Artificial intelligence method for early warning and health management of traditional chinese medicine preventive treatment based on multi-modal data fusion

Abstract

The application discloses a multimodal data fusion AI traditional Chinese medicine disease prevention and health management method, which comprises the following steps: firstly, a unique ID is allocated to a user, and three types of heterogeneous data with time stamps, i.e., western medicine physiology, traditional Chinese medicine characteristics and user behaviors, are collected through a multi-source module and are stored in a classified manner; secondly, the three types of data are subjected to targeted pretreatment and standardization, and are converted into unified numerical characteristic data; then, a hierarchical CNN-LSTM model based on an attention mechanism is adopted, a weight is allocated in combination with an analytic hierarchy process and a traditional Chinese medicine theory, and data are deeply fused to obtain a comprehensive health characteristic vector; the vector is input into a trained evaluation model, nine types of constitution recognition and disease prevention low, medium and high risk grading early warning are realized; finally, an integrated intervention scheme is generated based on constitution and risk, a health status is dynamically tracked in combination with real-time monitoring data, the scheme is iteratively optimized when preset conditions are met, effective fusion of multi-dimensional heterogeneous data is realized, disease prevention early warning accuracy is improved, and a full-closed-loop dynamic health management system is constructed.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and traditional Chinese medicine health management, and in particular to an AI-based method for disease prevention and early warning and health management using multimodal data fusion. Background Technology

[0002] Preventive medicine is one of the core ideas of traditional Chinese medicine (TCM) theory, focusing on the conditioning of sub-health states and early warning of potential disease risks, which is of great significance for maintaining human health. Traditional TCM preventive medicine relies on the clinical experience of physicians, using methods such as tongue diagnosis and pulse diagnosis to identify constitution and assess risks. This approach suffers from strong subjectivity and inconsistent diagnostic standards, making it difficult to achieve standardized and large-scale health management.

[0003] With the development of smart wearable devices, medical sensing technology, and artificial intelligence, data-driven intelligent health management technology is gradually being applied to the field of traditional Chinese medicine (TCM), enabling the digital collection and analysis of physiological indicators and TCM vital signs data. However, existing technologies still have many shortcomings: First, the data collection dimensions are limited, often focusing only on Western medical theory data or single TCM vital signs data, failing to effectively integrate three types of heterogeneous data: Western medical theory, TCM characteristics, and user behavior. This data bias leads to insufficient accuracy in constitution identification and risk warning. Second, the methods for fusion of heterogeneous data lack specificity. The continuous data from Western medical theory, the characteristic data from TCM, and the qualitative data from user behavior have different data types. Traditional fusion methods cannot highlight the weight of core health data, easily leading to fusion results that deviate from clinical reality. Third, the health management system lacks a dynamic closed loop, mostly consisting of static constitution identification and program recommendations. It cannot update warning results and optimize intervention programs in real time according to changes in the user's health status, significantly reducing the intervention effect.

[0004] Therefore, we propose an AI-based TCM disease prevention and early warning and health management method based on multimodal data fusion to address the aforementioned issues. Summary of the Invention

[0005] This invention overcomes the shortcomings of existing technologies and provides an AI-based method for disease prevention and early warning and health management in traditional Chinese medicine based on multimodal data fusion.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an AI-based method for disease prevention and health management using multimodal data fusion in Traditional Chinese Medicine, comprising the following steps: S1. Three-dimensional multi-source data collection: Each user is assigned a unique identity ID. The three-dimensional heterogeneous data of the user associated with the ID is obtained through the multi-source data collection module. The three-dimensional heterogeneous data includes Western medicine data, traditional Chinese medicine characteristic data, and user behavior data. Among them, Western medicine physiological data is collected in real time through smart wearable devices to collect continuous physiological indicators such as heart rate, blood pressure, and blood oxygen. Traditional Chinese medicine characteristic data are collected by a tongue image acquisition device and a pulse sensor to collect characteristic physical signs of tongue and pulse. User behavior data is collected through standardized questionnaires, including information on diet, exercise, emotions, and daily routines. All three types of data are associated with collection timestamps and collected and stored independently. S2. Heterogeneous Data Preprocessing and Standardization: Through the data preprocessing and standardization module, the alignment and association of three types of heterogeneous data are achieved based on user ID and collection timestamp, and then the collected three-dimensional heterogeneous data is processed in a targeted manner. For Western medicine data, a sliding window filtering method was used to remove noise, the 3σ principle was used to remove outliers, and the data was mapped to the [0,1] interval by min-max normalization; We used a CNN model to extract visual features of the tongue and wavelet transform to extract tactile features of the pulse from TCM characteristic data, thus completing the digital transformation of visual and tactile features. Establish 1-3 level quantization coding rules for user behavior data to perform quantization coding and standardized mapping, and unify continuous, feature-based, and qualitative heterogeneous data into standardized numerical feature data that can be fused and analyzed. S3, Hierarchical Multimodal Data Fusion: Through the multimodal data fusion module, a hierarchical CNN-LSTM deep learning model based on attention mechanism is used to deeply fuse standardized feature data. The process sequentially executes bottom-level concatenation of similar data, mid-level differentiated attention weight allocation, and high-level cross-dimensional comprehensive feature output to obtain a comprehensive health feature vector that can fully represent the user's health status. Among them, the weight allocation of the middle layer differentiated attention is determined based on the analytic hierarchy process combined with the TCM theory of disease prevention. The highest weight coefficient of 0.25 is assigned to pulse and heart rate, the basic weight coefficient of 0.1 is assigned to tongue appearance, and the appropriate weight coefficient of 0.15 is assigned to diet, exercise and rest. The remaining weights are equally distributed to other Western medical physiological indicators according to the importance of physiological indicators to disease prevention warning. In the hierarchical CNN-LSTM deep learning model, the CNN uses 3 convolutional layers to extract local features, and the LSTM uses 2 hidden layers to capture temporal features. S4. TCM Pre-Disease Stage Differentiation and Early Warning: Input the comprehensive health feature vector into the TCM constitution and pre-disease risk assessment model of the AI ​​TCM differentiation and early warning module. This model is based on the classic TCM theory of pre-disease treatment and three-dimensional massive health data, and is trained using the cross-entropy loss function and Adam optimizer. The disease prevention risk assessment model identifies nine types of TCM constitutions through feature vector similarity matching. At the same time, it sets thresholds based on feature vector similarity to classify disease prevention into low, medium, and high risk levels. A similarity of ≥80% indicates high risk, 50%-80% indicates medium risk, and <50% indicates low risk. Targeted early warning prompts are pushed according to the risk level. S5. Closed-loop health management and dynamic optimization: The health management module calls the TCM disease prevention knowledge base, which is based on the classic TCM disease prevention theory and establishes a mapping table of nine constitutions, three types of disease risks and intervention programs. Based on body constitution type and disease risk level, a personalized health intervention plan integrating diet therapy, exercise, work and rest, and emotional regulation is generated; Simultaneously, based on real-time dynamic monitoring data from the multi-source data acquisition module, the comprehensive health feature vector is continuously updated to achieve dynamic tracking of the user's physical condition and pre-disease risk. When the change in physical fitness matching degree is ≥10%, the risk level of pre-disease rises or falls, or the core physiological indicators deviate from the normal range by ≥15%, the intervention plan is triggered for real-time iterative optimization.

[0007] In a preferred embodiment of the present invention, the multi-source data acquisition module includes a Western medicine pharmacology data acquisition unit, a traditional Chinese medicine characteristic data acquisition unit, and a user behavior data acquisition unit that are independent of each other. The three units are respectively adapted to the acquisition requirements of continuous, characteristic, and qualitative data and are each configured with an independent data transmission interface. The Western medicine data acquisition unit uses a Bluetooth Low Energy interface, the traditional Chinese medicine data acquisition unit uses a USB 3.0 interface, and the user behavior data acquisition unit uses a network interface, enabling parallel acquisition, classified storage, and efficient retrieval of three types of heterogeneous data.

[0008] In a preferred embodiment of the present invention, the high-level cross-dimensional comprehensive feature output of the hierarchical CNN-LSTM deep learning model is achieved by concatenating the three types of feature matrices of physiology, vital signs and behavior after the bottom layer is concatenated and the middle layer is weighted into a 128-dimensional comprehensive health feature vector through a fully connected layer, thereby realizing deep coupling and feature fusion of cross-dimensional data.

[0009] In a preferred embodiment of the present invention, the training dataset of the TCM constitution and disease risk assessment model is three-dimensional correlation data of sub-healthy people professionally annotated by TCM physicians. The dataset selection criteria follow the "Clinical Guidelines for Sub-health in TCM", excluding user data with diagnosed diseases. The annotation content includes constitution type, disease risk level and corresponding feature data labels.

[0010] In a preferred embodiment of the present invention, the standardized questionnaire covers four core dimensions: dietary structure, exercise frequency, daily routine, and emotional state, and includes 20 quantitative questions to achieve standardized and normalized collection and subsequent quantitative coding of user behavior data.

[0011] In a preferred embodiment of the present invention, the intervention scheme optimized in real time needs to be verified by model effectiveness. The matching degree between the optimized scheme and the user's current comprehensive health feature vector is calculated. The scheme with a matching degree ≥ 85% is used as the final execution scheme, and the scheme that does not meet the standard is regenerated iteratively.

[0012] In a preferred embodiment of the present invention, the TCM disease prevention knowledge base is constructed with a precise mapping table of nine constitutions, three types of disease prevention risks and intervention programs. For each type of unbalanced constitution, a special syndrome differentiation diet therapy formula and suitable exercise method are matched. For medium and high risk disease prevention groups, a special module for emotional regulation is added.

[0013] In a preferred embodiment of the present invention, the visual features of the tongue include four core features: tongue color, tongue coating, tongue shape, and tongue posture; the tactile features of the pulse include four core features: pulse rate, pulse shape, pulse strength, and pulse position. After extraction and digital transformation, the various features form fixed-dimensional TCM characteristic feature data, which are incorporated into a standardized numerical feature data system.

[0014] This invention addresses the shortcomings of the prior art and has the following beneficial effects: (1) This invention integrates three types of heterogeneous data: Western medicine, traditional Chinese medicine characteristics, and user behavior. It achieves precise association and alignment of data through unique identity ID and collection timestamp, breaking through the defects of the existing technology in data collection being one-sided and single-dimensional. It fully depicts the user's health status from three dimensions: physiological indicators, traditional Chinese medicine signs, and lifestyle behavior, providing a comprehensive and systematic data source for subsequent constitution identification and disease risk warning, and improving the comprehensiveness and accuracy of health assessment.

[0015] Meanwhile, targeted preprocessing and standardization schemes were developed for the different attributes of the three types of data, realizing the transformation of continuous, characteristic, and qualitative heterogeneous data into unified standardized numerical characteristic data. This eliminated the fusion barriers caused by differences in data types, enabling health data of different dimensions and types to be fusionable, and laying the technical foundation for subsequent deep fusion of multimodal data.

[0016] (2) The present invention uses a hierarchical CNN-LSTM deep learning model based on the attention mechanism, combined with the hierarchical analysis method and the TCM theory of disease prevention to determine the differential weight coefficients, highlighting the value of core health data such as pulse and heart rate in constitution identification and disease prevention warning. It not only gives full play to the deep learning model's ability to fuse and analyze complex data, but also makes the fusion results fit the TCM clinical diagnosis and treatment logic, avoiding the problem of pure algorithm fusion being divorced from clinical practice, greatly improving the accuracy and practicality of multimodal data fusion, and the output comprehensive health feature vector can more realistically and comprehensively represent the user's health status.

[0017] Meanwhile, the constructed TCM constitution and disease risk assessment model, trained on massive amounts of labeled sub-health data, achieves standardized identification of nine TCM constitutions and quantitative classification of disease risk through feature vector similarity matching. It has established clear risk judgment thresholds, breaking through the limitations of traditional TCM disease prevention treatment which relies on physician experience, is highly subjective, and has inconsistent diagnostic standards. This makes the disease prevention warning results more scientific, objective, and operable, facilitating large-scale health screening and early warning. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments; Figure 1 This is a flowchart of a preferred embodiment of the AI-based TCM disease prevention and health management method based on multimodal data fusion. Figure 2 This is a flowchart illustrating the execution of a preferred embodiment of the AI-based TCM disease prevention and health management method based on multimodal data fusion. Figure 3 This is a diagram of a preferred embodiment of the three-dimensional multi-source data acquisition architecture of the present invention; Figure 4 This is a diagram of a hierarchical multimodal data fusion architecture according to a preferred embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or it can be fixed through another intermediate component. When a component is said to be "connected to" another component, it can be directly connected to the other component or it may be fixed through another intermediate component. When a component is said to be "set on" another component, it can be set directly on the other component or it may be set through another intermediate component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0021] 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 invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] Example 1: Risk warning and health management for the general sub-healthy population during the period of no disease occurrence. This example focuses on a 35-year-old urban white-collar worker in a sub-healthy state, implementing TCM-based disease prevention and health management. This individual exhibits lifestyle habits such as prolonged sleep deprivation, infrequent exercise, and irregular eating habits, but has no clearly diagnosed disease. Figure 1-4 As shown, the specific implementation steps are as follows: S1. Three-dimensional multi-source data acquisition: A unique identifier ID, JM2026001, was assigned to the user. Three types of heterogeneous data associated with this ID were collected through a multi-source data acquisition module: Western medicine theory, traditional Chinese medicine characteristics, and user behavior. All data were associated with the collection timestamp (March 1, 2026 - March 7, 2026, 7 consecutive days of collection) and stored independently. Western medicine physiological data: Real-time collection of three continuous physiological indicators through smart bracelets (wearable data collection devices): heart rate (resting 68-92 beats / minute, average 75 beats / minute), blood pressure (average 126 / 82 mmHg), and blood oxygen (96%-98%, average 97%). Traditional Chinese Medicine (TCM) characteristic data: Tongue image characteristics (pale red tongue color, thin white tongue coating, moderate tongue shape, normal tongue posture) are collected through a tongue image acquisition device (high-definition image acquisition equipment), and pulse characteristics (pulse rate 72 beats / minute, thin pulse shape, slow pulse force, deep pulse position) are collected through a pulse sensor (pressure sensor acquisition device). User behavior data was collected through a standardized questionnaire (covering 4 core dimensions: diet, exercise frequency, daily routine, and emotional state, with a total of 20 quantitative questions). The results showed that users tended to eat raw and cold foods and frequently ordered takeout (diet dimension), exercised ≤ once a week and mainly sedentary (exercise dimension), fell asleep at ≥23:30 every day and slept for about 6 hours (daily routine dimension), and experienced high work pressure and occasional anxiety and irritability (emotional dimension).

[0023] S2. Heterogeneous data preprocessing and standardization: Through data preprocessing and standardization modules, based on user ID JM2026001 and collection timestamp, three types of heterogeneous data collected over 7 days are aligned and correlated along the time dimension to form a unified user health dataset. Then, targeted processing is carried out for different types of data: Western medicine physiological data: The sliding window filtering method was used to denoise the continuously collected heart rate and blood pressure data to eliminate abnormal values ​​caused by equipment vibration and human activity; the abnormal heart rate value of 92 beats / minute and the abnormal blood pressure value of 135 / 88 mmHg were removed by the 3σ principle; min-max normalization was used to map all physiological index data to the [0,1] interval to achieve data dimension unification. Traditional Chinese Medicine (TCM) characteristic data: A 3-layer CNN convolutional neural network model is used to extract features from high-definition tongue images to obtain digital vectors of four visual features: tongue color, tongue coating, tongue shape, and tongue posture; wavelet transform is used to decompose and extract features from pressure signals collected by pulse sensors to obtain digital vectors of four tactile features: pulse rate, pulse shape, pulse momentum, and pulse position, thus completing the full digital transformation of TCM characteristic features. User behavior data: The questionnaire results were quantified and coded according to the 1-3 level quantification coding rules (level 1 is good, level 2 is average, and level 3 is poor). Among them, diet was level 3, exercise was level 3, daily routine was level 3, and mood was level 2. The qualitative life habit information was transformed into standardized numerical feature data. Finally, the three types of processed data are uniformly transformed into standardized numerical feature data that can be fused and analyzed, forming a standardized health feature dataset for users.

[0024] S3, Hierarchical multimodal data fusion: The standardized numerical feature data is deeply fused using a multimodal data fusion module and a hierarchical CNN-LSTM deep learning model based on an attention mechanism. The specific steps are as follows: The underlying data of the same type is spliced ​​together: the heart rate, blood pressure and blood oxygen feature vectors of Western medicine are spliced ​​together, the tongue and pulse feature vectors of traditional Chinese medicine are spliced ​​together, and the user behavior feature vectors of diet, exercise, rest and mood are spliced ​​together to form three types of feature matrices of the same type. Mid-level differentiated attention weight allocation: Based on the analytic hierarchy process and combined with the TCM theory of disease prevention, the weight coefficients of each feature are determined. The pulse and heart rate are each assigned the highest weight coefficient of 0.25, the tongue is assigned the basic weight coefficient of 0.1, the diet, exercise and rest are each assigned the appropriate weight coefficient of 0.15, and the remaining 0.05 weight is equally distributed to the blood oxygen and blood pressure indicators according to the importance of disease prevention warning. The three feature matrices are weighted. High-level cross-dimensional comprehensive feature output: The physiological, physical signs and behavioral feature matrices of the bottom layer and the middle layer weighted by the fully connected layer are cross-dimensionally spliced ​​to generate a 128-dimensional comprehensive health feature vector. This vector fully represents the user's overall health status and physical characteristics. Among them, the CNN part of the hierarchical CNN-LSTM deep learning model uses 3 convolutional layers to extract local key information of each feature, and the LSTM part uses 2 hidden layers to capture the temporal change features of user health data in the 7-day time dimension, thereby improving the comprehensiveness and accuracy of the fused features.

[0025] S4. Early warning system based on syndrome differentiation in the pre-disease stage of Traditional Chinese Medicine: The above 128-dimensional comprehensive health feature vector is input into the TCM constitution and pre-disease risk assessment model of the AI ​​TCM syndrome differentiation and early warning module. The training dataset of this model is three-dimensional correlation data of sub-healthy people professionally annotated by TCM physicians. The screening criteria follow the "Clinical Guidelines for Sub-health in TCM", excluding user data with diagnosed diseases. The annotation content includes constitution type, pre-disease risk level and corresponding feature data labels. The model is trained to convergence using the cross-entropy loss function and Adam optimizer. The model uses a feature vector similarity matching algorithm to match the user's comprehensive health feature vector with the standard feature vectors of the nine constitutions in the model, and determines that the user has a Qi deficiency constitution with a similarity of 76%. At the same time, based on the threshold standard set by the feature vector similarity, the user is determined to be at medium risk of being healthy (50%-80% is considered medium risk). Based on the above assessment, a targeted warning message is sent to the user: You have a Qi deficiency constitution according to Traditional Chinese Medicine, which is a risk factor before illness. Long-term sleep deprivation, lack of exercise, and irregular diet are the main causes. You are prone to sub-health symptoms such as fatigue, dizziness, shortness of breath, and weakness in the limbs. It is recommended to replenish Qi and improve your lifestyle habits in a timely manner.

[0026] S5. Closed-loop health management and dynamic optimization: The health management module accesses the Traditional Chinese Medicine (TCM) knowledge base for disease prevention, which contains a precise mapping table of nine body constitutions, three types of disease prevention risks, and intervention plans. Based on the user's Qi deficiency constitution and the risk assessment for disease prevention, a personalized health intervention plan integrating diet therapy, exercise, rest, and emotional regulation is generated. Specifically: Dietary therapy: It is recommended to eat foods that replenish qi, such as astragalus and black chicken soup, yam and millet porridge, and red date and longan tea. Reduce the intake of raw, cold, greasy and spicy foods, and ensure a balanced intake of whole grains, vegetables and high-quality protein every day. Exercise: We recommend gentle exercises such as Tai Chi, Baduanjin, and slow walking, 3-4 times a week for 30-40 minutes each time, and avoid high-intensity and strenuous exercise; Daily Routine: Establish a regular daily routine, go to bed at 11:00 PM or later, and ensure 7-8 hours of sleep. You can do 5 minutes of deep breathing exercises in the morning. Emotional state: It is recommended to meditate for 10 minutes every day to relax, listen to soothing music, reduce work anxiety, and regulate emotions through activities such as gardening and calligraphy; Subsequently, the user's real-time dynamic health data is continuously collected through the multi-source data acquisition module, and the 128-dimensional comprehensive health feature vector is updated weekly to achieve dynamic tracking of the user's physical condition and disease risk. Two weeks after the user implemented the intervention plan, the body fitness matching degree changed by 12% (≥10%) after data collection and feature vector updates, triggering real-time iterative optimization of the intervention plan. The optimized intervention plan, based on the user's actual implementation, adjusted the exercise frequency to 4 times a week, 40 minutes each time, and added spleen-strengthening and qi-tonifying drinks such as Poria and Coix Seed Water. The matching degree of the optimized plan with the user's current comprehensive health feature vector was calculated, and the matching degree was 91% (≥85%). It was determined to be an effective plan and used as the final implementation plan to continuously provide health management guidance to the user.

[0027] Example 2: Early warning and health management of high-risk groups for pre-disease status in sub-healthy populations This embodiment focuses on TCM disease prevention and health management for a 50-year-old high-risk individual with sub-health conditions. This individual has a family history of hypertension, a long-term high-salt diet, lack of exercise, occasional dizziness and chills, and no clearly diagnosed cardiovascular or cerebrovascular disease. The specific implementation process is the same as in Embodiment 1. The core implementation results and intervention plan are as follows: Data Acquisition and Processing: After acquisition and processing through steps S1-S3, a 128-dimensional comprehensive health feature vector for the user (ID: JM2026002) is generated. The Western medicine data shows an average resting heart rate of 82 beats / minute and an average blood pressure of 132 / 86 mmHg. The traditional Chinese medicine data shows a pale tongue, a thick and greasy tongue coating, a pulse rate of 68 beats / minute, and a deep and slow pulse shape. The user behavior data shows a high-salt diet, 0 times of exercise per week, sleep time ≥23:00, and long-term low mood.

[0028] Early warning based on syndrome differentiation: After S4 step determination, this user has a Yang deficiency constitution with a feature vector similarity of 89%, corresponding to a high risk of pre-disease (≥80% is high risk). The following warning is pushed: You have a Yang deficiency constitution according to Traditional Chinese Medicine, with a high risk of pre-disease. High salt diet, lack of exercise, and long-term staying up late are the main causes. Combined with a family history of hypertension, you are prone to cardiovascular and cerebrovascular sub-health problems, accompanied by symptoms such as aversion to cold, dizziness, and soreness in the waist and knees. You need to start warming and regulating Yang immediately and strictly improve your lifestyle habits.

[0029] Closed-loop management and optimization: Based on the assessment results of Yang deficiency constitution and high risk of pre-disease, a personalized intervention plan is generated, focusing on warming Yang and dispelling cold, low-salt diet, and gentle exercise. Specifically: dietary therapy recommends warming Yang foods such as ginger and mutton soup, longan and red date porridge, and eucommia tea, and strictly controls daily salt intake to <5g; exercise recommendations include gentle exercises such as walking, jogging, and Wuqinxi (Five Animal Frolics), 4 times a week, 20-30 minutes each time, until the body feels slightly warm; the daily sleep requirement is to fall asleep before 22:30, and soak feet in hot water before bed; for emotional support, 15 minutes of mindfulness training and communication with family and friends are recommended daily to alleviate low mood.

[0030] One month after the user implemented the intervention plan, the core physiological indicator of blood pressure deviated from the normal range by 18% (≥15%), triggering the plan iteration and optimization. The optimized plan added a traditional Chinese medicine physiotherapy suggestion of massaging Yongquan and Zusanli acupoints every morning to warm the Yang. The optimized plan matched the user's current feature vector with a degree of 89% (≥85%), and was used as the final implementation plan. After three months of continuous tracking, the user's physical similarity dropped to 72%, the risk level of no disease dropped to medium risk, and symptoms such as aversion to cold and dizziness were significantly improved.

[0031] Example 3: Physical conditioning and health maintenance for low-risk sub-health populations This embodiment focuses on TCM disease prevention and health management for a 28-year-old young person with low risk of sub-health. This person has generally good lifestyle habits, occasionally stays up late, has mild spleen deficiency, and no obvious sub-health symptoms. After S1-S4 steps, the user is determined to have mild spleen deficiency constitution, with a feature vector similarity of 45%, corresponding to low risk of disease (<50% is low risk). Based on the assessment results, the health management module generates a lightweight, personalized conditioning plan, focusing on strengthening the spleen and stomach and maintaining regular sleep patterns, without requiring high-intensity exercise or dietary adjustments. Subsequently, the user's health data is continuously and dynamically tracked. Within 6 months, the user's physical condition matching degree changes by less than 10%, core physiological indicators are normal, and no intervention plan optimization is triggered, thus achieving normalized health maintenance for low-risk sub-health individuals.

[0032] All data acquisition devices in this invention are general-purpose smart wearable devices, tongue image acquisition instruments, pulse sensors, etc., which do not require customized hardware, reducing the cost and threshold of technology implementation and facilitating large-scale promotion and application. The standardized questionnaire, TCM disease prevention knowledge base, and constitution-risk-intervention program mapping table described in this invention can all be flexibly adjusted according to the constitution characteristics and lifestyles of different regions and populations, and have good adaptability and scalability. The hierarchical CNN-LSTM deep learning model and the TCM constitution and disease risk assessment model of the present invention can be continuously trained and optimized based on newly added user health data. With the accumulation of data, the model’s recognition accuracy and early warning accuracy can be continuously improved. The fully closed-loop health management process of this invention can be automated through a software system. From data collection, preprocessing, and fusion to early warning, solution generation, and iterative optimization, the entire process requires no manual intervention. Only TCM physicians need to regularly maintain the knowledge base and model, which greatly improves the management efficiency of TCM disease prevention.

[0033] The technical solution of this invention can be integrated with intelligent medical platforms, health management apps, community medical service systems, etc., and can be applied to various scenarios such as family health management, community sub-health screening, and enterprise employee health management, with broad engineering application prospects and social value.

[0034] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. These are all equivalent modifications and improvements made to the above embodiments based on the essential technology of the present invention, and all of these fall within the protection scope of the present invention.

Claims

1. A multimodal data fusion-based AI-powered TCM disease prevention and early warning and health management method, characterized in that, Includes the following steps: S1. Three-dimensional multi-source data collection: Each user is assigned a unique identity ID. The three-dimensional heterogeneous data of the user associated with the ID is obtained through the multi-source data collection module. The three-dimensional heterogeneous data includes Western medicine data, traditional Chinese medicine characteristic data, and user behavior data. Among them, Western medicine physiological data is collected in real time through smart wearable devices to collect continuous physiological indicators such as heart rate, blood pressure, and blood oxygen. Traditional Chinese medicine characteristic data are collected by a tongue image acquisition device and a pulse sensor to collect characteristic physical signs of tongue and pulse. User behavior data is collected through standardized questionnaires, including information on diet, exercise, emotions, and daily routines. All three types of data are associated with collection timestamps and collected and stored independently. S2. Heterogeneous Data Preprocessing and Standardization: Through the data preprocessing and standardization module, the alignment and association of three types of heterogeneous data are achieved based on user ID and collection timestamp, and then the collected three-dimensional heterogeneous data is processed in a targeted manner. For Western medicine data, a sliding window filtering method was used to remove noise, the 3σ principle was used to remove outliers, and the data was mapped to the [0,1] interval by min-max normalization; We used a CNN model to extract visual features of the tongue and wavelet transform to extract tactile features of the pulse from TCM characteristic data, thus completing the digital transformation of visual and tactile features. Establish 1-3 level quantization coding rules for user behavior data to perform quantization coding and standardized mapping, and unify continuous, feature-based, and qualitative heterogeneous data into standardized numerical feature data that can be fused and analyzed. S3, Hierarchical Multimodal Data Fusion: Through the multimodal data fusion module, a hierarchical CNN-LSTM deep learning model based on attention mechanism is used to deeply fuse standardized feature data. The process sequentially executes bottom-level concatenation of similar data, mid-level differentiated attention weight allocation, and high-level cross-dimensional comprehensive feature output to obtain a comprehensive health feature vector that can fully represent the user's health status. Among them, the weight allocation of the middle layer differentiated attention is determined based on the analytic hierarchy process combined with the TCM theory of disease prevention. The highest weight coefficient of 0.25 is assigned to pulse and heart rate, the basic weight coefficient of 0.1 is assigned to tongue appearance, and the appropriate weight coefficient of 0.15 is assigned to diet, exercise and rest. The remaining weights are equally distributed to other Western medical physiological indicators according to the importance of physiological indicators to disease prevention warning. In the hierarchical CNN-LSTM deep learning model, the CNN uses 3 convolutional layers to extract local features, and the LSTM uses 2 hidden layers to capture temporal features. S4. TCM Pre-Disease Stage Differentiation and Early Warning: Input the comprehensive health feature vector into the TCM constitution and pre-disease risk assessment model of the AI ​​TCM differentiation and early warning module. This model is based on the classic TCM theory of pre-disease treatment and three-dimensional massive health data, and is trained using the cross-entropy loss function and Adam optimizer. The disease prevention risk assessment model identifies nine types of TCM constitutions through feature vector similarity matching. At the same time, it sets thresholds based on feature vector similarity to classify disease prevention into low, medium, and high risk levels. A similarity of ≥80% indicates high risk, 50%-80% indicates medium risk, and <50% indicates low risk. Targeted early warning prompts are pushed according to the risk level. S5. Closed-loop health management and dynamic optimization: The health management module calls the TCM disease prevention knowledge base, which is based on the classic TCM disease prevention theory and establishes a mapping table of nine constitutions, three types of disease risks and intervention programs. Based on body constitution type and disease risk level, a personalized health intervention plan integrating diet therapy, exercise, work and rest, and emotional regulation is generated; Simultaneously, based on real-time dynamic monitoring data from the multi-source data acquisition module, the comprehensive health feature vector is continuously updated to achieve dynamic tracking of the user's physical condition and pre-disease risk. When the change in physical fitness matching degree is ≥10%, the risk level of pre-disease rises or falls, or the core physiological indicators deviate from the normal range by ≥15%, the intervention plan is triggered for real-time iterative optimization.

2. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The multi-source data acquisition module includes three independent data acquisition units: Western medicine pharmacology data acquisition unit, traditional Chinese medicine characteristic data acquisition unit, and user behavior data acquisition unit. These three units are respectively adapted to the acquisition requirements of continuous, characteristic, and qualitative data and are each configured with an independent data transmission interface. The Western medicine data acquisition unit uses a Bluetooth Low Energy interface, the traditional Chinese medicine data acquisition unit uses a USB 3.0 interface, and the user behavior data acquisition unit uses a network interface, enabling parallel acquisition, classified storage, and efficient retrieval of three types of heterogeneous data.

3. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The high-level cross-dimensional comprehensive feature output of the hierarchical CNN-LSTM deep learning model is achieved by concatenating the three types of feature matrices of physiology, vital signs and behavior after the bottom layer is concatenated and the middle layer is weighted into a 128-dimensional comprehensive health feature vector through a fully connected layer, thereby realizing deep coupling and feature fusion of cross-dimensional data.

4. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The training dataset for the TCM constitution and disease risk assessment model consists of three-dimensional correlation data of sub-healthy individuals professionally annotated by TCM physicians. The dataset selection criteria follow the "Clinical Guidelines for Sub-health in TCM", excluding user data with diagnosed diseases. The annotation content includes constitution type, disease risk level, and corresponding feature data labels.

5. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The standardized questionnaire covers four core dimensions: diet, exercise frequency, daily routine, and emotional state. It includes 20 quantitative questions to achieve standardized and regulated collection and subsequent quantitative coding of user behavior data.

6. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The intervention plan optimized in real time needs to be verified by model effectiveness. The matching degree between the optimized plan and the user's current comprehensive health feature vector is calculated. The plan with a matching degree ≥ 85% is used as the final execution plan. The plan that does not meet the standard is regenerated iteratively.

7. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The TCM knowledge base for preventing disease contains a precise mapping table of nine body constitutions, three types of disease risks, and intervention programs. It matches exclusive syndrome differentiation and dietary therapy formulas and suitable exercise methods for each type of unbalanced body constitution. A dedicated module for emotional regulation has been added for people with medium to high risk of disease.

8. The AI-based TCM disease prevention and health management method based on multimodal data fusion according to claim 1, characterized in that: The visual features of the tongue include four core features: tongue color, tongue coating, tongue shape, and tongue posture. The tactile features of the pulse include four core features: pulse rate, pulse shape, pulse strength, and pulse position. After extraction and digital transformation, these features form fixed-dimensional TCM characteristic feature data, which are then incorporated into a standardized numerical feature data system.

Images