Gynecological traditional Chinese medicine data processing method and system based on four diagnosis data weight analysis

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a gynecological traditional Chinese medicine data processing method and system based on four diagnosis data weight analysis. The method comprises the following steps: acquiring four diagnosis data; map mapping is carried out according to the four-diagnosis data to obtain four-diagnosis map mapping data; obtaining expert experience data and historical case data according to the four-diagnosis map mapping data; performing decision tree construction according to the expert experience data to obtain an expert experience judgment tree model; performing hierarchical weighted regression according to the historical case data to obtain historical case hierarchical weighted data; performing four-diagnosis feature weighting processing according to the expert experience judgment tree model and the historical case level weighting data to obtain four-diagnosis data weight analysis data. According to the method, accurate butt joint of four diagnosis features and a traditional Chinese medicine knowledge system is realized, and a complete link from four diagnosis original feature collection, semantic mapping and rule data collaborative modeling to context adaptive reasoning is constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and in particular, to a method and system for processing traditional Chinese medicine data for gynecology based on the weight analysis of four diagnostic data. Background Art

[0002] The clinical diagnosis and treatment process of traditional Chinese medicine relies on the comprehensive diagnosis of "inspection, auscultation and olfaction, interrogation, and palpation" to obtain physical signs information such as the tongue image, voice, main complaint symptoms, and pulse condition of the patient, and then assist the doctor in syndrome differentiation analysis. However, traditional four diagnostic information mainly relies on artificial perception and subjective experience, and has problems such as being difficult to quantify, inconsistent standards, and poor repeatability, which seriously restricts the structured management, objective analysis, and intelligent application of traditional Chinese medicine information.

[0003] In recent years, with the development of artificial intelligence technologies such as image processing, speech recognition, natural language processing, and physiological signal analysis, some studies have begun to attempt to perform multi-modal acquisition and computational modeling on the key data in the four diagnostic processes, in order to improve the objectivity and computational performance of traditional Chinese medicine data processing. However, most of the existing methods focus on the modeling and classification of a single modality (such as tongue image recognition or pulse condition analysis), and how to implement a multi-modal clinical diagnosis and treatment assistance system for traditional Chinese medicine has become a problem. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention proposes a method and system for processing traditional Chinese medicine data for gynecology based on the weight analysis of four diagnostic data to solve at least one of the above technical problems.

[0005] The present application provides a method for processing traditional Chinese medicine data for gynecology based on the weight analysis of four diagnostic data, including the following steps: Step S1: Obtain four diagnostic data; Step S2: Perform atlas mapping according to the four diagnostic data to obtain four diagnostic atlas mapping data; Step S3: Obtain expert experience data and historical case data according to the four diagnostic atlas mapping data; construct a decision tree according to the expert experience data to obtain an expert experience judgment tree model; perform hierarchical weighted regression according to the historical case data to obtain historical case hierarchical weighted data; Step S4: Perform four diagnostic feature weighted processing according to the expert experience judgment tree model and the historical case hierarchical weighted data to obtain four diagnostic data weight analysis data.

[0006] In the present invention, by performing a structured mapping of the four diagnostic data with the traditional Chinese medicine knowledge graph, combining the expert experience judgment tree model and the hierarchical weighted results of historical cases, a four diagnostic feature weighting mechanism integrating knowledge and data-driven is constructed, which can effectively improve the accuracy and personalization of the identification of traditional Chinese medicine gynecological syndromes, avoid the deviation caused by a single data source, improve the adaptability and interpretability of the model to complex clinical situations, and at the same time have good scalability and clinical practical value.

[0007] Preferably, step S1 is specifically as follows: Step S11: Obtain the inspection image data through an image acquisition device, and perform inspection image analysis based on the inspection image data to obtain inspection image feature data; Step S12: Collect the auscultation audio data through an audio acquisition device, and perform auscultation audio analysis based on the auscultation audio data to obtain auscultation audio feature data; Step S13: Obtain the interrogation text data through a text input interface, and perform interrogation text analysis based on the interrogation text data to obtain interrogation text feature data; Step S14: Collect the palpation pulse data through a pulse sensor, and perform palpation pulse analysis based on the palpation pulse data to obtain palpation pulse feature data; Step S15: Perform four diagnostic feature description processing based on the inspection image feature data, auscultation audio feature data, interrogation text feature data, and palpation pulse feature data to obtain four diagnostic data.

[0008] In the present invention, relevant data for inspection, auscultation, interrogation, and palpation are respectively obtained through a multi-modal acquisition device, and structured features are extracted by using technologies such as image analysis, audio processing, natural language understanding, and pulse signal analysis, and fused into unified four diagnostic data. This not only realizes the objective quantification and standardized processing of the four diagnostic information of traditional Chinese medicine, but also improves the computability and interoperability of multi-source data, provides a high-quality and fusible input basis for map mapping and intelligent diagnosis, and has strong technical adaptability and application extensibility.

[0009] Preferably, step S11 is specifically as follows: The image acquisition device is used to perform combined illumination operation with white light and ultraviolet light for shooting operation to obtain inspection image data, where the inspection image data includes the user's tongue image data and the user's facial color image data; The preset tongue template data is used to perform three-zone processing on the user's tongue image data to obtain tongue three-zone data, where the tongue three-zone data includes tip-of-tongue zone data, middle-of-tongue zone data, and root-of-tongue zone data; Extract color features from the three-zone data of the tongue image to obtain the color feature data of the tongue image zones, where the color feature data of the tongue image zones is the pixel proportion of red, white, and purple in each zone of the three-zone data of the tongue image; Calculate the ratio of the tongue area to the tongue thickness for the user's tongue image data to obtain the ratio data of the tongue area to the tongue thickness; Perform surface crack calculation and texture feature encoding on the three-zone data of the tongue image to obtain the tongue image surface crack data and the tongue coating area texture data respectively; Perform five-zone processing on the user's facial color image data using the preset facial template data to obtain the five-zone facial data, where the five-zone facial data includes the forehead zone data, cheek zone data, nasal bridge zone data, perioral zone data, and chin zone data; Perform redness calculation, paleness calculation, and freckle texture feature extraction based on the five-zone facial data to obtain the redness data of the facial zones, the paleness data of the facial zones, and the freckle texture feature data of the facial zones respectively; Vectorize according to the color feature data of the tongue image zones, the ratio data of the tongue area to the tongue thickness, the tongue image surface crack data, the tongue coating area texture data, the redness data of the facial zones, the paleness data of the facial zones, and the freckle texture feature data of the facial zones to obtain the visual inspection image feature data.

[0010] In the present invention, combined illumination of white light and ultraviolet light is adopted, which enhances the imaging contrast of the surface texture of the tongue coating, pigment distribution, and skin color levels during the shooting process, effectively improving the detail clarity of the image under complex lighting conditions. By introducing preset tongue and facial area templates, automatic segmentation of the three zones (tip of the tongue, middle of the tongue, root of the tongue) of the tongue image and the five zones (forehead, cheeks, nasal bridge, perioral area, chin) of the face is realized respectively, significantly improving the positioning accuracy and consistency of area recognition. Color proportion statistics, morphological structure modeling (such as the ratio of area to thickness), crack detection, and texture encoding are respectively carried out within each area to generate a multi-dimensional vector composed of seven types of detailed image features. The present invention has the characteristics of multi-areas, quantifiable, and repeatable acquisition, effectively improving the accuracy, standardization, and algorithm compatibility of image features in traditional Chinese medicine visual perception modeling.

[0011] Preferably, the auscultation audio feature data includes spectral feature data, frequency jitter feature data, and respiratory rhythm feature data, and step S12 is specifically: Collect auscultation audio data through an audio acquisition device; Extract frame-level acoustic features from the auscultation audio data to obtain frame-level acoustic feature data; Perform clustering calculation based on the frame-level acoustic feature data to obtain frame-level acoustic feature clustering data; Segment the auscultation audio data according to the frame-level acoustic feature clustering data to obtain the segmented auscultation audio data; Extract the spectral features, frequency jitter features, and respiratory rhythm features from the segmented auscultation audio data to obtain spectral feature data, frequency jitter feature data, and respiratory rhythm feature data, respectively.

[0012] In the present invention, through audio acquisition and frame-level acoustic feature extraction, combined with a clustering algorithm, dynamic segmentation of continuous audio signals is realized, effectively separating physiological sound segments with analytical value such as exhalation, inhalation, and other characteristics different from breathing sounds, solving the problems of mixed sound types and blurred boundaries in traditional Chinese medicine auscultation. On this basis, multi-dimensional acoustic indexes such as spectral features, frequency jitter features, and respiratory rhythm are extracted, enabling the originally subjective auscultation process to have quantification ability. The present invention enhances the usability and expression accuracy of audio data in Chinese medicine feature analysis, providing high-quality and low-noise sound feature input for constructing a structured and multi-modal Chinese medicine perception system.

[0013] Preferably, step S14 is specifically as follows: Control the pressure sensor array and the feedback controller to adjust the pressing granularity on the three positions of cun, guan, and chi and synchronously collect the pulse wave signals to obtain the pulse-taking data of pulse diagnosis; Extract the pulse rate features from the pulse-taking data of pulse diagnosis to obtain the pulse rate feature data; Perform rhythmic analysis on the pulse rate feature data to obtain the pulse rate rhythmic feature data; Extract the waveform morphology features from the pulse-taking data of pulse diagnosis to obtain the waveform morphology feature data; Perform waveform symmetry analysis and smoothness analysis on the waveform morphology feature data to obtain the waveform symmetry feature data and the waveform smoothness feature data, respectively; Vectorize the pulse rate feature data, the pulse rate rhythmic feature data, the waveform morphology feature data, the waveform symmetry feature data, and the waveform smoothness feature data to obtain the pulse-taking data features of pulse diagnosis.

[0014] In the present invention, by controlling the pressure sensor array to collect at multiple pressing intensities on the three positions of cun, guan, and chi respectively, the pulse wave signals at different pulse positions and different pressing depths are effectively separated, improving the perception and discrimination ability of typical Chinese medicine pulse conditions such as floating pulse and sinking pulse; combined with rhythmic analysis and waveform symmetry analysis, not only can the conventional pulse rate changes be identified, but also the Chinese medicine description features such as "slippery", "uneven", and "taut" of the pulse condition can be quantitatively judged; the vectorization processing of multiple key pulse condition features significantly enhances the usability and interpretability of the pulse condition data in the machine learning model, solving the problems of difficult quantification and weak interpretability in traditional Chinese medicine pulse diagnosis.

[0015] Preferably, step S15 is specifically as follows: Perform modal unified spatial embedding based on the visual inspection image feature data, auscultation and olfaction audio feature data, interrogation text feature data, and palpation pulse condition feature data to obtain four-diagnosis modal unified spatial data; Perform sub-modal precision allocation on the four-diagnosis modal unified spatial data to obtain four-diagnosis sub-modal precision data; Perform modal interaction attention calculation on the four-diagnosis sub-modal precision data to obtain four-diagnosis feature interaction data; Perform trainable gating setting on the four-diagnosis feature interaction data to obtain four-diagnosis modal gating data; Perform sparsity constraint on the four-diagnosis modal gating data to obtain sparsity constraint data, and perform redundant channel skipping processing on the sparsity constraint data to obtain four-diagnosis channel pruning data; Perform lightweight semantic label reasoning based on the four-diagnosis channel pruning data to obtain four-diagnosis data.

[0016] In the present invention, by embedding four types of features of visual inspection, auscultation and olfaction, interrogation, and palpation into a unified modal space, the fusion problem of multi-source traditional Chinese medicine features with inconsistent dimensions and large distribution differences is solved; a sub-modal precision allocation strategy is introduced to achieve 4-bit low-precision processing of visual features and 16-bit high-precision processing of text semantics, effectively reducing the model calculation complexity while maintaining the diagnostic label matching / recognition accuracy. Through the modal interaction attention mechanism, cross-modal focusing on key diagnostic information labels is realized, and combined with trainable gating and sparsity regularization constraints, redundant feature channels are automatically identified during the training stage and non-critical branches are skipped during the inference stage, significantly reducing the number of parameters and inference latency. The generated four-diagnosis data not only retains high-value diagnostic information but also has advantages such as lightweight deployability, structural interpretability, and semantic traceability, providing an efficient, concise, and interpretable inference basis for traditional Chinese medicine intelligent diagnosis systems.

[0017] Preferably, step S2 is specifically as follows: Step S21: Perform entity standardization on the four-diagnosis data to obtain four-diagnosis entity standardized data, and use a preset traditional Chinese medicine knowledge graph to perform node attribute matching on the four-diagnosis entity standardized data to obtain node matching data; Step S22: Screen relationship paths for the node attribute matching data to obtain relationship path data; Step S23: Perform modal perception mapping adjustment on the traditional Chinese medicine knowledge graph according to the relationship path data and the node attribute matching data to obtain four-diagnosis graph mapping data.

[0018] In the present invention, through entity standardization processing of the four diagnostic data, multimodal features from images, audio, text, pulse conditions, etc. are uniformly transformed into structured tags, and attribute matching is performed with standard nodes in a preset traditional Chinese medicine knowledge graph, improving the docking accuracy of unstructured four diagnostic data in the graph. Through the relationship path screening technology, concept chains highly relevant to the current features are screened out from a large number of possible paths, avoiding interference from invalid paths. Further, by combining modal source information to perform modal perception mapping adjustment, the edge weights and node activation responses in the graph can be optimized according to the relative weights of the four diagnostic features of inspection, auscultation and olfaction, interrogation, and palpation, generating a graph mapping result with modal weight injection and path credibility adjustment. This method significantly enhances the pertinence and structure perception ability of graph reasoning, providing a graph input structure with clearer semantics and better modal coupling for subsequent reasoning models.

[0019] Preferably, step S3 is specifically as follows: Step S31: Retrieve a preset knowledge base and a historical experience base according to the four diagnostic graph mapping data to obtain expert experience data and historical case data; Step S32: Perform structural analysis according to the expert experience data to obtain decision tree structure data; Step S33: Optimize the sorting of tree nodes for the decision tree structure data according to the expert experience data to obtain an expert experience judgment tree model; Step S34: Perform four diagnostic label mapping according to the historical case data to obtain historical case four diagnostic mapping data; Step S35: Perform hierarchical weighted regression according to the historical case four diagnostic mapping data to obtain historical case hierarchical weighted data.

[0020] In the present invention, through semantic-driven retrieval of the four diagnostic graph mapping data, expert experience data and historical case data are respectively retrieved from the knowledge base to ensure that the obtained content has a high degree of relevance to the current feature context. On this basis, structural analysis is performed on the expert experience data, a decision tree structure is constructed, and the path accuracy and judgment consistency of the tree model are improved through a node sorting optimization algorithm, which helps to stabilize the reasoning output of the rule path. For historical case data, first perform four diagnostic label-level mapping to unify the representation form, and then use a hierarchical weighted regression model to distinguish and model the feature weights according to the label levels to achieve hierarchical expression of the feature influencing factors. At the knowledge fusion level, the structural alignment and fusion modeling of expert rules and historical data are realized, enhancing the pertinence, adjustability and structural interpretability of the feature weighting mechanism.

[0021] Preferably, step S4 is specifically as follows: Step S41: Perform model fusion according to the expert experience judgment tree model and the historical case hierarchical weighted data to obtain a preliminary fusion model; Step S42: Extract context features from the mapped data of the four diagnostic spectra to obtain four diagnostic context feature data; Step S43: Adjust the weights of the preliminary fusion model according to the four diagnostic context feature data to obtain four diagnostic data weight analysis data.

[0022] In the present invention, by fusing the expert experience judgment tree model with the hierarchical weighted data of historical cases, a preliminary fusion model with both rule interpretability and data-driven nature is constructed, giving full play to the structural advantages of expert knowledge and the statistical accuracy of case data. On this basis, context features, including information such as modal source, node association density, and feature sparsity, are extracted in combination with the mapped data of the four diagnostic spectra for dynamically adjusting the weight ratios of various sources in the model. The generated four diagnostic data weight analysis data has context awareness and adaptive weighting characteristics, which can effectively improve the robustness and expression consistency of feature evaluation in different situations, enhance the model's response ability to individual feature differences, and provide a more targeted and controllable weight basis for the subsequent reasoning process.

[0023] Preferably, the present application also provides a traditional Chinese medicine data processing system for gynecology based on four diagnostic data weight analysis, which is used to execute the traditional Chinese medicine data processing method for gynecology based on four diagnostic data weight analysis as described above. The traditional Chinese medicine data processing system for gynecology based on four diagnostic data weight analysis includes: A four diagnostic data acquisition module, which is used to obtain four diagnostic data; A four diagnostic knowledge graph mapping module, which is used to perform graph mapping according to the four diagnostic data to obtain four diagnostic graph mapping data; A four diagnostic knowledge fusion and modeling module, which is used to obtain expert experience data and historical case data according to the four diagnostic graph mapping data; construct a decision tree according to the expert experience data to obtain an expert experience judgment tree model; perform hierarchical weighted regression according to the historical case data to obtain historical case hierarchical weighted data; A four diagnostic feature weighted analysis module, which is used to perform four diagnostic feature weighting processing according to the expert experience judgment tree model and the historical case hierarchical weighted data to obtain four diagnostic data weight analysis data.

[0024] The beneficial effects of the present invention are as follows: The present invention not only realizes the structured acquisition and feature extraction of the four diagnostic information, but further accurately docks the unstructured perception data into the semantic space of traditional Chinese medicine knowledge through the atlas mapping mechanism, breaking through the problems that the four diagnostic data are difficult to be semantically associated and knowledge-based in the past. In terms of fusion modeling, the present invention creatively introduces a dual-track mechanism of "expert experience judgment tree + historical case hierarchical regression", which not only retains the interpretable structural rules of traditional Chinese medicine knowledge, but also excavates the influence intensity of different four diagnostic features in the multi-layer label system through hierarchical weighted modeling, realizing the multi-layer expression and context differentiation of feature influence factors, and solving the problems of "static weight, single rule, and insufficient feature generalization" existing in traditional methods.

[0025] The present invention introduces a context awareness mechanism in the weighted analysis process, which can dynamically adjust the fusion model parameters according to information such as the distribution state of current individual features and the depth of the atlas path, so that the output weight analysis result has higher personalized adaptability and scenario transferability. The overall solution constructs a complete link from the acquisition of four diagnostic original features, semantic mapping, rule data collaborative modeling to context-adaptive reasoning, forming a new paradigm of traditional Chinese medicine intelligent analysis that is different from traditional single recognition methods and has data-driven + rule-driven + semantic alignment + reasoning enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Other features, objects, and advantages of the present application will become more apparent by reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 The flowchart showing the steps of a method for processing gynecological traditional Chinese medicine data based on the weighted analysis of four diagnostic data in an embodiment; Figure 2 The flowchart showing the steps of a method for collecting four diagnostic data in an embodiment; Figure 3 The flowchart showing the steps of a method for mapping four diagnostic knowledge graphs in an embodiment; Figure 4 The flowchart showing the steps of a method for fusing four diagnostic knowledge models in an embodiment; Figure 5 The flowchart showing the steps of a method for weighted analysis of four diagnostic features in an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The technical methods of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0029] It should be understood that although terms such as "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly, the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0030] Please refer to Figures 1 to 5 , this application provides a method for processing gynecological traditional Chinese medicine data based on the weight analysis of four diagnostic data, including the following steps: Step S1: Obtain four diagnostic data; Specifically, the system includes a tongue diagnosis image collector, a pulse diagnosis sensor module, a speech recognition and semantic extraction module, and a comprehensive input and interaction platform for four diagnostic information, which is used to uniformly manage the four diagnostic data collection tasks. An image acquisition device is used to take pictures of the user's tongue image and facial color, and combined with image processing algorithms, features such as tongue coating color, tongue body color, tongue body moisture, tongue crack distribution, facial redness, and facial freckle texture are extracted. After the above image features are processed by template partitioning and image segmentation, they are converted into structured color and texture indicators. The user's voice and breathing sounds are collected by an audio acquisition device, and their spectral features, frequency stability (such as Jitter), and rhythmic information are extracted. Odor data usually provides descriptive information by manual input into the platform, and the system extracts relevant olfactory features by combining keyword matching and knowledge graph node recognition. The user states the main symptoms by voice or text input, and the system uses speech recognition (ASR), Chinese word segmentation, named entity recognition and other semantic analysis technologies to convert natural language into standardized disease labels, and further constructs a multi-dimensional symptom vector for atlas mapping and model analysis. The system collects the pulse conditions at the cun, guan, and chi positions through a pressure sensor array, and extracts indicators such as pulse rate, pulse waveform amplitude, rise time, fall time, waveform symmetry, and smoothness.

[0031] Step S2: Perform atlas mapping based on the four diagnostic data to obtain four diagnostic atlas mapping data; Specifically, in the four diagnostic methods mapping step, the system is based on the preset traditional Chinese medicine knowledge graph, maps the standardized four diagnostic features obtained from the collection to the nodes in the graph, and then forms structured four diagnostic methods mapping data. The traditional Chinese medicine knowledge graph has a clear ontology structure, and the graph includes the following types of nodes and semantic relationships: The node types include symptom nodes, syndrome types nodes, etiology nodes, and treatment method nodes; the relationship types include semantic edges such as "manifested as", "caused by", "suggested medication", "attributed to syndrome type", etc., all of which are directed edge relationships. The system processes the four diagnostic structured data, converts each symptom expression (such as "white and thick tongue coating", "abdominal distension and pain") into a vector form recognizable by the model, and uses a multi-label text matching model (such as a combination of BERT and BiGRU) to calculate the similarity between the symptom semantics and the graph node names. The specific calculation formula is as follows: , is the similarity score, ranging from [0,1], is the th symptom feature as input, is the th node in the knowledge graph (such as "liver qi stagnation"), is the vector representation of symptom , is the vector representation of graph node , is the Euclidean norm of the vector representation of symptom , is the Euclidean norm of the vector representation of graph node . To ensure the matching quality, the system sets a similarity threshold (such as 0.75), and only retains the symptom-node pairs whose similarity with the node label is not lower than this threshold, filtering out irrelevant nodes with low semantic relevance. For example, when the input symptom is "abdominal distension and pain", the system will match the "liver qi stagnation" node in the knowledge graph; when the input features are a combination of "pale and white tongue" and "slow pulse", the system can infer the path related to "yang deficiency" and complete the corresponding mapping in the graph structure.

[0032] Step S3: Obtain expert experience data and historical case data according to the four diagnostic methods mapping data; construct a decision tree based on the expert experience data to obtain an expert experience judgment tree model; perform hierarchical weighted regression on the historical case data to obtain historical case hierarchical weighted data; Specifically, the system utilizes the rule knowledge base organized by existing traditional Chinese medicine experts to construct a symbolic diagnostic decision tree model. The rule base is organized in the form of "if condition... then conclusion...", for example, if the tongue coating is thick and greasy and the pulse condition is slippery → it is determined as "phlegm-dampness obstructing the interior"; if the tongue is dull and the pulse is stagnant → it is determined as "blood stasis blocking the circulation". The above rules can be structurally represented as a tree structure composed of multi-level nested conditional judgment nodes and conclusion nodes, and its basic coding form is: if tongue_coating == "thick and greasy" and pulse["slippery"] > 0.6: return "phlegm-dampness obstructing the interior". To improve the consensus and reliability of the diagnostic model, the system allows multiple experts to review the rule base and sets a weight score for each rule. The system uses the weighted average of expert scores to sort and prune the nodes of the tree structure to form an expert experience judgment tree model; The system collects structured historical case data, which includes the four diagnostic characteristics of users (including standardized labels of inspection, auscultation and olfaction, inquiry, and palpation) and the corresponding traditional Chinese medicine diagnosis result labels. The system divides the characteristics into multiple levels according to dimensions (such as the tongue diagnosis layer, pulse diagnosis layer, inquiry layer, etc.), and constructs a multi-layer weighted ridge regression model to predict the probability score of each diagnostic label. The specific model calculation is as follows: , is the label vector corresponding to the historical case, is the matrix transpose of, is the weight matrix of the sample (similarity weight for different cases), is the feature matrix of the historical case (such as the four diagnostic vectors of multiple patients), is the ridge regression penalty factor (controlling overfitting), is the identity matrix (with the same column dimension as ). The above two modules respectively output the expert rule path judgment result and the historical case statistical weighted result.

[0033] Step S4: Perform weighted processing on the four diagnostic characteristics according to the expert experience judgment tree model and the historical case hierarchical weighted data to obtain the weighted analysis data of the four diagnostic data.

[0034] Specifically, for each traditional Chinese medicine syndrome type label , the system calculates its score according to the following linear weighted formula: , is the weighted analysis data of the four diagnostic data of the traditional Chinese medicine syndrome type label , is the empirical credibility factor, which is used to adjust the fusion ratio of the expert model and the historical model. The system can be automatically set through cross-validation, etc., and the set value is 0.65, is the label score output by the expert decision tree model (rule path matching score), It is the label score based on the historical case weighted regression model; After the fusion score is generated, the system further performs reverse weight enhancement on the original four diagnostic features. That is, for the label with a higher score (such as the score of "liver qi stagnation" is 0.81), the system will search for the closely related four diagnostic features (such as "dark red tongue", "abdominal distension and pain", "string-like pulse", etc.); assign a higher feature importance factor to the dimension of the above features in the four diagnostic feature vector; this operation forms a set of weighted four diagnostic feature data, which can be used for diagnostic reasoning based on data processing, result interpretability output or user display. This process is regarded as a "semantic label-driven feature attention adjustment", which helps to construct an interpretable connection path between features and labels.

[0035] Preferably, step S1 is specifically as follows: Step S11: Obtain the inspection image data through an image acquisition device, and perform inspection image analysis based on the inspection image data to obtain inspection image feature data; Specifically, use a high-resolution image acquisition device (resolution ≥ 8MP) to perform standardized shooting operations on users, and collect the following two types of images. For the tongue image, if the user opens their mouth naturally, collect it frontally under a white light background with a color temperature of 5500K to ensure that the image is clear and there are no interfering shadows; for the facial color image, if the user is frontal and unobstructed, remove the mask, hair, glasses and other obstructions, and use a standard light source environment for image acquisition. Apply a semantic segmentation model (such as U-Net or DeepLabV3+) to accurately segment the tongue area or facial area in the image, and extract the region of interest (ROI) for analysis. The tongue features include the color of the tongue body (such as light red, dark red, purple), which is identified based on the average value of the pixels in the tongue body area in the Lab space; the tongue coating coverage, which represents the coverage ratio of the tongue coating area on the tongue body; the moisture level of the tongue body, which is estimated based on the distribution of the highlight reflection area; the morphological features of the tongue body, such as the degree of fatness and thinness, the depth of tooth marks, etc., which are calculated based on edge contour extraction and geometric ratios. The facial color uniformity is judged by the standard deviation of the Lab color distribution in the facial skin area; the yellowing / greening index is extracted based on the average values of the a channel (red-green axis) and b channel (yellow-blue axis) in the Lab space; the system extracts the average color temperature value of the "U area" (including both cheeks and the perioral area) of the face. If the color temperature is on the high side, it indicates the sign of upper heat. If the average Lab value of the tongue body satisfies 、 、 ,the system will label the tongue image as a red tongue, , , , is the average brightness value of the tongue body area, is the number of pixels participating in the statistics (i.e., the number of pixels in the tongue body area), is the pixel sequence term, is the The brightness value of a pixel (L channel in the Lab color space), is the average color value of the red-green difference in the tongue area, is the red-green difference of the pixel, and is the average color value of the yellow-blue difference in the tongue area, is the yellow-blue difference of the

[0036] Step S12: Collect auscultation audio data through an audio acquisition device, and perform auscultation audio analysis based on the auscultation audio data to obtain auscultation audio feature data; Specifically, through an electret microphone or the built-in recording module of a smart terminal, guide the user to perform voice reading and natural breathing processes in a quiet environment, and collect their vocalization and breathing audio signals. It is recommended that the sampling rate be set to 16 kHz, the quantization bit depth be 16-bit, and the audio format be mono PCM encoding to ensure that the signal quality meets the requirements of subsequent analysis. Use a Wiener filter to denoise the audio signal, reduce the interference of background stationary noise on speech features, and improve the signal-to-noise ratio. Extract the following acoustic features from the denoised audio: MFCC, which is used to model the voice timbre feature; short-time energy, which reflects the change in voice intensity; pitch, which is used to analyze the change in pitch and tone features; and breath break point detection, which can identify the breathing interruption points based on the change in short-time zero-crossing rate or voice activity detection (VAD) method.

[0037] Combined with traditional Chinese medicine clinical experience, design the following acoustic feature recognition logic: If the voice signal has the characteristics of frequent breath interruption and unstable tone fluctuation, the system can initially determine it as the manifestation of "qi deficiency"; if the voice is overall low and weak and the frequency is low, it can be classified as the relevant feature of "yang deficiency"; the above rules can be used as the label basis for model training or for enhancing the interpretability of model output.

[0038] Specifically, extract the short-time energy to reflect whether the voice is strong or not, , is the short-time energy of the frame, representing the loudness or intensity of this frame, is the in-frame sample point index, is the window length, is the original audio signal value corresponding to the th sample point in the frame,

[0039] Step S13: Obtain the inquiry text data through the text input interface, and perform inquiry text analysis based on the inquiry text data to obtain the inquiry text feature data; Specifically, the doctor or patient manually inputs the chief complaint and medical history through the system interface, for example: "Menstruation has been delayed in the past two months, accompanied by abdominal distension and pain, and aversion to cold with cold limbs". The text processing process is Chinese word segmentation + named entity recognition (NER). A custom dictionary is constructed to include gynecological specific symptoms (such as "metrorrhagia and metrostaxis", "vaginal bleeding") and syndrome entities (such as "kidney yang deficiency"). Use BERT (or RoBERTa) for multi-label classification → map the text symptoms to standardized syndrome labels, and enhance the keyword frequency statistics or logical rules: the appearance of "aversion to cold" and "cold limbs" → increases the possibility of yang deficiency, and the appearance of "distending pain" and "chest distress" → activates the qi stagnation related path.

[0040] Step S14: Collect the pulse diagnosis pulse data through the pulse sensor, and perform pulse diagnosis pulse analysis based on the pulse diagnosis pulse data to obtain the pulse diagnosis pulse feature data; Specifically, use FFT to calculate the frequency spectrum of the pulse signal for analyzing the pulse periodicity. , is the frequency domain coefficient (spectrum), is the time sampling point number, is the total number of signal sampling points, is the pulse time series signal, is the base of the natural logarithm, is the imaginary unit ( ), is the constant term of pi, is the frequency serial number. Numerical judgment is performed according to the spectrum through a preset rule to obtain the pulse diagnosis pulse feature data.

[0041] Specifically, the system uses a micro pressure sensor array (such as a MEMS type array sensor), which is arranged at the cun, guan, and chi positions on the wrist for pulse diagnosis, and is fixed by a silicone bracket or a shaped wristband. The sensor continuously collects the pulse waveform data generated by the contact pressure changing with time, and digitally transmits it to the backend system through the ADC module. The system sampling frequency is set to not less than 100 Hz to ensure sufficient sampling of the pulse waveform details, and the single acquisition duration is 5 - 10 seconds. The three-channel structure respectively collects the pulse signals at the cun, guan, and chi positions, and the system automatically performs channel position binding and number marking. The dimensions of pulse feature analysis include waveform frequency, such as judging the pulse rate by period detection or FFT calculation, with the unit of bpm. Waveform amplitude, which represents the energy and strength of the pulse wave, and commonly uses the maximum amplitude or RMS value. Rise time / fall time, which calculates the interval between the inflection points of the rising edge and the falling edge using the first derivative of the waveform, and is used to identify pulse types such as "slippery" and "unsmooth". The system uses Fourier transform (FFT) to calculate the position of the spectral peak to determine the main frequency, and combines the autocorrelation function to determine the waveform periodicity. The system inputs the continuous pulse waveforms into a deep learning model that combines convolutional neural network (CNN) and attention mechanism (Self-Attention) to achieve pulse classification, including but not limited to common traditional Chinese medicine pulse types such as slippery pulse, stringy pulse, thin pulse, and slow pulse.

[0042] Step S15: Perform four-diagnosis feature description processing based on the visual inspection image feature data, auscultation and olfaction audio feature data, interrogation text feature data, and palpation pulse feature data to obtain four-diagnosis data.

[0043] Specifically, the features of each diagnosis method are normalized and structured to form a unified four-diagnosis feature data structure, and qualitative labels (such as "light red") or quantitative values (such as the tongue coating area ratio is 0.36) are assigned to each sub-feature.

[0044] Preferably, step S11 is specifically: Using a combination of white light and ultraviolet light for illumination operation through an image acquisition device to perform shooting operation to obtain visual inspection image data, where the visual inspection image data includes the user's tongue image data and the user's facial color image data; Specifically, the system is configured with a dual-channel high-color-rendering lighting device, including a white-light LED lighting source with a color rendering index (CRI) ≥ 90, which is used to truly restore the natural colors of structures such as the tongue body and face under visible light conditions, and is used to extract structural features such as color, area, and shape; an ultraviolet-light LED lighting source with a wavelength range of 365 nm to 400 nm, which can stimulate the reflection feature changes of tissue surface pigments, moisture distribution, and fine cracks, and is used to enhance the response perception of humidity, roughness, and hidden textures. Each shooting process includes the following two synchronous channels. The white-light image channel is used for extracting the main color and structural form; the ultraviolet image channel is used for humidity perception and texture enhancement analysis. The two images are taken by a unified imaging module at the same angle and fixed distance to ensure that the image perspectives and sizes are consistent and to avoid alignment errors caused by angle deviations.

[0045] The user's tongue image data is processed into three partitions using the preset tongue template data to obtain tongue image three-partition data, where the tongue image three-partition data includes tip-of-tongue partition data, middle-of-tongue partition data, and root-of-tongue partition data. Specifically, the template partitioning rule is based on the relative position and contour of the tongue body, setting the ratio. Tip-of-tongue area: the first 1 / 4 of the length; middle-of-tongue area: the middle 1 / 2 of the length; root-of-tongue area: the last 1 / 4 of the length. Use a tongue segmentation model (such as U-Net) to extract the tongue body area; cut it into three sub-areas according to the longitudinal ratio on the segmentation contour, and name them tip-of-tongue partition data, middle-of-tongue partition data, and root-of-tongue partition data respectively.

[0046] Color feature extraction is performed on the tongue image three-partition data to obtain tongue image partition color feature data, where the tongue image partition color feature data is the pixel proportion of red, white, and purple in each partition of the tongue image three-partition data. Specifically, pixel-level color statistics are performed on the three-partition images respectively; after counting the number of red, white, and purple pixels in each area and normalizing them, calculate the color proportion: , is the color (such as red, white, purple) in the area (tip, middle, root) pixel proportion, is the color in the area pixel count, is the area The total number of pixels in it. The HSV or CIELab color space is used for color classification, and the pixel color type is judged according to the following thresholds: Red pixels are recognized (in the HSV space) as Hue ∈ [0°, 30°] and Saturation (S) > 0.6, purple pixels are recognized (in the HSV space) as Hue ∈ [270°, 310°], and white pixels are recognized (in the HSV space) as Saturation (S) < 0.2 and Value (V, brightness) > 0.8. The above Hue angle unit is degree (°), and the value ranges of Saturation and Value are [0, 1]. The recognition threshold for the CIELab space can be further set according to experimental experience and optimized according to the image source and acquisition conditions.

[0047] Calculate the ratio of the tongue body area to the thickness for the user's tongue image data to obtain the ratio data of the tongue body area to the thickness; Specifically, the tongue image is segmented by a semantic segmentation network (such as U-Net or DeepLabV3+), and the complete tongue body contour area is extracted and denoted as , that is, the pixel area of the tongue body area in the image. The system further identifies the tongue coating coverage area inside the tongue body area. The recognition of the tongue coating area combines color features (such as lightness and b channel in the Lab color space) and texture features (such as Local Binary Pattern LBP) for fusion analysis, and separates the tongue coating area through a feature enhancement segmentation network to obtain the pixel area of the tongue coating area, denoted as . , is the tongue coating thickness ratio, is the tongue coating area (the number of pixels in the segmented area), is the total tongue body area (the total number of pixels after segmentation). The system can set a threshold to judge the thickness of the tongue coating. For example, if <0.3, it is judged as thin coating; if >0.6, it is judged as thick coating; the intermediate value is medium coating.

[0048] Calculate the surface cracks and texture feature encoding for the three-partition data of the tongue image to obtain the tongue image surface crack data and the tongue coating area texture data respectively; Specifically, after converting the original RGB tongue image to a grayscale image, the Laplacian operator is used for edge detection to enhance the boundary response of the dark crack area in the image. After smoothing the background interference by median filtering, combined with connected component analysis or skeleton extraction algorithm, the continuous crack path area is extracted to form a crack set, and the crack set is summed to obtain . , is the crack density, is the total crack length (pixel-level measurement), is the total area of the tongue body region; Use a semantic segmentation model (such as U-Net) to divide the tongue image into regions and extract the tongue coating coverage area , which is used as the target for texture analysis. Texture encoding includes extracting local neighborhood gray-scale relationships, constructing a gray-level co-occurrence matrix, and extracting contrast, entropy, and correlation.

[0049] Use the preset facial template data to perform five-partition processing on the user's facial color image data to obtain facial five-partition data, where the facial five-partition data includes forehead partition data, cheek partition data, nasal bridge partition data, perioral partition data, and chin partition data; Specifically, use a face detection and key point localization model (such as Mediapipe) to process the input face image, extract the two-dimensional coordinate points of key structures such as eyes, nasal bridge, lips, and jawline, and use them as the reference points for subsequent region division. Based on the above key points, according to the anatomical structure and color diagnosis requirements, divide the face image into the following five functional regions: forehead partition data, cheek partition data, nasal bridge partition data, perioral partition data, and chin partition data.

[0050] Calculate the redness, paleness, and freckle texture features based on the facial five-partition data to obtain the facial partition redness data, facial partition paleness data, and facial partition freckle texture feature data respectively; Specifically, the redness calculation , is the redness of the facial partition (such as forehead, nasal bridge), is the average value of the component (red-green axis) in the Lab color space; The judgment condition for facial pallor (based on the HSV space) (not a formula but with logical expressiveness): and , is the average saturation of the region , is the average brightness of the region . Extract the freckle distribution through color clustering (such as K-means, K = 3); count the total area and total number of all spots, calculate the spot area ratio and density; calculate the texture contrast based on the gray-level co-occurrence matrix (GLCM): , where is the texture contrast, that is, the facial partition freckle texture feature data, is the row index in the GLCM, representing the gray value of the first pixel in the pixel pair, is the column index in the GLCM, representing the gray value of the first pixel in the pixel pair, For grayscale pairs The probability of occurrence.

[0051] Vectorize according to the color feature data of tongue image regions, the ratio data of tongue body area to thickness, the crack data on the surface of tongue images, the texture data of the tongue coating area, the redness data of face regions, the paleness data of face regions, and the texture feature data of skin patches in face regions to obtain the visual inspection image feature data.

[0052] Specifically, pack the data obtained above into a data group.

[0053] Preferably, the auscultation audio feature data includes spectral feature data, frequency jitter feature data, and respiratory rhythm feature data. Step S12 is specifically as follows: Collect auscultation audio data through an audio acquisition device; Specifically, use a high-sensitivity electret microphone or a mobile phone integrated microphone to record in a quiet room (noise < 30 dB).

[0054] Extract frame-level acoustic features from the auscultation audio data to obtain frame-level acoustic feature data; Specifically, segment the audio into short-time frames (frame length 20 ms, frame shift 10 ms), and extract feature vectors for each frame, such as MFCC (Mel Frequency Cepstral Coefficients), which is used to describe the timbre characteristics of speech, and usually takes 13 dimensions; energy, representing the sound intensity; fundamental frequency, reflecting the pitch; zero-crossing rate, used to judge whether the breath is rapid.

[0055] Perform clustering calculation according to the frame-level acoustic feature data to obtain frame-level acoustic feature clustering data; Specifically, use K-means or GMM (Gaussian Mixture Model) to cluster the feature vectors of each frame; the clustering dimensions include MFCC, F0, Energy, and the number of clusters K is set to 3 - 6 (e.g., speaking voice, breathing sound, pause).

[0056] Perform segmentation processing on the auscultation audio data according to the frame-level acoustic feature clustering data to obtain auscultation audio segmented data; Specifically, merge the parts with the same clustering label in consecutive frames into an audio segment; filter out segments with a duration shorter than 200 ms (noise segments).

[0057] Extract spectral feature, frequency jitter feature, and respiratory rhythm feature from the auscultation audio segmented data to obtain spectral feature data, frequency jitter feature data, and respiratory rhythm feature data respectively.

[0058] Specifically, use STFT (Short-Time Fourier Transform) to calculate the spectral energy map for each segment of speech signal, , is the spectral centroid position, is the frequency point sequence term, is the th frequency point, is the complex spectrum value of the audio at this frequency point, is the amplitude (modulus) of this frequency point, , is the spectrum bandwidth, is the th frequency point, is the frequency point sequence term, is the amplitude (modulus) of this frequency point, and the spectral feature data includes the spectral centroid position and the spectrum bandwidth; , is the feature of sound stability (i.e., frequency jitter feature data), and the larger it is, the more unstable it is, is the total number of cycles, is the cycle sequence term, is the duration of the th cycle (estimated from the fundamental frequency), duration of the , is the respiratory rhythm feature data, is the standard deviation of the respiratory cycle, is the mean of the respiratory cycle.

[0059] Preferably, step S14 is specifically: Control the pressure sensor array and the feedback controller to adjust the pressing granularity of the cun, guan, and chi positions and synchronously collect the pulse wave signals to obtain the pulse-taking data; Specifically, use a 3-channel pressure sensor array to locate the three positions of cun (proximal part of the wrist), guan (middle part), and chi (distal part) respectively; dynamically adjust the pressure application intensity of the three channels through a closed-loop feedback controller to form a three-stage acquisition pressure range. Light pressure stage: 30 - 50 mmHg, for initial perception; medium pressure stage: 60 - 80 mmHg, to strengthen the main waveform signal; heavy pressure stage: 90 - 120 mmHg, for deep pulse detection. Each stage lasts for 2 - 3 seconds and automatically transitions in sequence to ensure uniform sampling of data in each stage. Collect 10 seconds of pulse wave signals for each position, and the sampling frequency ≥ 100Hz.

[0060] Extract the pulse rate feature according to the pulse-taking data to obtain the pulse rate feature data; Specifically, perform peak detection on the pulse wave of each position; calculate the time difference between consecutive peaks and infer the pulse rate (unit: bpm): , is the pulse rate feature data, is the average time interval between consecutive pulse wave peaks (unit: second).

[0061] Performing rhythmic analysis based on the pulse rate characteristic data to obtain pulse rate rhythmic characteristic data; Specifically, the RR interval ( ) refers to the time difference between two consecutive pulse wave peak values, , is the pulse rate rhythmic characteristic data, is the standard deviation of the RR interval, is the average value of the RR interval.

[0062] Performing waveform morphology feature extraction based on the palpation pulse condition data to obtain waveform morphology feature data; Specifically, the system extracts the following morphological feature parameters from the original pulse wave signal based on time-domain signal analysis, such as the main peak amplitude, the maximum amplitude within the pulse wave period; the valley amplitude, the minimum amplitude within the period; the main wave width, the time span of the complete main wave; the rise time, the time taken for the waveform to rise from the starting point to the main peak; the fall time, the time taken for the waveform to fall from the main peak to the end point. The system performs a first-order difference processing on the pulse wave signal to obtain the derivative sequence of the waveform, and marks the wave valleys, wave peaks, and rising starting points of each period through an extreme point detection method.

[0063] Performing waveform symmetry analysis and smoothness analysis on the waveform morphology feature data to obtain waveform symmetry feature data and waveform smoothness feature data respectively; Specifically, calculating the comparison of the rise and fall times: , is the waveform symmetry feature data, is the time interval from the wave valley (starting point) to the rise of the main peak, is the time interval from the main peak to the next wave valley (ending point). The closer the symmetry value is to 1, the more symmetrical the pulse condition indicates; a slippery pulse is represented as < 1, and a stringy pulse is represented as > 1. The smoothness can be judged by the derivative change rate and waveform continuity. The smoothness index is defined as the total change of the continuous first-order derivative of the curve: , is the waveform smoothness feature data, is the total number of discrete time points within the current pulse wave period, is the time point number processed in the current loop, is the signal at moment, approximated by the difference method, is the signal at moment, approximated by the difference method, is the moment, is the moment. The smaller the smoothness value, the smoother the waveform, which is used to assist in identifying the "slippery pulse"; on the contrary, the larger the smoothness value, the more drastic the change in the waveform slope, corresponding to the states of "unsmooth pulse" or blood stasis.

[0064] Vectorize the pulse rate feature data, pulse rate rhythm feature data, waveform morphology feature data, waveform symmetry feature data, and waveform smoothness feature data to obtain the palpation pulse condition feature data.

[0065] Specifically, structure the above-obtained data into a vector input.

[0066] Preferably, step S15 is specifically: Perform modality-unified spatial embedding based on the inspection image feature data, auscultation audio feature data, interrogation text feature data, and palpation pulse condition feature data to obtain the four-diagnosis modality-unified spatial data; Specifically, each modality is mapped to a shared semantic space of a unified dimension through an exclusive linear transformation or a feed-forward network: , is the unified modality vector after projection, is the transformation weight matrix, is the original feature vector (for each modality, corresponding to one of the inspection image feature data, auscultation audio feature data, interrogation text feature data, and palpation pulse condition feature data), is the bias vector. The above four modality embedding vectors will be used as a unified structure input and sent to the modality fusion module or the atlas alignment module to achieve multi-modal semantic reasoning, diagnostic assistance, and result interpretation.

[0067] Perform sub-modal accuracy allocation on the four-diagnosis modality-unified spatial data to obtain the four-diagnosis sub-modal accuracy data; Specifically, the system inputs each embedding vector into an exclusive or shared accuracy scoring function , which can be implemented using a multi-layer perceptron (MLP) structure, outputs a scalar score, and converts the above scoring result through the function into an accuracy weighted distribution, , is the accuracy distribution weight of the modal, is the exclusive or shared accuracy scoring function, is the four-diagnosis modality-unified spatial data.

[0068] Perform modal interaction attention calculation on the four-diagnosis sub-modal accuracy data to obtain the four-diagnosis feature interaction data; Specifically, for any two modalities , the The modality is used as a query, and the modality is used as a key and a value for interactive attention modeling: , is the attention weight between different modalities, is a weight normalization function that normalizes the attention scores into a probability distribution for weighted combination vectors, is the query matrix, which is obtained by linearly mapping the current modality through to get, , is to map the current modality vector to the query vector of the trainable weight matrix, is the key vector (Key), the semantic index vector of other modalities, and is used to perform similarity matching with , , is the key transformation matrix (Key weight matrix), which maps the other modality vector to the key vector of the trainable matrix, is the vector dimension scaling factor, taking or the dimension of the vector (such as 128), is the value vector, the actual information vector carried by other modalities, and is used for weighted aggregation, , is the value transformation matrix (Value weight matrix), which maps the other modality vector to the value vector of the trainable matrix, is the cross-modal interaction condition, indicating that the current modality is different from the target modality, and realizing the reading of heterogeneous information such as image → text, text → audio, etc. The system performs attention reading and weighted fusion on multiple other modalities: , is the enhanced vector of the modality after fusing the information of other modalities, is the interaction weight between modalities (which can be a fixed average or determined by the precision module), is the attention weight between different modalities. The representation after fusing all modalities together constitutes the four-diagnosis feature interaction data of the system.

[0069] Perform trainable gating settings on the four-diagnosis feature interaction data to obtain the four-diagnosis modality gating data; Specifically, each modality adds a gating weight to control its influence on the fused representation: , , is the gated modal feature vector, is the modal gating weight (scalar or vector), is the -th interactive feature vector of the modality, is the Sigmoid function, is the gating weight matrix, a trainable parameter that determines the retention weight of each modality.

[0070] Perform sparsity constraint on the four diagnostic modality gating data to obtain sparsity-constrained data, and perform redundant channel skipping processing on the sparsity-constrained data to obtain four diagnostic channel pruning data; Specifically, for the multi-modal gating channels composed of four types of diagnostic information: inspection, auscultation and olfaction, interrogation, and palpation, define the channel weight vector , and by introducing the L1 norm constraint into the model training objective function, encourage some channel weights to converge to zero, realizing automatic sparse selection at the feature level. Its sparse regularization term is defined as follows: , is the sparsity loss, which is used to suppress the activation of redundant channels during model training, is the four diagnostic modality channel sequence term, is the total number of four diagnostic modality channels, is the -th gating activation value of the channel, and its value range is [0,1]. Under the action of the sparsity constraint, if the gating value of a certain channel is lower than the preset pruning threshold , that is, if , it is considered that the contribution of this channel to the determination of the diagnostic label is limited, and it is skipped and does not participate in the subsequent feature fusion process. This operation realizes the dynamic selection and channel pruning of the four diagnostic modality information, thereby simplifying the model structure and improving the generalization performance.

[0071] Perform lightweight semantic label inference based on the four diagnostic channel pruning data to obtain four diagnostic data.

[0072] Specifically, the system performs weighted fusion on the four diagnostic feature channels retained after pruning and screening to obtain the representation vector , and its calculation method is as follows: , is the diagnostic feature representation after weighted fusion, is the -th effective channel (such as inspection, interrogation, tongue image, pulse condition), is the channel attention weight, calculated by the attention module or importance evaluation mechanism, is the embedded feature vector output by this channel, which has been retained through feature pruning. The fused representation vector Input into a small neural network classifier to generate the prediction results of the semantic labels of the four diagnostic methods , and the calculation formula is as follows: , is the diagnostic feature representation after weighted fusion, is the data of the four diagnostic methods, is the multi-class output activation function, which is used to convert the linear output into a probability distribution, is the weight matrix of the output layer, which is the weight parameter in the classifier and is used to map the fused features to the label space, is the bias term of the output layer, which is the bias parameter in the classifier and enhances the model fitting ability. The classifier output represents the prediction probability of each semantic label of the four diagnostic methods (such as "red tongue", "late menstruation", "slippery pulse", "pale complexion", etc.). The system will select the corresponding label as the output result according to the maximum probability term or the confidence threshold to obtain the data of the four diagnostic methods.

[0073] Preferably, step S2 is specifically as follows: Step S21: Perform entity standardization on the data of the four diagnostic methods to obtain the standardized data of the four diagnostic entities, and use the preset traditional Chinese medicine knowledge graph to perform node attribute matching on the standardized data of the four diagnostic entities to obtain the node matching data; Specifically, perform lexical cleaning, spelling unification, simplified and traditional conversion, and abbreviation restoration on the original text of the four diagnostic features (such as "red tongue", "late menstruation", "slippery pulse"); use the standard word library of traditional Chinese medicine terms (such as "Ontology of Traditional Chinese Medicine Symptoms") to map the original entities to standard entity codes; in the preset traditional Chinese medicine knowledge graph, search for semantic nodes such as symptoms, syndromes, constitutions, and etiologies that match the standardized entities; the matching method adopts Exact match (string exact match), or word vector cosine similarity matching (such as using SimCSE + BERT): , is the cosine similarity between the two (range 0~1), is the text vector of the four diagnostic entities, is the label vector of the graph nodes, is the Euclidean norm (L2 norm) of the text vector of the four diagnostic entities, is the Euclidean norm (L2 norm) of the label vector of the graph nodes.

[0074] Step S22: Screen the relationship path data of the node attribute matching data to obtain the relationship path data; Specifically, the knowledge graph is a directed attribute graph, including nodes such as symptoms, syndrome types, etiologies, treatments, acupoints, and drugs. The relationship types include "manifested as", "attributed to", "accompanied by", "suitable for use", "should be avoided", etc. For multiple matching nodes, a semantic path screening strategy is executed: using a graph traversal algorithm (such as DFS / BFS), searching for semantic valid paths with a maximum length ≤ 3; screening paths where the starting point and the ending point have logical traditional Chinese medicine semantics (for example, symptom → syndrome type → treatment method); eliminating paths without semantic value or duplicate loop paths. For each path that meets the conditions , calculate its total semantic value score as the basis for priority: , is the path (such as "purple tongue → blood stasis syndrome → promoting blood circulation and removing stasis method"), is the th edge (relationship type) in the path, is the number of relationship types, is the semantic importance weight of this relationship (empirically set or learned through training). For example, the path "red tongue → manifested as → yin deficiency and excessive fire → suitable for use → anemarrhena" will be preferentially retained due to semantic continuity and high relationship weight.

[0075] Step S23: Perform modal perception mapping adjustment on the traditional Chinese medicine knowledge graph according to the relationship path data and the node attribute matching data to obtain the four diagnostic methods knowledge graph mapping data.

[0076] Specifically, for different modal sources in the four diagnostic methods (such as images, audio, text, pulse conditions), a perception factor is given. In the graph path , each edge has an original weight . If the key nodes or relationships in the path mainly come from a certain modality , then the path score is weighted and corrected as follows: , is the path score after modal weighting, represents each edge of the path , is the perception factor corresponding to the dominant modality of this path, is the weight factor of the path-associated modality (for example, 1.0 for images and 1.2 for pulse conditions). If the path of slippery pulse → spleen deficiency syndrome mainly comes from the palpation modality , then the score of this path is increased. For example, a certain path points from "slippery pulse" to "spleen deficiency syndrome", and this path mainly depends on the original data from the palpation modality. If the system sets = 1.2, then the score of this path will be multiplied by 1.2, thereby enhancing its importance in diagnostic reasoning and reflecting its higher credibility.

[0077] Preferably, step S3 is specifically as follows: Step S31: Retrieve a preset knowledge base and a historical experience base according to the four diagnostic atlas mapping data to obtain expert experience data and historical case data; Specifically, the knowledge base includes structured syndrome types, symptoms, diagnostic rules, prescription matching, etc., which are derived from "Traditional Chinese Medicine Diagnosis" and "Clinical Pathway Specifications", etc. It adopts an atlas structure, and semantic relationships such as "belonging", "triggering", and "adapting" are established between nodes to support multi-hop reasoning and rule invocation; the experience base is a real historical case dataset, including four diagnostic features + diagnostic labels (syndrome types) + treatment plans. Among them, the "four diagnostic features" include subjective and objective information obtained from inspection, auscultation and olfaction, interrogation, and palpation. The diagnostic syndrome type is the entity coding in the standard term system, and the treatment plan covers traditional Chinese medicine prescriptions, acupuncture, recuperation suggestions, etc. After the user inputs the features to complete entity standardization and atlas mapping, the system performs retrieval operations on the knowledge base and the experience base according to the feature label combination (such as "tongue texture: purple", "pulse condition: slippery", "lower abdominal cold pain"). The specific methods are as follows: label matching query, using the standardized entity label to perform explicit query in the knowledge base to extract the corresponding syndrome type, prescription, and rule content. Vector similarity recall, embed the label as a vector (using pre-trained models such as SimCSE, BERT, RoFormer, etc.), and calculate its semantic similarity with the samples in the experience base. The following cosine similarity formula is used: According to the labels obtained from the atlas mapping, such as "tongue texture purple", "pulse slippery", "lower abdominal cold pain"; the samples with the top K similarity rankings form the Top-K matching candidates, which are used as the recall results of similar cases.

[0078] Step S32: Perform structure analysis on the expert experience data to obtain decision tree structure data; Specifically, the knowledge base contains a large number of expert experience rules manually input or accumulated historically, and their expression forms are uniformly in the form of condition-conclusion structure, that is, if [condition 1] ∧ [condition 2] ∧..., then [diagnostic conclusion]. The system parses the above rules into a tree structure, and the structure specification is as follows: each non-leaf node represents a condition judgment (for example, "is the tongue coating thick and greasy?"), each leaf node represents a diagnostic result label (that is, the traditional Chinese medicine syndrome type, such as "phlegm-dampness obstruction" and "blood stasis syndrome"), each path represents a complete reasoning link from the top-level judgment to the diagnostic label, and the node structure can be recursively nested to support multi-level combined judgment; Parse the "if... then..." rules in the knowledge base into a symbolic tree structure: each non-leaf node represents a condition judgment, and each leaf node represents a diagnostic feature or traditional Chinese medicine syndrome type. [Is the tongue coating thick and greasy?] → Yes → [Is the pulse slippery?] → Yes → Phlegm-dampness obstruction: [Is the pulse slippery?] → No → Spleen deficiency with dampness obstruction: [Is the tongue coating thick and greasy?] → No → [Is the tongue texture dull?] → Yes → Blood stasis syndrome.

[0079] Step S33: Optimize the sorting of tree nodes for the decision tree structure data based on expert experience data to obtain an expert experience judgment tree model; Specifically, the input structure is an expert experience judgment tree constructed manually, where each non-leaf node represents a conditional judgment (such as symptom appearance, sign matching), and the leaf node corresponds to a specific syndrome or diagnosis result label; the system performs data-driven optimization on the judgment tree structure based on "expert scoring", "case support degree", and "symptom statistical frequency". To improve the judgment efficiency and path accuracy of the tree structure, the system adopts the following strategies for node re-sorting and conditional pruning: Sort by symptom frequency; count the occurrence frequency of all conditional nodes in historical samples; arrange high-frequency conditions preferentially near the root node to divert as early as possible. If nodes are given different confidence weights by experts (such as credibility scores, experience levels), then preferentially lift the path branches with high weights; support the fusion of expert scores and data indicators to construct a weighted sorting scoring function. If a certain node or subtree has a support degree lower than the set threshold in the training sample, or the classification gain is not obvious (such as information gain less than ε), it is automatically marked as redundant and structural pruning is performed; the pruning method can be pre-pruning (such as support degree threshold filtering) or post-pruning (such as minimum error rate correction). The optimization of the judgment tree structure can adopt algorithms such as ID3 / C4.5, which divide nodes based on information gain or information gain rate and are suitable for global structure reconstruction; the incremental editing algorithm, which supports local adjustment and dynamic expansion of the tree structure and is suitable for structure refinement and update maintenance of existing expert trees; the path gain weighted algorithm, which considers feature distribution, weight, and node depth in path scoring to improve the overall path decision efficiency and interpretability.

[0080] Use expert scoring or case support degree statistics to re-sort and optimize the pruning of conditional nodes; the node sorting strategy is to preferentially match the conditions with high symptom occurrence frequencies; preferentially place the branches with high weights on the upper-level nodes; the decision tree structure optimizes the structural stability through ID3 / C4.5 or the incremental editing algorithm.

[0081] Step S34: Perform four diagnostic label mapping based on historical case data to obtain historical case four diagnostic mapping data; Specifically, the input is the free text fields in each historical case, mainly including the chief complaint information (such as "menstrual cycle delayed in the past two months"); the current medical history and past medical history (such as "usually prone to fatigue"); the tongue and pulse findings recorded by the doctor (such as "pale tongue" and "deep and thready pulse"); and the clinical observation descriptions (such as "sallow complexion"). The system uses entity recognition and a label normalization mapping rule library to complete label extraction and standardized classification: uses a domain-trained Chinese NER model (such as BERT+CRF) to identify entities related to the four diagnostic methods; establishes a feature word normalization mapping rule library to uniformly map the original expressions (such as "sallow" and "pale and bluish") to standard labels (such as "yellow" and "white", etc.); the system automatically classifies the extracted labels into one of the four diagnostic modules of "inspection, auscultation and olfaction, inquiry, and palpation"; the natural language description fields in each historical case record are mapped to standard four-diagnostic feature labels; entity recognition + label normalization mapping rule library is adopted; for example: "sallow complexion" → inspection label: "complexion: yellow"; "delayed menstruation, scanty menstrual flow, pale tongue" → are respectively classified into the inquiry and tongue diagnosis modules.

[0082] Step S35: Perform hierarchical weighted regression based on the four-diagnostic mapping data of the historical case to obtain the hierarchical weighted data of the historical case.

[0083] Specifically, construct a multi-level feature and label relationship for the historical case data. The first layer (input dimension): is the four-diagnostic features, including tongue diagnosis features, pulse diagnosis features, complexion features, and inquiry features respectively. Each category is a sub-vector, which together form the full feature vector of the patient; the second layer (output label): is the traditional Chinese medicine diagnosis label, such as qi deficiency, yang deficiency, blood stasis, phlegm dampness, etc., constituting a multi-label space. The goal of the model is, based on the feature vector of the current patient , combined with the similar samples in the historical case, calculate the confidence score of each diagnostic label , and form a weight prediction vector . The regression target calculates the weighted prediction score for each label, which is used to assist in judgment. , is the label vector corresponding to the historical case, is the transpose of the matrix , is the weight matrix of the samples (the similarity weights for different cases), is the feature matrix of the historical case (such as the four-diagnostic vectors of multiple patients), is the ridge regression penalty factor (controlling overfitting), is the identity matrix (with the same column dimension as ). The model can be independently trained for each type of label and supports the simultaneous output of prediction scores for multiple syndromes.

[0084] Preferably, step S4 is specifically as follows: Step S41: Perform model fusion based on the expert experience judgment tree model and the historical case hierarchical weighted data to obtain a preliminary fusion model; Specifically, based on the rule judgment tree model constructed by experts, after inputting the standardized four diagnostic features, the diagnostic advice labels are output along the preset decision path, and the output form is the symbolic decision scores of each candidate label; use the weighted regression model trained from real case data, combine the sample similarity and label distribution, and calculate the numerical probability or weight value of the candidate label; to balance the interpretability of expert knowledge and the objectivity of historical data, adopt a weighted linear fusion strategy to merge the above two outputs into a unified diagnostic score. , For the fusion score of the th syndrome type / label, is the fusion coefficient (such as 0.6 - 0.8, representing the credibility of expert experience), is the output score of the expert decision tree (such as the score after logistic regression),

[0085] Step S42: Extract context features from the four diagnostic atlas mapping data to obtain four diagnostic context feature data; Specifically, this module receives the set of matched nodes output from the atlas mapping module, such as tongue image type nodes: "thick tongue coating", "red tongue"; pulse condition type nodes: "slippery pulse", "slow pulse"; chief complaint type nodes: "cold pain in the lower abdomen", "frequent nocturia"; these nodes serve as the central nodes of the regions of interest in the atlas and participate in context structure mining and embedding calculation. The system performs adjacent node context structure extraction. In the medical knowledge atlas, based on the first-order adjacent relationship of each node, a node context set is constructed, and the semantic types and structural paths of its connecting edges (such as "manifested as", "attributed to", "suitable for use", etc.) are collected. If multi-order structure analysis is enabled, the system can be further extended to a 2-hop neighborhood to construct a deeper semantic context subgraph. Use a graph neural network (such as a graph convolutional network GCN) to model the context structure. For each node, its embedding vector is obtained by aggregating the embedding values of its neighbor nodes, and the specific formula is as follows: , is the context representation (feature vector) of node , is the activation function (such as ReLU), is node , is node , is the set of adjacent nodes of node , that is, the node context set, is the adjacent node Embedding. This structure realizes the semantic propagation ability among "symptoms - syndrome types - etiologies - treatments" through adjacency aggregation, enabling node embeddings to have context semantic capabilities.

[0086] Step S43: Adjust the weights of the preliminary fusion model based on the four diagnostic context feature data to obtain the four diagnostic data weight analysis data.

[0087] Specifically, to reflect the dynamic adjustment of the current node score by context influence, a context awareness factor is introduced, which represents the semantic consistency enhancement value of node in its path. The value range of is [-0.2, +0.2]; a positive value indicates semantic context reinforcement (such as the co-occurrence enhancement of adjacent nodes), and a negative value indicates context conflict or incoordination. The system applies the context adjustment factor to the initial score to calculate the weight: where comes from the context semantic enhancement value of the path where node is located, with a range of [-0.2, +0.2]. This method can maintain the continuity of the original score structure while introducing the non-linear influence of semantic enhancement. The value of can be obtained from any of the following strategies: path semantic consistency scoring, based on indicators such as label similarity, synonym co-occurrence, and medical logical coherence between node and other nodes on its path; adjacency density enhancement, if node is strongly correlated with multiple high-weight adjacent nodes (such as appearing in multiple case paths), then its is increased; expert semantic weight annotation, certain specific combinations are considered to have specific reasoning value by experts, and can be enhanced through rules .

[0088] Preferably, the present application also provides a traditional Chinese medicine data processing system for gynecology based on four diagnostic data weight analysis, which is used to execute the traditional Chinese medicine data processing method for gynecology based on four diagnostic data weight analysis as described above. The traditional Chinese medicine data processing system for gynecology based on four diagnostic data weight analysis includes: A four diagnostic data acquisition module for obtaining four diagnostic data; A four diagnostic knowledge graph mapping module for performing graph mapping according to the four diagnostic data to obtain four diagnostic graph mapping data; A four diagnostic knowledge fusion and modeling module for obtaining expert experience data and historical case data according to the four diagnostic graph mapping data; constructing a decision tree based on the expert experience data to obtain an expert experience judgment tree model; performing hierarchical weighted regression according to the historical case data to obtain historical case hierarchical weighted data; The four diagnostic feature weighted analysis module is used to perform four diagnostic feature weighting processing based on the expert experience judgment tree model and the historical case hierarchical weighted data to obtain the four diagnostic data weight analysis data.

[0089] Therefore, from any perspective, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended application documents rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the application documents are intended to be encompassed within the present invention.

[0090] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features invented herein.

Claims

1. A traditional Chinese medicine data processing method for gynecology based on the weight analysis of four diagnostic data, characterized in that, It includes the following steps: Step S1: Obtain the four diagnostic data; Step S2: Perform atlas mapping based on the four diagnostic data to obtain four diagnostic atlas mapping data; Step S3: Obtain expert experience data and historical case data according to the four diagnostic atlas mapping data; construct a decision tree based on the expert experience data to obtain an expert experience judgment tree model; perform hierarchical weighted regression according to the historical case data to obtain historical case hierarchical weighted data; Step S4: Perform four diagnostic feature weighting processing according to the expert experience judgment tree model and the historical case hierarchical weighted data to obtain four diagnostic data weight analysis data.

2. The method according to claim 1, wherein Specifically, step S1 is as follows: Step S11: Obtain inspection image data through an image acquisition device, and perform inspection image analysis according to the inspection image data to obtain inspection image feature data; Step S12: Collect auscultation audio data through an audio acquisition device, and perform auscultation audio analysis according to the auscultation audio data to obtain auscultation audio feature data; Step S13: Obtain interrogation text data through a text input interface, and perform interrogation text analysis according to the interrogation text data to obtain interrogation text feature data; Step S14: Collect palpation pulse data through a pulse sensor, and perform palpation pulse analysis according to the palpation pulse data to obtain palpation pulse feature data; Step S15: Perform four diagnostic feature description processing according to the inspection image feature data, auscultation audio feature data, interrogation text feature data, and palpation pulse feature data to obtain four diagnostic data.

3. The method according to claim 2, characterized in that Specifically, step S11 is as follows: Use the image acquisition device to perform combined illumination operation with white light and ultraviolet light for shooting operation to obtain inspection image data, where the inspection image data includes user tongue image data and user facial color image data; Use the preset tongue template data to perform three-zone processing on the user tongue image data to obtain tongue three-zone data, where the tongue three-zone data includes tip of tongue zone data, middle of tongue zone data, and root of tongue zone data; Extract color features from the tongue three-zone data to obtain tongue zone color feature data, where the tongue zone color feature data is the pixel ratio of red, white, and purple in each zone of the tongue three-zone data; Calculate the ratio of tongue body area to thickness for the user tongue image data to obtain tongue body area thickness ratio data; Perform surface crack calculation and texture feature encoding on the tongue three-zone data to obtain tongue surface crack data and tongue coating area texture data respectively; Use the preset face template data to perform five-zone processing on the user facial color image data to obtain face five-zone data, where the face five-zone data includes forehead zone data, cheek zone data, nose bridge zone data, perioral zone data, and chin zone data; Perform redness calculation, paleness calculation, and freckle texture feature extraction according to the face five-zone data to obtain face zone redness data, face zone paleness data, and face zone freckle texture feature data respectively; Vectorize according to the color feature data of tongue image partitions, the ratio data of tongue body area and thickness, the crack data on the surface of tongue images, the texture data of the tongue coating area, the redness data of face partitions, the paleness data of face partitions, and the texture feature data of face partition freckles to obtain the visual inspection image feature data.

4. The method according to claim 2, wherein Among them, the auscultation audio feature data includes spectral feature data, frequency jitter feature data, and respiratory rhythm feature data. Step S12 is specifically as follows: Collect auscultation audio data through an audio acquisition device; Extract frame-level acoustic features from the auscultation audio data to obtain frame-level acoustic feature data; Perform clustering calculations on the frame-level acoustic feature data to obtain frame-level acoustic feature clustering data; Segment the auscultation audio data according to the frame-level acoustic feature clustering data to obtain segmented auscultation audio data; Extract spectral features, frequency jitter features, and respiratory rhythm features from the segmented auscultation audio data to obtain spectral feature data, frequency jitter feature data, and respiratory rhythm feature data respectively.

5. The method according to claim 2, wherein Step S14 is specifically as follows: Control the pressure sensor array and the feedback controller to adjust the pressing granularity on the cun, guan, and chi positions and synchronously collect pulse wave signals to obtain palpation pulse data; Extract pulse rate features from the palpation pulse data to obtain pulse rate feature data; Perform rhythmic analysis on the pulse rate feature data to obtain pulse rate rhythmic feature data; Extract waveform shape features from the palpation pulse data to obtain waveform shape feature data; Perform waveform symmetry analysis and smoothness analysis on the waveform shape feature data to obtain waveform symmetry feature data and waveform smoothness feature data respectively; Vectorize the pulse rate feature data, pulse rate rhythmic feature data, waveform shape feature data, waveform symmetry feature data, and waveform smoothness feature data to obtain palpation pulse feature data.

6. The method according to claim 2, characterized in that, Step S15 is specifically as follows: Perform modal unified space embedding on the visual inspection image feature data, auscultation audio feature data, interrogation text feature data, and palpation pulse feature data to obtain four-diagnosis modal unified space data; Perform sub-modal accuracy allocation on the four-diagnosis modal unified space data to obtain four-diagnosis sub-modal accuracy data; Perform modal interaction attention calculation on the four-diagnosis sub-modal accuracy data to obtain four-diagnosis feature interaction data; Perform trainable gating settings on the four-diagnosis feature interaction data to obtain four-diagnosis modal gating data; Perform sparsity constraint on the four-diagnosis modal gating data to obtain sparsity constraint data, and perform redundant channel skipping processing on the sparsity constraint data to obtain four-diagnosis channel pruning data; Perform lightweight semantic label reasoning based on the four-diagnosis channel pruning data to obtain four-diagnosis data.

7. The method according to claim 1, wherein Step S2 is specifically as follows: Step S21: Standardize the entities of the four-diagnosis data to obtain four-diagnosis entity standardized data, and use the preset traditional Chinese medicine knowledge graph to match the node attributes of the four-diagnosis entity standardized data to obtain node matching data; Step S22: Screen the relationship paths of the node attribute matching data to obtain relationship path data; Step S23: Perform modal perception mapping adjustment on the traditional Chinese medicine knowledge graph according to the relationship path data and the node attribute matching data to obtain four-diagnosis graph mapping data.

8. The method according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Retrieve the preset knowledge base and historical experience base according to the four diagnostic atlas mapping data to obtain expert experience data and historical case data; Step S32: Perform structural analysis based on the expert experience data to obtain decision tree structure data; Step S33: Optimize the sorting of tree nodes in the decision tree structure data according to the expert experience data to obtain an expert experience judgment tree model; Step S34: Perform four diagnostic label mapping based on the historical case data to obtain historical case four diagnostic mapping data; Step S35: Perform hierarchical weighted regression based on the historical case four diagnostic mapping data to obtain historical case hierarchical weighted data.

9. The method according to claim 4, wherein Step S4 specifically includes: Step S41: Perform model fusion based on the expert experience judgment tree model and the historical case hierarchical weighted data to obtain a preliminary fusion model; Step S42: Extract context features from the four diagnostic atlas mapping data to obtain four diagnostic context feature data; Step S43: Adjust the weights of the preliminary fusion model according to the four diagnostic context feature data to obtain four diagnostic data weight analysis data.

10. A traditional Chinese medicine data processing system for gynecology based on the weight analysis of four diagnostic data, characterized in that A gynecological traditional Chinese medicine data processing system for implementing the gynecological traditional Chinese medicine data processing method based on four diagnostic data weight analysis as claimed in claim 1, the system includes: A four diagnostic data acquisition module for acquiring four diagnostic data; A four diagnostic knowledge graph mapping module for performing graph mapping according to the four diagnostic data to obtain four diagnostic atlas mapping data; A four diagnostic knowledge fusion and modeling module for obtaining expert experience data and historical case data according to the four diagnostic atlas mapping data; constructing a decision tree based on the expert experience data to obtain an expert experience judgment tree model; performing hierarchical weighted regression based on the historical case data to obtain historical case hierarchical weighted data; A four diagnostic feature weighted analysis module for performing four diagnostic feature weighted processing based on the expert experience judgment tree model and the historical case hierarchical weighted data to obtain four diagnostic data weight analysis data.

Images