Multi-modal traditional Chinese medicine diagnosis data fusion verification method
By constructing a third-order tensor and using a non-negative tensor decomposition algorithm, the syndrome components of concurrent syndromes in traditional Chinese medicine are decoupled, solving the problem of quantitative analysis of concurrent syndromes in traditional Chinese medicine diagnosis, and realizing the precision of traditional Chinese medicine diagnosis and individualized treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-03
AI Technical Summary
The complexity of concurrent syndromes in TCM diagnosis makes it impossible to quantify the specific composition ratio of each individual syndrome component within a complex syndrome, affecting the accuracy of drug combination and adjustment of treatment principles during the treatment phase.
A multimodal TCM diagnostic data fusion and verification method was adopted. By collecting tongue appearance, pulse appearance and consultation information, a third-order tensor was constructed. Then, a non-negative constraint tensor decomposition algorithm was used to decouple the data into syndrome components with non-negative characteristics and calculate the quantitative composition ratio of each syndrome component.
It enables structured and quantitative analysis of concurrent syndromes in Traditional Chinese Medicine, improving the accuracy of diagnosis and the formulation of individualized treatment principles, reducing the black-box nature of traditional artificial intelligence methods, and enhancing the credibility of clinical results.
Smart Images
Figure CN121786607A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the technical field of TCM diagnostic information processing, and discloses a multimodal TCM diagnostic data fusion and verification method. Background Technology
[0002] In TCM clinical diagnosis, complex syndromes are prevalent, essentially involving the simultaneous presence and combination of multiple individual syndrome elements. Current modern research in TCM diagnosis largely focuses on using artificial intelligence (AI) to classify and identify single syndromes or diseases, or on feature splicing and fusion at the multimodal data level. For complex syndromes, only qualitative, holistic labeling can be provided, without further quantifying the specific proportions of each individual syndrome component within the complex syndrome. The essence of TCM complex syndromes is the positive superposition of multiple syndrome components. Without applying non-negative constraints, current tensor decomposition methods may exhibit negative load strengths, contradicting the pathological logic of TCM syndromes. For example, while a diagnosis of liver qi stagnation and spleen deficiency can be made, the relative proportions of liver qi stagnation and spleen deficiency cannot be determined. This ambiguity in qualitative diagnosis directly leads to insufficient precision in medication combinations during treatment, making it difficult to dynamically adjust treatment principles and dosages based on the primary and secondary relationships of the syndrome components, thus reducing the precision of TCM diagnosis and treatment. Summary of the Invention
[0003] The purpose of this section is to outline some aspects of the embodiments of this application and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents, and such simplifications or omissions should not be construed as limiting the scope of this application.
[0004] To address the aforementioned technical issues, this application provides a method for fusing and verifying multimodal TCM diagnostic data.
[0005] On the one hand, this application provides a multimodal TCM diagnostic data fusion and verification method, including S1 collecting multimodal TCM diagnostic data containing tongue appearance, pulse appearance and consultation information; for each patient, converting each modality data into a feature vector; using the feature vector as the basic unit, constructing a third-order tensor for all patients' feature data according to the patient dimension, diagnostic modality dimension and feature index dimension; S2 decomposes the third-order tensor into a combination of a set of basis components by non-negative constraint tensor decomposition; the third-order tensor decomposition produces a patient factor matrix, each column of which represents a decoupled syndrome component with non-negative properties, and each value in the matrix represents the patient's loading intensity on a specific component.
[0006] S3 extracts the load intensity values of the target patient's corresponding row vector in the patient factor matrix on each potential syndrome component column; and calculates the quantitative composition ratio of various syndrome components in the patient's concurrent syndrome based on the relative magnitude of each intensity value in the row vector.
[0007] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: The statistical features, texture features of the tongue body region, and geometric features of the tongue shape in the color space of the tongue image are extracted to form the tongue image feature vector; The time-domain features, frequency-domain features, and nonlinear dynamic features of the waveforms in the pulse data are extracted to form the pulse feature vector; The consultation information is processed by natural language processing to extract bag-of-words vectors of key symptoms to form the consultation feature vector.
[0008] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: The third-order tensor is a three-dimensional array data structure with dimensions I×J×K, where: The first dimension I is the patient dimension, and the value of I is equal to the total number of patient samples; The second dimension J is the diagnostic modality dimension. The value of J is equal to the number of diagnostic modalities used, and they are arranged in a preset order. The third dimension K is the feature index dimension, and the value of K is equal to the total length of all modal feature vectors.
[0009] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: For the i-th patient, the feature vector of its j-th examination modality is used as a modality slice of the third-order tensor with patient dimension index i and modality dimension index j; By traversing all patients and all modalities, the feature vectors are filled into the three-dimensional array, which is a three-dimensional vector.
[0010] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: The third-order tensor is processed using a nonnegative tensor decomposition algorithm; The nonnegativity constraint is to force all elements in the factor matrices and core tensors generated during the decomposition process to be nonnegative. The non-negative tensor decomposition algorithm is the non-negative CP decomposition.
[0011] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method of this application, the third-order tensor is decomposed into the sum of multiple rank tensors during non-negative CP decomposition, wherein each rank tensor is composed of the outer product of the patient factor vector, the modality factor vector, and the feature factor vector. The patient factor matrix is formed by arranging all the patient factor vectors together.
[0012] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: Each column of the patient factor matrix represents a decoupled syndrome component with non-negative properties. After the non-negative tensor decomposition is completed, the Rth column of the patient factor matrix, R, is defined as the Rth independent potential syndrome component decoupled from the original mixed data, with values of 1, 2, 3... Each value in the matrix represents the patient's loading intensity on a specific component, including the specific value in the i-th row and R-th column of the patient factor matrix. This value represents the quantitative contribution or weight of the i-th patient to the R-th potential syndrome component in the overall syndrome presentation. The intensity values of all patients in the same syndrome component column constitute the distribution of the current syndrome component in the population.
[0013] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: The step of extracting the load intensity values of the row vectors on each potential syndrome component column includes: obtaining the corresponding specific row vectors from the patient factor matrix through the identifier index of the target patient; The specific row vector is a non-negative array containing multiple elements, where the position index of each element corresponds to a specific potential syndrome component column; Read all elements in the specific row vector in positional order to obtain the numerical sequence of load intensity of the target patient on a series of potential syndrome components.
[0014] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: Normalize all elements in the load intensity numerical sequence; The normalization process involves dividing the value of each element in the sequence by the sum of the values of all elements in the current load intensity value sequence, and scaling the values of all elements in the sequence to a preset sum reference. After the normalization process, the value of each element is converted into a ratio value, and the set of ratio values constitutes the quantitative composition ratio.
[0015] As a preferred embodiment of the multimodal TCM diagnostic data fusion and verification method proposed in this application, wherein: Based on the aforementioned quantitative composition ratio, a structured diagnostic report is generated; The structured diagnostic report at least lists the various syndrome components contained in the concurrent syndrome of the target patient, the quantitative composition ratio of each syndrome component, and the primary and secondary syndrome relationship determined based on the quantitative composition ratio.
[0016] The beneficial effects of this application are as follows: This application achieves a structured and mathematical representation of complex TCM diagnostic information by unifying heterogeneous data such as tongue, pulse, and questioning into a third-order tensor of patient modal features.
[0017] This application decomposes the third-order tensor into a patient factor matrix and a base component, thereby decoupling the clinically qualitative concurrent syndrome into multiple independent potential syndrome components with non-negative properties and obtaining the specific loading intensity of each component in the patient, thus solving the problem of the inability to quantify the internal structure of the syndrome.
[0018] This application normalizes the loading intensity of target patients in the patient factor matrix, thereby achieving a specific quantitative composition ratio of various syndrome components in the output of concurrent syndromes. This improves the accuracy of TCM syndrome differentiation from qualitative judgment to quantitative analysis, provides data support for individualized treatment principles and drug dosage matching, reduces the black box characteristics of traditional artificial intelligence methods in TCM diagnosis, and enhances the credibility and acceptability of clinical results. Attached Figure Description
[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained through these drawings without creative effort. Wherein: Figure 1 A flowchart of a multimodal TCM diagnostic data fusion and verification method provided in this application; Figure 2 This application provides a two-layer correction loop for a multimodal TCM diagnostic data fusion and verification method. Figure 3 A schematic diagram of the construction and filling of a third-order tensor for a multimodal TCM diagnostic data fusion and verification method provided in this application; Figure 4 The overall flowchart of a multimodal TCM diagnostic data fusion and verification method provided in this application is shown. Detailed Implementation
[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.
[0021] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.
[0022] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.
[0023] Example 1 like Figure 1 As shown, a multimodal TCM diagnostic data fusion and verification method includes: S1 collects multimodal TCM diagnostic data including tongue appearance, pulse appearance, and consultation information; for each patient, each modal data is converted into a feature vector; using the feature vector as the basic unit, the feature data of all patients are constructed into a third-order tensor according to the patient dimension, diagnostic modality dimension, and feature index dimension; Specifically, tongue image data is acquired through digital image acquisition equipment under standard lighting conditions, pulse image data is acquired through a high-precision pulse sensor, recording the pressure pulsation waveform sequence at the radial artery, and consultation information is entered through a structured electronic medical record system or human-computer interaction interface, recording the patient's chief complaint, symptoms, medical history, and other text or structured data.
[0024] Tongue and pulse diagnosis data are obtained using existing equipment: tongue diagnosis instrument, pulse diagnosis instrument, and traditional Chinese medicine tongue image processing and analysis software.
[0025] It should be noted that the tongue diagnosis instrument, pulse diagnosis instrument, and TCM tongue surface image processing and analysis software have all obtained medical device registration certificates.
[0026] The statistical features, texture features of the tongue body region, and geometric features of the tongue shape in the color space of the tongue image are extracted to form the tongue image feature vector; The time-domain features, frequency-domain features, and nonlinear dynamic features of the waveforms in the pulse data are extracted to form the pulse feature vector; The consultation information is processed by natural language processing to extract bag-of-words vectors of key symptoms to form the consultation feature vector.
[0027] Specifically, the acquired digital images of the tongue are first preprocessed, including image calibration and tongue region segmentation to eliminate interference from the background and other parts of the face. Three types of features are extracted from the segmented tongue region to form a tongue image feature vector: Furthermore, in color spaces such as RGB and HSV, color statistics of the tongue body and tongue coating areas are calculated, such as the mean, standard deviation, and distribution skewness of each channel, to quantify tongue color (e.g., pale red, crimson, bluish-purple) and coating color (e.g., white, yellow, grayish-black).
[0028] Furthermore, the texture features of the tongue region are calculated using a gray-level co-occurrence matrix to determine the roughness, contrast, and uniformity of the tongue surface, which are used to describe the thickness, dryness, and putridity of the tongue coating.
[0029] It should be noted that the gray-level co-occurrence matrix is a classic texture analysis method. By statistically analyzing the co-occurrence frequency of pixel gray levels in an image at specific directions and distances, it quantifies the spatial relationship between pixels. Those skilled in the art can combine the known techniques of the gray-level co-occurrence matrix with the tongue surface data disclosed in this application to obtain the texture features of the tongue region.
[0030] Furthermore, the outline of the tongue is extracted, and its geometric attributes are calculated, such as the length-to-width ratio of the tongue, the degree of serration of the outline, the presence or absence of teeth marks or cracks, and their morphological measurements, in order to determine the age, size, and presence or absence of teeth marks on the tongue.
[0031] All the calculated feature values are arranged in a predetermined order to form a numerical tongue image feature vector that comprehensively represents tongue image information.
[0032] Specifically, the acquired raw pulse waveform sequences are preprocessed, including denoising, period segmentation, and alignment, to obtain a series of representative single-period pulse waves. From these waveforms, the following three types of features are extracted to construct the pulse feature vector: Furthermore, morphological parameters, such as the height of the main wave, the height of the diphthoplast, the height of the tidal wave, the time interval corresponding to each wave, and the area ratio, are directly extracted from the time-domain pulse waves to describe the depth of the pulse position, the strength of the pulse, and the smoothness or roughness of the pulse shape.
[0033] Spectral analysis of pulse signals is performed to extract features such as energy distribution and dominant frequency of their power spectrum in different frequency bands, providing supplementary information for traditional time-domain analysis. Nonlinear dynamic analysis methods are used to calculate pulse sequence, entropy value or fractal dimension, etc.
[0034] The nonlinear dynamics include bifurcation theory, chaos theory, and nonlinear time series analysis.
[0035] By sequentially combining time-domain, frequency-domain, and nonlinear features, a numerical pulse feature vector representing pulse information is constructed.
[0036] The acquired consultation text information is automatically analyzed using natural language processing techniques to construct feature vectors: First, the text is segmented and standardized, and then matched with a pre-built TCM symptom lexicon.
[0037] Then, the bag-of-words model is used to statistically analyze specific key symptom entries, such as existence Boolean identifiers for symptoms like rib pain, loss of appetite, and fatigue.
[0038] Finally, the statistical values of all keyword entries are arranged in a fixed order to form a numerical consultation feature vector, which reflects the presentation pattern of the patient's symptoms.
[0039] The third-order tensor is a three-dimensional array data structure with dimensions I×J×K, where: The first dimension I is the patient dimension, and the value of I is equal to the total number of patient samples; The second dimension J is the diagnostic modality dimension. The value of J is equal to the number of diagnostic modalities used, and they are arranged in a preset order. The third dimension K is the feature index dimension, and the value of K is equal to the total length of all modal feature vectors.
[0040] After converting each patient's tongue, pulse, and medical history information into standardized feature vectors, the dispersed and modally heterogeneous feature vectors are statistically transformed into a unified and structured third-order tensor, such as... Figure 3 As shown.
[0041] The left side of the figure lists the original feature vectors of three patients. For example, the tongue, pulse, and consultation feature vectors of patient 1 are V11, V12, and V13, respectively.
[0042] By traversing all patients (i) and all diagnostic modalities (j), each feature vector Vij is filled into the modal slice position in the tensor, which is uniquely determined by its coordinates (i,j).
[0043] The right side of the diagram abstractly represents the final generated third-order tensor data structure, where: Patient dimension (I): Size is 3, corresponding to three patients.
[0044] Diagnostic modality dimension (J): 3 in size, arranged in a fixed order (e.g., tongue, pulse, questioning).
[0045] Feature metric dimension (K): The length is the sum of the dimensions of all feature vectors.
[0046] The figure clearly marks instances of modal slices, such as slice T(i=1,j=1,:), which fully stores the tongue image feature vector V11 of patient 1.
[0047] Specifically, the first dimension is the patient dimension, with index i, used to represent different individual patients.
[0048] Furthermore, suppose there are I patients in total, then the length of the dimension is I, and the index i (1≤i≤I) points to all the data of the i-th patient.
[0049] The second dimension is the diagnostic modality dimension, index j, which is used to represent different diagnostic methods, arranged in a preset fixed order, such as: tongue appearance, pulse appearance, and questioning modality.
[0050] If a total of J modalities are used, then the dimension length is J, and the index j points to the j-th diagnostic modality. For example, j=1 represents the tongue image modality.
[0051] The third dimension is the feature index dimension, index k, which is used to represent the specific feature entry of each diagnostic modality feature vector. The total length K of the feature index dimension is equal to the sum of the lengths of all modality feature vectors, that is, the sum of the dimension numbers of the tongue image, pulse image, and consultation feature vectors. Index k points to a globally unique feature index position.
[0052] The third-order tensor T is mathematically defined as a three-dimensional array of size I×J×K.
[0053] Using the aforementioned feature vectors as basic units, the three-dimensional array T is filled according to the following rules: For the i-th patient, the entire tongue image feature vector is filled into a tensor into a slice with index (i, j = fixed sequence number of tongue image modality, k = corresponding range of tongue image feature position).
[0054] The pulse feature vector of the i-th patient is filled into a slice of the tensor with index (i, j = fixed sequence number of pulse mode, k = corresponding pulse feature position range).
[0055] Fill the entire consultation feature vector of the i-th patient into a slice of the tensor with index (i, j = fixed sequence number of consultation modality, k = corresponding consultation feature position range).
[0056] Iterate through all patients (i from 1 to I).
[0057] Through the above construction, any specific element T(i,j,k) in tensor T has physical meaning. For example, it represents the quantified value of the k-th feature index of the i-th patient under the j-th diagnostic modality.
[0058] An example of a preferred specific element T(i,j,k) includes: T(5,1,3) may represent the third color statistical feature value of the tongue image of the 5th patient (modal 1).
[0059] By integrating the originally scattered data into a data cube with a defined coordinate system, the position of each element is uniquely determined by the three coordinates of its patient, modality, and feature, thus fully encapsulating all the information and internal structure of the multimodal diagnostic data.
[0060] For the i-th patient, the feature vector of its j-th examination modality is used as a modality slice of the third-order tensor with patient dimension index i and modality dimension index j; By traversing all patients and all modalities, the feature vectors are filled into the three-dimensional array, which is a three-dimensional vector.
[0061] In this application, a preferred method for implementing modal slicing includes: For a defined third-order tensor T, after fixing its patient dimension index as i and its examination modality dimension index as j, the resulting vector expanded along the feature index dimension K is the modality slice of the tensor at position (i,j).
[0062] Specifically, the modal slice is represented as T(i,j,:), which is a one-dimensional vector with a length equal to the total length K of the feature index dimension. The modal slice is used to store all feature information of the i-th patient under the j-th diagnostic modality.
[0063] A preferred method for slicing feature vectors includes filling the generated feature vectors into their corresponding modality slices, with the following specific rules: For the i-th patient, obtain the feature vector generated in the j-th examination modality, for example, j=1 corresponds to the tongue image modality.
[0064] All elements of the feature vector are sequentially assigned to the segment with index (i,j,k_start:k_end) in the third-order tensor T. Here, k_start and k_end are pre-allocated, continuous index ranges of the modal feature vector on the global feature index dimension K, realizing the use of the feature vector as a modal slice T(i,j,:) of the tensor T.
[0065] A preferred method for constructing tensors by traversing and filling them includes: Initialize a three-dimensional array data structure of size I×J×K, with all elements either empty or zero, as a container for tensor T.
[0066] Iterate through patient dimension index i (from 1 to I), and for each patient i, iterate through examination modality dimension index j (from 1 to J).
[0067] In each inner and outer loop, the slice filling operation in point 2 above is performed, and the corresponding feature vector is filled into the modal slice determined by the current index (i,j).
[0068] After all modalities of all patients have been traversed and processed, all positions in the three-dimensional array are filled with the corresponding feature data. Thus, the three-dimensional array constitutes the third-order tensor T defined in this application, which carries all multimodal diagnostic information.
[0069] This application discloses an implementable technical means to integrate multi-source heterogeneous TCM feature vectors into a third-order tensor data structure through clearly defined modal slicing and traversal filling rules.
[0070] S2 decomposes the third-order tensor into a combination of a set of basis components by non-negative constraint tensor decomposition; the third-order tensor decomposition produces a patient factor matrix, each column of which represents a decoupled syndrome component with non-negative properties, and each value in the matrix represents the patient's loading intensity on a specific component.
[0071] The third-order tensor is processed using a nonnegative tensor decomposition algorithm; The nonnegativity constraint is to force all elements in the factor matrices and core tensors generated during the decomposition process to be nonnegative. The non-negative tensor decomposition algorithm is the non-negative CP decomposition.
[0072] Specifically, through non-negative CP decomposition, tensors containing mixed information are decoupled into several basis components with clear interpretable meanings, and key patient factor matrices are extracted from them.
[0073] The nonnegative CP decomposition is used for factor analysis of higher-order arrays. It approximates a given tensor as the sum of a finite number of rank tensors and applies a nonnegativity constraint to the third-order tensor T.
[0074] The nonnegativity constraint ensures that every element in all factor vectors obtained during the entire decomposition and optimization process must be greater than or equal to zero. This ensures that each base component obtained from the decomposition represents a pure and positive contribution pattern. Furthermore, different components are superimposed with nonnegative weights, resulting in data representation that is both additive and interpretable. This generates a combination of TCM syndrome elements to form a mixed syndrome.
[0075] Specifically, the nonnegative CP decomposition is defined as a constrained optimization problem to find a set of nonnegative factor vectors {aᵣ,bᵣ,cᵣ} such that the tensor formed by the sum of their outer products has the smallest difference from the original third-order tensor T.
[0076] Constrained optimization problems are solved using alternating least squares. An example of a preferred least squares method is to fix two sets of factor vectors in each iteration, for example, fix all bᵣ and cᵣ, simplifying the multivariate optimization problem to a least squares problem for only the other set of factor vectors, i.e. aᵣ, and then solve it. Furthermore, the other groups are fixed and solved alternately in turn.
[0077] Specifically, ensuring nonnegativity occurs after solving each simplified subproblem. For the currently optimized factor vector, such as aᵣ, the theoretical solution is calculated according to the least squares principle. Since the theoretical solution may contain negative elements, a deterministic operation of nonnegative projection is immediately performed, checking each element of the solution vector one by one: If the element value is ≥0, it is retained.
[0078] If the element value is less than 0, it is forced to be set to 0.
[0079] The solution after non-negative projection is used as the valid update value for this iteration.
[0080] Furthermore, the algorithm fixes the other groups of factor vectors and repeats the steps for the next group of factor vectors, alternating in this way until the update changes of all factor vectors are less than a preset threshold.
[0081] It should be noted that this application can also set and adjust the iteration termination threshold and the maximum number of iterations of the optimized non-negative CP decomposition algorithm parameters; and adjust the weight configuration of the feature vector by adjusting the correlation coefficient between each modal feature and the syndrome component. The higher the correlation coefficient, the greater the weight ratio.
[0082] By using nonnegativity as a domain constraint for the optimization problem and employing the nonnegative projection enforcement step embedded in the iterative algorithm, it is ensured that in the final convergent solution, each element of the patient factor vector, modality factor vector, and feature factor vector must be zero or a positive number.
[0083] In this application, a preferred embodiment of obtaining a matrix component combination through non-negative CP decomposition includes: The specific process and results of performing nonnegative CP decomposition on the third-order tensor T are as follows: The decomposition approximates the third-order tensor T as a summation of R rank-tensors, i.e., T≈Σ[aᵣ*bᵣ*cᵣ], where r ranges from 1 to R, R is the preset number of decomposition components, representing the number of potential symptom components expected to be decoupled from the data, the symbol * denotes the cross product operation of vectors, and aᵣ, bᵣ, and cᵣ are three non-negative factor vectors, corresponding to the three dimensions of tensor T respectively. The rank-tensor generated by each cross product aᵣ, bᵣ, and cᵣ is defined as a basis component constituting the original tensor. Where: The patient factor vector aᵣ is used to describe the distribution intensity of the r-th basic component across all I patients.
[0084] The modal factor vector bᵣ is used to describe the performance weight of the r-th base component in the J diagnostic modalities (tongue, pulse, question).
[0085] The feature factor vector cᵣ is used to describe the loading pattern of the r-th basis component on the K global feature indices.
[0086] The combination of the basis components refers to the set of R rank tensors (aᵣ, bᵣ, cᵣ). The entire decomposition process, under non-negative constraints, optimizes and solves all R sets {aᵣ, bᵣ, cᵣ} through an iterative algorithm, so that the combination can approximate the original tensor T.
[0087] An example of a preferred choice of R: The value of the number of decomposition components R can be in the range of 3-8, with a default value of 5, corresponding to the five common core syndrome components in traditional Chinese medicine, such as liver stagnation, spleen deficiency, blood stasis, damp heat, and qi deficiency.
[0088] The third-order tensor is decomposed into the sum of multiple rank tensors during non-negative CP decomposition, where each rank tensor is composed of the outer product of the patient factor vector, the modality factor vector, and the feature factor vector. The patient factor matrix is formed by arranging all the patient factor vectors together.
[0089] In this application, a preferred method for constructing and establishing a patient factor matrix includes: The patient factor matrix is derived directly from the results of non-negative CP decomposition.
[0090] The patient factor matrix A is a two-dimensional matrix of size I×R, i.e., I rows and R columns. The R patient factor vectors aᵣ obtained from the decomposition of the patient factor matrix are arranged side by side.
[0091] Furthermore, the r-th column of the matrix is the patient factor vector aᵣ corresponding to the r-th base component. The column vector contains the loading intensity of all I patients on the r-th potential syndrome component.
[0092] The i-th row of the matrix corresponds to the i-th patient. The row vector is an R-dimensional vector, and its first to R-th elements represent the loading intensity of the patient on all R decoupled potential syndrome components.
[0093] Specifically, by executing the non-negative CP decomposition algorithm, we solve and output R sets of factor vectors {a1,b1,c1}, {a2,b2,c2},...,{a_R,b_R,c_R}.
[0094] Furthermore, the R patient factor vectors a1, a2, ..., a_R obtained from the solution are used as column vectors and arranged sequentially from left to right. A preferred method for arranging them sequentially includes corresponding component numbers and arranging them from left to right according to the number order.
[0095] Furthermore, the side-by-side column vectors form an I-row, R-column matrix, which is the patient factor matrix A.
[0096] This application decouples a third-order tensor reflecting multimodal information of a population into interpretable basis components and generates a patient factor matrix. Each column of the patient-i factor matrix represents the distribution of an independent potential syndrome component in the population, and each row is a syndrome fingerprint of a patient across all components.
[0097] Each column of the patient factor matrix represents a decoupled syndrome component with non-negative properties. After the non-negative tensor decomposition is completed, the Rth column of the patient factor matrix, R, is defined as the Rth independent potential syndrome component decoupled from the original mixed data, with values of 1, 2, 3... Each value in the matrix represents the patient's loading intensity on a specific component, including the specific value in the i-th row and R-th column of the patient factor matrix. This value represents the quantitative contribution or weight of the i-th patient to the R-th potential syndrome component in the overall syndrome presentation. The intensity values of all patients in the same syndrome component column constitute the distribution of the current syndrome component in the population.
[0098] It should be noted that the core tensor is an important diagnostic result for the patient. For example, when examining a patient's eyes to see if there is inflammation, if the patient also has floaters, then when examining the patient's eyes, floaters are a non-core tensor, while inflammation is a core tensor.
[0099] Specifically, the patient factor matrix A is obtained through non-negative CP decomposition, an interpretive mapping rule is established, and the elements in the matrix are associated with semantic concepts in traditional Chinese medicine diagnostics, thus completing the transformation from mathematical output to medical knowledge.
[0100] Column vectors are used to represent potential symptom components; mapping is performed on each column vector of the patient factor matrix A: Specifically, the R-th column vector (R=1, 2, ..., R) of the patient factor matrix is defined as a potential symptom component. For example, the first column vector is defined as component C1, the second column vector is defined as component C2, and so on.
[0101] A preferred example of decoupling includes: each potential syndrome component Cᵣ is considered as an independent pathological information pattern separated from the original third-order tensor T, i.e., the original data mixed with multi-patient, multimodal information.
[0102] The potential need is for medical semantic labeling.
[0103] A preferred medical interpretation of the nonnegativity property includes: since the decomposition is subject to nonnegativity constraints, all elements in each column vector are nonnegative, ensuring that the contribution of each potential symptom component to any patient is positive or zero, which can be interpreted as a pure, unidirectional pathological influence factor whose numerical value represents the strength of the factor's influence.
[0104] After defining the overall meaning of the column vectors, the medical interpretation of each specific element in the matrix was also determined: Mapping of element A(i,R): The specific value in the i-th row and R-th column of matrix A is mapped and interpreted as: the quantitative contribution or weight of the R-th potential syndrome component Cᵣ in the overall clinical manifestation of the i-th patient. This value is called the loading intensity of patient i on component Cᵣ.
[0105] The intensity value in the load intensity is a relative scalar. The higher the value, the more significant the portion of the patient's syndrome presentation explained by the Rth potential pathological pattern. A value of zero or close to zero indicates that the patient's syndrome presentation basically does not include this pathological pattern. For example, if C1 is subsequently labeled as the liver stagnation factor, then A(5,1)=0.85 means that the pathological element of liver stagnation contributes a high weight to the syndrome presentation of patient No. 5.
[0106] The R-th column of matrix A, i.e., the vector [A(1,R),A(2,R),…,A(I,R)]^T, is extracted separately, where ^T is the transpose of the matrix. This vector represents the intensity distribution of the R-th potential symptom component in the entire patient population (sample size I). By analyzing this intensity, the prevalence and variability of this component in a specific population can be assessed.
[0107] A preferred method for analyzing the intensity includes, for example, calculating the mean, variance, and observing its distribution pattern.
[0108] A preferred implementation method for optimizing the parameters of the nonnegative tensor decomposition algorithm and the weight configuration of the eigenvectors includes: The clinical efficacy evaluation results of the target patient at preset time points after treatment based on the structured diagnostic report are obtained; the clinical efficacy evaluation results are correlated with the primary and secondary syndromes determined in the structured diagnostic report; if the efficacy does not meet expectations, a correction signal is generated based on the correlation analysis results, and the correction signal is used to adjust the parameters of the non-negative tensor decomposition algorithm and the weight configuration of the feature vector; based on the adjusted parameters or weights, the original multimodal diagnostic data or the third-order tensor of the target patient is reprocessed to generate an updated quantitative composition ratio and structured diagnostic report.
[0109] Specifically, the key parameter of the nonnegative tensor decomposition algorithm is the number of decomposition components R, i.e., the preset number of potential symptom components. The process is as follows: Figure 2 As shown: If the curative effect fails to meet expectations, for example, if the diagnosis is mainly liver depression but the treatment of soothing the liver is ineffective, analysis is initiated. A typical situation is that the poor curative effect may stem from the over-simplification of the model, with too small R value and failure to isolate key but secondary syndromes, or from the over-decomposition with too large R value, introducing noise components. <0000,290><0000,291>Furthermore, compare the original diagnosis proportions [p1, p2,..., p_R] of cases with poor curative effects with the diagnosis proportion distribution of similar cases with definite curative effects. If a significant difference is found, generate a correction signal for the R value. <0000,292><0000,293>Furthermore, adjust the strategy upward, R’ > R, which is triggered when the analysis believes that important syndrome components may be missing. For example, if the curative effect feedback indicates that there may be unrecognized blood stasis or damp-heat components, increase the R value by one step, such as R’ = R + 1. <0000,294><0000,295>Furthermore, adjust the strategy downward, R’ < R, which is triggered when the analysis believes that there are redundant or noise components in the current decomposition. By analyzing the load stability of each component in all cases with poor curative effects, identify components with low contribution and instability, and merge or remove them to reduce the R value. <0000,296><0000,297>After adjustment, use the new R’ value to re-execute the S2 tensor decomposition and S3 proportion calculation steps on the data of the affected patient group to obtain an updated patient factor matrix and quantitative composition ratio. <0000,298>[[ID=I4]]<0000,299>An example of adjusting the weight configuration of a preferred eigenvector includes: <0000,300>The weight configuration is used to adjust the relative importance of different examination modalities, such as tongue, pulse, and inquiry, in the integrated diagnosis. <0000,301><0000,302>Specifically, assign an initial weight vector W = [w_tongue, w_pulse, w_inquiry,...] to J examination modalities. Usually, the initial values are equal weights. Before constructing the third-order tensor, multiply the modal eigenvector of each patient by its corresponding modal weight wj. <0000,303><0000,304>Collect cases with definite curative effect feedback. For each case with poor curative effect, trace back the diagnostic decision: check the components judged as the main syndromes, and compare the modal factor vector bᵣ of the patient components with the standard modal factor vector. <0000,305><0000,306>If the diagnosis dominated by a modality with high weight dependence, for example, tongue image, is negated by the curative effect feedback, reduce the weight wj of the modality; if it is found that in cases with good curative effects, the characteristic pattern of a previously underestimated modality, for example, inquiry, has potential similarity with the current case with poor curative effect, increase the weight wj of the modality. <0000,307><0000,308>Generate a new modal weight vector W’ according to the above rules. <0000,309><0000,310>The original feature vectors are reweighted using W', and then the third-order tensor is reconstructed from this.
[0120] For this new tensor, perform nonnegative tensor decomposition and all subsequent calculations with the current optimal R value to generate an updated diagnostic report that incorporates knowledge of treatment efficacy.
[0121] This application establishes a mapping channel for clinical efficacy, transforming the abstract concept of poor efficacy into specific and operable adjustment strategies for model parameters (R value) and data weights (W vector), thereby achieving diagnostic self-optimization. By utilizing continuously generated clinical feedback data, it automatically optimizes the internal model, further verifying and ensuring the clinical effectiveness of its fusion diagnosis.
[0122] This application transforms a numerical matrix into a series of potential syndrome components with medical diagnostic semantics and their intensity on an individual basis, generating a syndrome component intensity spectrum for each patient, with corresponding row vectors [A(i, 1), A(i, 2), ..., A(i, R)]. These vectors are used to describe the internal composition of the patient's syndrome, indicating which components it contains and the relative strength of each component. Each row of the patient factor matrix A becomes a representation of the quantitative composition of the patient's individual syndrome, which can be used to analyze the patient's symptoms and their respective weights.
[0123] S3 extracts the load intensity values of the target patient's corresponding row vector in the patient factor matrix on each potential syndrome component column; and calculates the quantitative composition ratio of various syndrome components in the patient's concurrent syndrome based on the relative magnitude of each intensity value in the row vector.
[0124] The step of extracting the load intensity values of the row vectors on each potential syndrome component column includes: obtaining the corresponding specific row vectors from the patient factor matrix through the identifier index of the target patient; The specific row vector is a non-negative array containing multiple elements, where the position index of each element corresponds to a specific potential syndrome component column; Read all elements in the specific row vector in positional order to obtain the numerical sequence of load intensity of the target patient on a series of potential syndrome components.
[0125] Individual syndrome intensity extraction based on patient factor matrix By generating and assigning medical interpretations to a patient factor matrix, individualized quantitative data of syndrome components are extracted for a specific target patient.
[0126] The patient factor matrix has been interpreted, with columns defined as potential symptom components and elements defined as loading strengths.
[0127] A preferred method for extracting numerical sequences of individual load intensity includes: The patient factor matrix is retrieved based on the unique identifier index of the target patient, such as the patient's ID i in the database.
[0128] An example of a preferred retrieval method includes: when constructing the patient factor matrix A, a correspondence has been established between the row order and the patient identifier index.
[0129] Specifically, the i-th row of the matrix (i=1,2,…,I) is pre-assigned and uniquely allocated to store the load intensity of all syndrome components of the patient identified by index i in the database. The correspondence is the basic rule when constructing the matrix, ensuring that the index i is directly used as the row subscript of the matrix.
[0130] When the system needs to retrieve the target patient with the identifier index 'it', it uses the input patient identifier index 'it' as the row subscript and directly accesses the 'it'-th row of the patient factor matrix A through the index access mechanism of standard arrays or matrices.
[0131] It should be noted that the index access mechanism is computationally equivalent to locating the exact location where all data in the it-th row is stored by calculating a fixed offset from the starting memory address or data storage location.
[0132] Once the target row is located via addressing, a read operation is performed: Specifically, all R numerical elements constituting the row are read continuously from storage, each corresponding to the load intensity of one of the R potential syndrome components.
[0133] The read values are combined into a one-dimensional, ordered array of values according to their column order in the matrix.
[0134] The numerical array represents the search results: the specific row vector corresponding to the target patient.
[0135] From the patient factor matrix, obtain the i-th row corresponding to the identifier index i. The row data is the specific row vector corresponding to the target patient.
[0136] Furthermore, the specific row vector is a one-dimensional non-negative array containing R elements, where R is the total number of potential symptom components.
[0137] The position index (1,2,...,R) of the first element, second element, ... up to the Rth element of the specific row vector corresponds to the first column, second column, ..., Rth column of the patient factor matrix, and corresponds to the first, second, ..., Rth potential syndrome components.
[0138] Furthermore, each element value in the specific row vector is read in order from the 1st position to the Rth position.
[0139] The first element value read is the load intensity of the target patient on the first potential syndrome component.
[0140] The second element value read is the load intensity of the target patient on the second potential syndrome component.
[0141] Similarly, after all R elements have been read sequentially, an ordered sequence of load intensity values of length R is obtained. The sequence is represented in array form [intensity_1, intensity_2, ..., intensity_R] and is used to record the quantitative contribution of the target patient to all R decoupled syndrome components.
[0142] Normalize all elements in the load intensity numerical sequence; The normalization process involves dividing the value of each element in the sequence by the sum of the values of all elements in the current load intensity value sequence, and scaling the values of all elements in the sequence to a preset sum reference. After the normalization process, the value of each element is converted into a ratio value, and the set of ratio values constitutes the quantitative composition ratio.
[0143] In this application, a preferred method for normalization processing includes: The process involves normalizing all elements in the load intensity numerical sequence.
[0144] Specifically, the sum of sequence elements is calculated by performing a summation operation on all R numerical elements of the load intensity sequence, denoted as [s1, s2, ..., sR], to obtain the sum Statotal. That is: Statotal = s1 + s2 + ... + sR. The sum Statotal represents the total load intensity of the target patient across all syndrome components.
[0145] Furthermore, each element in the sequence is scaled proportionally, transforming the sequence into a proportional value.
[0146] In a preferred embodiment, the value of each element in the sequence is divided by the sum Total, i.e., for the r-th element (r from 1 to R), its proportion value pr is calculated: This scales all elements of the entire numerical sequence to a base whose sum is 1.
[0147] Furthermore, through the calculation of the proportional values, the original load intensity sequence [s1, s2, ..., sR] is converted into a new proportional value sequence [p1, p2, ..., pR].
[0148] Each proportion value pr is a value between 0 and 1 (inclusive), used to represent the weight or proportion of the r-th potential syndrome component in the overall syndrome manifestation of the target patient.
[0149] The entire sequence of proportion values [p1, p2, ..., pR] is defined as the quantitative composition proportion, which clearly reveals the internal component structure of the patient's concurrent syndrome.
[0150] This application transforms the potentially vastly different absolute load intensities among different patients into a consistent relative proportion, eliminating the interference of overall intensity differences between individuals on comparative analysis and making the composition proportions comparable.
[0151] The output proportion value pr can be directly interpreted as the syndrome component Cr accounting for approximately pr*100%. For example, if p1=0.7 and p2=0.3, it can be directly concluded that in the patient's syndrome composition, component one accounts for approximately 70% and component two accounts for approximately 30%, providing accurate data basis for the primary and secondary judgment.
[0152] Based on the aforementioned quantitative composition ratio, a structured diagnostic report is generated; The structured diagnostic report at least lists the various syndrome components contained in the concurrent syndrome of the target patient, the quantitative composition ratio of each syndrome component, and the primary and secondary syndrome relationship determined based on the quantitative composition ratio.
[0153] A preferred example of the primary and secondary syndrome relationship determined based on the quantitative composition ratio includes: a quantitative composition ratio ≥ 50% indicates a primary syndrome, 20%-50% indicates a secondary syndrome, and < 20% indicates a mild syndrome.
[0154] The structured diagnostic report is a standard document automatically generated by the system, which includes a list of syndrome costs, quantitative composition ratios, and judgments on the relationship between primary and secondary syndromes.
[0155] The report explicitly lists all potential syndrome components decoupled from the target patient data in the syndrome component list. (C1, C2, ..., The syndrome components may include liver stagnation factors, spleen deficiency factors, etc.
[0156] The quantitative composition ratio is simultaneously labeled next to each listed syndrome component, with its corresponding quantitative composition ratio value (p1, p2, ..., p2) marked accordingly. The proportions are presented as percentages, such as C1: 65.2%; C2: 34.8%, etc., to show the weight of each component in the patient's overall syndrome.
[0157] The primary and secondary syndrome relationship is determined by analyzing the proportion values and classifying the syndrome components into primary and secondary categories according to preset logical rules. For example, the component with the highest proportion is judged as the primary syndrome, and other components with proportions exceeding a certain threshold are judged as secondary syndromes.
[0158] The report is generated through an automated, deterministic data processing and population process. Specifically, the input to the process is a quantitative composition ratio sequence [p1, p2, ..., pR], and predefined labels for each component.
[0159] Furthermore, a pre-set structured report template is invoked, which contains fixed chapters and fields, such as diagnostic summary, syndrome decomposition, quantitative analysis, and primary and secondary judgment.
[0160] Furthermore, the ingredient labels will be matched with the corresponding proportions. Fill in the section on syndrome decomposition and quantitative analysis.
[0161] In the primary / secondary judgment section, the system performs sorting and comparison: identifying the proportion value. The component with the largest value is marked as the primary syndrome; the remaining proportions are iterated through, and components that are greater than a preset threshold are marked as secondary syndromes; the judgment results, such as primary syndrome: liver stagnation; secondary syndrome: spleen deficiency, are formatted and entered into the report.
[0162] Furthermore, the system generates complete structured documents, such as PDFs, HTML documents, or specific formats integrated into the electronic medical record system, containing text descriptions and data tables, to complete the entire analysis process. Figure 4 The diagram shows the overall flowchart of TCM diagnostic data fusion and verification. It involves collecting patient tongue and pulse data, constructing a third-order tensor to analyze and diagnose the patient's syndrome, and finally outputting the diagnostic results.
[0163] This application transforms the abstract concept of concurrent syndromes into a clear list of components and percentages, enabling TCM diagnosis to move from qualitative description to quantitative expression. Through automatic judgment based on fixed rules, it reduces the discrepancies that may arise from relying solely on the doctor's subjective experience, providing a clear and objective data reference for determining medication principles during treatment.
[0164] It is important to note that the constructions and arrangements of this application shown in several different exemplary embodiments are merely illustrative. Although only two embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible without substantially departing from the novel teachings and advantages of the subject matter described in this application. These modifications may include, for example, changes in the size, dimensions, structure, shape, and proportions of various elements, as well as parameter values (e.g., temperature, pressure, etc.), installation arrangements, the use of materials, colors, orientations, etc. For example, an element shown as integrally formed may be composed of multiple parts or elements, the position of elements may be inverted or otherwise altered, and the nature or number or position of discrete elements may be changed or altered. Therefore, all such modifications are intended to be included within the scope of this application. The order or sequence of any process or method steps may be changed or rearranged by alternative embodiments. Any "apparatus plus function" clause is intended to cover, and not only structurally equivalent but also equivalent structures, the structures performing the functions described herein. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of this application. Therefore, this application is not limited to a particular embodiment, but extends to various modifications that still fall within the scope of the appended claims.
[0165] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of performing this application as currently considered, or those features that are not relevant to implementing this application) may be omitted.
[0166] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those of ordinary skill in the art who benefit from this disclosure, the development effort will be a routine task in design, manufacturing, and production without requiring extensive experimentation.
[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application, and all such modifications and substitutions should be covered within the scope of the claims of this application.
Claims
1. A method for fusing and validating multimodal TCM diagnostic data, characterized in that, include: S1 collects multimodal TCM diagnostic data including tongue appearance, pulse appearance, and consultation information; For each patient, the modal data are converted into feature vectors; using the feature vectors as basic units, the feature data of all patients are constructed into a third-order tensor according to the patient dimension, the diagnostic modality dimension, and the feature index dimension. S2 decomposes the third-order tensor into a combination of a set of basis components by non-negative constraint tensor decomposition; the third-order tensor decomposition produces a patient factor matrix, each column of which represents a syndrome component with decoupled non-negative properties, and each value in the matrix represents the patient's loading intensity on a specific component. S3 extracts the load intensity values of the target patient's corresponding row vector in the patient factor matrix on each potential syndrome component column; and calculates the quantitative composition ratio of various syndrome components in the patient's concurrent syndrome based on the relative magnitude of each intensity value in the row vector.
2. The multimodal TCM diagnostic data fusion and verification method as described in claim 1, characterized in that: The statistical features, texture features of the tongue body region, and geometric features of the tongue shape in the color space of the tongue image are extracted to form the tongue image feature vector; The time-domain features, frequency-domain features, and nonlinear dynamic features of the waveforms in the pulse data are extracted to form the pulse feature vector; The consultation information is processed by natural language processing to extract bag-of-words vectors of key symptoms to form the consultation feature vector.
3. The multimodal TCM diagnostic data fusion and verification method as described in claim 2, characterized in that: The third-order tensor is a three-dimensional array data structure with dimensions I×J×K, where: The first dimension I is the patient dimension, and the value of I is equal to the total number of patient samples; The second dimension J is the diagnostic modality dimension. The value of J is equal to the number of diagnostic modalities used, and they are arranged in a preset order. The third dimension K is the feature index dimension, and the value of K is equal to the total length of all modal feature vectors.
4. The multimodal TCM diagnostic data fusion and verification method as described in claim 3, characterized in that: For the i-th patient, the feature vector of its j-th examination modality is used as a modality slice of the third-order tensor with patient dimension index i and modality dimension index j; By traversing all patients and all modalities, the feature vectors are filled into the three-dimensional array, which is a three-dimensional vector.
5. The multimodal TCM diagnostic data fusion and verification method as described in claim 1, characterized in that: The third-order tensor is processed using a nonnegative tensor decomposition algorithm; The nonnegativity constraint is to force all elements in the factor matrices and core tensors generated during the decomposition process to be nonnegative. The non-negative tensor decomposition algorithm is a non-negative CP decomposition. By adjusting the parameters of the non-negative tensor decomposition algorithm, the parameter and feature vector weight configuration of the non-negative tensor decomposition algorithm are optimized.
6. The multimodal TCM diagnostic data fusion and verification method as described in claim 5, characterized in that: The third-order tensor is decomposed into the sum of multiple rank tensors during non-negative CP decomposition, where each rank tensor is composed of the outer product of the patient factor vector, the modality factor vector, and the feature factor vector. The patient factor matrix is formed by arranging all the patient factor vectors together.
7. The multimodal TCM diagnostic data fusion and verification method as described in claim 6, characterized in that: Each column of the patient factor matrix represents a decoupled syndrome component with non-negative properties. After the non-negative tensor decomposition is completed, the Rth column R of the patient factor matrix is set to 1, 2, 3..., and is defined as the Rth independent potential syndrome component decoupled from the original mixed data. Each value in the matrix represents the patient's loading intensity on a specific component, including the specific value in the i-th row and R-th column of the patient factor matrix. This value represents the quantitative contribution or weight of the i-th patient to the R-th potential syndrome component in the overall syndrome presentation. The intensity values of all patients in the same syndrome component column constitute the distribution of the current syndrome component in the population.
8. The multimodal TCM diagnostic data fusion and verification method as described in claim 1, characterized in that: The step of extracting the load intensity values of the row vectors on each potential syndrome component column includes: obtaining the corresponding specific row vectors from the patient factor matrix through the identifier index of the target patient; The specific row vector is a non-negative array containing multiple elements, where the position index of each element corresponds to a specific potential syndrome component column; Read all elements in the specific row vector in positional order to obtain the numerical sequence of load intensity of the target patient on a series of potential syndrome components.
9. The multimodal TCM diagnostic data fusion and verification method as described in claim 8, characterized in that: Normalize all elements in the load intensity numerical sequence; The normalization process involves dividing the value of each element in the sequence by the sum of the values of all elements in the current load intensity value sequence, and scaling the values of all elements in the sequence to a preset sum reference. After the normalization process, the value of each element is converted into a ratio value, and the set of ratio values constitutes the quantitative composition ratio.
10. The multimodal TCM diagnostic data fusion and verification method as described in claim 8, characterized in that: Based on the aforementioned quantitative composition ratio, a structured diagnostic report is generated; The structured diagnostic report at least lists the various syndrome components contained in the concurrent syndrome of the target patient, the quantitative composition ratio of each syndrome component, and the primary and secondary syndrome relationship determined based on the quantitative composition ratio.