplification circuit; the collected pulse wave parameters include: ① Amplitude = Peak amplitude value - Trough amplitude value (voltage signal difference collected by PPG sensor, unit mV); ② Up slope = (Peak amplitude value - Up slope starting point amplitude value) / (Peak time - Up slope starting point time) (voltage unit mV, time unit ms, which can reflect the blood vessel elasticity).

[0058] S15. Based on the sensor monitor, obtain physiological data including body temperature, blood oxygen, blood pressure, heart rate and arterial stiffness.

[0059] Specifically, the smart wearable device collects the body temperature, blood oxygen, blood pressure and heart rate values of the user in real time, and calculates the arterial stiffness, which is as follows:

[0060] Body temperature: collected by a contact infrared temperature sensor embedded in the inner side of the watchband, the sensor uses a gold-plated probe to improve heat conduction efficiency, and the outer layer is covered with a waterproof and breathable film to avoid the influence of sweat and water vapor on measurement;

[0061] Blood oxygen saturation (SpO2): calculated based on the reflectivity of red and infrared light of the PPG sensor, the formula is SpO2 = K1-K2*R, where K1, K2 are calibration constants calibrated by clinical data, and R = red light reflectivity / infrared light reflectivity;

[0062] Heart rate (HR): calculated by the pulse period collected by the PPG sensor, the formula is HR = 60 / average pulse period (seconds);

[0063] Blood pressure (systolic blood pressure SBP, diastolic blood pressure DBP): using a PPG+ECG combined sensor module (containing two metal electrode pieces with a diameter of 5mm, embedded in the side of the watch and the inner side of the watchband), calculated by PWT (the time difference between the R wave of ECG and the PPG peak, unit: seconds), the formula is SBP = a-b*PWT, DBP = c-d*PWT, where a, b, c, d are calibration constants calibrated by clinical data;

[0064] Arterial stiffness (core indicator PWV): based on the synchronous sampling data of PPG and ECG, calculate PTT (the same parameter as PWT, i.e. the time difference between the R wave of ECG and the PPG peak, unit: seconds), combined with the user's input height (unit: m) to convert, the formula is PWV = height / PTT.

[0065] S2. Based on the FHIR standard, the multi-modal data is encrypted and encoded to generate a multi-modal data structured report supporting cross-system sharing.

[0066] Specifically, before encryption and encoding, the images, audio and other data in the multi-modal data are uploaded to the cloud storage server; the cloud generates a resource URL server associated with the user's unique identifier, and automatically generates a URL. When calling the URL, the user needs to initiate a request through the cloud API gateway and carry the user identity Token (validity period of 2 hours); the gateway verifies the legality of the Token based on the JWT algorithm, and returns the resource access permission if the verification is passed, and refuses access if it is not passed.

[0067] The method for encrypting and encoding multi-modal data based on the FHIR standard comprises:

[0068] S21. Map the uploaded multi-modal data to FHIR resources to build a JSON structured report containing four diagnostic data, physiological indicators, etc.

[0069] Specifically, the mapping rules of the data and the corresponding FHIR resources are as follows:

[0070] The tongue, face, and nail microcirculation images are mapped to the Media resource.

[0071] The sound recording is mapped to the Media resource.

[0072] The results of the traditional Chinese medicine questionnaire are mapped to the QuestionnaireResponse resource.

[0073] The pulse wave waveform is mapped to the Observation resource.

[0074] The physiological index data (blood oxygen, heart rate, and body temperature) are mapped to the Observation resource.

[0075] The structured report is constructed in JSON format, including the patientId (hashed user identifier), cloud URL of the tongue / face / nail microcirculation images, cloud URL of the sound recording, results of the traditional Chinese medicine questionnaire, cloud URL of the pulse wave waveform, and cloud URL of the physiological index data. This ensures the standardized integration of multi-modal data, and the multi-modal data are called from the cloud through the URL.

[0076] S22. The multi-modal data are encrypted according to the FHIR standard.

[0077] Specifically, the public key and the private key are generated; the sender uses the public key of the receiver to encrypt the core sensitive information, and the encrypted information can only be decrypted by the private key of the receiver; at the same time, partial character replacement is performed on the user's name, ID number, and other information to balance the security and usability. (For example, the name "Li Hua" is desensitized to "Li *")

[0078] Dynamic desensitization is implemented according to the permission level of the data receiver: the internal system can obtain complete data, and the external system uses partial character replacement.

[0079] The API gateway's JWT authentication is used for the cloud resource URL, and HTTPS is used to ensure transmission security.

[0080] The whole process strictly follows the FHIR Security specification: the sender signs the encrypted fields and metadata with its own private key, and records the data signature and traceability information using the FHIR Provenance resource, which is associated with the original report to ensure the strong binding of the URL and the report content; all encryption processes are compatible with the FHIR JSON structure containing the URL field, realizing the security and standardized unification of cross-system interaction.

[0081] Step S23. Cross-system sharing is realized through OAuth 2.0 authentication, RESTful API, and HTTPS.

[0082] Specifically, when calling the URL, a request needs to be initiated through the cloud API gateway and carry the user identity Token (validity period of 2 hours); the gateway verifies the legality of the Token based on the JWT algorithm, which is used for instant authentication of resource access within a single system.

[0083] When accessing across systems, the calling party needs to apply for a Token from the authentication server through the OAuth 2.0 process; interaction is achieved through the FHIR RESTful API, HTTPS transmission is adopted, and synchronous response and asynchronous response are supported.

[0084] S3. Based on a plurality of learning models, the multi-modal data structured report is parsed to obtain a traditional Chinese medicine blood and qi analysis result.

[0085] Specifically, referring to Figure 4 , the method of step S3 comprises:

[0086] S31. Based on the multi-modal data structured report, the multi-modal data is aligned and structured, and an aligned feature matrix is obtained.

[0087] Specifically, the method of step S31 comprises:

[0088] S311. The tongue image photo is processed by improving the U-Net model.

[0089] Preprocessing enhancement: first, color space conversion is performed to strengthen the chroma difference between the tongue and the background; through adaptive histogram equalization by limiting contrast, the problem of local uneven illumination is improved; geometric normalization is used to correct the shooting posture deviation; then, the denoising technology based on the generative adversarial network is used to eliminate the interference of complex background, providing high-quality input data for subsequent processing.

[0090] The method for improving the U-Net segmentation comprises:

[0091] The encoder uses the ResNet50 / 101 backbone network to extract features, and embeds the hybrid level serial dilated convolution (HCDC) module to expand the receptive field;

[0092] The ASPP module is introduced at the end of the encoder to fuse multi-scale tongue features, and the FPN is used to connect the low-level contour features and high-level semantic features horizontally;

[0093] The decoder uses the PSA module to extract multi-scale edge features, and the MAG module optimizes the tongue edge, and the residual soft connection is used instead of the traditional jump connection to suppress noise;

[0094] Finally, the CRF is used to smooth the boundary, and the morphological operation is used to repair the continuity of the tongue.

[0095] Feature vector extraction: Extract features such as texture, color, and shape from the segmented tongue image to form a feature vector, providing data support for preliminary judgment of TCM constitution and blood stasis type.

[0096] S312. Face photo processing by improving U-Net model.

[0097] Color parameter extraction: After standardization preprocessing, the face image is converted to Lab color space, and the luminance (L), red-green chroma (a), and blue-yellow chroma (b) are extracted. The derived parameters such as color difference index ΔE, color saturation C, and hue angle h are calculated, and the features are mapped in detail according to TCM theory.

[0098] Improved U-Net segmentation: The encoder contains 4 levels of downsampling, each level extracts features through two 3x3 convolution layers (ReLU activation), and adopts 2x2 max pooling to compress the spatial dimension, with the number of channels gradually increasing from 64 to 512. The 4 levels of upsampling recover the resolution through transposed convolution, and fuse the low-level texture features of the encoding path through jump connection. The spatial dynamic multi-head self-attention module is introduced, where DMSA (deformable multi-head self-attention) learns the deformation features of irregular facial regions, and NMSA (neighborhood multi-head self-attention) extracts local neighborhood features with a 7x7 pixel window, balancing global deformation modeling and local texture analysis through a dynamic alternating mechanism. The output layer generates 9 segmentation results of TCM face regions.

[0099] Texture feature vector output: Output quantized texture feature map, including skin roughness, spot density, and glossiness, to form a feature vector and provide a multi-dimensional data basis for constitution determination.

[0100] S313. Nail fold microcirculation processing by improving U-Net model.

[0101] Enhanced blood vessel details: Through adaptive histogram equalization, the blood vessel details in the nail fold microcirculation image are enhanced, highlighting the blood vessel texture.

[0102] U-Net segmentation: A U-Net pre-trained model based on the generator structure in the GAN framework is used. The input image is standardized (mean standardization: [0.585, 0.256, 0.136]; standard deviation normalization: [0.229, 0.124, 0.095]), and the input size is automatically adjusted to a multiple of 32. The encoder gradually extracts high-level semantic features through convolution and downsampling operations, and the decoder recovers the spatial resolution through upsampling and transposed convolution, introducing jump connection to fuse low-level detail features. The output layer generates a single-channel probability map, which is binarized by a 0.5 threshold to output the blood vessel segmentation mask.

[0103] Morphological statistical analysis: count the number of vascular loops, calculate the vascular tortuosity, and classify the vascular morphology. These indicators directly reflect the microcirculation disorder.

[0104] S314. Sound signal processing through MFCC model.

[0105] Audio processing: pre-process the collected sound recording, remove noise and other interference factors.

[0106] MFCC feature optimization extraction: optimize the extraction of Mel Frequency Cepstral Coefficient (MFCC) features through feature engineering, generate normalized attention heat map, capture frequency-energy and timing features of speech, and form feature vector.

[0107] Blood stasis determination application: input the optimized MFCC features and time distribution features into the machine learning model pre-trained for "speech pattern-blood stasis syndrome" recognition, identify blood stasis syndrome and non-blood stasis syndrome through the timing characteristics of speech features, and provide the basis for diagnosis of blood stasis type.

[0108] S315. Survey questionnaire processing.

[0109] Scoring rules: use 5-level scoring system (1-5 points), and reverse score the items (1→5, 2→4, 3→3, 4→2, 5→1). The original score is the sum of the item scores, and the cumulative results of the positive and negative scores are recorded accurately.

[0110] Score conversion: calculate the conversion score according to the formula [(original score-item number) / (item number*4)]*100, and convert the T score by combining gender, age and other demographic characteristics, eliminate the influence of individual differences, and form comparable standard scores.

[0111] Constitution determination: according to the standard of "Classification and Determination of Traditional Chinese Medicine Constitution", the normal constitution needs to meet the conversion score ≥60 points, and the conversion scores of other 8 kinds of biased constitutions are all <30 points; the conversion score ≥60 points and the other scores are all <40 points are "basically yes", otherwise "no". The conversion score of biased constitution (qi deficiency constitution and other 8 types) is ≥40 points, 30-39 points is "tendency yes", and <30 points is "no". If multiple constitution characteristics are met at the same time, the determination results need to be marked in parallel.

[0112] S316. Pulse wave processing through ITM model.

[0113] Preprocessing: smooth the pulse wave signal using weighted Henderson moving average algorithm, remove noise outside the range of 0.005-30Hz through Butterworth filter; complete baseline correction using cubic spline fitting, and accurately extract feature points such as starting point and peak value of pulse wave through cross-cutting method (ITM).

[0114] Parameter calculation:

[0115] EPA (End Point Angle): Based on the intersection tangent method (ITM), the first derivative maximum point on the left side and the second derivative zero point on the right side of the peak value are determined as the starting points of the intersection tangent, a fitting line is generated, the intersection point of the tangent line at the peak value and the two intersection tangent lines is found, and the EPA is the angle formed by the two end points and the peak point, reflecting the steepness of the local shape of the peak region.

[0116] VH (Virtual Height): The vertical distance between the horizontal lines of the intersection points of the tangent line at the peak value and the two intersection tangent lines is calculated, and after standardization, the VH is obtained, which makes up for the limitation of EPA only focusing on the local region of the peak value and reflects the overall waveform characteristics of the pulse wave.

[0117] PSI (Pulse Sharpness Index): Derived by the formula (PSI(degree)^(-1)=1000 / [EPA×(1+VH)]), PSI≥14.0 is mild vascular aging, PSI between 11.0 and 13.9 is moderate vascular aging, and PSI≤10.9 is severe vascular aging.

[0118] a1 / a2 ratio: According to the human quantum resonant cavity model, the inch, Guan, and Chi pulses correspond to different zang-fu organs. By analyzing the deviation of a1 / a2 (the amplitude ratio of the first wave and the second wave) in the Gaussian wave packet parameters, the imbalance state of the yin and yang properties of the three pulses is quantified. Under normal conditions, the ratio is close to 1.8, and deviation indicates abnormal function of the corresponding zang-fu organs.

[0119] S317. Physiological index processing.

[0120] Body temperature: Collected by a contact infrared temperature sensor embedded in the inner side of the watchband, the sensor uses a gold-plated probe to improve heat conduction efficiency, and the outer layer is covered with a waterproof and breathable film to avoid the influence of sweat and water vapor on measurement.

[0121] Oxygen saturation (SpO2): Based on the red light and infrared light reflection intensity calculation of the PPG sensor, the formula is SpO2=K1-K2*R, where K1, K2 are calibration constants calibrated by clinical data, and R=red light reflection intensity / infrared light reflection intensity.

[0122] Heart rate (HR): Calculated by the pulse period collected by the PPG sensor, the formula is HR=60 / average pulse period (seconds).

[0123] Blood pressure (systolic blood pressure SBP, diastolic blood pressure DBP): Using a PPG+ECG combined sensor module, it is obtained by calculating PWT (the time difference between the R wave of ECG and the peak of PPG, unit: seconds), the formula is SBP=a-bPWT, DBP=c-dPWT, where a, b, c, d are calibration constants calibrated by clinical data.

[0124] Arterial stiffness (core indicator PWV): Based on the PPG and ECG synchronous sampling data, calculate PTT (the same parameter as PWT), combined with the user input height (unit m) conversion, the formula is PWV = height / PTT. These physiological indicators combined with TCM theory analysis provide physiological evidence for constitution analysis, and arterial stiffness directly quantifies the degree of vascular aging.

[0125] The model performs feature alignment operation on the features extracted from the above seven dimensions to ensure the comparability of the numerical values input into the fusion model:

[0126] Alignment method: adaptive normalization is adopted for different types of features:

[0127] Image features (tongue, face, and nail fold): numerical features are normalized to the [0, 1] interval by minimum-maximum normalization, and the formula is: Eliminate the scale difference between pixel value and physical indicator.

[0128] Time series signal features (sound, pulse wave): vectorized features are standardized by Z-Score, and the formula is: (μ is the mean, σ is the standard deviation), which solves the amplitude fluctuation problem in frequency and time domain.

[0129] Questionnaires and physiological indicators: quantile normalization is performed on scalar values such as conversion points and blood pressure to force them to conform to a uniform distribution and reduce individual baseline bias.

[0130] Alignment logic: the feature alignment module automatically matches the feature dimensions to generate an aligned feature matrix, ensuring the weighted comparability of multi-modal features in the fusion model, and avoiding the dominance of a single modality in decision-making.

[0131] Step S32. Based on the aligned feature matrix, a gated attention fusion mechanism is constructed.

[0132] Specifically, referring to Figure 5 , the method of step S32 includes:

[0133] Step S321. Input preparation

[0134] Input data: receive the aligned feature matrix output by step S31. The matrix contains seven modalities of aligned feature vectors, represented as:

[0135] x1: tongue photo feature vector (dimension d1)

[0136] x2: face photo feature vector (dimension d2)

[0137] x3: nail fold microcirculation photo feature vector (dimension d3)

[0138] x4: sound signal feature vector (dimension d4)

[0139] x5: traditional Chinese medicine questionnaire feature vector (dimension d5)

[0140] x6: pulse wave signal feature vector (dimension d6)

[0141] x7: physiological index feature vector (dimension d7)

[0142] Feature connection: in order to facilitate the processing of the gating neuron, all the modal feature vectors are concatenated to form a global feature vector:

[0143] x all =[x1;x2;x3;x4;x5;x6;x7] (1);

[0144] Wherein, the dimension of x all is This vector captures the global information across modalities for subsequent gating decisions.

[0145] Step S322. Feature encoding

[0146] For each modality i (i = 1, 2, …, 7), use a fully connected neuron with a tanh activation function to encode the input feature x i into a fixed-dimensional internal representation h i .

[0147] Encoding formula:

[0148]

[0149] Where: is the weight matrix, with a size of D h ×d i , used to map x i of dimension d i to a unified dimension D h . The D h of all modalities is set to the same value to ensure the comparability of subsequent fusion; is the bias vector, with a size of D h ; the tanh activation function (output range [-1, 1]) introduces nonlinearity and enhances the representation ability of the feature.

[0150] Output: h i represents the internal encoding feature of modality i, with a dimension of D h .

[0151] Step S323. Gating mechanism:

[0152] Deploy one gating neuron for each modality i, which takes the global feature vector x all Outputs one scalar gating value g for input i i Dynamically quantifies the contribution of modality i to the current input sample, makes decisions based on all modality features, and realizes sample-adaptive attention allocation.

[0153] Gating formula:

[0154] i = σ(U i x all + c i ) (3);

[0155] Where: U i is a weight vector (size 1 x D, where D = dim(x all )) used to learn the decision boundary of modality i; c i is a scalar bias term; σ is a sigmoid activation function (output range [0, 1]) representing the contribution probability: when g i ≈ 1, modality i contributes significantly; when g i ≈ 0, modality i contributes weakly or not at all.

[0156] Step S324. Modulate the output

[0157] Use the gating value g i to weight the internal encoding feature h i to generate the modality i's modulated contribution vector c i This is equivalent to a soft switch mechanism that only allows h i to significantly affect the overall output when g i is high.

[0158] Modulation formula:

[0159] i = g i · h i (4);

[0160] Where: · represents element-level multiplication (broadcast operation) of scalar and vector, and the output c i has dimension D h When g i = 0, c i is a zero vector and modality i has no contribution; when g i = 1, c i fully retains the features of h i .

[0161] Step S325. Fusion output​​​

[0162] The modulation contribution vectors of all modalities are aggregated to generate a final fusion feature representation z as the overall output of the gated attention unit.

[0163] Fusion formula:

[0164]

[0165] where: z has a dimension of D h is a dynamic weighted sum that highlights the features of important modalities.

[0166] Step S33. Based on the gated attention fusion mechanism, a TCM blood and qi classification is performed to obtain a TCM blood and qi classification result output, including:

[0167] Specifically, based on the fusion feature vector z generated by the gated attention fusion mechanism, the model performs a single task: TCM blood and qi classification. This task directly outputs a defined "TCM blood and qi analysis result", including types such as deficiency of blood and qi, blood stasis due to qi stagnation, blood dryness due to yin deficiency, blood cold due to yang deficiency, and damp-heat accumulation.

[0168] Step S331. Task definition and model architecture

[0169] Task definition: Output TCM blood and qi classification results, with each type determined independently (allowing multiple types to coexist). The specific types are as follows:

[0170] Deficiency of blood and qi (corresponding to qi deficiency or blood deficiency state)

[0171] Blood stasis due to qi stagnation (qi stagnation leading to blood stasis)

[0172] Blood dryness due to yin deficiency (yin deficiency leading to blood dryness)

[0173] Blood cold due to yang deficiency (yang deficiency leading to blood cold)

[0174] Damp-heat accumulation (damp-heat accumulation leading to blood stasis or dry heat)

[0175] Model architecture: Sigmoid multi-label classifier is used (as blood and qi types may coexist), with the input being the fusion feature vector z and the output layer having 5 independent neurons (each corresponding to a blood and qi type). Formula:

[0176] p 血气 =σ(W 血气 h+b 血气 ) (6);

[0177] where: h = ReLU(W h z + b h ) is the shared hidden layer feature (dimension d h ), which extracts cross-modal commonalities; W 血气is a weight matrix (size d h 血气 is a bias vector; σ is a Sigmoid function, output probability p i ∈ [0, 1] (i = 1 to 5, corresponding to 5 types). Decision rule: if p i > 0.5, activate the type (such as "Qi stagnation and blood stasis + dampness and heat stagnation").

[0178] Step S332. Training and loss function

[0179] Loss function: binary cross-entropy loss, formula:

[0180]

[0181] where y i is the true label.

[0182] Data source: based on intelligent wearable data annotation, types such as blood deficiency and qi stagnation and blood stasis, training set from clinical annotation of tongue and pulse and other multi-modal data.

[0183] Output logic and explainability

[0184] Original output processing: after the model outputs the probability vector, it is converted into a readable result. Example:

[0185] If p 气滞血瘀 = 0.85 > 0.5, p 湿热壅盛 = 0.72 > 0.5, output "qi stagnation and blood stasis, dampness and heat stagnation".

[0186] If all p i ≤ 0.5, output "no obvious blood qi abnormality".

[0187] Feature tracing: output back to the original features to ensure explainability.

[0188] S4. Based on the PICO framework and the Living systematic review technology, build an evidence graph RAG supporting incremental learning.

[0189] Specifically, referring to Figure 6 , the method of step S4 comprises:

[0190] S41. Based on the PICO framework, combined with the blood qi analysis results of traditional Chinese medicine, obtain the entities that the original evidence graph RAG needs to contain, and set the search terms.

[0191] ​Specifically, based on the PICO framework, combined with the information that needs to be presented in the TCM intelligent blood vessel health management report such as blood gas analysis results and dietary recommendations, the entities that need to be included in the original evidence graph are designed, such as setting the Patient in the PICO framework as "patients with low blood oxygen value".

[0192] According to the determined entity, the retrieval word is set, such as the above-mentioned Patient can be set as "Low blood oxygen" retrieval word, and Pubmed, Web of Science and other literature platforms that need to be considered are retrieved.

[0193] S42. Based on the evidence-based medicine evidence level principle and PICO matching logic, a five-dimensional hierarchical scoring system including research design scientificity, methodological rigor, PICO element adaptability, academic influence, and result practicality is constructed.

[0194] S43. The comprehensive score is calculated by weighted summation method, combined with total score threshold screening and single extremely low score forced exclusion rule, to obtain evidence scientific correlation target problem.

[0195] Specifically, the comprehensive score is calculated by weighted summation method, combined with total score threshold screening and single extremely low score forced exclusion rule, to ensure that the evidence scientific correlation target problem is ensured. For example, literature with a total score of less than 60 or a single score of less than 60 (assuming that the single and total scores are set to 100) is excluded.

[0196] S44. Knowledge extraction is performed on the screened high-quality literature, and entity alignment, relationship and attribute fusion operations are performed to construct a unified initial evidence graph.

[0197] Specifically, knowledge extraction is performed on the screened high-quality literature based on evidence-based medicine, including entity extraction and relationship extraction. Subsequently, entity alignment, relationship and attribute fusion operations are performed to eliminate redundancy and conflicts in the evidence graph RAG, and a unified initial evidence graph is constructed.

[0198] S45. Based on the initial evidence graph and TCM blood gas analysis results, design retrieval words that need to be updated.

[0199] Specifically, based on the existing entities in the initial evidence graph and the information that needs to be presented in the TCM blood gas analysis results, etc., the retrieval words that need to be updated are designed.

[0200] S46. Based on the API literature retrieval interface, the updated retrieval words are obtained, and named entity recognition, relationship extraction, literature scoring analysis, and conflict checking are performed to screen out the clinical evidence that needs to be updated;

[0201] S47. Update the updated clinical evidence into the initial evidence graph to obtain the evidence graph RAG.

[0202] Specifically, new clinical evidence is updated into the evidence graph RAG to realize incremental learning and obtain the evidence graph RAG supporting incremental learning.

[0203] S5. Input the TCM blood and qi analysis result into the pre-trained large language model, and combine the evidence graph RAG to obtain a TCM health management report.

[0204] Specifically, the method of pre-training a large language model includes:

[0205] Model design and meta-training:

[0206] Task definition and data preparation: The medical advice generation task is decomposed into six sub-tasks of diet therapy suggestion, seasonal conditioning suggestion, living conditioning suggestion, spiritual cultivation suggestion, exercise health preservation suggestion, and meridian health preservation suggestion. Each sub-task constructs a support set and a query set of 10 samples. Information such as patient symptoms and constitution and corresponding suggestions are extracted from electronic medical records, TCM classics and other multi-source data, and the task distribution is constructed according to season, disease type, etc.

[0207] Model architecture design: An encoder-decoder architecture is adopted, the encoder uses a BERT-like model to process patient text information, and the decoder designs a diet therapy suggestion head and multiple task heads. A specific prompt template is designed for each sub-task, such as: "Give diet therapy suggestions based on this blood and qi result", and a unified output format is adopted.

[0208] MAML meta-training process: First, set the initial parameters θ of the model, then randomly select a specific task T (such as "conditioning plan for summer qi deficiency constitution") from the task library. Update the model 1-2 times with a small number of samples (support set) of task T to obtain temporary parameters θ', which preliminarily adapt to task T. Verify the effect of parameter θ' on the test samples (query set) of task T, calculate the loss value L(θ'), and use it to measure the adaptability of the model to new tasks. Adjust the initial parameters θ in the reverse direction according to the loss value L(θ'), and repeat the above process until θ can quickly adapt to new tasks, i.e. obtain the ability to "learn how to learn".

[0209] Meta-testing and deployment:

[0210] For new patient cases, use the converged θ as the initial point to quickly fine-tune the parameters of each task head with a small amount of feature data, generate personalized medical health preservation suggestions, and integrate them into the intelligent medical system after evaluation by clinical experts.

[0211] Construct the CoT template input to the LLM to obtain a TCM health management report, for example:

[0212] First step: Contact the theory of "four diagnoses combined with reference" in traditional Chinese medicine, comprehensively analyze the core indicators of blood gas analysis, correlate the health preservation direction of different indicators, and clarify the core target of conditioning (such as eliminating dampness and tonifying yang), to ensure that the subsequent suggestions are developed around the target.

[0213] Second step: Generate suggestions by dimension. For example: diet matching constitution (selecting Chinese yam to tonify qi for qi deficiency), and seasonal conditioning combined with seasonal changes (keeping warm for yang deficiency in winter).

[0214] Third step: Check the logical consistency. Check whether each suggestion can be traced back to the blood gas indicators (such as avoiding high protein and high salt in diet therapy for acidosis), to ensure that there is no contradiction, and finally integrate into a personalized plan.

[0215] Call DeepSeek R1 large language model for analysis:

[0216] Input the user's blood gas analysis results into the DeepSeek R1 model, DeepSeek combines the blood gas analysis results, information in the evidence graph RAG, and the user's traditional Chinese medicine blood vessel health management plan based on the reasoning of the CoT template.

[0217] The present application integrates traditional Chinese medicine four diagnostic data, nail microcirculation images and modern physiological indicators, constructs a structured report, and solves the problem of heterogeneous data difficult to be saved together. The present application carries out data standardization conversion and cross-system sharing based on FHIR, breaks the data barriers between medical institutions. The present application combines machine learning and large language model, realizes traditional Chinese medicine blood gas analysis, and provides user health management suggestions based on evidence-based medicine. Through continuous data acquisition, the time trend curve of body temperature, blood pressure and other indicators is generated, and the health risk warning ability is improved.

[0218] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for constructing a TCM-based intelligent vascular health management model, characterized in that, include: S1. Based on smart wearable devices, acquire multimodal data including traditional Chinese medicine diagnostic data, image data, and various physiological indicators; S2. Encrypt and encode multimodal data based on the FHIR standard to generate a structured report of multimodal data that supports cross-system sharing; S3. The multimodal data structured report is parsed based on multiple learning models to obtain the results of TCM blood and qi analysis; wherein, the learning model includes the improved U-Net model. S4. Based on the PICO framework and Living systematic review technology, construct an evidence graph (RAG) that supports incremental learning; S5. Input the TCM blood and qi analysis results into the pre-trained large language model and combine them with the evidence graph RAG to obtain a TCM health management report.

2. The method according to claim 1, characterized in that, The method of step S1 includes: S11. Based on the camera component, acquire facial diagnostic data including tongue images, facial images, and nailfold microcirculation images; S12. Based on the recording component, acquire auscultation data; S13. Based on the results of the questionnaire on TCM constitution differentiation, obtain consultation data; S14. Based on the pulse wave waveform acquired by the pulse wave acquisition module, obtain palpation data; S15. Based on sensor monitoring devices, acquire physiological data including body temperature, blood oxygen, blood pressure, heart rate, and arterial stiffness.

3. The method according to claim 1, characterized in that, The method of step S3 includes: S31. Based on the structured report of multimodal data, perform feature alignment and structured representation of multimodal data, and obtain the aligned feature matrix; S32. Based on the aligned feature matrix, construct a gated attention fusion mechanism; S33. Based on the gating attention fusion mechanism, perform TCM blood and qi classification and output the TCM blood and qi classification results.

4. The method according to claim 3, characterized in that, The method of step S32 includes: S321. Construct a global feature vector based on the aligned feature matrix; S322. Encode the global feature vector to obtain the encoded features; S323. Based on the obtained encoding features and the gated neurons, obtain the gating mechanism; S324. Based on the gating mechanism, perform modulation output.

5. The method according to claim 2, characterized in that, In step S3, The results of the TCM blood and qi analysis include constitution type determination, blood stasis type, grading of vascular aging degree, and semi-quantitative assessment of organ function.

6. The method according to claim 5, characterized in that, The method of step S4 includes: S41. Based on the PICO framework and combined with the results of the TCM blood and qi analysis, obtain the entities that the original evidence map RAG needs to include, and set the search terms; S42. Based on the evidence-based medicine evidence level principle and PICO matching logic, a five-dimensional hierarchical scoring system is constructed, including the scientific nature of the research design, the rigor of the methodology, the fit of PICO elements, the academic influence, and the practicality of the results. S43. Calculate the comprehensive score by weighted summation, and combine the total score threshold screening and the rule of mandatory elimination of particularly low scores in a single item to obtain evidence-based scientific relevance to the target issue; S44. Extract knowledge from the selected high-quality documents and perform entity alignment, relationship and attribute fusion operations to construct a unified initial evidence graph; S45. Based on the initial evidence map and the results of the traditional Chinese medicine blood and qi analysis, design search terms that require knowledge updates; S46. Based on the API literature retrieval interface, obtain the updated search terms, and then perform named entity recognition, relation extraction, literature scoring and parsing, and conflict verification to filter out the clinical evidence that needs to be updated. S47. Update the initial evidence map with the updated clinical evidence to obtain the evidence map RAG.

7. The method according to claim 6, characterized in that, The TCM health management report includes blood and qi analysis results, dietary therapy suggestions, seasonal conditioning suggestions, daily life conditioning suggestions, mental cultivation suggestions, exercise and health preservation suggestions, and acupoint health preservation suggestions.

Images