A traditional chinese medicine multi-source combined reference health self-adaptive evaluation system and method

Abstract

The application relates to a traditional Chinese medicine multi-source combined health self-adaptive evaluation system and method, which comprises a data acquisition and preprocessing unit block, an initial evaluation model construction unit and a dynamic optimization and self-adaptive correction unit. The data acquisition and preprocessing unit continuously collects actual health change data of a user after intervention. The dynamic optimization and self-adaptive correction unit drives iterative correction of weights according to the difference between the actual health change data and the model prediction evaluation result, thereby forming a dynamic self-adaptive health evaluation system with a closed-loop feedback of "evaluation-simulation-feedback-optimization", and significantly improving creativity and scientificity. The comprehensive evaluation model can evolve autonomously with the time lapse of the health state of the user, and truly realizes individualized dynamic precise evaluation of traditional Chinese medicine "according to the person, time and local conditions", and effectively solves the technical problem that the prior art cannot simultaneously consider "group universality" and "individual specificity".

Description

Technical Field

[0001] This invention belongs to the field of comprehensive health evaluation technology, and relates to a health adaptive evaluation system and method that integrates multiple sources of traditional Chinese medicine. Background Technology

[0002] In the process of TCM diagnosis and treatment, the TCM diagnostic methods of "observation, auscultation, inquiry, and palpation" are digitized and presented. Based on existing TCM diagnostic instruments, infrared thermal imaging examinations, and the patient's basic information (present medical history and past medical history, etc.), computer technology and mathematical methods are used to fuse and process clinical data from multiple sources and multiple modalities to reflect the human body's health status from different dimensions and provide a health score.

[0003] Currently, preliminary explorations have been made in intelligent TCM diagnosis and health management systems based on artificial intelligence. For example, Reference 1 (CN120884255B: Intelligent TCM Diagnosis System Based on Multi-Source Data Fusion and AI Technology) discloses an intelligent TCM diagnosis system based on multi-source data fusion and AI technology. It uses the Analytic Hierarchy Process (AHP) to calculate the static weights of tongue diagnosis, facial diagnosis, and pulse diagnosis, and performs fusion inference through graph neural networks. Reference 2 (CN121687495A: A Method for Constructing a Health Service System Integrating Traditional Chinese and Western Medicine) discloses a method for constructing a health service system integrating traditional Chinese and Western medicine. It adopts a dual-channel deep learning architecture to integrate the characteristics of traditional Chinese and Western medicine and uses the Monte Carlo method for data simulation. In addition, Reference 3 (CN120853900A: A Multimodal Fusion Intelligent TCM Constitution Assessment System) also discloses a multimodal fusion intelligent TCM constitution assessment system, which uses reinforcement learning to dynamically adjust modal weights.

[0004] However, the aforementioned existing technologies still have the following common problems and defects: (1) Static or one-way weight allocation, lacking closed-loop feedback verification: The weights of most systems are determined once by expert experience, belonging to static AHP models. Although the scheme in Reference 3 introduces dynamic weights, its adjustment is mainly based on data quality (such as clarity and signal-to-noise ratio), lacking an independent, top-down verification and feedback mechanism to verify the scientific nature of the weight allocation and pass it back to the weight generation model. This results in the weights being fixed for a long time once determined, unable to adapt to the dynamic changes of individual differences and clinical practice.

[0005] (2) The evaluation dimensions are too simplistic, lacking a key assessment of the impact of "behavioral compliance" on health outcomes: Existing systems generally focus on physiological data (tongue appearance, pulse appearance, infrared radiation, etc.), neglecting the user's implementation of the health management plan (i.e., compliance). Traditional Chinese medicine health management emphasizes "three parts treatment and seven parts maintenance," and the user's self-management behavior is the core variable affecting the final health outcome. Existing technologies have failed to incorporate "compliance" as a dynamic and quantifiable evaluation factor into the comprehensive health scoring system, resulting in the evaluation results failing to truly reflect the user's expected recovery trend under intervention.

[0006] (3) Data quality and weight adjustment are mostly passive responses, lacking proactive simulation prediction and optimization: Existing technologies mostly adjust weights passively based on data quality after data collection. They lack a proactive simulation module that can use Monte Carlo simulation to stress test the existing weight system before or during evaluation, predict its performance under various data conditions, and proactively optimize the model structure accordingly. Summary of the Invention

[0007] To address the problems existing in the above-mentioned traditional methods, this invention proposes a health adaptive evaluation system and method based on multiple sources of traditional Chinese medicine.

[0008] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: On the one hand, a multi-source integrated health adaptive assessment system based on traditional Chinese medicine is provided, including: The data acquisition and preprocessing unit is used to acquire and preprocess multi-source heterogeneous TCM health data, which includes at least multimodal detection data, personal health record data, and time-series behavioral data reflecting intervention responses.

[0009] The initial evaluation model construction unit is used to generate initial weight allocations for each data source based on a preset multi-criteria initial weighting strategy, and to construct a comprehensive health evaluation model based on the initial weights to output health evaluation values; wherein, the initial weighting strategy is constrained by prior domain knowledge.

[0010] The dynamic optimization and adaptive correction unit is used to perform perturbation simulation and sensitive factor identification on the current comprehensive health assessment model based on the uncertainty propagation analysis strategy, and adaptively correct the initial weights according to the sensitive factor identification results. The corrected dynamic weights are then fed back to the initial assessment model construction unit to update the comprehensive health assessment model.

[0011] The data acquisition and preprocessing unit continuously collects actual health change data after users receive intervention, while the dynamic optimization and adaptive correction unit drives the iterative correction of weights based on the deviation between the actual health change data and the predicted evaluation results of the comprehensive health evaluation model, forming a closed-loop feedback dynamic adaptive mechanism between the comprehensive health evaluation model and the real data.

[0012] On the other hand, a TCM health adaptive assessment method based on multi-source data fusion is also provided, including the following steps: Step S1: Acquire and preprocess multi-source heterogeneous TCM health data. The TCM health data includes at least multimodal detection data, personal health record data, and time-series behavioral data reflecting intervention response.

[0013] Step S2: Generate initial weight allocation for each data source based on a preset multi-criteria initial weighting strategy, and construct a comprehensive health evaluation model based on the initial weights to output health evaluation values; wherein, the initial weighting strategy is constrained by prior domain knowledge.

[0014] Step S3: Based on the uncertainty propagation analysis strategy, the current comprehensive health assessment model is subjected to perturbation simulation and sensitive factor identification. The initial weights are adaptively corrected according to the sensitive factor identification results, and the corrected dynamic weights are fed back to update the comprehensive health assessment model.

[0015] Step S4: Continuously collect the user's actual health evolution data after receiving intervention measures. Based on the deviation between the actual health evolution data and the prediction results output by the comprehensive health evaluation model, drive the iterative correction of weights to form a closed-loop feedback dynamic adaptive adjustment.

[0016] One of the above technical solutions has the following advantages and beneficial effects: The aforementioned TCM multi-source adaptive health evaluation system and method comprises: a data acquisition and preprocessing unit, an initial evaluation model construction unit, and a dynamic optimization and adaptive correction unit. The data acquisition and preprocessing unit continuously collects actual health change data after the user receives intervention. The dynamic optimization and adaptive correction unit drives iterative correction of weights based on the difference between the actual health change data and the model's predicted evaluation results, thus forming a closed-loop feedback dynamic adaptive health evaluation system of "evaluation-simulation-feedback-optimization," significantly improving creativity and scientific rigor. The comprehensive evaluation model can autonomously evolve with the user's health status over time, truly realizing the individualized, dynamic, and precise evaluation of TCM based on "person, time, and place," effectively solving the technical challenge of existing technologies that struggle to balance "general applicability to the population" and "individual specificity." Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a block diagram of a health adaptive evaluation system based on multiple sources of traditional Chinese medicine in one embodiment. Figure 2 This is a schematic diagram illustrating the score calculation using piecewise linear interpolation in one embodiment; Figure 3 This is a schematic diagram of the probability distribution of the comprehensive health score in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0021] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] In one embodiment, such as Figure 1 As shown, a multi-source integrated health adaptive assessment system based on traditional Chinese medicine is provided, including: The data acquisition and preprocessing unit is used to acquire and preprocess multi-source heterogeneous TCM health data, which includes at least multimodal detection data, personal health record data, and time-series behavioral data reflecting intervention responses.

[0024] Specifically, the data acquisition and preprocessing unit is responsible for collecting raw data from various devices, questionnaires, wearable devices and user behavior records, and then performing standardized processing such as cleaning, transformation, labeling and alignment to output structured data in a unified format for subsequent analysis and use in the "digital human" model.

[0025] Multimodal TCM testing data includes, but is not limited to, infrared thermal imaging (whole body and head and face), tongue diagnosis images, pulse diagnosis waveforms, meridian electrical characteristics, psychological and emotional questionnaires, and infrared psychological assessment data.

[0026] (1) Acquisition of infrared thermal imaging (whole body and head and face) Data acquisition equipment: Medical-grade infrared thermal imager (resolution ≥320×240, temperature sensitivity ≤0.05℃) for data acquisition; Collection environment: constant temperature (22~25℃), no wind, no direct light, subject rests for 20 minutes before standing or sitting; Photo capture process: Full body: one photo each from the front, back, left side, and right side, at a distance of 2 meters. Head and face: one photo each from the front and left / right 45° angles, at a distance of 1 meter; Standardized output: Converted to a 16-bit floating-point temperature matrix (PNG / RAW format), including ambient temperature, humidity, acquisition time, and subject ID. Standardized pixel resolution and body position labeling (e.g., "Face_Front View").

[0027] (2) Acquisition of tongue diagnosis images Data collection equipment: Professional tongue diagnostic instrument (including D65 standard light source and color calibration card).

[0028] Data collection process: 1) Tongue image: naturally extend the tongue with the tip pointing downwards, and take pictures of the front of the tongue and the sublingual veins; 2) Facial examination: front, left cheek, and right cheek, with natural expression; 3) Take 3 pictures of each part in a row, and select the one with the highest clarity.

[0029] Standardized output: JPEG / PNG format, embedded ICC color profile, with additional tongue body section labels (such as "tongue root", "tongue tip", "tongue middle"), and a uniform resolution of 1920×1080.

[0030] (3) Pulse waveform acquisition Data acquisition equipment: Three pulse sensors (pressure type or strain type), sampling rate ≥100 Hz.

[0031] Data acquisition process: 1) Collect the pressure at three levels: “cun”, “guan”, and “chi”, and the pressure at three levels: light, medium, and heavy; 2) Collect a 30-second continuous waveform for each level and each pressure.

[0032] Standardized output: CSV / JSON format, including time series (milliseconds), pressure value (kPa), pressure depth (mm), and labeled with feature tags such as pulse position, pulse strength, and pulse momentum.

[0033] (4) Meridian electrical characteristics Data acquisition equipment: Meridian detector (four-electrode method, frequency 10~200 kHz).

[0034] Data collection process: 1) Detect the source points or well points of 24 meridians and measure resistance, capacitance, and impedance spectrum. 2) Repeat the data collection for each point 3 times and take the median.

[0035] Standardized output: tabular data (acupoint name, resistance value Ω, capacitance value pF, phase angle °), with baseline calibration values.

[0036] (5) Psychological and emotional questionnaires and infrared psychological assessment data 1) Psychological and Emotional Questionnaire: Tools: Traditional Chinese Medicine emotional scales (such as the simplified version of SCL-90, the Seven Emotions Self-Rating Scale).

[0037] Data collection method: Fill in the form on an electronic tablet or mobile phone. The answering time and options will be automatically recorded.

[0038] Standardized output: ratings for each dimension (joy, anger, worry, thought, grief, fear, and surprise) and total score, in structured JSON format.

[0039] 2) Infrared psychological assessment: Equipment: Shares equipment with infrared thermal imaging, and collects facial infrared images once before, during and after emotion-inducing tasks (such as recalling stressful events or viewing emotional images).

[0040] Standardized output: Temperature change matrix (ΔT), combined with temperature time history curves of ROIs such as nose tip, forehead, and eye socket.

[0041] Personal basic information and medical history mainly include basic personal information (age, gender, occupation, lifestyle habits, etc.), present illness (current chief complaint, course of illness, treatment process), past medical history (past illnesses, surgeries, injuries, allergies, etc.), and family history (health status of relatives). The data collection process needs to balance structured fields (for easy statistics) and free text (to retain the details required for TCM diagnosis), and standardized terms should be extracted using natural language processing technology.

[0042] (1) Data source Structured data entry: Collect information such as name, age, gender, height, weight, occupation, and contact information through electronic forms (Web / APP).

[0043] Free text entry: Doctors or patients fill in their current medical history, past medical history (such as hypertension, diabetes), family history (diseases of immediate family members), allergy history, etc.

[0044] (2) Data collection process Form design: Use hierarchical linkage (such as disease name dropdown search + onset time + treatment status) to reduce free text.

[0045] Text collection: Provide text fields and guide structured descriptions (e.g., "chief complaint: symptoms + duration + triggering factors").

[0046] Standardized output: 1) Basic information: stored in fields (e.g., age:45, gender:"M"). 2) Medical history text: entity recognition using NLP (disease, time, medication, family relationship), converted into ICD-11 encoding and a custom TCM syndrome terminology (e.g., "liver qi stagnation"), and the original text is retained as evidence.

[0047] Compliance and behavioral data: The implementation of intervention programs (such as exercise, diet, and medication) by users is recorded through APP check-in, wearable devices, etc., forming a "digital human" light-on / light-off model (the "digital human" is fully lit when the program is fully implemented, and the corresponding part is turned off when a project is not completed).

[0048] (1) Data source 1) App check-in: Users manually mark the completion status of tasks such as exercise, diet, medication, rest, and emotional regulation.

[0049] 2) Wearable devices: smart bracelets / watches (steps, heart rate, sleep quality, activity duration), smart pillboxes (medication records), smart body fat scales (weight, body fat percentage).

[0050] (2) Data collection process 1) Intervention plan distribution: The system generates a daily task list based on the user's physical condition and diagnosis (e.g., "Baduanjin for 30 minutes; take Xiaoyao San once in the morning and once in the evening; sit quietly for 10 minutes at noon").

[0051] 2) Execution Log: App check-in: Check "Completed" for each item in the task list, and attach the check-in timestamp and GPS location (optional anti-cheating measures).

[0052] Wearable devices: Automatically upload exercise duration, heart rate zones, sleep cycles, etc., and associate them with tasks.

[0053] 3) Missing records: If the system does not receive a completion signal by the deadline (e.g., 23:59 on the same day), it will automatically mark it as "incomplete".

[0054] (3) Data generation of digital human light-on / light-off model Model definition: Each intervention sub-item (such as exercise, diet A, diet B, medication A, medication B, sleep, emotional exercise) corresponds to a part of the "digital human" (head, hands, torso, legs, etc.); when all tasks are completed, the digital human lights up (green / gold); if a sub-item is not completed, the corresponding part turns off (gray / red), while the other parts remain lit.

[0055] Data collection and output: daily compliance record sheet, digital human status vector, and visualized data (which can be output as JSON or directly drive the front-end 3D model).

[0056] The initial evaluation model construction unit is used to generate the initial weight allocation of each data source based on the preset multi-criteria initial weighting strategy, and to construct a comprehensive health evaluation model based on the initial weights to output the health evaluation value; wherein, the initial weighting strategy is constrained by prior domain knowledge.

[0057] Specifically, the core of the initial evaluation model building unit is to build a closed-loop weight calculation engine of "analytic hierarchy process-Monte Carlo feedback", rather than a simple static AHP.

[0058] Experts were invited to conduct pairwise comparisons of criteria-level indicators such as "infrared norms," ​​"physical fitness," and "mental health," as well as specific data sources (such as online / offline tongue diagnosis and consultation) under each indicator, to construct a judgment matrix; the initial weight vectors of each level were calculated, and consistency checks were performed.

[0059] The initial evaluation model construction unit, as the core decision-making hub of this system, is connected to the data acquisition and preprocessing unit. It is used to receive multi-source heterogeneous data after standardization and to perform initial weight allocation and comprehensive score calculation.

[0060] The pre-defined multi-criteria initial weighting strategy refers to: for multiple evaluation dimensions (e.g., but not limited to, the four diagnostic methods, equipment testing indicators, subjective feelings, and behavioral compliance) contained in multi-source heterogeneous data, one or more multi-criteria decision analysis methods are used to calculate the initial importance coefficient (i.e., initial weight allocation) of each dimension in the overall health evaluation.

[0061] Multi-criteria initial weighting strategies include, but are not limited to, any one or more combinations of the following: (1) Subjective weighting methods based on expert experience, such as the Analytic Hierarchy Process (AHP), the Delphi method, or the Analytic Network Process (ANP). (2) Objective weighting methods based on the information content of the data itself, such as entropy weighting, standard deviation method or CRITIC method; (3) The subjective and objective weighting method is to combine the above subjective weights and objective weights to take into account both clinical experience and objective data laws.

[0062] After completing the weight allocation, the initial evaluation model construction unit performs a comprehensive operation on the initial weight vector and the standardized feature values ​​of each data source (preferably using linear weighted summation, weighted product, or fuzzy comprehensive evaluation synthesis operator) to output a quantitative initial health evaluation value.

[0063] To ensure that the output of the above multi-criteria initial weighting strategy conforms to the basic theories and clinical logic of traditional Chinese medicine, and to avoid abnormal weight combinations that contradict the holistic concept and syndrome differentiation and treatment of traditional Chinese medicine generated by pure mathematical operations, the initial weighting strategy is constrained by prior domain knowledge during the weight generation and modification process.

[0064] Prior domain knowledge constraints refer to a set of domain knowledge constraints that are pre-built and stored in the system. This set covers at least the deep domain rules such as the logic of etiology and pathogenesis in traditional Chinese medicine, the rules of association between viscera and meridians, the hierarchical relationship between constitution and syndrome, and the principle of monarch-minister-assistant-guide combination.

[0065] Prior domain knowledge constraints are preferably solidified in the system in the form of a TCM knowledge graph and a structured rule base. This graph is constructed based on standards such as "TCM Clinical Diagnosis and Treatment Terminology" and "TCM Constitution Classification and Determination". However, prior domain knowledge constraints are not limited to the form of knowledge graphs. They can also be expressed in alternative forms such as expert experience matrices, Bayesian prior distribution parameters, clinical path logic trees, or sets of mathematical inequalities. As long as they can impose non-data-dependent constraints based on TCM theory during the weight generation stage, they all fall within the scope of "prior domain knowledge constraints" described in this application.

[0066] The dynamic optimization and adaptive correction unit performs perturbation simulation and sensitive factor identification on the current comprehensive health assessment model based on the uncertainty propagation analysis strategy. It then adaptively corrects the initial weights based on the sensitive factor identification results and feeds the corrected dynamic weights back to the initial assessment model construction unit to update the comprehensive health assessment model.

[0067] Specifically, the core function of the dynamic optimization and adaptive correction unit is to quantitatively characterize the multi-source uncertainty in the comprehensive health evaluation model, track the propagation path of this uncertainty between different levels of the model, and guide the dynamic adjustment of weights based on the propagation analysis results, thereby realizing the evolution and upgrade of the evaluation model from "based on prior static assignment" to "based on data adaptive drive".

[0068] Uncertainty propagation analysis strategy refers to a systematic analysis method that uses mathematical means to quantitatively simulate how various input uncertainties (including but not limited to measurement errors of detection data, subjective response bias of users, estimation uncertainty caused by missing data, and intervention time-series fluctuations) in a comprehensive health evaluation model are transmitted and ultimately affect the stability of the output health evaluation value after weighted summation, nonlinear transformation and comprehensive synthesis operations within the model.

[0069] Perturbation simulation specifically refers to the active application of pre-defined amplitude and type of incentive changes (i.e., perturbations) to the input layer (i.e., standardized feature values ​​of each data source) and structural layer (i.e., the current initial weight allocation vector) of a comprehensive health assessment model within the framework of uncertainty propagation analysis, in order to simulate various non-ideal working conditions in real clinical scenarios, such as data collection errors, fluctuations in user compliance, and drift of physiological parameters over different time periods.

[0070] Uncertainty propagation analysis strategies include, but are not limited to, Monte Carlo simulation, probability box and interval analysis, or Latin hypercube sampling and quasi-Monte Carlo method.

[0071] Sensitive factor identification refers to using quantitative global sensitivity analysis to calculate and rank the contribution of each source of input uncertainty to the variability of the output health evaluation value based on a large number of input-output data pairs generated by the above-mentioned perturbation simulation. This allows for the precise identification of the key data sources and weight dimensions in the current evaluation model that have the most significant impact on the final score or that deviate the most from the prior TCM theory expectations.

[0072] For the identification of sensitive factors, the preferred method is global sensitivity analysis based on variance (such as the Sobol index method). Alternatively, a sensitivity analysis based on regression or a preliminary screening analysis based on the Morris method can be used.

[0073] The dynamic optimization and adaptive correction unit transmits the corrected dynamic weight vector to the initial evaluation model construction unit in real time or at a preset update cycle (such as before each follow-up visit or after every 5 sets of new data). Upon receiving this feedback, the initial evaluation model construction unit replaces the existing weight allocation stored in the current model with the dynamic weight vector, and re-executes the comprehensive health evaluation calculation based on the new weights, thereby generating an updated comprehensive health evaluation model.

[0074] It is important to emphasize that this feedback is not a one-time operation, but a recursive iterative process triggered successively as the data acquisition and preprocessing unit continuously collects actual user health evolution data. Each time the model is updated, it is again fed into the dynamic optimization and adaptive correction unit for a new round of uncertainty propagation analysis and sensitive factor identification, thus forming a continuous cycle of "evaluation → simulation → identification → correction → re-evaluation". As the number of iterations increases, the weight vector of the comprehensive health evaluation model gradually converges to the optimal stable solution that fits the current individual user's health status change pattern, achieving a precise transfer from the group prior model to the individual dynamic model.

[0075] By actively simulating various scenarios with missing and noisy data using Monte Carlo simulation and pre-determining weight correction strategies based on sensitivity analysis results, the evaluation system of this invention can still provide stable and reliable health scores when faced with complex and imperfect data in the real world, greatly reducing the dependence on perfect data input.

[0076] The corrected dynamic weights W adaptive The results are then re-introduced into the stratified comprehensive health scoring model (comprehensive score = weight 1 × score1 + weight 2 × score2 + weight 3 × score3) to calculate the final comprehensive health score, health ranking, key risk warnings, and various sub-scores (overall assessment, comprehensive assessment, individual assessment, chronic disease management, and key population management). This result accurately reflects the most objective and robust health status under current data conditions and individual adherence.

[0077] The data acquisition and preprocessing unit continuously collects actual health change data after users receive intervention, while the dynamic optimization and adaptive correction unit drives the iterative correction of weights based on the deviation between the actual health change data and the predicted evaluation results of the comprehensive health evaluation model, forming a closed-loop feedback dynamic adaptive mechanism between the comprehensive health evaluation model and the real data.

[0078] The initial evaluation model construction unit in the TCM multi-source adaptive health evaluation system provides a reasonable starting point constrained by prior knowledge, offering a feasible benchmark domain that does not deviate from the basic TCM paradigm for subsequent optimization. The dynamic optimization and adaptive correction unit, based on this reasonable starting point, performs efficient local perturbation and sensitivity optimization, achieving faster convergence and lower computational resource consumption compared to blind search starting from completely random weights. The real-time feedback data provided by the data acquisition and preprocessing unit continuously verifies and corrects the prior assumptions of the previous two units, constantly feeding objective clinical facts back into the model's underlying logic. The synergistic effect of the data acquisition and preprocessing unit, the initial evaluation model construction unit, and the dynamic optimization and adaptive correction unit enables the system to simultaneously possess the rigor of being "derived from theory (prior constraints)," the objectivity of being "derived from data (perturbation simulation)," and the self-evolving capability of being "derived from practice (closed-loop feedback)."

[0079] The aforementioned TCM multi-source adaptive health evaluation system comprises: a data acquisition and preprocessing unit, an initial evaluation model construction unit, and a dynamic optimization and adaptive correction unit. The data acquisition and preprocessing unit continuously collects actual health change data after the user receives intervention. The dynamic optimization and adaptive correction unit drives iterative correction of weights based on the difference between the actual health change data and the model's predicted evaluation results, thus forming a closed-loop feedback dynamic adaptive health evaluation system of "evaluation-simulation-feedback-optimization," significantly improving creativity and scientific rigor. The comprehensive evaluation model can autonomously evolve with the user's health status over time, truly realizing the individualized, dynamic, and precise evaluation of TCM based on "person, time, and place," effectively solving the technical challenge of existing technologies that struggle to balance "general applicability to the population" and "individual specificity."

[0080] In one embodiment, the multimodal detection data includes: whole-body and head and face infrared thermal imaging, tongue diagnosis images, pulse diagnosis waveforms, meridian electrical properties, psychological and emotional questionnaires, and infrared psychological assessment data.

[0081] Personal health record data includes: present medical history, past medical history, family history, and textual data of basic personal information.

[0082] Time-series behavioral data reflecting intervention response: User execution of the intervention plan is recorded through app check-ins and / or wearable devices, forming a digital human "lights-on" and "lights-off" model.

[0083] In one embodiment, the process of constructing a hierarchical health comprehensive scoring model in the initial evaluation model construction unit includes: calculating the basic information score; the basic information score is used to characterize the completeness of personal basic information, medical history information, and public health status; calculating the 360° assessment score; the 360° assessment score is dynamically calculated based on the initial weights generated by the analytic hierarchy process (AHP) to obtain the TCM digital score and the efficacy and constitution score; wherein, the TCM digital score refers to the individual scores of tongue diagnosis, pulse diagnosis, inquiry, infrared, meridians, and psychological emotions, and each score is assigned weight according to different collection scenarios such as online, offline, and in-person; the efficacy and constitution score refers to the constitution assessment score determined by referring to the "Standards for TCM Health Management Services" based on the change rate of TCM syndrome symptom scores; calculating the compliance score and intervention performance score; the compliance score is: The intervention performance score is calculated by setting minimum, baseline, and effort values ​​for quantitative key performance indicators and using piecewise linear interpolation. The basic information score, 360° assessment score, compliance score, and intervention performance score are weighted and summed based on the initial weight vector to generate a preliminary comprehensive score.

[0084] Specifically, the stratified comprehensive health scoring process includes: Basic Information Score (Score1): Calculates the completeness score of personal basic information, medical history information, and public health status.

[0085] 360° Assessment Score (Score2): This is the core evaluation layer of this application. Scores for each item are dynamically calculated based on the initial weights in the Application Performance Hierarchy Process (AHP).

[0086] Traditional Chinese Medicine Digital Score: It integrates individual scores from tongue diagnosis, pulse diagnosis, inquiry, infrared, meridians, and psychology / emotions. Each score is further subdivided into different scenarios such as online, in-person, and offline, and assigned weights accordingly.

[0087] Therapeutic effect and physical condition score: Calculate the rate of change of TCM syndrome / symptom score, and the physical condition assessment score with reference to the "Standards for TCM Health Management Services".

[0088] Compliance and Intervention Performance Rating (Score 3): 1) Compliance score: Number of completed check-in items / Total number of completed check-in items × 100.

[0089] 2) Target Achievement Scoring: Set minimum, baseline, and effort values ​​for quantitative KPIs, and calculate scores using piecewise linear interpolation (e.g., ...). Figure 2 (As shown).

[0090] Preliminary overall score = weight 1 × score 1 + weight 2 × score 2 + weight 3 × score 3.

[0091] In one embodiment, prior domain knowledge constraints include: prior constraints built into the TCM knowledge graph and rule base; specifically, they include: a holistic constraint rule on cold / heat and deficiency / excess: based on the TCM Eight Principles of Differentiation, the sum of the weights of cold-related indicators and heat-related indicators shall not be lower than a preset threshold, and the difference in weights between cold-related indicators and heat-related indicators shall not exceed a preset range; wherein, cold-related indicators include: aversion to cold, cold limbs, and preference for hot drinks; heat-related indicators include: fever, thirst, and preference for cold drinks; and a constraint rule on the transmission of organ connections, used to define strongly correlated indicator groups based on preset mutual generation and mutual restraint mapping relationships and organ exterior-interior relationships, and automatically generate joint constraint equations through the association rule base. The weight lower limit protection mechanism is used to set an absolute weight lower limit for indicators that reflect the core pathogenesis of TCM and whose weights are clinically recognized as not being too low. When the weight of an indicator falls below the absolute weight lower limit due to Monte Carlo simulation or missing data, the weight is forcibly pulled back to the absolute weight lower limit, and a backup weight transfer is triggered to transfer the importance of the missing part to other observable indicators that are most correlated with the syndrome of that indicator. The compliance-syndrome response rule, in view of the TCM characteristic of "three parts treatment and seven parts maintenance", predefines the sensitivity of different syndrome types to compliance. It is used as an adjustment coefficient for the compliance sub-dimension under the criterion layer when setting the initial weights in the analytic hierarchy process, so that the weight allocation conforms to the individualized diagnosis and treatment logic of TCM.

[0092] Specifically, the TCM knowledge graph and rule base serve as prior constraints, and the built-in core constraint rules include: (1) Holistic constraint of cold and heat, deficiency and excess: Based on the theory of "eight principles of syndrome differentiation" in traditional Chinese medicine, it is stipulated that the sum of the weights of the indicators related to "cold syndrome" (such as aversion to cold, cold limbs, and preference for hot drinks) and the indicators related to "heat syndrome" (such as fever, thirst, and preference for cold drinks) shall not be lower than the set threshold (such as 0.15), and the difference between the two weights shall not exceed the preset range, so as to avoid misjudgment of syndrome due to missing data or excessive weakening of one side by sensitivity analysis.

[0093] (2) Constraints on the Interrelationship between Organs: Based on the pre-defined relationships of mutual generation and mutual restraint and the internal-external relationship between organs, a group of strongly correlated indicators is defined. For example, when the weight of the "Liver Qi Stagnation" indicator is adjusted, the weight of the "Spleen Deficiency" indicator should increase accordingly (Liver Wood overacting on Earth); when the weight of "Heart Fire" increases, the weight of "Kidney Water" should not fall below the lower limit (Water and Fire in Harmony). The system automatically generates joint constraint equations through the association rule base.

[0094] (3) Weight lower limit protection mechanism: For indicators that reflect the core pathogenesis of TCM and whose weight is clinically recognized as not too low (such as tongue body and tongue coating in "tongue appearance", and pulse position and pulse strength in "pulse diagnosis"), an absolute weight lower limit (e.g. ≥0.05) is set. If the Monte Carlo simulation or missing data causes its weight to be lower than the lower limit, the constraint module will force it to be pulled back to the lower limit value and trigger "backup weight transfer" - transferring the importance of the missing part to other observable indicators that are most relevant to the syndrome (e.g., when pulse diagnosis data is missing, the weight of the corresponding symptoms in the consultation will be automatically increased according to the principle of "pulse and symptoms combined").

[0095] (4) Compliance-Syndrome Response Rule: In view of the TCM characteristic of "three parts treatment and seven parts maintenance", the sensitivity of different syndrome types to compliance is predefined. For example, the marginal contribution of the exercise check-in completion rate (compliance sub-item) of users with "Qi deficiency" syndrome to the final health score should be higher than that of users with "damp-heat" syndrome. This rule is used as an adjustment coefficient for the compliance sub-dimension under the criterion layer when setting the initial weights in AHP, ensuring that the weight allocation conforms to the TCM individualized diagnosis and treatment logic.

[0096] In one embodiment, the dynamic optimization and adaptive correction unit is further used to use the comprehensive health assessment model as the simulation object; set key uncertainty variables; wherein the key uncertainty variables include, but are not limited to, data quality perturbation, weight parameter perturbation, and compliance scenario perturbation; data quality perturbation is used to simulate different data sources under different missing rates and noise levels; weight parameter perturbation is used to apply random perturbation on the basis of the initial weight vector to simulate expert opinion disagreement or changes in clinical scenarios; compliance scenario perturbation is used to simulate the impact of different compliance levels on the final score; Monte Carlo simulation is performed based on the key uncertainty variables to generate the probability distribution of the comprehensive health score; global sensitivity analysis is performed on the probability distribution of the comprehensive health score to obtain the results of sensitive factor identification.

[0097] Specifically, incorporating "compliance" as a dynamic adjustment factor into the weighted adjustment closed loop is more in line with the realities of TCM health management: This application not only assesses compliance but, more importantly, treats compliance as a key perturbation variable and a sensitivity analysis object. The simulation engine can predict "which evaluation dimension's score will decrease most sensitively if user compliance drops from 80% to 60%?", thus allowing for early intervention. This transforms health evaluation from "a summary of past states" to "a prediction and guidance of future trends."

[0098] In one embodiment, the dynamic optimization and adaptive correction unit is further used to perform posterior constraint verification after calculating the sensitivity index using the Sobol index method or the Morris method on the Monte Carlo simulation results, obtaining the first-order sensitivity index and the total sensitivity index of each input variable, as well as the variable ranking in descending order of the total sensitivity index. Specifically, the posterior constraint verification step includes: if the difference in sensitivity index of a group of strongly correlated TCM variables exceeds a preset multiple, an association consistency check is automatically triggered to investigate whether the anomaly is caused by sampling noise or model nonlinearity, and the group sensitivity index is recalculated using grouped sensitivity analysis; for indicators with lower limit protection set by prior constraints built into the TCM knowledge graph and rule base, if they exhibit low sensitivity in sensitivity analysis, their weights are not reduced; if the weight sensitivity of a certain modality is less than a first preset threshold, its weight is reduced or its feature extraction method is optimized; if the weight sensitivity of a certain modality is greater than a second preset threshold, adaptive weight correction is triggered to generate dynamic weights.

[0099] In one embodiment, the principle of adaptive weight adjustment includes: constructing a nonlinear weight adjustment function based on the sensitivity analysis results; for modes with high sensitivity and high missing risk, the system automatically assigns a backup weight or transfers importance to other highly relevant modes; for modes with low sensitivity, the weight is appropriately reduced to simplify the model.

[0100] Specifically, the key innovation of this application, which distinguishes it from existing technologies, is based on Monte Carlo simulation and feedback optimization, and is used to solve the problem of static weights.

[0101] The specific process of Monte Carlo simulation and perturbation analysis includes: (1) The preliminary comprehensive scoring model generated in the initial evaluation model construction unit is used as the simulation object; (2) Define key uncertainty variables, which include, but are not limited to: Data quality perturbation: Simulate different data sources (such as infrared and tongue images) under different missing rates (10%-50%) and noise levels; Weight parameter perturbation: at the initial weight W initial Based on this, random perturbations (e.g., ±10%~±30%) are applied to simulate disagreements among experts or changes in clinical scenarios; Compliance situational perturbation: Simulate the impact of different compliance levels (e.g., high / medium / low compliance) on the final score.

[0102] Run Monte Carlo simulations (e.g., 10,000 iterations) to generate a probability distribution of the comprehensive health score (e.g., Figure 3 (As shown).

[0103] (3) Sensitivity analysis and adaptive weight correction. Specifically, a global sensitivity analysis is performed on the simulation results (e.g., using the Sobol index method or the Morris method) to quantify the contribution of each input variable (especially the AHP weights of each data source) to the variance of the final health score output.

[0104] To accurately quantify the contribution of each input variable (especially the AHP weights of each data source) to the variance of the comprehensive health score output, this invention uses Sobol global sensitivity analysis as the default method and Morris screening as an alternative for high-dimensional variable scenarios. The specific implementation is as follows: (1) Sobol exponent method (main method) The Sobol method, based on variance decomposition, decomposes the total variance of the model output Y=f(X) into the variance contributions of each input variable Xi and their interactions. The steps include: Define the input variable space: Select key variables for applying perturbations in the Monte Carlo simulation, including: AHP weights of each data source (such as tongue diagnosis weights). w tongue Pulse diagnosis weight w pulse Infrared weights w ir , etc.), data missing rate, noise level, compliance score, etc. Each variable has a defined probability distribution (such as normal distribution, uniform distribution), and the range is set based on clinical experience (e.g., the weight perturbation range is ±20% of the initial value).

[0105] Generate the sampling matrix: Use Sobol sequences or Latin hypercube sampling to generate... N ×(2 k The sampling matrix of ), where k For the number of input variables, N The baseline sample size (usually N=500) 1000).

[0106] Calculate the first-order sensitivity index Si: This measures the independent contribution of a single variable Xi to the output variance. The formula is as follows:

[0107]

[0108] in, For a single variable Xi, the first-order sensitivity index is... The first-order sensitivity index of the output Y corresponding to a single variable Xi. This represents the expected value of other variables when Xi is fixed, and Var(Y) is the total variance of the output.

[0109] Calculate the total-order sensitivity index STi: This measures the total contribution of Xi, including all its interactions, and is estimated by the following formula:

[0110]

[0111] in, The total order sensitivity index for a single variable Xi. The total order sensitivity index of the output Y corresponding to a single variable Xi.

[0112] Actual numerical calculations: The extended sampling method proposed by Saltelli was adopted, using two sets of independent sampling matrices. A and B Construct the cross matrix AB ( i )(Will A The i Column replacement B The i (columns), calculate the model outputs separately, and then perform an unbiased estimate of S. i and S Ti .

[0113] (2) Morris screening method (for high-dimensional or computationally limited scenarios) When there are more than 20 input variables and the number of Monte Carlo iterations is limited, the Morris method is used as a fast global sensitivity screening method to identify highly sensitive variables. Elementary Effect: Discretize the range of each input variable Xi into p levels, and change only one variable Δ at a time to calculate the change in output.

[0114]

[0115] Sampling Trajectories: R random trajectories are generated (usually r=10). 50), each variable in each trajectory is moved once in sequence to ensure global coverage.

[0116] Sensitivity indicators: 1) Mean μi 1) The mean of the absolute values ​​of Ei, representing the average strength of the variable's influence on the output; 2) Standard deviation σi The standard deviation of Ei reflects the degree of interaction or nonlinearity among variables; Filtering threshold: μi Variables exceeding a set threshold (such as 0.1 × global output range) are identified as highly sensitive variables and included in the subsequent adaptive correction process.

[0117] (3) Enhanced sensitivity analysis under specific constraints of traditional Chinese medicine To ensure that the sensitivity analysis results do not contradict traditional Chinese medicine theory, this invention adds a posterior constraint check after calculating the sensitivity index: If the difference in sensitivity index of a group of strongly correlated variables in traditional Chinese medicine (such as "aversion to cold" and "cold limbs") exceeds a preset multiple (e.g., 5 times), the system will automatically trigger a correlation consistency check to investigate whether the abnormality is caused by sampling noise or model nonlinearity. If necessary, group sensitivity analysis (treating the correlated variables as a group) will be used to recalculate the group sensitivity index.

[0118] For indicators with lower limit protection set by the TCM prior knowledge module (see the aforementioned "Constraint Rules for Cold, Heat, Deficiency and Excess"), even if they show low sensitivity in sensitivity analysis, their weight will not be reduced, so as to maintain the integrity of the diagnostic logic.

[0119] (4) Output and Application The final output of the sensitivity analysis is the first-order sensitivity index for each input variable (especially the weights of each AHP). Si and total sensitivity index STi and according to STi The variables are ranked in descending order. The system makes decisions based on the following rules: Highly sensitive variables ( STi >0.1): This is considered a key influencing factor and will be adjusted preferentially in the adaptive weighting adjustment.

[0120] Low sensitivity variables ( STi <0.01): Consider fixing its weight or removing it to simplify the model.

[0121] Interaction dominant variable ( STi Si ): This indicates that the variable mainly affects the output through interaction with other variables, and adjustments need to be made accordingly when correcting it.

[0122] The above analysis results will be used as input to the subsequent "nonlinear weight correction function" to achieve data-driven dynamic weight adaptation that conforms to the constraints of traditional Chinese medicine theory.

[0123] Key judgment: If the weight sensitivity of a certain modality (such as "pulse diagnosis") is extremely low (i.e., changing its weight has little impact on the final score), it indicates that the modality contributes redundancy to the model. The system can suggest using an adaptive weight correction method to reduce its weight or optimize its feature extraction method.

[0124] If the weight of a certain modality is highly sensitive and its data is often missing in real-world scenarios (such as difficulties in offline consultations), the system will trigger adaptive correction.

[0125] The adaptive weight adjustment specifically includes: Based on the sensitivity analysis results, a nonlinear weight correction function is constructed. The input to this function is the initial weight W. initial Given the data quality / compliance vector under the current simulation scenario, the output is the corrected dynamic weight W. adaptive .

[0126] Correction principles: For modalities with high sensitivity and high risk of missing data, the system automatically assigns a "backup weight" or transfers importance to other highly relevant modalities (for example, when facial diagnosis data is missing, based on the TCM theory of "face and tongue correspondence," the system automatically increases the weight of features related to facial color in tongue diagnosis and questioning). For modalities with low sensitivity, their weights are appropriately reduced to simplify the model.

[0127] In one embodiment, the nonlinear weight correction function is a double-nested nonlinear mapping function used to adjust the initial weights. Data quality vector Q (missing rate / signal-to-noise ratio for each modality), compliance scalar C, and sensitivity index. Mapped to dynamic weights .

[0128] The design process of the nonlinear weight correction function includes: in: It represents the element-wise exponentiation. Global contraction factor This is used to control the steepness of the weight distribution. This is the Hadamard product (element-by-element multiplication). These are learnable parameters for adjusting sensitivity, data quality, and compliance, respectively.

[0129] If the prior constraints built into the TCM knowledge graph and rule base are enabled, a projection operator is added after the softmax activation function. ,Will Project onto a convex set that satisfies all TCM constraints The output is a dynamic weight that conforms to traditional Chinese medicine theory; the projection problem is solved by solving a quadratic programming problem; the quadratic programming problem is as follows:

[0130]

[0131] in, Dynamically generated from a TCM constraint rule base. This is the final dynamically fused weight vector, whose dimensions correspond to the various data sources participating in the fusion (such as tongue diagnosis, pulse diagnosis, infrared thermal imaging, etc.). Each component represents the allocation coefficient of that modality in the final comprehensive decision. This differs from purely data-driven generation. Unlike the others, W is the final weight value that strictly satisfies the compliance of traditional Chinese medicine theory. It is achieved by forcibly applying the a priori rules of traditional Chinese medicine (such as the strong and weak constraints of "tongue and pulse combined assessment") through quadratic programming on the basis of the original weights fitted by least squares.

[0132] Specifically, this application constructs a two-layer nested nonlinear mapping function, which sets the initial weights... Data quality vector Q (missing rate / signal-to-noise ratio for each modality), compliance scalar C, and sensitivity index. Mapped to dynamic weights The function is designed as follows: in: It represents the element-wise exponentiation. Global contraction factor This is used to control the steepness of the weight distribution. This is the Hadamard product (element-by-element multiplication). Learnable parameters or empirical constants (typical values: Each factor adjusts the impact of sensitivity, data quality, and compliance. Softmax ensures normalization and that all weights are non-negative.

[0133] If the prior constraints built into the TCM knowledge graph and rule base are invoked, a projection operator is added after softmax. ,Will Project onto a convex set that satisfies all TCM constraints Above (e.g., the sum of cold and heat weights ≥ 0.15, tongue diagnosis weight ≥ 0.05, etc.). The projection problem can be efficiently solved by solving the following quadratic programming problem:

[0134]

[0135] in, It is dynamically generated from a TCM constraint rule base.

[0136] The actual iterative calculations for the three modalities of pulse diagnosis, tongue diagnosis, and infrared are shown in Table 1.

[0137] Table 1 Example of actual iterative calculation

[0138] After applying softmax and TCM constraints (tongue diagnosis weight lower limit of 0.05, cold and heat balance not disrupted), the pulse diagnosis weight is significantly reduced, while the weights of tongue diagnosis and infrared diagnosis are increased, and the final output conforms to TCM theory.

[0139] The above method ensures that this application does not rely on general algorithms, but embeds a series of feature optimization and weight correction mechanisms specifically designed for TCM evaluation, effectively overcoming the risk of being identified as "simple application of general algorithms".

[0140] In one embodiment, the process of reducing weights includes: when the total sensitivity index of a certain modality remains below a set threshold, a weight reduction strategy is adopted to reduce the corresponding weights; the weight reduction strategy is a single or combination of the exponential decay factor method, weight redistribution, and adaptive penalty function method; the exponential decay factor method is used to reduce the original weights. Multiply by an attenuation factor ,in The weights are determined by the sensitivity index and the preset lower limit; the reduced weights are:

[0141] in, For the maximum attenuation rate, It is the largest total sensitivity index among all modalities. To indicate the first i The original weights of each highly sensitive mode before assignment Indicates the first i The new weights of a highly sensitive mode after weight redistribution.

[0142] Weight reallocation is used to proportionally redistribute the weights saved from redundant modes to highly sensitive and reliable modes, and then compliance verification is performed based on prior constraints of the TCM knowledge graph and rule base; the weight reallocation process is as follows:

[0143]

[0144] in, To represent the total amount of weights that can be reallocated from redundant modes (low-sensitivity modes), The transfer ratio is given, and high represents the set of highly sensitive modes. This represents the new weight of the j-th highly sensitive mode after weight redistribution. This represents the original weight of the j-th highly sensitive mode before assignment. To represent the kth k The total sensitivity index for each modality is used to normalize the allocation ratio.

[0145] The adaptive penalty function method treats low sensitivity as a penalty signal by embedding a regularization term in the nonlinear weight correction function; the final weights are: in, To determine the severity of the punishment, This represents the dynamic weights obtained after correction using the adaptive penalty function method (corresponding to the final weights output to the evaluation module). This represents the initial weights of the mode (usually derived from the corresponding components in the initial weight vector calculated by the AHP method). To represent the threshold for judging low sensitivity, a value of 0.01 to 0.01 is set (i.e., a penalty is triggered when the total sensitivity index is below 0.01 to 0.01). When STi <Sthreshold STi < When the weights are exponentially decayed, they remain unchanged otherwise.

[0146] Specifically, when the total sensitivity index of a certain modality (such as "pulse diagnosis") Continuously below the set threshold (e.g.) A score <0.01 indicates that this modality contributes redundantly to the overall health score. The system employs the following weight reduction strategies (which can be selected individually or in combination): (1) Exponential decay factor method For the original weights Multiply by an attenuation factor ,in Determined by the sensitivity index and the preset lower limit:

[0147] In the formula Maximum attenuation rate (empirical value) ), This represents the largest total sensitivity index across all modalities. This approach ensures that the weights do not drop to zero, preserving the basic contribution.

[0148] (2) Weight Re-allocation The weights saved from redundant modes are proportionally redistributed to highly sensitive and reliable modes. For example:

[0149]

[0150] in The transfer ratio is 0.6, and "high" represents the set of highly sensitive modalities. After the transfer, the TCM prior knowledge constraint module needs to be called for compliance verification (such as the balance of cold and heat weights, the linkage of internal organs, etc.) to ensure that the adjustment does not disrupt the diagnostic logic.

[0151] The weights saved from redundant modes are adjusted based on the weights built into the TCM knowledge graph and rule base.

[0152] (3) Adaptive penalty function method Treating excessively low sensitivity as a "penalty signal," a regularization term is embedded in the nonlinear weight correction function. The final weights are: In the formula The penalty intensity (typical value 2.0). This approach is smooth and differentiable, facilitating end-to-end optimization.

[0153] The process of optimizing feature extraction includes: feature reconstruction based on traditional Chinese medicine theory; calculating the contribution of fine-grained features for each modality using game theory-based SHAP or LIME methods; and removing features with a contribution below an absolute threshold, retaining only the top-performing features. k The model is input with high-contribution features; an autoencoder is used to extract the retained features. k High-contribution features are compressed to obtain compressed features; an SE module is inserted into the deep learning feature extraction network to automatically learn the activation weights of each feature channel of the compressed features.

[0154] Specifically, the autoencoder employs a fully connected neural network, consisting of three parts: an encoder, a bottleneck layer, and a decoder. The encoder performs dimensionality reduction through multiple layers (e.g., from the input dimension k). k Reduced to 256, 128, and 64), extracting abstract representations of high-contribution features; the bottleneck layer outputs low-dimensional compressed features z. z (Dimension d) k d k The decoder symmetrically converts z z Reconstruct back to the original dimensions. During training, minimize the reconstruction loss using the mean squared error. After training, retain only the encoder portion for feature compression. For low-sensitivity modalities, instead of simply reducing weights, we optimize the feature extraction process to uncover potentially useful information or reduce computational redundancy. Examples of pulse diagnosis, tongue diagnosis, and infrared spectroscopy are given below.

[0155] (1) Feature reconstruction based on traditional Chinese medicine theory Taking "pulse diagnosis" as an example: If global sensitivity analysis finds that traditional pulse parameters (such as pulse location, pulse strength, and pulse rate) have low sensitivity, the system automatically switches to the TCM "pulse and symptom integration" mode. Instead of using pulse diagnosis features in isolation, the system concatenates pulse diagnosis features with strongly correlated symptoms such as "cold and heat, sweating, and bowel movements" in the consultation to form a joint feature vector. Then, the sensitivity of the combined features is recalculated. If the combined sensitivity increases, the pattern is retained.

[0156] Taking "tongue diagnosis" as an example: if the sensitivity of tongue color and tongue coating morphology is low, the system introduces tongue image partition weights (such as tongue tip, tongue middle, tongue root, and tongue side). Based on the TCM organ partition theory, only high-contribution partition features are retained (such as the tongue tip reflecting the heart and lungs, and the tongue middle reflecting the spleen and stomach), while low-contribution partitions are discarded, thus achieving feature dimensionality reduction and enhancement.

[0157] (2) Selection of interpretability features The contribution of fine-grained features for each modality is calculated using game theory-based SHAP or LIME methods. Features with a contribution below an absolute threshold (e.g., 0.05) are discarded, and only the top k high-contribution features are input into the model. For example, in pulse diagnosis, "pulse amplitude" may have a low contribution, while "pulse width" may have a high contribution; therefore, only the pulse width feature is retained.

[0158] (3) Feature compression of the autoencoder For high-dimensional redundant modalities (such as hundreds of pixel regions in infrared thermal imaging), a sparse autoencoder is trained with a bottleneck layer dimension much smaller than the input layer, forcing the network to learn the most informative latent features. Subsequently, only the bottleneck layer features are used in weight calculation, indirectly reducing the unstable impact of this modality on weight allocation.

[0159] A sparse autoencoder is a neural network architecture that adds sparsity constraints to a standard autoencoder. Its encoder compresses a high-dimensional input (such as hundreds of pixels in infrared thermal imaging) into bottleneck layer latent features much smaller than the input dimension, and the decoder attempts to reconstruct the original input from these latent features. Sparse constraints (such as KL divergence penalties or L1 regularization) force most neurons in the bottleneck layer to be in an "inactive" state, with only a few neurons activated, thus compelling the network to learn the most informative and compact feature representations. After training, the decoder is discarded, and only the low-dimensional latent features from the bottleneck layer output are used in subsequent weight calculations, effectively reducing the unstable impact of high-dimensional redundant modalities on weight allocation.

[0160] (4) Dynamic feature recalibration (Squeeze-and-Excitation) Inserting an SE module into a deep learning feature extraction network automatically learns the activation weights for each feature channel. Channels with low sensitivity are automatically assigned extremely low activation coefficients by the network, which is equivalent to achieving "soft removal" at the feature level without manually reducing modality weights.

[0161] The dynamic feature recalibration (SE module) process includes: (1) Squeeze: Perform global average pooling on each channel of the input feature map to compress the spatial information of each channel into a scalar and obtain the channel descriptor.

[0162] (2) Excitation: The channel descriptor is fed into two fully connected layers (first dimensionality reduction and then dimensionality increase), and the channel weights between 0 and 1 are generated by the Sigmoid activation function to represent the importance of each channel.

[0163] (3) Recalibration: Multiply each channel of the original feature map by the corresponding learned weight to achieve feature recalibration. Channels with low sensitivity automatically receive weights close to 0, which is equivalent to "soft removal" and does not require manual adjustment of modality weights.

[0164] In one specific embodiment, a sub-healthy user, "Zhang San," uses the TCM multi-source integrated health adaptive assessment system proposed in this application for the first time. The specific implementation steps include: Step 1: Data Collection Zhang San completed the following through an app and offline health stations: Online: Complete a psychological questionnaire, upload a photo of your tongue, and answer medical questions; Offline: Complete infrared thermal imaging, pulse diagnosis, and meridian testing at the health kiosk, and have a face-to-face consultation with a doctor; Compliance: Based on all the projects Zhang San completed, the system generates a complete "digital person" (all lights are on). The system generates an intervention plan for him that includes diet and exercise, and sets a requirement to retest certain indicators after 2 weeks.

[0165] Step 2: Initial Evaluation The system calls the initial evaluation model construction unit and calculates based on the initial weights of AHP. Zhang San's comprehensive health score is 78 points, of which the weight of "physical fitness" is not high, and the infrared indicator shows "disorder of cold and heat in the triple burner".

[0166] Step 3: Simulation and Correction Dynamic optimization and adaptive correction unit activated: (1) Scenario simulation: The system found through Monte Carlo simulation that the weight of the "infrared" mode ranked first in the sensitivity analysis. At the same time, the system predicted that Zhang San's compliance probability of completing the "offline infrared retest" within the next 2 weeks was only 40% (because offline appointments are required).

[0167] (2) Adaptive correction: According to the preset rules, when “offline infrared” data with “high sensitivity” and “high risk of missing” cannot be obtained, the system automatically activates the “backup weight” scheme. According to the theory of traditional Chinese medicine, some of the importance of “internal organs cold and heat” in “infrared” data is transferred to the weight of related issues such as “fear of cold, cold limbs, and preference for hot drinks” in “online consultation”.

[0168] Step 4: Dynamic Review and Final Output Two weeks later, Zhang San failed to complete the offline infrared retest, but he completed the online medical consultation check-in and exercise goals on time.

[0169] Real-time evaluation: During the review, the system used the revised dynamic weights. Wadaptive We also collected Zhang San's latest online consultation data and exercise check-in data (high compliance score).

[0170] The output includes: Overall score: Updated to 82 points (due to good compliance and symptom improvement).

[0171] Important note: The system prompts "Compliance is good, but to ensure accurate monitoring of the 'Triple Burner Cold and Heat' state, it is recommended to complete an offline infrared examination in the near future, or increase the frequency of online weekly reports on symptoms such as aversion to cold and cold limbs." "Digital Human" Status: Due to incomplete infrared inspection, the torso of the "digital human" is in a "half-lit" state, reminding the user of the data gap.

[0172] As can be seen from this embodiment, the system of the present invention not only provides a score, but more importantly, it proactively adapts to data gaps through a closed-loop feedback mechanism and provides forward-looking and personalized health management suggestions.

[0173] In one embodiment, a TCM health adaptive assessment method based on multi-source data fusion is also provided, comprising the following steps: Step S1: Acquire and preprocess multi-source heterogeneous TCM health data. The TCM health data includes at least multimodal detection data, personal health record data, and time-series behavioral data reflecting intervention response.

[0174] Step S2: Generate initial weight allocation for each data source based on a preset multi-criteria initial weighting strategy, and construct a comprehensive health evaluation model based on the initial weights to output health evaluation values; wherein, the initial weighting strategy is constrained by prior domain knowledge.

[0175] Step S3: Based on the uncertainty propagation analysis strategy, the current comprehensive health assessment model is subjected to perturbation simulation and sensitive factor identification. The initial weights are adaptively corrected according to the sensitive factor identification results, and the corrected dynamic weights are fed back to update the comprehensive health assessment model.

[0176] Step S4: Continuously collect the user's actual health evolution data after receiving intervention measures. Based on the deviation between the actual health evolution data and the prediction results output by the comprehensive health evaluation model, drive the iterative correction of weights to form a closed-loop feedback dynamic adaptive adjustment.

[0177] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0178] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and all such modifications and improvements fall within the scope of protection of this application.

Claims

1. A health adaptive assessment system based on multiple sources of traditional Chinese medicine, characterized in that, include: The data acquisition and preprocessing unit is used to acquire and preprocess multi-source heterogeneous TCM health data, wherein the TCM health data includes at least multimodal detection data, personal health record data, and time-series behavioral data reflecting intervention response; The initial evaluation model construction unit is used to generate initial weight allocations for each data source based on a preset multi-criteria initial weighting strategy, and to construct a comprehensive health evaluation model based on the initial weights to output a health evaluation value; wherein, the initial weighting strategy is constrained by prior domain knowledge; The dynamic optimization and adaptive correction unit is used to perform perturbation simulation and sensitive factor identification on the current comprehensive health evaluation model based on the uncertainty propagation analysis strategy, and adaptively correct the initial weights according to the sensitive factor identification results, and feed the corrected dynamic weights back to the initial evaluation model construction unit to update the comprehensive health evaluation model. The data acquisition and preprocessing unit continuously collects actual health change data after the user receives intervention. The dynamic optimization and adaptive correction unit drives the iterative correction of weights based on the deviation between the actual health change data and the predicted evaluation results of the comprehensive health evaluation model, forming a closed-loop feedback dynamic adaptive mechanism between the comprehensive health evaluation model and the real data.

2. The health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 1, characterized in that, The multimodal detection data includes: whole body and head and face infrared thermal imaging, tongue diagnosis images, pulse diagnosis waveforms, meridian electrical characteristics, psychological and emotional questionnaires, and infrared psychological assessment data; The personal health record data includes: text data of present medical history, past medical history, family history, and basic personal information; The time-series behavioral data reflecting the intervention response are recorded by using an app to check in and / or wearable devices to record the user's execution of the intervention plan, forming a digital human light-on-light-off model.

3. The health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 1, characterized in that, The process of constructing a comprehensive health assessment model includes: Calculate a basic information score; the basic information score is used to characterize the completeness of personal basic information, medical history information, and public health status. The 360° assessment score is calculated by dynamically calculating a TCM numerical score and a efficacy and constitution score based on initial weights generated using the analytic hierarchy process (AHP). The TCM numerical score refers to a comprehensive score integrating tongue diagnosis, pulse diagnosis, medical history taking, infrared imaging, meridian analysis, and psychological and emotional factors, with each score assigned weight according to different data collection scenarios (online, in-line, and offline). The efficacy and constitution score refers to a constitution assessment score determined based on the rate of change in TCM syndrome symptom scores, referencing the "Standards for TCM Health Management Services." Calculate the compliance score and the intervention performance score; the compliance score is: The intervention performance score is calculated by setting minimum values, benchmark values, and effort values ​​for quantitative key performance indicators and using piecewise linear interpolation. The basic information score, 360° assessment score, compliance score, and intervention performance score are weighted and summed based on the initial weight vector to generate a preliminary comprehensive score.

4. The health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 1, characterized in that, Prior domain knowledge constraints include: the prior constraints built into the TCM knowledge graph and rule base; specifically including: Holistic constraint rules for cold / heat and deficiency / excess: Based on the theory of differentiation of syndromes according to the eight principles of traditional Chinese medicine, it is stipulated that the sum of the weights of the indicators related to cold syndrome and the indicators related to heat syndrome shall not be lower than a preset threshold, and the difference between the weights of the indicators related to cold syndrome and the indicators related to heat syndrome shall not exceed a preset range; wherein, the indicators related to cold syndrome include: aversion to cold, cold limbs, and preference for hot drinks; the indicators related to heat syndrome include: fever, thirst, and preference for cold drinks; The organ-related transmission constraint rules are used to define strongly correlated index groups based on the preset mutual generation and mutual restraint mapping relationships and the internal-external relationship of the organs, and automatically generate joint constraint equations through the association rule base; The weight lower limit protection mechanism is used to set an absolute weight lower limit for indicators that reflect the core pathogenesis of TCM and whose weight is clinically recognized as not being too low. When the weight of an indicator falls below the absolute weight lower limit due to Monte Carlo simulation or missing data, the weight is forcibly pulled back to the absolute weight lower limit and a backup weight transfer is triggered to transfer the importance of the missing part to other observable indicators that are most correlated with the syndrome of the indicator. The compliance-syndrome response rule, based on the TCM principle of "three parts treatment and seven parts maintenance," predefines the sensitivity of different syndrome types to compliance. It is used as an adjustment coefficient for the compliance sub-dimension under the criterion layer when setting the initial weights in the analytic hierarchy process, so that the weight allocation conforms to the logic of individualized diagnosis and treatment in TCM.

5. The TCM multi-source adaptive health evaluation system according to claim 1, characterized in that, Based on the uncertainty propagation analysis strategy, the current comprehensive health assessment model is subjected to perturbation simulation and sensitive factor identification, including: The comprehensive health assessment model is used as the simulation object; Key uncertainty variables are defined; these include, but are not limited to, data quality perturbation, weight parameter perturbation, and compliance scenario perturbation; data quality perturbation is used to simulate different data sources under different missing rates and noise levels; weight parameter perturbation is used to apply random perturbation to the initial weight vector to simulate expert opinion disagreement or changes in clinical scenarios; and compliance scenario perturbation is used to simulate the impact of different compliance levels on the final score. Monte Carlo simulations were performed based on key uncertainty variables to generate a probability distribution of the comprehensive health score; A global sensitivity analysis was performed on the probability distribution of the comprehensive health score to obtain the results of sensitive factor identification.

6. The health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 1, characterized in that, The initial weights are adaptively adjusted based on the results of the sensitivity factor identification, including: After calculating the sensitivity index using the Sobol index method or the Morris method on the Monte Carlo simulation results, posterior constraint verification is performed to obtain the first-order sensitivity index and total sensitivity index of each input variable, as well as the variable ranking in descending order of the total sensitivity index. Specifically, the posterior constraint verification step includes: if the difference in sensitivity index of a group of strongly correlated TCM variables exceeds a preset multiple, an association consistency check is automatically triggered to investigate whether the anomaly is caused by sampling noise or model nonlinearity, and the group sensitivity index is recalculated using grouped sensitivity analysis; for indicators with lower limit protection set by the prior constraints built into the TCM knowledge graph and rule base, even if they show low sensitivity in sensitivity analysis, their weights are not reduced. If the weight sensitivity of a certain modality is less than the first preset threshold, then its weight is reduced or its feature extraction method is optimized. If the weight sensitivity of a certain modality is greater than the second preset threshold, then the weight adaptive correction is triggered to generate dynamic weights.

7. A health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 6, characterized in that, The principles of adaptive weight adjustment include: Based on the results of the sensitivity analysis, a nonlinear weight correction function is constructed. For modalities with high sensitivity and high risk of missing data, the system automatically assigns a "backup weight" or transfers the importance to other highly relevant modalities. For modes with low sensitivity, the weights are appropriately reduced to simplify the model.

8. A health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 6, characterized in that, The nonlinear weight correction function is a double-nested nonlinear mapping function used to adjust the initial weights. Data quality vector Q, compliance scalar C, and sensitivity index Mapped to dynamic weights ; The design process of the nonlinear weight correction function includes: in: It represents the element-wise exponentiation. Global contraction factor This is used to control the steepness of the weight distribution. This is the Hadamard product (element-by-element multiplication). These are learnable parameters for adjusting sensitivity, data quality, and compliance, respectively. If prior constraints based on the TCM knowledge graph and rule base are enabled, a projection operator is added after the softmax activation function. ,Will Project onto a convex set that satisfies all TCM constraints The output is a dynamic weight that conforms to traditional Chinese medicine theory; the projection problem is solved by solving a quadratic programming problem; the quadratic programming problem is as follows: in, Dynamically generated from a TCM constraint rule base. This is the final weight vector for dynamic fusion.

9. A health adaptive evaluation system based on multiple sources of traditional Chinese medicine as described in claim 6, characterized in that, The process of reducing weights includes: When the total sensitivity index of a certain mode is consistently lower than the set threshold, a weight reduction strategy is adopted to reduce the corresponding weight; the weight reduction strategy can be a single or combination of the exponential decay factor method, weight redistribution, and adaptive penalty function method. The exponential decay factor method is used to adjust the original weights. Multiply by an attenuation factor ,in The weights are determined by the sensitivity index and the preset lower limit; the reduced weights are: in, For the maximum attenuation rate, It is the largest total sensitivity index among all modalities. To indicate the first i The original weights of each highly sensitive mode before assignment Indicates the first i The new weights of each highly sensitive mode after weight redistribution; The weight reallocation is used to proportionally redistribute the weights saved from redundant modes to highly sensitive and reliable modes, and then perform compliance verification based on the prior constraints built into the TCM knowledge graph and rule base; the weight reallocation process is as follows: in, To represent the total amount of weights that can be reallocated after being extracted from redundant modes, The transfer ratio is given, and high represents the set of highly sensitive modes. Indicates the first j The new weights of the highly sensitive modes after weight redistribution Indicates the first j The original weights of each highly sensitive mode before assignment To indicate the first Tk The total sensitivity index of each modality is used to normalize the allocation ratio; The adaptive penalty function method treats low sensitivity as a penalty signal, embedding a regularization term in the nonlinear weight correction function; the final weights are: in, To determine the severity of the punishment, This represents the dynamic weights obtained after correction using the adaptive penalty function method. This represents the initial weights of the mode. This represents the threshold for determining low sensitivity.

10. A TCM health adaptive evaluation method based on multi-source data fusion, characterized in that, Includes the following steps: Step S1: Acquire and preprocess multi-source heterogeneous TCM health data, wherein the TCM health data includes at least multimodal detection data, personal health record data, and time-series behavioral data reflecting intervention response; Step S2: Generate initial weight allocation for each data source based on a preset multi-criteria initial weighting strategy, and construct a comprehensive health evaluation model based on the initial weights to output a health evaluation value; wherein, the initial weighting strategy is constrained by prior domain knowledge. Step S3: Based on the uncertainty propagation analysis strategy, the current comprehensive health assessment model is subjected to perturbation simulation and sensitive factor identification, and the initial weights are adaptively corrected according to the sensitive factor identification results. The corrected dynamic weights are then fed back to update the comprehensive health assessment model. Step S4: Continuously collect the user's actual health evolution data after receiving intervention measures. Based on the deviation between the actual health evolution data and the prediction results output by the comprehensive health evaluation model, drive the iterative correction of weights to form a closed-loop feedback dynamic adaptive adjustment.

Images