Traditional Chinese medicine constitution differentiation weight loss intervention method and system based on cross-modal knowledge distillation

Abstract

The application discloses a traditional Chinese medicine constitution differentiation weight loss intervention method and system based on cross-modal knowledge distillation, and belongs to the technical field of intelligent medical treatment. The system comprises a mobile phone APP software end and an intelligent weight loss intervention bracelet hardware end. The method comprises the following steps: the mobile phone APP software end collects tongue images and palm images, and adopts ResNet+ViT double-branch coding to extract local texture and global semantic features; the bracelet end collects pulse time sequence signals, and adopts LSTM+Transformer double-branch coding to extract short-term rhythm and long-range association features. The pulse features are used as teacher features, and the tongue and hand features are used as student features. Cross-modal knowledge distillation alignment is realized through a joint loss function. Based on a modal difference coefficient and an exponential decay formula, a fusion weight is dynamically adjusted, a 0-100 constitution bias score is output, obesity constitution types are distinguished, and exclusive electrical stimulation intervention parameters are matched. Intervention is performed throughout the whole process of pulse negative feedback parameter adjustment, and seven-day cycle iteration optimization is performed. Ablation experiments show that the differentiation accuracy of the application reaches 92.8%, which is 23.4% higher than that of the existing conventional combination.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular to a method and system for weight loss intervention based on TCM constitution differentiation based on cross-modal knowledge distillation. Background Technology

[0002] With the deep integration and application of artificial intelligence, big data, and deep learning technologies in smart healthcare and traditional Chinese medicine (TCM) health management, my country's TCM digitalization and disease prevention and constitution conditioning industry has entered a new stage of rapid development. Relying on AI technology to promote the objectification of TCM constitution identification, standardization of syndrome differentiation, and intelligent intervention has become a core and important direction for implementing the Healthy China strategy and upholding the principles of TCM innovation. Obesity, as a highly prevalent chronic metabolic disease, is confirmed by modern medical research to be closely related to metabolic disorders and endocrine imbalances. In the TCM theoretical system, obesity is essentially a syndrome of deficiency in the root and excess in the branch, caused by long-term imbalances in constitution, such as phlegm-dampness accumulation, internal damp-heat, spleen and stomach qi deficiency, and yang deficiency with dampness obstruction. Simply relying on external weight loss methods such as dieting and exercise can only treat the symptoms. Only by addressing the root cause of TCM constitution through syndrome differentiation and conditioning, strengthening the body's foundation, eliminating dampness, and invigorating the spleen and stomach can the long-term goal of healthy weight loss and preventing rebound be achieved. Tongue diagnosis, hand diagnosis, and pulse diagnosis are the three core objective diagnostic dimensions of the TCM diagnostic system. Tongue diagnosis indicates the pathogenesis of cold, heat, phlegm, and dampness, while hand diagnosis indicates the abundance or deficiency of qi, blood, and body fluids. Pulse diagnosis indicates the root cause of the core diseases of the internal organs. The combined assessment of these three signs is the core diagnostic basis for the accurate differentiation and treatment of obesity in TCM clinical practice, and it is also an indispensable data foundation for realizing intelligent AI weight loss and symptomatic intervention.

[0003] Traditional Chinese medicine (TCM) constitution identification relies entirely on the clinical experience and subjective judgment of licensed TCM practitioners. This leads to inconsistent diagnostic standards, a shortage of qualified personnel, low efficiency, and difficulty in large-scale application, failing to meet the public's urgent need for intelligent, routine home-based weight loss and health management. With the rapid development of machine learning, deep learning, and multimodal data fusion technologies, TCM constitution identification is gradually moving from traditional manual experience-based diagnosis to a new stage of digitalization, objectification, and intelligence. Multimodal data fusion modeling has become the mainstream research direction for AI-based TCM constitution identification. Existing research explores related technologies in areas such as tongue image feature extraction, hand diagnosis constitution model construction, and pulse time-series signal modeling. Among these, constitution identification methods based on multi-level feature fusion of tongue images, hand diagnosis constitution discrimination models based on machine learning, and multimodal data fusion constitution identification systems have been gradually implemented in trials, laying a solid technical foundation for intelligent three-modal TCM diagnosis.

[0004] Currently, ResNet and ViT visual dual-branch networks have significant technical advantages in extracting local texture details and global semantic features from TCM tongue and hand images. LSTM and Transformer temporal dual-branch networks can accurately capture pulse cycle fluctuations, rhythm changes, and long-term temporal correlation features, providing reliable algorithmic support for deep feature encoding of heterogeneous physical signs data. Meanwhile, cross-modal knowledge distillation technology is rapidly emerging. By constructing a knowledge transfer architecture of a high-precision multimodal teacher model and a lightweight on-device student model, complex deep TCM diagnostic knowledge can be efficiently compressed and distilled into a small mobile model. While ensuring high accuracy in diagnostic identification, this significantly reduces model computational overhead and inference latency, providing a new feasible path for the local offline deployment of TCM multimodal constitution identification technology on mobile devices.

[0005] Currently, existing TCM intelligent constitution identification technologies are mainly based on single-modal or tongue and pulse dual-modal research. Related studies mostly adopt single network feature extraction and simple feature fusion methods. Although they have achieved basic constitution classification and recognition effects, they generally have not constructed a three-in-one trimodal collaborative acquisition architecture of tongue-hand-pulse, have not configured a dedicated dual-branch coding network for the heterogeneous characteristics of static image modality and temporal dynamic modality, and have not introduced cross-modal knowledge distillation to achieve high-precision model lightweight edge deployment. They also lack a syndrome differentiation weighted fusion mechanism that conforms to the clinical diagnosis and treatment rules of TCM and a closed-loop intervention system for weight loss linked by constitution differentiation. The overall integrity of the technology, the accuracy of syndrome differentiation, and its practicality all have obvious shortcomings, making it difficult to meet the actual clinical application and market popularization needs of TCM intelligent constitution weight loss conditioning.

[0006] Specifically, existing TCM intelligent constitution identification technology still has the following defects and shortcomings: 1. The lack of a three-modal collaborative acquisition architecture, incomplete constitution differentiation dimensions, and incomplete TCM diagnostic basis. Existing TCM AI constitution identification systems are mostly limited to single tongue image, single pulse image, or tongue and pulse dual-modal acquisition applications. They generally lack the key diagnostic dimension of verifying the state of Qi and blood through hand diagnosis, and do not follow the TCM three-in-one diagnostic principle of pulse as the basis of diagnosis, tongue to determine cold, heat, phlegm and dampness, and hand to assess the state of Qi and blood. Some devices attempt to collect tongue and hand images by integrating a camera into a wristband. However, due to limitations in hardware size, lighting environment, and shooting angle, the image quality is poor and effective features are severely lost. A scientific hardware architecture of separate acquisition of high-definition images from a mobile phone and layered and accurate acquisition of pulse images from a wristband has not been formed. This results in inconsistent quality of basic acquisition data, incomplete constitution differentiation basis, low accuracy of obesity constitution differentiation, and a lack of TCM clinical diagnostic support for weight loss and conditioning.

[0007] 2. The AI ​​feature encoding network lacks adaptability, and the extraction of deep features from heterogeneous modalities is insufficient. Existing deep learning models for TCM constitution identification generally use a single traditional CNN network to uniformly process two types of heterogeneous data: static images of the tongue and hands and dynamic time-series pulse data. They do not design specific encoding structures for the characteristics of different modalities. The image data does not adopt a dual-branch architecture that combines ResNet local fine-grained feature extraction with ViT global semantic feature extraction. The pulse time-series data does not have a dual-branch network that combines LSTM short-term temporal memory with Transformer long-term temporal association encoding. As a result, the modal heterogeneity cannot be effectively bridged, the extraction of core diagnostic features of deep constitution is incomplete, subtle differences in constitution are difficult to distinguish accurately, and the overall accuracy of constitution identification is low.

[0008] (3) The lack of a cross-modal knowledge distillation mechanism makes it impossible to balance the accuracy of model differentiation with the universal deployment on mobile devices. Existing TCM multimodal constitution identification models either use complex multimodal models with high computational power for high differentiation accuracy, but rely on cloud computing power, have a large amount of computation, high inference latency, and cannot be deployed offline on mobile devices; or use lightweight small models for convenient deployment, but have insufficient deep differentiation feature learning and a significant decrease in identification accuracy. Existing technologies generally do not construct a cross-modal knowledge distillation training architecture that combines high-precision feature learning of teacher models with lightweight knowledge transfer of student models. This makes it impossible to achieve efficient compression and transfer of differentiation knowledge, and it is difficult to simultaneously meet the dual requirements of high accuracy of TCM constitution differentiation and universal deployment on mobile devices. The market adaptability is poor.

[0009] (4) The multimodal fusion method is simple and crude, without the priority weighting ratio of TCM syndrome differentiation, and the fusion logic is divorced from clinical practice. Existing multimodal TCM constitution identification technology all adopt the simple fusion method of direct feature splicing or equal weighting, without setting up a scientific weighted fusion system based on the primary and secondary relationship of TCM clinical syndrome differentiation, with pulse as the main factor, tongue as the auxiliary factor, and hand diagnosis as the verification factor; without configuring a fixed syndrome differentiation weighting ratio in combination with the TCM disease prevention and treatment practice guidelines, the pure data-driven fusion is divorced from the core theory of TCM syndrome differentiation and treatment, the synergistic and complementary advantages of multimodal signs cannot be brought into play, the AI ​​syndrome differentiation results have low consistency with the real TCM clinical diagnosis, and the targeting and professionalism of weight loss constitution syndrome differentiation are seriously insufficient.

[0010] (5) The diagnosis and intervention are disconnected from each other, and there is no adaptive closed-loop regulation mechanism for body constitution. The weight loss intervention effect is poor and easy to rebound. The existing TCM AI body constitution identification system only stays at the level of body constitution classification and text suggestion push. It has not established a closed-loop linkage system between the body constitution diagnosis results and the hardware of acupoint microcurrent electrical stimulation for weight loss intervention. All intervention parameters are fixed and uniformly output. It does not match the exclusive intervention waveform, frequency, current intensity and intervention duration according to different obesity bias body constitutions such as phlegm-dampness constitution, damp-heat constitution and qi deficiency constitution. It lacks a step-by-step intensity adaptive adjustment of body constitution score, real-time negative feedback parameter adjustment of pulse and seven-day cycle iterative long-term correction mechanism. It is a one-size-fits-all approach and does not treat the root cause of the disease. The weight loss is only a symptom and not a root cause. The effect of body constitution improvement is weak and the weight is very easy to rebound after the intervention is stopped.

[0011] (6) The model has weak generalization and anti-interference ability, poor individual and environmental adaptability, and is difficult to scale up for mass production. The existing TCM multimodal constitution model does not have a multimodal feature normalization preprocessing and cross-domain adaptive optimization mechanism. It is easily affected by multiple sources of interference such as light intensity, pulse collection pressure, and individual differences in user constitution. The data distortion rate is high and the model has poor generalization stability across scenarios and users. At the same time, the algorithm model has not completed embedded lightweight solidification and hardware unified calibration, resulting in insufficient mass production adaptability and difficulty in meeting the needs of large-scale popularization and application of smart home weight loss and health management for the general public. Summary of the Invention

[0012] The purpose of this invention is to overcome the technical problems existing in the prior art, such as unreasonable hardware architecture for TCM constitution identification, insufficient extraction of AI heterogeneous physical characteristics encoding, lack of core logic for TCM syndrome differentiation integration, and lack of closed-loop iterative regulation in intelligent weight loss intervention. It provides a TCM constitution differentiation weight loss intervention method and system based on cross-modal knowledge distillation.

[0013] The objective of this invention is achieved through the following technical solution: A first aspect of the present invention provides a method for weight loss intervention based on TCM constitution differentiation based on cross-modal knowledge distillation, comprising the following steps: S1. Simultaneously collect multi-modal vital sign data of users during the meridian flow period, and perform time-series alignment on the multi-modal vital sign data; the multi-modal vital sign data includes tongue image, hand image and pulse time-series signal; S2. Preprocess the pulse time sequence signal, tongue image and hand image respectively, extract the traditional artificial medical features and AI deep features corresponding to each modality of vital signs data after preprocessing, and perform unified dimension normalization processing to obtain pulse normalization features, tongue normalization features and hand normalization features respectively. S3. Using pulse normalization features as teacher features, and tongue and hand normalization features as student features, the student features are aligned with the teacher features through training with a joint loss function to obtain aligned tongue and hand features; the pulse normalization features, aligned tongue and hand features are then fed into a lightweight MLP perceptron to output pulse single-modal scores, tongue single-modal scores, and hand diagnosis single-modal scores, respectively. S4. Calculate the modality difference coefficient based on the score output in step S3, and dynamically adjust the modality weight parameters according to the modality difference coefficient, the correlation between each modality's physical characteristics and TCM constitution differentiation. S5. Normalize the adjusted modal weight parameters and calculate the comprehensive obesity constitution bias score; and determine the obesity TCM constitution type based on the comprehensive obesity constitution bias score; S6. Based on the aforementioned obesity TCM constitution classification and comprehensive obesity constitution bias score, adaptively and stepwise match exclusive weight loss electrical stimulation intervention parameters; S7. Based on the exclusive weight loss electrical stimulation intervention parameters, initiate microcurrent electrical stimulation intervention on the targeted acupoints; collect pulse dynamic data in real time during the intervention process, and adjust the exclusive weight loss electrical stimulation intervention parameters in real time according to the pulse dynamic data.

[0014] In some embodiments, the extraction of AI deep features in step S2 includes: The preprocessed tongue and hand images are input into the ResNet+ViT dual-branch image encoding architecture. The ResNet branch extracts local texture detail features, and the ViT branch extracts global body semantic features. The outputs of the two branches are concatenated to form the tongue and hand feature vectors. The preprocessed pulse time series signal is input into the LSTM+Transformer pulse time series dual-branch encoding architecture, where the LSTM branch extracts short-term time series fluctuation features and the Transformer branch extracts long-term time series correlation features. The outputs of the two branches are concatenated to form the pulse feature vector.

[0015] In some embodiments, the preprocessing of the tongue image or hand image in step S2 includes: Image size is normalized to 224×224×3 RGB standard size, white balance is calibrated for color temperature and gamma value, Gaussian filter kernel 7×7 for reflection removal, adaptive histogram equalization for shadow correction, and pixels are normalized to the 0-255 standard range.

[0016] In some embodiments, the joint loss function in step S3 is as follows: in, Indicates joint loss, Represents the classification cross-entropy loss. This represents the knowledge distillation loss. These are the weighting coefficients.

[0017] In some embodiments, the modal difference coefficient in step S4 is calculated using the following formula: in, This represents the modal difference coefficient, where d is the feature dimension. Let be the mean of the k-th dimension feature of the current individual. The mean of the k-th dimension feature of the training baseline data; The dynamic adjustment of modal weight parameters based on the modal difference coefficient, the correlation between modal characteristics and TCM constitution differentiation includes: when When adjusting the modal weights, use the following formula: in, To adjust the weights of the k-th mode, Let these be the initial weights for the k-th mode. This is the weighting adjustment factor. The correlation between the characteristics of various modalities and TCM constitution differentiation.

[0018] In some embodiments, the comprehensive obesity body constitution bias score in step S5 is calculated using the following formula: in, This indicates a single-modal pulse score. This indicates a single-modal score for tongue image. This indicates a single-modal score for hand diagnosis. , , To adjust the weight parameters of each mode; The method of determining the TCM constitution type of obesity based on the comprehensive obesity constitution bias score includes: Divided into neutral and normal types; It is divided into spleen and stomach qi deficiency constitution; Divided into Yang deficiency and dampness obstruction constitution; Divided into damp-heat internal constitution; It is divided into two types: one with phlegm and dampness accumulation.

[0019] In some embodiments, step S6, adaptive stepwise matching of specific weight-loss electrical stimulation intervention parameters, includes: Based on the TCM constitution classification of obesity, basic electrical stimulation intervention parameters are matched, and then the intensity is adjusted in a stepwise manner according to the comprehensive obesity constitution bias score: <30 points uses baseline electrical stimulation intervention parameters; Intensity increases by 20% for scores with a strength of S > 60; intensity increases by 40% for scores with a strength of S > 60.

[0020] In some embodiments, the following steps are also included: S8. After a single intervention is completed, all data is archived and stored, and a personalized treatment plan is pushed out. S9. Periodic review and summary of physical fitness score changes; adjust the baseline electrical stimulation intervention parameters based on changes in physical fitness scores within the period.

[0021] A second aspect of the present invention provides a TCM constitution differentiation and weight loss intervention system based on cross-modal knowledge distillation for the intervention method, comprising a mobile APP software terminal and a smart weight loss intervention bracelet hardware terminal, which are connected wirelessly via Bluetooth. The mobile APP software includes: A tongue image standardization acquisition module is used to acquire tongue images during the meridian flow period and to preprocess the tongue images. A standardized hand image acquisition module is used to acquire hand images during the meridional period and to preprocess the hand images. The ResNet+ViT image dual-branch coding module is used to extract AI depth features from the preprocessed tongue and hand images, respectively. The trimodal AI weighted fusion diagnostic module is used to align student and teacher features using normalized pulse characteristics as teacher features and normalized tongue and hand characteristics as student features through training with a joint loss function, resulting in aligned tongue and hand features. It also feeds the normalized pulse characteristics, aligned tongue and hand features into a lightweight MLP perceptron, outputting single-modal scores for pulse, tongue, and hand diagnosis, respectively. Furthermore, it calculates modality difference coefficients based on the scores and dynamically adjusts the modality weight parameters according to the modality difference coefficients, the correlation between each modality's physical characteristics and TCM constitution differentiation. Finally, it normalizes the adjusted modality weight parameters to calculate a comprehensive obesity constitution bias score and determines the TCM constitution type based on this comprehensive obesity constitution bias score. The intervention parameter matching and distribution module is used to adaptively and hierarchically match exclusive weight loss electrical stimulation intervention parameters based on the obesity TCM constitution classification and comprehensive obesity constitution bias score, and distribute the exclusive weight loss electrical stimulation intervention parameters to the smart weight loss intervention bracelet hardware. The hardware components of the smart weight loss intervention bracelet include: The PPG+FSR multimodal pulse layered acquisition module is used to acquire pulse time sequence signals during the meridian flow period; The motion posture noise reduction preprocessing module is used to preprocess the pulse timing signal; The LSTM+Transformer pulse timing coding module is used to extract AI deep features from the preprocessed pulse timing signal. The microcurrent electrical stimulation weight loss intervention module is used to receive the exclusive weight loss electrical stimulation intervention parameters and activate the microcurrent electrical stimulation intervention at the targeted acupoints; it is also used to collect pulse dynamic data in real time during the intervention process and adjust the exclusive weight loss electrical stimulation intervention parameters in real time based on the pulse dynamic data.

[0022] In some embodiments, the mobile APP software also includes: The data archiving and personalized treatment push module is used to archive and store all data after a single intervention and push personalized treatment plans; it is also used to periodically review and summarize the changes in physical fitness scores, and adjust the basic electrical stimulation intervention parameters according to the changes in physical fitness scores within the period.

[0023] It should be further noted that the technical features corresponding to the above-mentioned options and embodiments can be combined or substituted with each other to form new technical solutions without conflict.

[0024] This invention addresses the core shortcomings of existing TCM-based obesity constitution identification and intelligent weight loss technologies, including significant hardware data acquisition interference, limited AI feature extraction, lack of quantitative standards for TCM diagnosis, one-size-fits-all weight loss interventions, lack of safety feedback, and easy rebound after treatment. Based on a split-collaborative hardware architecture, dual-branch AI coding, a TCM fixed-weight fusion diagnosis system, and a closed-loop iterative intervention system, this invention represents substantial innovation and significant technological progress compared to similar technologies. It boasts several core advantages, including accurate diagnosis, targeted intervention, safety and stability, non-rebound weight loss, low mass production cost, and ease of home use. Specific beneficial effects are as follows: 1. This invention adopts a separate collaborative acquisition architecture for the mobile phone and wristband, with clear hardware division of labor and no data interference, significantly improving acquisition accuracy. Existing technologies often use wristbands with integrated cameras for acquisition, resulting in mutual interference between image and pulse data, misalignment of time sequences, large data errors, and distorted diagnostic results. This invention uses a mobile phone to separately acquire tongue and hand images, while the wristband independently acquires pulse time-series signals. Precise timestamp synchronization is achieved via Bluetooth 5.0, ensuring standardized and highly reliable acquisition of trimodal vital signs data from the hardware source. No customized dedicated hardware is required; professional acquisition results can be achieved using general-purpose mass-produced equipment, resulting in low mass production costs and high applicability.

[0025] 2. This invention features a unique dual-branch heterogeneous AI-specific feature encoding mode, adapting to two types of heterogeneous modal attributes to achieve multi-dimensional refined feature representation, significantly improving feature extraction coverage and discriminative power. Traditional techniques only use a single general-purpose CNN network to process image and temporal data in parallel, which cannot adapt to the heterogeneous differences between static image spatial features and dynamic pulse temporal features. This results in defects such as loss of local details, lack of long-range correlation capture, and single representation of physical characteristics, as well as poor ability to distinguish subtle physical biases. This invention builds a dedicated dual-branch encoding architecture for modal characteristics. The image modality adopts a ResNet+ViT dual-branch structure to simultaneously extract fine-grained features of local texture and color and global physical semantic features; the pulse temporal modality adopts an LSTM+Transformer dual-branch structure to take into account both short-term pulsation rhythm fluctuation features and long-range hierarchical pressure correlation features; all modal features are mapped to the 0-1 standard range through unified dimensional normalization, completely eliminating data bias caused by hardware differences, environmental interference, and physiological baseline shifts, providing a high-quality, high-discriminative feature data base for subsequent cross-modal alignment and accurate identification.

[0026] 3. This invention introduces a cross-modal knowledge distillation alignment mechanism to solve the problem of heterogeneous modal domain offset, achieving deep fusion of multimodal features and significantly improving diagnostic accuracy and model generalization. Existing technologies generally directly and crudely splice and fuse heterogeneous modal features without considering the differences in feature space distribution of images and time-series signals, which easily leads to problems such as feature redundancy, feature conflict, and modal information misalignment, resulting in poor diagnostic stability, weak generalization ability, and inability to adapt to different users and different acquisition scenarios. This invention pioneers a pulse-guided cross-modal knowledge distillation (CMKD) mechanism, using pulse features as the teacher features and tongue and hand visual features as student features. Through joint constraint training using classification cross-entropy loss and KL divergence distillation loss, it achieves accurate cross-domain alignment of multimodal features, unifying the mapping of three types of heterogeneous features to the same feature space, effectively bridging modal heterogeneity differences, and fully leveraging the complementary and synergistic advantages of pulse, tongue, and hand signs, significantly improving the consistency between AI constitution diagnosis and clinical real-person diagnosis.

[0027] 4. This invention employs a dynamic adaptive weight fusion mechanism guided by traditional Chinese medicine (TCM) a priori knowledge, abandoning the traditional fixed weight model to achieve individualized quantitative diagnosis tailored to each individual, thus completely resolving the industry pain point of "one weight for all." Traditional TCM multimodal quantitative diagnosis often uses manually preset fixed weight ratios. These weight parameters are fixed and unadjustable, failing to adapt to the differentiated characteristics of pulse quality, tongue appearance, complexion, and individual constitution among different users. This easily leads to the suppression of dominant discriminative features and interference from ineffective features, violating the core principle of individualized diagnosis and treatment in TCM. This invention, based on the modal difference coefficient, detects the degree of feature domain shift in real time. Combining the correlation between each modality and constitution diagnosis, it dynamically adjusts the fusion weights of pulse, tongue, and hand diagnosis modalities through an exponential adaptive weight algorithm. While retaining the TCM priori logic of "pulse as the primary factor, tongue and hand as secondary factors," it dynamically adapts to individual differences in physical signs, accurately outputting a 0-100 point quantitative score for constitution bias. The diagnostic logic is more clinically relevant, more adaptable, and more accurate.

[0028] 5. This invention employs a real-time negative feedback parameter adjustment based on pulse readings, coupled with a seven-day iterative closed-loop mechanism, ensuring safe and controllable intervention and preventing weight regain. Traditional electrical stimulation for weight loss involves fixed parameters and unidirectional output, lacking physiological monitoring and parameter optimization, posing safety risks, failing to address the root cause of body constitution, and making weight regain highly likely. This invention monitors the pulse readings in real-time throughout the intervention process, automatically reducing the intervention intensity by 10%-20% in case of abnormalities, ensuring safe use. It reviews and optimizes intervention parameters every seven days, dynamically iterating and adjusting as body constitution improves, correcting obesity-related imbalances at their source, and achieving long-term, healthy weight loss without rebound.

[0029] 6. This invention is compatible with general-purpose hardware, requires no cloud computing power, is simple to deploy, convenient for home use, and is suitable for a wide range of people. The system can run locally offline with regular mobile phones and smartwatches, without the need for professional medical personnel. Users can independently complete body constitution identification and weight loss intervention at home, which is suitable for the long-term conditioning needs of various obese groups, in line with the concept of preventive medicine in traditional Chinese medicine, and has extremely high value for industrial mass production and market promotion. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a TCM constitution-based weight loss intervention method based on cross-modal knowledge distillation, as shown in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the overall workflow of the trimodal TCM constitution AI fusion syndrome differentiation and intervention system according to an embodiment of the present invention; Figure 3 This is a structural diagram of the tongue and hand image + pulse AI constitution identification model based on cross-modal knowledge distillation, as shown in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the principle of the adaptive weighting algorithm in an embodiment of the present invention. Detailed Implementation

[0031] The technical solution 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, not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. 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.

[0032] It should be noted that the defects in the solutions in the prior art are all the results of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of this application in the following text should be the inventors' contributions to this application in the process of invention and creation, and should not be understood as technical content known to those skilled in the art.

[0033] In view of the technical problems pointed out in the background art, the present invention provides the following embodiments: The core architecture of this invention adopts a separate collaborative deployment mode of mobile APP software and intelligent weight loss intervention bracelet hardware. Both ends achieve low-latency, highly stable bidirectional wireless data communication and command interaction via Bluetooth 5.0 modules. This abandons the unreasonable integrated hardware structure of traditional smart bracelets that combine a camera for image acquisition, implementing separate and independent operation for professional image acquisition and precise pulse diagnosis. The core technology of this invention has a fixed closed-loop chain as follows: standardized synchronous acquisition of vital signs → modality-specific dual-branch AI deep feature encoding → TCM fixed-weight mathematical weighted fusion differentiation → adaptive matching of constitution-based step-by-step intervention parameters → execution of microcurrent physical acupoint intervention → real-time dynamic negative feedback safety parameter adjustment of pulse diagnosis → seven-day cycle data review and iterative optimization.

[0034] This invention addresses the differences in characteristics between two types of heterogeneous data: static image modal and temporal dynamic modal. It configures a dedicated and adapted fixed AI dual-branch coding network structure: tongue and hand image modalities uniformly adopt a ResNet+ViT dual-branch coding structure, which is specifically used to extract local texture details, color distribution features, and global physical semantic representation features of tongue and hand images; pulse temporal modalities uniformly adopt an LSTM+Transformer dual-branch coding structure, which is specifically used to extract short-term pulse temporal fluctuation periodic features and long-term hierarchical pressure correlation global temporal features. This invention strictly adheres to the priority principles of TCM clinical diagnosis and treatment, setting fixed quantitative fusion weight ratios: pulse diagnosis as the core dominant weight (W1) = 60%, tongue diagnosis as the auxiliary correction weight (W2) = 25%, and hand diagnosis as the verification weight (W3) = 15%. A comprehensive constitution imbalance score of 0-100 points is calculated using a fixed mathematical quantitative formula, accurately classifying obesity into five TCM constitution types: phlegm-dampness accumulation, spleen and stomach qi deficiency, yang deficiency with dampness obstruction, damp-heat accumulation, and balanced / normal. Based on the constitution type and degree of imbalance, customized electrical stimulation quantitative intervention parameters are adaptively matched. This invention forms a complete technical closed loop from four dimensions: hardware physical architecture, AI model coding structure, TCM mathematical fusion algorithm, and physical intervention closed-loop control. This effectively improves the accuracy of quantitative diagnosis of obesity constitution and the targeted nature of weight loss intervention, addressing the pain points of traditional intelligent weight loss diagnosis, such as crudeness, a one-size-fits-all approach, weak intervention effects, and easy rebound after weight loss.

[0035] The mobile app software is configured with six core functional units, dedicated to standardized acquisition of tongue and hand images, image quality preprocessing and correction, ResNet+ViT image dual-branch feature encoding, trimodal AI weighted fusion for constitution differentiation, individualized intervention parameter matching and Bluetooth transmission, and full data archiving and conditioning suggestion push. This provides standardized image data and algorithmic computing power support for the quantitative differentiation of obesity. The smart weight loss intervention bracelet hardware is configured with six core hardware functional units, dedicated to PPG+FSR multimodal pulse layered temporal acquisition, motion posture noise reduction preprocessing, LSTM+Transformer pulse temporal dual-branch feature encoding, microcurrent electrical stimulation for weight loss intervention physical output, real-time pulse negative feedback parameter tuning, and device power management and display interaction. This provides reliable hardware support for accurate pulse temporal signal acquisition and the physical implementation of weight loss intervention. The overall system features a clear division of labor between software and hardware, with no interference between modal acquisitions and strict alignment of data timing.

[0036] In an exemplary embodiment, a TCM constitution-based weight loss intervention method is provided, which relies entirely on hardware operation combined with quantitative software calculation in a closed loop. The core mechanism utilizes dual-branch AI-specific coding, TCM fixed-weight mathematical fusion, tiered parameter adaptive control, real-time pulse feedback adjustment, and a seven-day cycle of quantitative iterative optimization. The entire process requires no manual intervention from professional medical personnel; the system operates fully automatically and intelligently. Figure 1 As shown, it includes the following steps: S1. Simultaneously collect multi-modal vital sign data of users during the meridian flow period, and perform time-series alignment on the multi-modal vital sign data; the multi-modal vital sign data includes tongue image, hand image and pulse time-series signal; S2. Preprocess the pulse time sequence signal, tongue image and hand image respectively, extract the traditional artificial medical features and AI deep features corresponding to each modality of vital signs data after preprocessing, and perform unified dimension normalization processing to obtain pulse normalization features, tongue normalization features and hand normalization features respectively. S3. Using pulse normalization features as teacher features, and tongue and hand normalization features as student features, the student features are aligned with the teacher features through training with a joint loss function to obtain aligned tongue and hand features; the pulse normalization features, aligned tongue and hand features are then fed into a lightweight MLP perceptron to output pulse single-modal scores, tongue single-modal scores, and hand diagnosis single-modal scores, respectively. S4. Calculate the modality difference coefficient based on the score output in step S3, and dynamically adjust the modality weight parameters according to the modality difference coefficient, the correlation between each modality's physical characteristics and TCM constitution differentiation. S5. Normalize the adjusted modal weight parameters and calculate the comprehensive obesity constitution bias score; and determine the obesity TCM constitution type based on the comprehensive obesity constitution bias score; S6. Based on the aforementioned obesity TCM constitution classification and comprehensive obesity constitution bias score, adaptively and stepwise match exclusive weight loss electrical stimulation intervention parameters; S7. Based on the exclusive weight loss electrical stimulation intervention parameters, initiate microcurrent electrical stimulation intervention on the targeted acupoints; collect pulse dynamic data in real time during the intervention process, and adjust the exclusive weight loss electrical stimulation intervention parameters in real time according to the pulse dynamic data.

[0037] Specifically, in combination Figure 2 The specific implementation process of each step is described. Step S1 specifically includes: 1. Wristband fitting and alignment + overall hardware power-on self-test initialization calibration. Users precisely wear and fix the smart weight loss intervention bracelet on the wrist at the dedicated pulse collection point of the radial artery, ensuring that the sensor detection area at the bottom of the bracelet fits tightly against the skin without loosening, displacement, or obstruction by clothing. After the bracelet is powered on, it automatically completes self-checks for sensor unit faults, Bluetooth communication handshake pairing, and zeroing and calibration of pulse collection reference parameters. Simultaneously, the mobile APP completes image acquisition light white balance, shooting distance and angle reference calibration, and after verifying that the two-way device communication link is connected and normal, the entire system enters the vital sign collection standby ready state.

[0038] 2. During the meridian flow period, the mobile phone and wristband work together to collect standardized data synchronously in three modalities: tongue, hand, and pulse. The optimal timeframes for spleen and stomach function according to the TCM meridian flow (7:00-9:00 and 19:00-21:00) were selected to ensure that the constitution identification data conformed to the TCM chronomedicine diagnostic standards. A pop-up window on the mobile app guided users to cooperate with the data collection process, sequentially completing standardized acquisition and basic image quality correction of tongue and palm images. Simultaneously, the smart bracelet initiated PPG+FSR pulse stratified temporal continuous acquisition, recording multi-cycle pulse pulsation data. Data from all three modalities were uniformly labeled with the same timestamp for temporal alignment. The raw vital signs data were temporarily cached and stored separately, constructing a standardized three-modal raw vital signs dataset.

[0039] Step S2 specifically includes: For three types of heterogeneous modal data—pulse time-series signals, tongue images, and hand images—independent preprocessing and feature extraction are performed. First, noise removal, data standardization, and distortion correction are completed through hardware and software joint preprocessing with precise quantization parameters. Then, traditional artificial medical features and AI deep features are extracted simultaneously. Finally, unified dimensional normalization processing is performed. The parameters are reproducible and the features are quantifiable throughout the process, completely eliminating various interference biases and providing a standard data foundation for subsequent fusion and diagnosis.

[0040] 1. Dedicated preprocessing for pulse timing signals (including precise quantization parameters) The original pulse signal is a dual-mode time-series signal of PPG photoelectric + FSR pressure, which is susceptible to limb movement, power frequency interference, and skin contact error. It adopts dual preprocessing of hardware filtering and software adaptive noise reduction. All parameters are fixed and quantized, with no manually adjustable deviations. ① Hardware preprocessing (second-order low-pass active filter) Filter type: Second-order Butterworth low-pass active filter; Cutoff frequency: f0=30Hz; Stopband attenuation: -40dB; Sampling frequency: 100Hz; Removes high-frequency electromyography interference above 30Hz, power frequency noise, and environmental electromagnetic noise, retains the effective physiological waveform of the pulse, and repairs waveform distortion and noise.

[0041] ② Software preprocessing (adaptive motion posture correction) Based on triaxial acceleration sensing data, dynamic baseline drift correction and waveform outlier removal algorithms are adopted; the outlier threshold is: waveform amplitude fluctuation > ±0.8V is judged as invalid noise and directly removed; sliding timing window: 256 points / window; overlapping window step size: 64 points; to complete the correction of limb swing and jitter interference, output a stable, continuous and distortion-free pure pulse timing signal.

[0042] 2. Pulse temporal signal feature extraction (parallel application of traditional features and AI deep features) ① Artificial characteristics of pulse diagnosis in traditional Chinese medicine (including complete calculation formulas) Simultaneously extracting the quantitative features of classic TCM pulse diagnosis, and assisting AI in deep feature extraction to improve diagnostic accuracy, a total of 8 core traditional physiological features are included: Pulse peak value (2) The trough value of the main pulse wave (3) Pulse rate characteristics (beats / minute) (4) Pulse pressure difference amplitude (5) Rhythmic regularity (6) Rising slope, falling slope, time-domain energy entropy.

[0043] Formula explanation: -Pure pulse time sequence -Sampling frequency, -Number of sampling points per cycle -Single cycle duration, - Timing window cycle number.

[0044] ② AI Deep Feature Encoding (Dual-Branch Network) LSTM branch: extracts short-term temporal fluctuations of the pulse, peak rhythm of the pulsation, continuous memory of the intensity of the pulse, and deep temporal features of the fast and slow rhythm; Transformer branch: extracts deep features of long-term temporal correlation of the pulse, layered pressure coupling of the pressure, and global waveform rhythm correlation; the two types of deep features are concatenated and fused with traditional features to obtain the complete pulse feature vector.

[0045] 3. Preprocessing of tongue and hand images (including precise quantization parameters) Tongue and hand images undergo completely consistent standardized image preprocessing, with fixed parameters and uniform image specifications. ① Image size normalization: Unify the size to the 224×224×3 RGB standard size; ② White balance calibration: Color temperature 5500K, gamma value γ=1.2; ③ Reflection removal: Gaussian filter kernel 7×7, reflection threshold normalization > 0.95 area passivation; ④ Shadow correction: Adaptive histogram equalization (CLAHE), block size 8×8; ⑤ Pixel normalization: Gray value mapping to the 0-255 standard range to eliminate differences in lighting and shooting angle.

[0046] 4. Feature extraction from tongue and hand images (parallel application of traditional features and AI deep features) ① Traditional Chinese Medicine tongue and hand visual artificial features: all traditional features are quantified and numericalized, aligned with the dimensions of AI deep features, and jointly enhanced feature recognition.

[0047] Traditional characteristics of tongue appearance: mean RGB value of tongue body, color of tongue coating, texture roughness, color entropy, local contrast, mean saturation, and thickness of tongue coating; Traditional characteristics of hand features: average palm color, energy of palm lines, contrast of light and dark, color variance, depth of veins, and evenness of skin tone.

[0048] ②AI deep feature encoding: Two-branch deep features are concatenated and traditional visual features are integrated to obtain complete feature vectors for tongue and hand images respectively.

[0049] ResNet residual branch: Extract local texture, color distribution, tongue coating color blocks, palm lines, complexion brightness, and fine-grained local features. ViT Visual Transformer Branch: Image blocks are encoded in 16×16 to extract global constitution semantics, overall complexion bias, and global visual features related to blood and qi throughout the body.

[0050] 5. Unified Dimensional Normalization of Multimodal Features The maximum-minimum normalization algorithm is used to uniformly map all features of the pulse, tongue, and hand modalities to the standard quantization range of [0,1], completely eliminating hardware differences, acquisition differences, and physiological baseline data offsets. Normalization calculation formula: Formula explanation: -Original feature values, -Eigenminimum, -Eigenvalue, - Normalized standard features; the final output is homogeneous, standard, and highly distinguishable features, which are directly used for subsequent trimodal TCM weighted fusion diagnosis.

[0051] Furthermore, such as Figure 3 As shown, step S3 differs from existing technologies that rely on simple CNN single feature extraction and crude feature concatenation. It employs a four-level AI fusion architecture: Cross-Modal Knowledge Distillation (CMKD) feature alignment, heterogeneous modality-specific encoding, feature vector-level weighted fusion, and multilayer perceptron classification inference. Combined with a fixed weighted ratio based on the priority of TCM clinical diagnosis, it achieves deep fusion of features from three heterogeneous modalities and accurate constitution differentiation. The specific process, model structure, and quantitative calculations are as follows: 1. Cross-modal knowledge distillation (CMKD): Achieving trimodal heterogeneous feature alignment This step addresses the core issue of the difficulty in bridging the heterogeneity of the three modalities (pulse time-series signal, tongue image, and hand image). Through teacher-student modality setting and joint loss function training, the features of the three modalities are mapped to the same feature space, eliminating modal differences and laying the foundation for subsequent weighted fusion. Specific implementation steps are as follows: ① Teacher-student modality setting: based on extracted pulse pattern normalized features Teacher characteristics (with pulse diagnosis as the core basis in Traditional Chinese Medicine, providing physiological characteristic constraints); and extracted tongue image normalized features. Hand-shaped normalization features For student features (which need to be aligned with teacher features to bridge the heterogeneous differences between images and time-series signals).

[0052] ② Joint Loss Function Design: Construct a total loss function to align student features with teacher features while ensuring the accuracy of single-modality physical fitness scores. The formula is as follows: Total loss (physical meaning: the total error of model training, reflecting the effect of feature alignment and the accuracy of single-modality physical fitness scores); : Classification cross-entropy loss (physical meaning: classification error of single-modality physical fitness score, ensuring that each modality feature has a clear ability to distinguish physical fitness), the formula is: ,in This is the true label for the constitution category (one-hot encoding, such as [1,0,0,0,0] for phlegm-dampness accumulation constitution). Let K be the probability of classifying a feature as the i-th type of constitution, where K=5 (five types of constitutions). Knowledge distillation loss (physical meaning: the difference in distribution between student and teacher features, used to achieve cross-modal feature alignment), the formula is as follows: ,in The probability distribution of teacher characteristics (pulse pattern) Let KL represent the probability distribution of student features (tongue / hand images), and KL represent the KL divergence (which measures the difference between two distributions; the smaller the value, the better the alignment). Weighting coefficient (value is 0.3, physical meaning: balances classification loss and distillation loss, ensures that feature alignment and physical condition discrimination accuracy are taken into account, and avoids physical condition discrimination bias caused by over-alignment).

[0053] ③ Feature alignment implementation: Minimize the total loss through the built-in software program in the mobile APP. Training tongue features Hand features Characteristics of Teachers with Pulse Alignment is performed to obtain the aligned tongue features. Hand features The three types of modal features are uniformly mapped to the same 256-dimensional feature space, completely bridging the heterogeneous differences between images and time-series signals.

[0054] 2. Four-level multimodal AI fusion inference architecture (ResNet+ViT+LSTM+Transformer) The fusion architecture strictly follows the priority of TCM clinical diagnosis and combines cross-modal aligned features to complete AI inference and body composition scoring in four levels. The specific descriptions and functions of each level of the model are as follows: Level 1: Image Dual-Branch Encoding Layer (ResNet+ViT): Continuing with the dual-branch visual network from step S2, the ResNet residual network (lightweight ResNet18, 3×3 convolutional kernels, 128 hidden layer dimensions) is responsible for extracting fine-grained local features of the tongue and hand images (tongue color patches, palm texture, etc.); the ViT visual Transformer network (16×16 image blocks, 3 encoding layers, 4 multi-head attention heads) is responsible for extracting global visual semantic features (overall complexion, physical constitution bias related features). The outputs of the two branches are concatenated, normalized, and aligned to obtain aligned tongue and hand image features. , .

[0055] Level 2: Temporal Dual-Branch Encoding Layer (LSTM + Transformer): Continuing with the dual-branch temporal network from step S2, the LSTM network (256 input points, 128 hidden layer dimensions, 30 iterations) is responsible for extracting short-term temporal features of the pulse (peak rhythm, intensity, and strength of the pulse). The Transformer temporal network (2 encoding layers, 4 multi-head attention heads, 64 temporal windows) is responsible for extracting long-term correlation features of the pulse (global waveform rhythm, pressure coupling relationship). The outputs of the two branches are concatenated and used as the teacher features. It is used for cross-modal alignment training.

[0056] Level 3: Unimodal physical fitness score inference layer (lightweight MLP perceptron): This layer aligns the three types of modal features ( , , The data is fed into three independent lightweight MLP perceptrons (input dimension 256, hidden layer 1, dimension 64, output dimension 1) to independently complete the inference of unimodal constitution bias scores and output pulse unimodal scores respectively. Tongue image monomodal scoring Hand diagnosis single-modal scoring The scoring range is 0-100 points (the higher the score, the more severe the physical imbalance).

[0057] Next, in steps S4 and S5, the fourth level of reasoning is completed. Fixed weight allocation is abandoned, and a dynamic cross-domain adaptive optimization logic guided by EEG is adopted. Combined with the core principles of TCM clinical diagnosis (pulse diagnosis as the foundation, tongue diagnosis and hand diagnosis as supplementary aids), adaptive weight adjustment is achieved through three steps: "modal difference detection - dynamic weight adjustment - feature adaptation optimization," balancing the standardization of diagnosis with individualized adaptation. Figure 4 As shown, the details are as follows: ① Modal difference coefficient detection: The software program analyzes the characteristics of the three modalities after cross-modal alignment. , , ) and corresponding single-modal scores ( , , The modality difference coefficient is calculated based on the distribution difference between the training baseline data and the training baseline data. (Physical meaning: The distribution distance between the current individual modal features and the training baseline features) The larger the value, the more severe the modal distribution shift (the formula is as follows): Formula explanation: d is the feature dimension (256 dimensions); Let be the mean of the k-th dimension feature of the current individual; The mean of the k-th dimension feature of the training baseline data; set a threshold. ,when When a significant modal distribution shift is detected, the weights need to be dynamically adjusted; when At this time, the initial weights are used without any additional adjustments.

[0058] ② Dynamic adaptive weight adjustment (core formula): based on modal difference coefficient Correlation between various modalities and TCM constitution differentiation Dynamically adjust the weight parameters of the three modes. (Pulse) (Tongue appearance) (Hand diagnosis), weight constraints are as follows: The initial weights are set according to the priority of TCM syndrome differentiation. , , The formula is adjusted as follows: Formula explanation: The weights of the k-th mode after adjustment; Let be the initial weights for the k-th mode; This is the weight adjustment coefficient (with a value of 0.5, physical meaning: the intensity of weight adjustment, controlling the magnitude of weight change during offset). The correlation between the k-th modal feature and TCM constitution differentiation (values ​​range from 0 to 1, with higher correlation values ​​indicating higher correlation). The closer to 1, the smaller the weight adjustment, preserving the discriminative advantage of the core modality.

[0059] ③ Feature Distribution Adaptation and Weight Normalization: Adaptive Batch Normalization (ANR) is used to perform secondary adaptation on the three modal features, dynamically adjusting the feature mean and variance to ensure consistency with the training baseline feature distribution, thus mitigating weight bias caused by modality shift. Subsequently, the adjusted weights are... , , Normalization is performed to ensure the weights sum to 1, and the final output is a dynamic weight adapted to the individual's physical characteristics. The comprehensive score calculation formula is as follows: Formula explanation: A comprehensive obesity constitution bias score ranging from 0 to 100 points. , , The scores were based on single-modal assessments of pulse, tongue, and hand appearance, respectively. , , The weights are dynamically and adaptively adjusted (not fixed values), which can be adapted in real time according to the individual modal distribution shift and the correlation of physical characteristics. This not only follows the core logic of TCM syndrome differentiation, but also solves the problem that fixed weights cannot adapt to individual differences.

[0060] Based on the overall physical fitness bias score Based on the correlation between aligned features and physical constitution, five types of standard obese TCM constitutions are accurately identified, with fixed and quantified identification thresholds: Divided into neutral and normal types; It is divided into spleen and stomach qi deficiency constitution; Divided into Yang deficiency and dampness obstruction constitution; Divided into damp-heat internal constitution; The body is classified into two types: one with phlegm and dampness accumulation. The results of the diagnosis (body type, comprehensive score, and single-modal score) are stored in real time in the local database of the mobile phone and visualized through the human-computer interaction interface of the APP, providing a direct basis for matching parameters for subsequent individualized weight loss intervention.

[0061] Step S6 specifically includes: The mobile app adaptively and progressively matches personalized weight-loss electrical stimulation intervention parameters to the wristband based on the trimodal comprehensive body constitution type and risk score. The app incorporates a database of body constitution, intervention parameters, and acupoints from the Traditional Chinese Medicine (TCM) guidelines for preventing disease in overweight individuals. The core matching logic is as follows: (1) Phlegm-dampness constitution: The corresponding treatment method is to strengthen the spleen and promote diuresis, resolve phlegm and eliminate turbidity. The core acupoints for electrical stimulation are Neiguan + Zusanli + Fenglong. The pulse waveform is sparse-dense wave, the basic frequency is 2Hz / 10Hz alternating, the current intensity is 0.5-1.2mA, the duration of a single intervention is 20 minutes, and it is done twice a day. (2) Damp-heat constitution: The corresponding treatment method is to eliminate dampness and turbidity and clear and drain latent fire. The core acupoints for electrical stimulation are Neiguan + Quchi + Yinlingquan. The pulse waveform is a continuous wave with a basic frequency of 5Hz and a current intensity of 0.4-1.0mA. The duration of a single intervention is 18 minutes, twice a day. (3) Qi deficiency constitution / Yang deficiency and dampness obstruction constitution: corresponding to the treatment of nourishing the original qi and strengthening the spleen. The core acupoints for electrical stimulation are Neiguan + Zhongwan (bracelet adapted to the mapping of adjacent acupoints) + Qihai (bracelet adapted to the mapping of adjacent acupoints). The pulse waveform is intermittent wave, the basic frequency is 1Hz, the current intensity is 0.3-0.8mA, the single intervention duration is 25 minutes, once a day; (4) Balanced constitution: Only basic conditioning, the core acupoint for electrical stimulation is Neiguan, the pulse waveform is a low-frequency continuous wave, the basic frequency is 1Hz, the current intensity is 0.2-0.5mA, the duration of a single intervention is 10 minutes, once a day.

[0062] The app matches basic intervention parameters based on body type, and then adjusts the intensity in a stepwise manner according to the comprehensive body imbalance score (0-100 points): the basic intensity is used when the score is <30 points (mild imbalance); the intensity is increased by 20% when the score is 30-60 points (moderate imbalance); and the intensity is increased by 40% when the score is >60 points (severe imbalance). At the same time, the basic parameters can be automatically fine-tuned by ±1Hz and ±1mA according to the slight fluctuation of the pulse on the day to achieve individualized and precise adaptation. The parameters are sent to the wristband via Bluetooth for execution.

[0063] Step S7 specifically includes: The intelligent weight loss intervention bracelet, based on the received individualized and quantified weight loss electrical stimulation intervention parameters, activates a microcurrent constant current source circuit to stably output the corresponding waveform, frequency, and current intensity, targeting and stimulating corresponding acupoints for weight loss physical intervention. Throughout the intervention, the bracelet continuously collects dynamic pulse data in real time, monitoring changes in the body's physiological state. If abnormal pulse fluctuations are detected, the system automatically reduces the intervention current intensity by 10%-20%, and the abnormal feedback data is transmitted back to the mobile app in real time, achieving dynamic negative feedback closed-loop safety parameter adjustment during the intervention process.

[0064] Furthermore, after a single intervention is completed, the wristband transmits the actual intervention duration, real output parameters, and pulse feedback data back to the mobile app. The mobile app integrates the original trimodal data, AI constitution identification results, biased quantitative scores, and all intervention execution data into a unified archive, forming a personalized weight loss health record for each user. Based on the constitution type, the system automatically pushes suitable dietary recommendations (e.g., for phlegm-dampness constitution, recommends spleen-strengthening and dampness-removing ingredients such as Job's tears and Poria cocos; for damp-heat constitution, recommends heat-clearing and dampness-removing ingredients such as winter melon and mung beans), personalized lifestyle and exercise recommendations, forming an integrated conditioning model combining hardware physical intervention and software lifestyle guidance.

[0065] Furthermore, the system uses a fixed seven-day review and iteration cycle for TCM constitution conditioning, periodically summarizing changes in constitution imbalance scores, body fat and weight indicators, physical characteristics, and intervention compliance data. If the constitution score decreases by less than 5 points within the cycle, the baseline parameters for strengthening the intervention are fine-tuned (frequency +0.5Hz / intensity +0.1mA); if the score decreases by 5 points or more, the current parameters are maintained. Through this seven-day cycle quantitative iteration long-term correction mechanism, the system continuously adapts to the dynamic changes in the user's constitution, correcting the imbalance of obesity-related constitutions at its root, and achieving the long-term closed-loop weight loss conditioning goal of healthy fat reduction and less rebound.

[0066] In another exemplary embodiment, based on the same inventive concept as the method embodiment, a TCM constitution differentiation and weight loss intervention system based on cross-modal knowledge distillation is provided, including a mobile APP software terminal and a smart weight loss intervention bracelet hardware terminal, which are connected wirelessly via Bluetooth. The overall hardware architecture of this invention requires no additional development or customization of specialized TCM pulse diagnosis equipment, nor does it require the configuration of medical-grade professional image capturing equipment. It relies entirely on two types of universal standard hardware terminals—existing mature, mass-produced consumer smartphones and low-power smart wearable weight-loss intervention bracelets—to collaboratively network and fully realize the core functions of trimodal vital sign acquisition, AI feature encoding calculation, TCM constitution differentiation, and individualized weight-loss intervention and regulation. The hardware structure of this invention requires no structural modification, hardware upgrade, or additional dedicated peripherals, and is fully compatible with existing commercially available civilian smart terminal hardware systems. This results in low mass production costs, strong versatility, and convenient deployment.

[0067] The system's overall hardware interaction and connection logic is clear and well-defined, employing a separate and collaborative hardware deployment architecture: The smartphone terminal integrates an image acquisition unit, a main control processing unit, and memory and solid-state storage units via an internal system bus, primarily responsible for the core tasks of standardized tongue and hand image acquisition and storage, local inference calculation of lightweight AI models, multimodal fusion diagnostic calculation, and intervention parameter distribution control; The smart weight loss intervention bracelet integrates a main control MCU unit, pulse sensor acquisition unit, motion posture sensor unit, microcurrent intervention output unit, and Bluetooth wireless communication unit via an internal hardware bus, primarily responsible for pulse time-series physiological signal acquisition, data noise reduction preprocessing, and physical execution output of acupoint microcurrent weight loss intervention; The smartphone terminal and the smart bracelet terminal establish a stable two-way data interaction channel via a Bluetooth 5.0 wireless communication link, enabling synchronous interaction of uplink transmission of raw vital sign data, downlink transmission of intervention control parameters, and real-time feedback parameter adjustment data return. All hardware units work together to support the stable operation of the system's fully closed-loop intelligent workflow.

[0068] The specific core hardware quantitative configuration and the composition of the built-in hardware functional units are as follows: 1. Hardware main body and built-in functional units of mobile APP software The processor uses a Snapdragon 778G or higher, or a Dimensity 810 or higher mobile main control chip of equivalent specifications, to meet the local offline inference computing power requirements of lightweight dual-branch deep learning models on mobile devices; the device has ≥8GB of RAM for temporary caching of image feature data and intermediate parameters for model inference, and ≥128GB of solid-state storage for permanent storage of system software programs, AI model parameters, and full user health records; the rear high-definition camera has ≥48MP resolution and supports automatic white balance calibration, image reflection suppression correction, and size normalization cropping; it has a built-in standard Bluetooth 5.0 communication module, and the overall algorithm inference response latency is ≤800ms, fully adapting to the real-time health diagnosis and parameter adjustment requirements of mobile devices.

[0069] The mobile app software integrates six hardware adaptation software functional units, specifically including: (1) Tongue image standardization acquisition module: guides users to acquire tongue images under uniform lighting and angle, and automatically completes reflection removal, shadow correction and white balance calibration; (2) Hand image standardization acquisition module: guides users to acquire a full palm image and automatically completes palm print area positioning and image size normalization; (3) ResNet+ViT image dual-branch coding module: Lightweight deployment of ResNet residual network and ViT visual Transformer network to extract local texture and global semantic features of tongue and hand images; (4) Trimodal AI weighted fusion diagnostic module: Built-in TCM fixed weight fusion algorithm to complete multimodal feature weighted calculation and constitution identification; (5) Intervention parameter matching and distribution module: Match exclusive electrical stimulation parameters according to body type and distribute them to the wristband via Bluetooth; (6) Data archiving and personalized conditioning push module: Stores all collected and intervention data, and pushes dietary and lifestyle conditioning suggestions that are suitable for the individual's constitution.

[0070] 2. Hardware main body and built-in functional units of the smart weight loss intervention bracelet The smart weight loss intervention bracelet's hardware utilizes a low-power wearable-grade MCU as its main control core, comprehensively managing the logic control and data preprocessing of all hardware units. It is equipped with a dual-mode pulse layering acquisition unit featuring PPG photoelectric sensing and FSR pressure sensing, with a continuous pulse signal sampling rate ≥100Hz, meeting the precise acquisition requirements of traditional Chinese medicine pulse layering pulsation signals. A built-in three-axis acceleration motion posture sensing module is used to identify limb movement status in real time and perform motion noise reduction and correction processing of pulse signals. It features a medical-grade microcurrent constant current source output circuit, supporting precise adjustment of intervention waveform, operating frequency, and current intensity at multiple levels. It includes a standard lithium battery power management module and a small visual display screen, and a built-in Bluetooth 5.0 communication module. The overall electrical performance and safety indicators fully comply with the mandatory safety standards for civilian wearable devices. The bracelet integrates six core hardware function execution units, specifically including: (1) Main control module: a low-power MCU is used to control the overall logic and data processing of the wristband; (2) PPG+FSR multimodal pulse layer acquisition module: The pulse pulsation signal is acquired by the PPG photoelectric sensor and the pulse layer acquisition is realized by the FSR pressure sensor. (3) Motion posture noise reduction preprocessing module: Built-in three-axis accelerometer to identify limb movement state and complete pulse signal noise reduction; (4) LSTM+Transformer pulse temporal coding module: a lightweight LSTM and Transformer network is embedded to extract pulse temporal features; (5) Microcurrent electrical stimulation weight loss intervention module: built-in constant current source circuit, outputting microcurrent with adjustable waveform, frequency and intensity to stimulate targeted acupoints; (6) Power management and display interaction module: responsible for the power management of the wristband and displaying information such as intervention status and physical fitness score.

[0071] 3. Supporting hardware: This invention does not require any additional dedicated auxiliary hardware or peripheral devices. It can complete basic human-computer operations such as program startup, parameter viewing and data management by relying on the native touch screen and button interaction module of the mobile phone. All supporting hardware are existing general-purpose mass-produced equipment, which do not require additional customization and processing. It has strong scalability and low difficulty in market popularization.

[0072] Furthermore, the main body of the software in this invention is a computer program stored in the storage units of a mobile phone and a smart band. It is compatible with conventional mobile operating systems such as Android 10.0 and above, and iOS 14.0 and above. Developed using the Python programming language and a lightweight mobile deep learning framework, the model undergoes quantization compression and pruning, supporting local offline inference without cloud computing power dependency. Each module of the software strictly corresponds to the entire business process of data collection, preprocessing, encoding, fusion, identification, intervention, and archiving. The modular division of labor is clear, and the logic is tightly coupled. The specific core software module functions are as follows: 1. Tongue and Hand Image Acquisition and Image Quality Preprocessing Module: Based on the native image acquisition interface of the mobile terminal, it realizes standardized shooting of tongue and hand images, automatically completes reflection removal, shadow correction, white balance calibration and size normalization preprocessing, and provides standardized image input for image dual-branch encoding.

[0073] 2. Pulse Acquisition and Motion Noise Reduction Preprocessing Module: Embedded inside the wristband, it realizes continuous acquisition of PPG+FSR pulse timing signals, integrates hardware filtering matching software noise reduction and motion posture correction algorithms, and outputs clean pulse timing waveform data.

[0074] 3. ResNet+ViT image dual-branch and LSTM+Transformer pulse temporal coding module: The core AI feature extraction unit, with a lightweight deployment of a dual-branch network structure, completes the encoding and extraction of local and global features of the image and short-term and long-term temporal features of the pulse, and outputs a standardized modality depth feature vector.

[0075] 4. TCM Fixed-Weight Trimodal Weighted Fusion and Constitution Classification Module: It has built-in fixed-weight fusion calculation formula and four-level AI inference code, automatically completes single-modal score calculation and comprehensive bias score calculation, and realizes accurate TCM constitution classification and local data storage for five types of obesity.

[0076] 5. Adaptive Intervention Parameter Matching and Bluetooth Communication Distribution Module: It has a built-in constitution-intervention parameter database, which automatically matches and quantifies intervention parameters based on constitution type and bias score, and completes real-time parameter distribution and feedback data transmission from mobile phone to wristband via Bluetooth protocol.

[0077] 6. Pulse negative feedback adjustment and seven-day cycle iterative review module: Real-time monitoring of pulse fluctuations to complete safe dynamic parameter adjustment, periodic summarization of constitution and intervention data, automatic iterative optimization of subsequent intervention benchmark parameters, and realization of long-term closed-loop conditioning and optimization.

[0078] 7. Data archiving and personalized conditioning push module: All collected and intervention data are uniformly archived to establish health records. Based on body constitution classification, dietary, rest, and exercise recommendations are automatically pushed to realize integrated hardware and software conditioning services.

[0079] For example, based on the above methods and systems, a specific TCM method for differentiating obesity constitutions is provided, focusing on three core obese constitutions that are prevalent in clinical practice: phlegm-dampness accumulation, damp-heat retention, and spleen-stomach qi deficiency with yang deficiency and dampness obstruction. The balanced and normal constitution serves only as a baseline for basic health control, and can be flexibly expanded to subdivide imbalanced subtypes according to clinical treatment needs. The physical sign collection environment covers two common home scenarios: natural light in a typical home environment and soft lighting in a fixed indoor environment, verifying the system's stability and environmental adaptability in daily home data collection. The test subjects are selected from individuals of different ages, genders, and degrees of obesity to verify the system's universality and population suitability. The entire process of physical sign collection, AI inference, syndrome differentiation calculation, and intervention regulation is completed locally offline, without relying on cloud computing power, ensuring user privacy data security and device response speed. Specifically, this includes: 1. Hardware and software basic deployment and parameter initialization This implementation method eliminates the need for repetitive architectural design; it only requires basic assembly and alignment according to the hardware specifications of this invention: the smart weight loss intervention bracelet is worn at the radial artery pulse acquisition point on the wrist, the mobile phone is fixed in the shooting position, and the two ends establish stable communication via Bluetooth 5.0. The device's self-test functions for sensor acquisition, AI inference, and microcurrent output are considered normal. The mobile phone and bracelet each have their corresponding embedded software and lightweight inference program installed. The model is pre-quantized, pruned, and adapted for local offline operation. Basic configurations such as the system's sampling rate, filtering parameters, image specifications, feature dimensions, and TCM fusion weights are uniformly and permanently set, meeting the conditions for direct implementation and operation.

[0080] 2. Trimodal synchronous acquisition, preprocessing, and AI-based dialectical classification execution. During the golden period for collecting data on the spleen and stomach meridian flow in Traditional Chinese Medicine (TCM), the system automatically and synchronously completes the standardized acquisition of tongue and hand images and the acquisition of pulse time-series signals. After unifying the timestamps and aligning the time sequences, the device automatically performs image quality correction and pulse motion noise reduction preprocessing. The system calls a lightweight dual-branch AI model to automatically complete the encoding of image and time-series features. After normalization processing, the data is input into a fixed-weight TCM formula to automatically calculate a comprehensive score of 0-100 points for constitution imbalance, quickly identifying three types of core obese constitutions: phlegm-dampness accumulation, damp-heat retention, and spleen-stomach qi deficiency with yang deficiency and dampness obstruction. The balanced constitution serves as a healthy baseline. The entire process of syndrome differentiation and classification is completed automatically without manual intervention.

[0081] 3. Implementation of simultaneous acquisition and data preprocessing of three modalities during the meridional period ① Standardized synchronous vital sign acquisition: The system acquisition process is started during the golden period of the TCM spleen and stomach meridian flow (7:00-9:00 or 19:00-21:00). The mobile APP pop-up guides the subjects to cooperate in the standardized acquisition of tongue image and full palm image. The system automatically completes reflection removal, shadow correction, white balance calibration (color temperature 5500K, gamma value γ=1.2) and size normalization (224×224×3 RGB) preprocessing. At the same time, the smart bracelet starts PPG+FSR dual-mode pulse layered continuous acquisition (sampling frequency 100Hz), and records multi-cycle pulse pulsation time series data. The three types of modal acquisition data are uniformly marked with timestamps to achieve accurate time series alignment. The raw data are temporarily cached and stored to form a standardized three-modal raw vital sign dataset.

[0082] ② Modal-specific noise reduction and image quality enhancement preprocessing: The original pulse time series data from the wristband undergoes hardware second-order low-pass active filtering (cutoff frequency 30Hz, stopband attenuation -40dB) and software motion posture adaptive dual noise reduction (abnormal amplitude ±0.8V removal, 256-point sliding window baseline drift correction) to remove limb tremors, baseline drift, and power frequency noise interference, outputting clean and regular pulse time series waveform data; the tongue and hand images from the mobile phone undergo image quality enhancement and defect correction (Gaussian filtering reflection removal, CLAHE adaptive histogram equalization shadow correction), outputting standardized and compliant image data, providing a high-quality input foundation for subsequent AI dual-branch feature encoding and cross-modal knowledge distillation feature alignment.

[0083] 4. Implementation of Constitution Differentiation by Integrating AI Dual-Branch Feature Encoding with Traditional Chinese Medicine Adaptive Weighting ① Dual-branch deep feature encoding and cross-modal knowledge distillation alignment: Preprocessed tongue and hand images are input into a ResNet+ViT dual-branch encoding model (ResNet18 extracts local texture and color features, ViT divides into 16×16 blocks to extract global semantic representation features), and concatenated to form deep feature vectors for tongue and hand images; denoised pulse time-series data are input into an LSTM+Transformer dual-branch encoding model (LSTM extracts short-term temporal fluctuation cycle features, Transformer extracts long-term hierarchical correlation features), and concatenated to form deep feature vectors for pulse images; the three types of modal features are uniformly normalized to the 0~1 standard range to complete feature homogenization preprocessing. Then, the cross-modal knowledge distillation (CMKD) process is started, using pulse features as teacher features and tongue and hand features as student features. Through training with a joint loss function (Ltotal=LCE+0.3·LKD), the three types of modal features are aligned, and the aligned tongue and hand features and the original pulse features are output, bridging the modal heterogeneity differences.

[0084] ② Level 4 AI Inference and Three Types of Obesity Core Body Mass Classification: The aligned three modal features are fed into a lightweight MLP perceptron for independent inference, outputting pulse single-modal scores S1, tongue single-modal scores S2, and hand diagnosis single-modal scores S3; an adaptive weight adjustment process is initiated to calculate the modal difference coefficient Ddiff (threshold θ=0.2), and the weights of the three modalities are dynamically adjusted according to the formula Wk=Winit,k·exp(-0.5·Ddiff·sk) (initial weights Winit1=0.6, Winit2=0.25, Winit3=0.15). After normalization, the weights are input into the fusion formula S=S1×W1+S2×W2+S3×W3 to calculate the comprehensive body mass bias score S. Based on the score threshold, the system automatically identifies three common types of obese TCM constitutions: phlegm-dampness accumulation, damp-heat retention, and spleen-stomach qi deficiency with yang deficiency and dampness obstruction. The balanced constitution serves as a baseline control for the classification. The syndrome differentiation results, quantitative scores, and dynamic weights are stored locally and visualized on mobile devices, providing a direct classification basis for subsequent precise matching of exclusive electrical stimulation intervention parameters.

[0085] 5. Matching and distributing physical fitness-based intervention parameters and implementing microcurrent intervention. ① Precise matching and tiered parameter adjustment of the three types of body constitution-specific intervention parameter database: The mobile APP is built with a dedicated database based on Yang Guizhou's (2016) guidelines for the prevention of disease in overweight patients using traditional Chinese medicine, which provides a one-to-one correspondence between body constitution, intervention parameters, and acupoints. The specific treatment methods, targeted acupoints, pulse waveforms, base frequencies, current intensities, intervention durations, and daily intervention frequency for the three core obese body constitutions are all solidified into standardized quantitative parameters. The specific basic intervention baseline parameter configurations for the three body constitutions are as follows: First, the phlegm-dampness accumulation constitution corresponds to the treatment method of strengthening the spleen and eliminating dampness, resolving phlegm and eliminating turbidity. The core acupoints for electrical stimulation are set as Neiguan + Zusanli + Fenglong. The pulse waveform adopts a sparse-dense wave, the base frequency is set to alternate between 2Hz and 10Hz, the base current intensity is 0.5-1.2mA, the standard intervention duration is 20 minutes, and the daily intervention frequency is 20 minutes. The treatment plan is divided into three categories: 1) Pre-treatment: 2 sessions; 2) Damp-heat constitution: corresponding to the treatment of eliminating dampness and clearing latent heat, with the core acupoints for electrostimulation set as Neiguan + Quchi + Yinlingquan, using continuous wave pulses, a fixed base frequency of 5Hz, a base current intensity of 0.4-1.0mA, a standard intervention duration of 18 minutes per session, and 2 sessions per day; 3) Spleen and stomach qi deficiency with yang deficiency and dampness obstruction constitution: corresponding to the treatment of tonifying primordial qi and strengthening the spleen, with the core acupoints for electrostimulation set as Neiguan + Zhongwan (with adjacent acupoint mapping on the wristband) + Qihai (with adjacent acupoint mapping on the wristband), using intermittent wave pulses, a fixed base frequency of 1Hz, a base current intensity of 0.3-0.8mA, a standard intervention duration of 25 minutes per session, and 1 session per day; 4) Balanced normal constitution: used as a basic control, only for daily maintenance and strengthening intervention. The system first matches the corresponding set of basic benchmark intervention parameters based on the automatically identified body type. Then, it implements a three-tiered adaptive intensity adjustment according to the comprehensive body imbalance score of 0-100 points: the original basic intensity remains unchanged for scores <30 points (mild imbalance), the overall intervention intensity is increased by 20% for scores 30-60 points (moderate imbalance), and the overall intervention intensity is increased by 40% for scores >60 points (severe imbalance). At the same time, it automatically makes small adjustments to the basic frequency and current intensity by ±1Hz and ±0.1mA, respectively, based on the real-time fluctuation of the pulse on the same day. This achieves triple individualized and precise adaptation of body type, imbalance degree, and daily physiological state. All the final optimized intervention quantitative parameters are sent to the smart bracelet in real time via Bluetooth 5.0, and the bracelet completes the parameter loading and is ready to execute.

[0086] ② Targeted acupoint microcurrent intervention negative feedback execution: The wristband activates the microcurrent constant current source circuit according to the received quantification parameters, and outputs the corresponding waveform, frequency and intensity microcurrent to target and stimulate the corresponding acupoints for weight loss and conditioning in traditional Chinese medicine. Throughout the intervention, the wristband continuously collects dynamic pulse data in real time. If abnormal pulse fluctuations are detected, the intervention intensity is automatically reduced by 10%-20%. Abnormal data is transmitted back to the mobile APP in real time, realizing dynamic negative feedback closed-loop safe parameter adjustment during the intervention process, taking into account both intervention effect and safety of use.

[0087] 6. Data archiving and storage, and implementation of a seven-day cycle iterative long-term management system. After a single intervention, the wristband transmits all data back to the mobile app, including the actual intervention duration, real output waveform frequency and current intensity parameters, and real-time pulse feedback fluctuations. The app integrates the original trimodal data, AI-based syndrome differentiation results, quantitative scores of body constitution imbalances, and all intervention execution records for unified archiving, creating a personalized body constitution weight loss and health conditioning profile for each user. Based on the three body constitution types, the system automatically pushes personalized weight loss conditioning suggestions, including suitable dietary restrictions, sleep patterns, and exercise. The system uses a fixed seven-day review and iteration cycle for TCM body constitution conditioning, periodically summarizing changes in body constitution imbalance scores, body fat and weight indicators, changes in trimodal physical characteristics, and intervention compliance data. If the body constitution score decreases by less than 5 points within the cycle, the intervention baseline parameters are slightly adjusted to strengthen the intervention; if the score decreases by ≥ 5 points, the current intervention parameters remain unchanged. Through seven-day cycle iterations and long-term corrections, the system continuously adapts to the dynamic changes in the user's body constitution, correcting the imbalance of obesity-related body constitutions from the root, achieving a long-term closed-loop conditioning effect of healthy weight loss and resistance to rebound.

[0088] 7. Comparison Experiment of System Implementation Effects To fully verify the differences in actual diagnostic accuracy, constitution improvement efficiency, fat and weight loss effect, and intervention safety of the present invention's split-body collaborative three-modal AI diagnostic intervention system compared to existing conventional single weight loss devices, single pulse conditioning devices, and single image constitution identification devices, this implementation method sets up three groups of controlled experimental tests under the same test conditions. The test sample uniformly selected 60 overweight and obese subjects, and divided them into three groups of 20 people each. The constitutions covered the three core obese constitutions of phlegm-dampness, damp-heat, and qi deficiency and yang deficiency. The home use environment, intervention period of 28 days, and basic control standards for work and rest and diet were uniform. The core indicators for comparison included five core dimensions: constitution diagnosis accuracy, average body fat reduction rate, average weight reduction, constitution bias score improvement value, and incidence of adverse intervention feedback. The measured data were objectively and quantitatively compared to fully demonstrate the technological progress and practical application advantages of the present invention. The specific implementation effect comparison experimental data are shown in Table 1 below.

[0089] Table 1 Comparison of experimental results with actual measurements The experimental data table above clearly shows that, compared with existing conventional weight loss devices and single-modal identification devices, this invention significantly improves the accuracy of constitution identification, achieves significantly better results in reducing body fat and weight than the control group, and multiplies the improvement in constitution bias by using a three-modal synchronous acquisition of split-body collaborative data, dual-branch AI-specific feature encoding, TCM fixed-weight three-category constitution-based quantitative fusion differentiation, constitution-specific acupoint parameter matching stepwise intervention, and pulse negative feedback closed-loop iterative optimization mechanism. Furthermore, there is no adverse feedback throughout the intervention process, resulting in a high intervention safety factor. This invention effectively solves the core problems of existing technologies, such as inaccurate differentiation, general but non-specific intervention, weak fat reduction effect, slow constitution improvement, and potential safety hazards. This fully verifies that the technical solution of this invention is feasible, has excellent intervention effect, and high clinical practical value.

[0090] 8. Multi-module synergistic ablation experiment To demonstrate that the combination of dual-branch heterogeneous coding, cross-modal knowledge distillation CMKD alignment, and exponential dynamic adaptive weight fusion technologies employed in this invention is not a simple superposition of conventional technologies in the field, but rather produces a significant and unexpected synergistic gain effect, this invention sets up multiple sets of module ablation control experiments. Under a unified dataset, unified training parameters, and unified testing environment, the core innovative modules are removed one by one, and the changes in the model's physical constitution identification accuracy are compared to verify the synergistic effect of the technology combination of this invention.

[0091] Five validation models were set up for the ablation experiment: a dual-branch encoding model only, a dual-branch encoding + CMKD alignment model, a dual-branch encoding + dynamic weight model, a basic model without CMKD and dynamic weights, and the complete full-module model of this invention. The experimental results are shown in Table 2 below.

[0092] Table 2 Ablation Experiment Results The above ablation experiments show that: (1) when the dual-branch encoding is retained alone and CMKD and dynamic weights are removed, the accuracy drops significantly by 19.2%; when any core innovation module is missing alone, there is a significant performance degradation, proving that the innovation modules of this invention cannot be deleted or replaced individually. (2) When the two core modules of CMKD feature alignment and dynamic adaptive weights are removed at the same time, the model accuracy drops sharply from 92.8% to 73.6%, a decrease of 19.2%, which is much greater than the sum of the performance loss of a single missing module, proving that the combination of technologies in this invention is not a simple superposition, but produces an unexpected synergistic effect of 1+1>2. (3) Compared with the conventional CNN+fixed weight combination in the field, the overall architecture accuracy of this invention is improved by 23.4%, which fully proves that this specific multimodal fusion system is not a conventional choice in the industry and has significant technological progress and creativity.

[0093] 9. Sensitivity Comparison Experiment of Dynamic Weight Algorithm This invention sets up a comparative experiment of multiple mainstream dynamic weighting algorithms under the same conditions, including four schemes: linear decay weighting, ordinary attention weighting, fixed weight weighting, and the exponential decay weighting of this invention. All schemes use the dual-branch encoding and CMKD aligned features of this invention, only replacing the weight fusion module. The three indicators of identification accuracy, feature robustness, and cross-scene generalization ability are compared. The weight comparison experiment results are shown in Table 3 below.

[0094] Table 3 Results of the weighted comparison experiment The algorithm sensitivity experiment shows that: (1) Although the conventional linear decay and ordinary attention weighting methods in the industry can achieve dynamic weighting, they have defects such as rigid correction, noise sensitivity, and suppression of modality advantage features, resulting in limited performance improvement. (2) The original modality difference coefficient Ddiff + physical sign correlation sk + exponential decay correction mechanism of this invention can achieve: the greater the modality shift, the faster the invalid feature weights will decay adaptively; the higher the physical sign correlation, the more the core diagnostic feature weights will be retained adaptively. The control logic fits the TCM diagnostic rules, and the nonlinear control fits the nonlinear change characteristics of human physical signs better. (3) Under the same feature input conditions, this algorithm improves the accuracy by 4.1% and reduces the generalization error by 4.3% compared with the conventional attention weighting, and improves the accuracy by 6.6% compared with the linear weighting.

[0095] This embodiment implements the entire process through standardized hardware deployment, software configuration, local inference of the AI ​​dual-branch model, trimodal synchronous acquisition and preprocessing, TCM fixed-weight quantitative diagnosis, tiered adaptive intervention execution, pulse negative feedback safety adjustment, and seven-day iterative optimization. It accurately reproduces the core technical solution of this invention, fully verifying the rationality of the system structure, the stability of the model operation, the accuracy of diagnosis and quantification, the safety and controllability of intervention, and the long-term and stable conditioning effect. This implementation relies on general-purpose mass-produced conventional hardware, eliminating the need for dedicated medical acquisition equipment and cloud computing power. The entire process of constitution diagnosis and weight loss intervention can be completed locally offline. The hardware mass production cost is low, the software deployment is simple, and home operation is convenient. The split-collaborative acquisition architecture avoids data interference from the source, dual-branch AI encoding improves feature extraction accuracy, TCM quantitative integration solves the problem of lack of standardized diagnosis, tiered intervention achieves precise conditioning for each individual, negative feedback parameter adjustment ensures safety, and iterative cycles achieve weight loss without rebound, effectively overcoming various core defects of existing intelligent weight loss technologies.

[0096] Without departing from the core innovative concept and technical solution of this invention, those skilled in the art may make detailed adaptive adjustments to the model feature dimensions, intervention parameter values, and subdivided body constitution types according to actual mass production and clinical treatment needs, and all such adjustments should fall within the protection scope of this invention.

Claims

1. A TCM constitution-based weight loss intervention method based on cross-modal knowledge distillation, characterized in that, Includes the following steps: S1. Simultaneously collect multi-modal vital sign data of users during the meridian flow period, and perform time-series alignment on the multi-modal vital sign data; the multi-modal vital sign data includes tongue image, hand image and pulse time-series signal; S2. Preprocess the pulse time sequence signal, tongue image and hand image respectively, extract the traditional artificial medical features and AI deep features corresponding to each modality of vital signs data after preprocessing, and perform unified dimension normalization processing to obtain pulse normalization features, tongue normalization features and hand normalization features respectively. S3. Using pulse normalization features as teacher features, and tongue and hand normalization features as student features, the student features are aligned with the teacher features through training with a joint loss function to obtain aligned tongue and hand features; the pulse normalization features, aligned tongue and hand features are then fed into a lightweight MLP perceptron to output pulse single-modal scores, tongue single-modal scores, and hand diagnosis single-modal scores, respectively. S4. Calculate the modality difference coefficient based on the score output in step S3, and dynamically adjust the modality weight parameters according to the modality difference coefficient, the correlation between each modality's physical characteristics and TCM constitution differentiation. S5. Normalize the adjusted modal weight parameters and calculate the comprehensive obesity constitution bias score; and determine the obesity TCM constitution type based on the comprehensive obesity constitution bias score; S6. Based on the aforementioned obesity TCM constitution classification and comprehensive obesity constitution bias score, adaptively and stepwise match exclusive weight loss electrical stimulation intervention parameters; S7. Based on the exclusive weight loss electrical stimulation intervention parameters, initiate microcurrent electrical stimulation intervention on the targeted acupoints; collect pulse dynamic data in real time during the intervention process, and adjust the exclusive weight loss electrical stimulation intervention parameters in real time according to the pulse dynamic data.

2. The TCM constitution-based weight loss intervention method according to claim 1, characterized in that, Step S2, the extraction of AI deep features, includes: The preprocessed tongue and hand images are input into the ResNet+ViT dual-branch image encoding architecture. The ResNet branch extracts local texture detail features, and the ViT branch extracts global body semantic features. The outputs of the two branches are concatenated to form the tongue and hand feature vectors. The preprocessed pulse time series signal is input into the LSTM+Transformer pulse time series dual-branch encoding architecture, where the LSTM branch extracts short-term time series fluctuation features and the Transformer branch extracts long-term time series correlation features. The outputs of the two branches are concatenated to form the pulse feature vector.

3. The TCM constitution-based weight loss intervention method according to claim 1, characterized in that, The preprocessing of the tongue or hand image in step S2 includes: Image size is normalized to 224×224×3 RGB standard size, white balance is calibrated for color temperature and gamma value, Gaussian filter kernel 7×7 for reflection removal, adaptive histogram equalization for shadow correction, and pixels are normalized to the 0-255 standard range.

4. The TCM constitution-based weight loss intervention method according to claim 1, characterized in that, The joint loss function in step S3 is as follows: in, Indicates joint loss, Represents the classification cross-entropy loss. This represents the knowledge distillation loss. These are the weighting coefficients.

5. The TCM constitution-based weight loss intervention method according to claim 1, characterized in that, In step S4, the modal difference coefficient is calculated using the following formula: in, This represents the modal difference coefficient, where d is the feature dimension. Let be the mean of the k-th dimension feature of the current individual. The mean of the k-th dimension feature of the training baseline data; The dynamic adjustment of modal weight parameters based on modal difference coefficients, the correlation between modal characteristics and TCM constitution differentiation includes: when When adjusting the modal weights, use the following formula: in, To adjust the weights of the k-th mode, Let these be the initial weights for the k-th mode. This is the weighting adjustment factor. The correlation between the characteristics of various modalities and TCM constitution differentiation.

6. The TCM constitution-based weight loss intervention method according to claim 5, characterized in that, In step S5, the comprehensive obesity body constitution bias score is calculated using the following formula: in, This indicates a single-modal pulse score. This indicates a single-modal score for tongue image. This indicates a single-modal score for hand diagnosis. , , To adjust the weight parameters of each mode; The method of determining the TCM constitution type of obesity based on the comprehensive obesity constitution bias score includes: Divided into neutral and normal types; It is divided into spleen and stomach qi deficiency constitution; Divided into Yang deficiency and dampness obstruction constitution; Divided into damp-heat internal constitution; It is divided into two types: one with phlegm and dampness accumulation.

7. The TCM constitution-based weight loss intervention method according to claim 6, characterized in that, Step S6 involves adaptive stepwise matching of specific weight-loss electrical stimulation intervention parameters, including: Based on the TCM constitution classification of obesity, basic electrical stimulation intervention parameters are matched, and then the intensity is adjusted in a stepwise manner according to the comprehensive obesity constitution bias score: <30 points uses baseline electrical stimulation intervention parameters; Intensity increases by 20% for scores with a strength of S > 60; intensity increases by 40% for scores with a strength of S > 60.

8. The TCM constitution-based weight loss intervention method according to claim 1, characterized in that, It also includes the following steps: S8. After a single intervention is completed, all data is archived and stored, and a personalized treatment plan is pushed out. S9. Periodic review and summary of physical fitness score changes; adjust the baseline electrical stimulation intervention parameters based on changes in physical fitness scores within the period.

9. A TCM constitution-based weight loss intervention system based on cross-modal knowledge distillation, used in the intervention method described in any one of claims 1-8, characterized in that, It includes a mobile app software and a smart weight loss intervention bracelet hardware, which are connected wirelessly via Bluetooth; The mobile APP software includes: A tongue image standardization acquisition module is used to acquire tongue images during the meridian flow period and to preprocess the tongue images. A standardized hand image acquisition module is used to acquire hand images during the meridional period and to preprocess the hand images. The ResNet+ViT image dual-branch coding module is used to extract AI depth features from the preprocessed tongue and hand images, respectively. The trimodal AI weighted fusion diagnostic module is used to align student and teacher features using normalized pulse characteristics as teacher features and normalized tongue and hand characteristics as student features through training with a joint loss function, resulting in aligned tongue and hand features. It also feeds the normalized pulse characteristics, aligned tongue and hand features into a lightweight MLP perceptron, outputting single-modal scores for pulse, tongue, and hand diagnosis, respectively. Furthermore, it calculates modality difference coefficients based on the scores and dynamically adjusts the modality weight parameters according to the modality difference coefficients, the correlation between each modality's physical characteristics and TCM constitution differentiation. Finally, it normalizes the adjusted modality weight parameters to calculate a comprehensive obesity constitution bias score and determines the TCM constitution type based on this comprehensive obesity constitution bias score. The intervention parameter matching and distribution module is used to adaptively and stepwise match exclusive weight loss electrical stimulation intervention parameters based on the obesity TCM constitution classification and comprehensive obesity constitution bias score, and distribute the exclusive weight loss electrical stimulation intervention parameters to the smart weight loss intervention bracelet hardware. The hardware components of the smart weight loss intervention bracelet include: The PPG+FSR multimodal pulse layered acquisition module is used to acquire pulse time sequence signals during the meridian flow period; The motion posture noise reduction preprocessing module is used to preprocess the pulse timing signal; The LSTM+Transformer pulse timing coding module is used to extract AI deep features from the preprocessed pulse timing signal. The microcurrent electrical stimulation weight loss intervention module is used to receive the exclusive weight loss electrical stimulation intervention parameters and activate the microcurrent electrical stimulation intervention at the targeted acupoints; it is also used to collect pulse dynamic data in real time during the intervention process and adjust the exclusive weight loss electrical stimulation intervention parameters in real time based on the pulse dynamic data.

10. The TCM constitution-based weight loss intervention system according to claim 9, characterized in that, The mobile app software also includes: The data archiving and personalized treatment push module is used to archive and store all data after a single intervention and push personalized treatment plans; it is also used to periodically review and summarize the changes in physical fitness scores, and adjust the basic electrical stimulation intervention parameters according to the changes in physical fitness scores within the period.

Images