Intelligent Diagnosis and Treatment Decision System of Traditional Chinese Medicine Based on Fusion of Knowledge Graph and Sequence Deep Learning

Abstract

The application discloses a traditional Chinese medicine intelligent diagnosis and treatment decision system fusing a knowledge graph and sequence deep learning. The system comprises a knowledge graph construction module, a data processing module, an intelligent evaluation module and a prescription optimization module. The knowledge graph construction module is used for constructing a traditional Chinese medicine multi-modal knowledge graph and generating a semantic aggregation vector of medicinal materials. The data processing module is used for fusing patient four-examination information, genotype and physiological indexes and other multi-source heterogeneous information into a personalized context vector. The intelligent evaluation module is based on an attention mechanism sequence model, and combines the semantic aggregation vector and the personalized context vector to quantitatively evaluate the risk of a traditional Chinese medicine prescription sequence. The prescription optimization module generates adjustment suggestions based on the knowledge graph and patient information when detecting the risk. The application solves the problems that traditional Chinese medicine prescription analysis relies on experience and lacks quantitative evaluation and personalized adaptation, realizes automatic and interpretable analysis of prescription risk, and improves the accuracy and practicability of intelligent auxiliary decision-making of traditional Chinese medicine.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligence, in particular to a traditional Chinese medicine intelligent diagnosis and treatment decision system fusing knowledge graph and sequence deep learning. BACKGROUND

[0002] As a unique medical resource in China, traditional Chinese medicine has unique advantages in chronic diseases, complex diseases and health conditioning. Traditional Chinese medicine compound is the core means of traditional Chinese medicine treatment, and its compatibility follows the theory of "jun, chen, za and shi", and achieves therapeutic effect through multi-component and multi-target synergistic effect. However, the traditional Chinese medicine diagnosis and treatment highly depends on the personal experience of doctors, and has the following significant limitations:

[0003] Traditional compatibility taboos (such as eighteen incompatibles) are mostly qualitative experience summaries, and it is difficult to quantitatively warn the synergistic or antagonistic effect of multiple drugs in complex compound, especially when the patient has special constitution or comorbidity, the potential risk is difficult to systematically evaluate. The modern research of chemical components, target pathways and other of medicinal materials and the traditional knowledge of nature, taste and meridian are scattered in different systems, lack of effective fusion and semantic association, leading to difficulty in fully calling multi-dimensional knowledge in diagnosis and treatment decision. The existing systems are mostly based on fixed rules, and it is difficult to fuse the multi-source heterogeneous data such as four diagnostic information, genotype and physiological indicators of patients, and it is impossible to realize the real individualized prescription risk prediction and optimization suggestion.

[0004] At present, although some individual studies try to combine knowledge graph or deep learning technology, most of them focus on single technical path, and cannot realize the deep fusion of knowledge structuring, risk quantification evaluation and individualization adaptation. Therefore, an intelligent diagnosis and treatment decision scheme is urgently needed to systematically solve the above problems. SUMMARY

[0005] In order to solve the problems existing in the prior art, the present application provides a traditional Chinese medicine intelligent diagnosis and treatment decision system fusing knowledge graph and sequence deep learning.

[0006] As one aspect of the present application, the present application provides a traditional Chinese medicine intelligent diagnosis and treatment decision system fusing knowledge graph and sequence deep learning, comprising:

[0007] A knowledge graph construction module is used for extracting entities and relationships between entities in the field of traditional Chinese medicine from structured and unstructured data sources, and constructing a multi-modal knowledge graph containing entities and relationships between entities, and performing embedding representation learning on the multi-modal knowledge graph to generate semantic aggregation vectors for each medicinal material; wherein the entities include any one or more of traditional Chinese medicinal materials, chemical components, biological targets, pathways, diseases and traditional Chinese medicine syndromes;

[0008] The data processing module is used to acquire multi-source heterogeneous information of the target object, vectorize the multi-source heterogeneous information separately and then fuse them to generate a personalized context vector of the target object.

[0009] The multi-source heterogeneous information includes one or more of the following: information from the four diagnostic methods, drug metabolism-related genotype information, and clinical physiological indicators.

[0010] The intelligent assessment module is used to perform risk quantification assessment on the input traditional Chinese medicine prescription sequence based on the semantic aggregation vector and the personalized context vector using a sequence deep learning model based on a gated attention mechanism.

[0011] The prescription optimization module is used to generate and output adjustment suggestions for the traditional Chinese medicine prescription sequence based on the multimodal knowledge graph and the personalized context vector when the result of the risk quantification assessment indicates that the traditional Chinese medicine prescription sequence has risks.

[0012] In one embodiment, the knowledge graph construction module performs embedding representation learning on the multimodal knowledge graph to generate a semantic aggregation vector for each medicinal herb, including:

[0013] The importance weight of each entity in the set of associated entities of a medicinal material is calculated to represent the semantics of the medicinal material using a gated attention mechanism. The embedding vector of each associated entity is multiplied by the calculated importance weight and then weighted and summed to generate the semantic aggregation vector of the medicinal material.

[0014] In one embodiment, the step of using a gated attention mechanism to calculate the importance weight of each entity in the set of associated entities of the medicinal material for representing the semantics of the medicinal material, and then multiplying the embedding vector of each associated entity with the calculated importance weight and performing a weighted summation to generate the semantic aggregation vector of the medicinal material includes:

[0015] Generate the semantic aggregation vector of the i-th medicinal material according to the following formula:

[0016] in, in, Indicates the i-th medicinal material semantic aggregation vector; This indicates the relationship between the i-th medicinal herb and the multimodal knowledge graph. The set of all directly connected entities; Represents a set One of the entities; Representing entities Vector representation of; Representing entities For characterizing medicinal materials Its importance in the current task; Indicates medicinal materials The initial vector; This represents a pre-trained importance score vector; This represents the pre-trained feature transformation matrix; This represents the pre-trained transformation bias vector; Symbol for vector concatenation; This represents the hyperbolic tangent activation function.

[0017] In one embodiment, the multi-source heterogeneous information is vectorized and then fused to generate the target object.

[0018] Personalized context vectors, including:

[0019] The four diagnostic methods information of the target object are mapped into TCM syndrome vectors through unique thermal coding or syndrome embedding models. ;

[0020] The drug metabolism-related genotype information of the target object is mapped into a genotype vector through one-hot encoding. ;

[0021] The clinical physiological indicators of the target subjects are mapped into physiological indicator vectors through standardization processing. ;

[0022] The TCM syndrome vector, the genotype vector, and the physiological indicator vector are concatenated to generate a personalized context vector for the target object. The calculation formula is as follows:

[0023] in, This represents a vector concatenation function.

[0024] In one embodiment, the step of performing risk quantification assessment on the input traditional Chinese medicine prescription sequence using a sequence deep learning model based on a gated attention mechanism, based on the semantic aggregation vector and the personalized context vector, includes:

[0025] Obtain the semantic aggregation vector corresponding to each herb in the input Chinese herbal prescription sequence;

[0026] Based on the semantic aggregation vector corresponding to each herb in the Chinese medicine prescription sequence, an initial vector sequence of the prescription sequence is constructed.

[0027] Based on the initial vector sequence and the personalized context vector of the target object, a gated attention mechanism is used to determine the personalized attention energy score of the i-th herb to the j-th herb in the traditional Chinese medicine prescription sequence. The calculation formula is as follows:

[0028] in, and The first Weihedi The query vector and key vector of medicinal herbs.

[0029] The dimension of the key vector; This refers to the personalized context vector of the target object; and The first Weihedi Semantic aggregation vectors of medicinal herbs; and For global personalized bias weight vector and drug pair gating weight vector; Use the Sigmoid activation function; As an indicator function, when medicinal materials In the aforementioned multimodal knowledge graph, it refers to medicinal materials. The function value is 1 when the neighbor is selected, and 0 otherwise.

[0030] All personalized attention energy scores Normalized to attention weights using the Softmax function ;

[0031] The attention weight of each other herb in the prescription sequence for each herb pair is compared with a preset weight threshold. When the comparison result shows that the attention weight of a herb pair is equal to or greater than the preset weight threshold, all herb pairs with attention weights equal to or greater than the preset weight threshold are identified as risky herb pairs, and the prescription sequence is identified as having a risk. When the comparison result shows that no herb pairs have attention weights equal to or greater than the preset weight threshold, the prescription sequence is identified as having no risk, and a no-risk message is output.

[0032] In one embodiment, when the risk quantification assessment result indicates that the traditional Chinese medicine prescription sequence has a risk, generating and outputting adjustment suggestions for the traditional Chinese medicine prescription sequence based on the multimodal knowledge graph and the personalized context vector includes:

[0033] When the results of the risk quantification assessment indicate that the traditional Chinese medicine prescription sequence carries a risk,

[0034] Obtain the aforementioned risky medicinal material pairs;

[0035] In the multimodal knowledge graph, candidate herbs with the same or similar therapeutic effects as the target herbs in the risky herb pair are sought, and there are no contraindications between them; based on the syndrome and physiological state of the target object represented by the personalized context vector, the most suitable herb is selected from the candidate herbs as a replacement suggestion; and / or, when the interaction of the risky herb pair is dose-related, a dose modification suggestion for at least one herb in the risky herb pair is generated based on the physiological indicators of the target object in the personalized context vector.

[0036] In one embodiment, the prescription optimization module is further configured to, when outputting the adjustment suggestion, also associate the adjustment suggestion with the reasoning basis from the multimodal knowledge graph.

[0037] In one embodiment, it also includes:

[0038] The feedback optimization module is used to perform the following operations:

[0039] Receive efficacy feedback data from clinical practice, wherein the efficacy feedback data includes at least a sequence of traditional Chinese medicine prescriptions, a corresponding patient-specific context vector, and evaluation labels characterizing the clinical efficacy of the prescription;

[0040] Based on the evaluation labels, prescriptions labeled as valid are used as positive samples, and prescriptions labeled as invalid or causing adverse reactions are used as negative samples, thus forming a supervised learning training set.

[0041] A system overall loss function is constructed, the objective of which is to minimize the risk assessment error of the sequence deep learning model for positive samples, while maximizing the risk assessment error for negative samples; the overall loss function is a binary cross-entropy loss function.

[0042] Based on the supervised learning training set and the overall loss function, gradient descent optimization is performed using the backpropagation algorithm, simultaneously updating the following two types of parameters:

[0043] All trainable parameters in the sequence deep learning model;

[0044] The embedding vectors and relation representation matrices of entities in the multimodal knowledge graph, as well as the trainable parameters in the semantic aggregation vector generation process.

[0045] The beneficial effects of the above-mentioned technical solutions provided in the embodiments of the present invention include at least the following:

[0046] The TCM intelligent diagnosis and treatment decision-making system provided in this invention solves the problems in the background art from the following three aspects:

[0047] First, addressing the problem of fragmented and difficult-to-integrate traditional Chinese medicine knowledge systems, the system constructs a multimodal knowledge graph to structurally integrate the properties, chemical components, and modern targets of medicinal materials with traditional knowledge of their properties, flavors, and meridian tropisms. It also achieves semantic association with individualized data such as patients' four diagnostic methods and genotypes, thereby breaking the previous situation where various types of knowledge existed in isolation and forming a basic framework to support the integrated processing of complex diagnostic and treatment information.

[0048] Secondly, addressing the issue of prescription risk assessment relying on experience and lacking objective quantitative evidence, the system has established a standardized prescription analysis workflow, covering the entire process from data preprocessing, feature vector fusion, sequence model evaluation to optimization suggestion generation. This workflow enables automated and repeatable analysis of traditional Chinese medicine prescription information, significantly improving the efficiency and standardization of prescription risk quantitative assessment, and providing a followable technical path for clinical decision-making.

[0049] Finally, addressing the issue that existing knowledge bases are static and difficult to integrate deeply with intelligent models, the system transforms entities and relationships in the knowledge graph into dynamic vector representations that can be used for machine learning. This enables traditional Chinese medicine knowledge to be directly recognized and calculated by deep learning models, thereby significantly expanding the application depth and practical value of the knowledge base in intelligent diagnosis and treatment scenarios.

[0050] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a schematic diagram of the structure of the TCM intelligent diagnosis and treatment decision system that integrates knowledge graphs and sequence deep learning, provided in an embodiment of the present invention. Detailed Implementation

[0054] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0055] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," "outer," "far," "near," "front," and "rear," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0056] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0057] like Figure 1 As shown, this embodiment of the invention provides a TCM intelligent diagnosis and treatment decision-making system that integrates knowledge graphs and sequence deep learning, including:

[0058] The knowledge graph construction module is used to extract entities and relationships between entities in the field of traditional Chinese medicine from structured and unstructured data sources, construct a multimodal knowledge graph containing entities and relationships between entities, and perform embedding representation learning on the multimodal knowledge graph to generate semantic aggregation vectors for each medicinal material; wherein, the entities include any one or more of the following: medicinal materials, chemical components, biological targets, pathways, diseases, and TCM syndromes.

[0059] The data processing module is used to acquire multi-source heterogeneous information of the target object, vectorize the multi-source heterogeneous information separately and then fuse them to generate a personalized context vector of the target object.

[0060] The multi-source heterogeneous information includes one or more of the following: information from the four diagnostic methods, drug metabolism-related genotype information, and clinical physiological indicators.

[0061] The intelligent assessment module is used to perform risk quantification assessment on the input traditional Chinese medicine prescription sequence based on the semantic aggregation vector and the personalized context vector using a sequence deep learning model based on a gated attention mechanism.

[0062] The prescription optimization module is used to generate and output adjustment suggestions for the traditional Chinese medicine prescription sequence based on the multimodal knowledge graph and the personalized context vector when the result of the risk quantification assessment indicates that the traditional Chinese medicine prescription sequence has risks.

[0063] The beneficial effects of the above technical solution are as follows:

[0064] The TCM intelligent diagnosis and treatment decision-making system provided by this invention solves the problems in the background technology from the following three aspects:

[0065] First, addressing the problem of fragmented and difficult-to-integrate traditional Chinese medicine knowledge systems, the system constructs a multimodal knowledge graph to structurally integrate the properties, chemical components, and modern targets of medicinal materials with traditional knowledge of their properties, flavors, and meridian tropisms. It also achieves semantic association with individualized data such as patients' four diagnostic methods and genotypes, thereby breaking the previous situation where various types of knowledge existed in isolation and forming a basic framework to support the integrated processing of complex diagnostic and treatment information.

[0066] Secondly, addressing the issue of prescription risk assessment relying on experience and lacking objective quantitative evidence, the system has established a standardized prescription analysis workflow, covering the entire process from data preprocessing, feature vector fusion, sequence model evaluation to optimization suggestion generation. This workflow enables automated and repeatable analysis of traditional Chinese medicine prescription information, significantly improving the efficiency and standardization of prescription risk quantitative assessment, and providing a followable technical path for clinical decision-making.

[0067] Finally, addressing the issue that existing knowledge bases are static and difficult to integrate deeply with intelligent models, the system transforms entities and relationships in the knowledge graph into dynamic vector representations that can be used for machine learning. This enables traditional Chinese medicine knowledge to be directly recognized and calculated by deep learning models, thereby significantly expanding the application depth and practical value of the knowledge base in intelligent diagnosis and treatment scenarios.

[0068] In one embodiment, the knowledge graph construction module performs embedding representation learning on the multimodal knowledge graph to generate a semantic aggregation vector for each medicinal herb, including:

[0069] The importance weight of each entity in the set of associated entities of a medicinal herb for representing the semantics of that herb is calculated using a gated attention mechanism. The embedding vector of each associated entity is then multiplied by the calculated importance weight and summed in a weighted manner to generate the semantic aggregation vector of the medicinal herb. Specifically, the semantic aggregation vector of the i-th medicinal herb can be generated according to the following formula:

[0070] in, in, Indicates the i-th medicinal material The semantic aggregation vector, with values ​​ranging from 1 to 2. ,in For the embedded dimension; This indicates the relationship between the i-th medicinal herb and the multimodal knowledge graph. The set of all directly connected entities; Represents a set One of the entities; Representing entities The vector representation of the vector can be obtained through pre-training using knowledge graph embedding techniques (such as TransR); Representing entities For characterizing medicinal materials The importance of a task in the current context is calculated using the Softmax function, with values ​​ranging from (0,1), and all... The sum is 1; Indicates medicinal materials The initial vector can be randomly initialized or obtained by training an independent embedding from its name, and its value range is [value range missing]. ; This represents a pre-trained importance score vector; This represents the pre-trained feature transformation matrix; This represents the pre-trained transformation bias vector; The symbol represents vector concatenation, indicating that two vectors are joined to form a longer vector. For example, if... , ,but ; This represents the hyperbolic tangent activation function, used to compress the output to the interval (-1, 1); This represents a temporary loop variable that iterates through every entity in the set, including the current entity e itself. The vector representation of .

[0071] Among them, the above parameters (Importance scoring vector) (Eigenvalue transformation matrix) and The (transformation bias vector) is the core trainable parameter in the model of this invention. Together, they form a feedforward neural network used to calculate and assign semantic importance weights to each associated entity of the medicinal materials. A detailed explanation follows:

[0072] (Eigenvalue transformation matrix) and (Transformation bias vector): These two parameters together constitute a feature extractor, whose function is to transform the input concatenated vector. A nonlinear feature space transformation is performed. The feature transformation matrix... It is responsible for linearly combining and scaling the different dimensions of the input vector, while the transformed bias vector... This is then used to adjust the activation threshold of the combination. The purpose of this transformation process is to fuse and extract high-level interaction features that are highly relevant to the current drug interaction prediction task from the original, separate medicinal material and entity features.

[0073] (Importance Score Vector): This parameter is an importance scorer. Its function is to map the high-dimensional feature vector after feature transformation and activation into a single scalar score representing the importance weight through a dot product operation. Each element in this vector determines the contribution of the corresponding dimension in the feature space to the final importance judgment.

[0074] The optimal values ​​for these parameters were obtained through supervised training on a large-scale traditional Chinese medicine dataset. The process is summarized as follows: 1. Initialization: Before model training begins, conventional techniques in the field (such as the Xavier uniform distribution initialization method) are used to initialize the parameters. , and 1. Assign initial random values. 2. Optimization Objective: The ultimate goal of training is to minimize the overall loss function (e.g., binary cross-entropy loss) of the entire intelligent medical decision-making system. This loss function measures the difference between the model's accuracy in predicting prescription risk and the actual clinical annotations. The optimization of the above parameters serves this macro-level objective. 3. Learning Process: The gradient of the overall loss with respect to these parameters is calculated using the backpropagation algorithm. Subsequently, an adaptive moment estimation optimizer is used to iteratively update these parameter values ​​based on the gradient. This process is repeated until the model performance converges on the validation set. 4. Final State: After training, the parameters are fixed and saved, embedded into the intelligent medical decision-making system. In practical applications, the system directly calls these trained parameter values ​​for semantic aggregation calculations without retraining.

[0075] Through the data-driven training method described above, the model can automatically learn how to measure the importance of associated entities of medicinal herbs when predicting drug interactions. For example, for the toxic herb "Aconitum carmichaelii," the model may learn to assign higher attention weights to its toxic component "aconitine"; while for the qi-tonifying herb "Astragalus membranaceus," it may assign higher weights to its immunomodulatory targets. This enables... , and Ultimately, these become intelligent parameters that encode domain knowledge, generating semantic aggregation vectors for medicinal materials. It is an intelligent representation that has been selected based on task orientation and is rich in semantic emphasis, laying a solid foundation for subsequent high-precision risk assessment.

[0076] The beneficial effects of the above technical solution are as follows: The gating attention mechanism adopted in this invention enables the system to automatically identify and strengthen the more critical parts of various related information of medicinal materials (such as chemical composition, target points, etc.) that have a greater impact on the current risk judgment, according to the needs of different analysis tasks. This allows for the construction of more refined and targeted feature representations for each medicinal material, effectively improving the efficiency and accuracy of subsequent analysis steps. This mechanism also endows the model with good adaptability: the same medicinal material can generate different feature expressions in different prescription combinations or patient backgrounds, and the model can intuitively display the importance of different related information through internal weights, thereby enhancing the transparency of the entire analysis process and providing understandable and traceable decision-making basis. In addition, this technical solution also realizes the effective transformation of the complex network relationship structure in the knowledge graph into a vector form that can be directly processed by standard machine learning models, breaking through the application bottleneck of graph structure data in traditional sequence models and expanding the system's technical capability to process multi-source information.

[0077] In one embodiment, the multi-source heterogeneous information is vectorized and then fused.

[0078] Generating a personalized context vector for the target object includes:

[0079] The four diagnostic methods information of the target object are mapped into TCM syndrome vectors through unique thermal coding or syndrome embedding models. ;

[0080] The drug metabolism-related genotype information of the target object is mapped into a genotype vector through one-hot encoding. ;

[0081] The clinical physiological indicators of the target subjects are mapped into physiological indicator vectors through standardization processing. ;

[0082] The TCM syndrome vector, the genotype vector, and the physiological indicator vector are concatenated to generate a personalized context vector for the target object. The calculation formula is as follows:

[0083] in, This represents a vector concatenation function.

[0084] The beneficial effects of the above technical solution are as follows: by mapping heterogeneous data of different scales and meanings, such as TCM syndromes, genotypes, and physiological indicators, to the same vector space and splicing them together, the technical problem of multi-source data fusion is solved, providing a consistent and efficient input interface for subsequent models and realizing the standardization and unified representation of multimodal patient information; this technical solution enables the system to transform abstract personalized concepts into concrete and computable mathematical vectors, thereby giving the data-based analysis model the technical basis for distinguishing and processing information from different individuals, and enhancing the system's ability to accommodate and process individual differences in data.

[0085] In one embodiment, the step of performing risk quantification assessment on the input traditional Chinese medicine prescription sequence using a sequence deep learning model based on a gated attention mechanism, based on the semantic aggregation vector and the personalized context vector, includes:

[0086] Obtain the semantic aggregation vector corresponding to each herb in the input Chinese herbal prescription sequence;

[0087] Based on the semantic aggregation vector corresponding to each herb in the Chinese medicine prescription sequence, an initial vector sequence of the prescription sequence is constructed.

[0088] Based on the initial vector sequence and the personalized context vector of the target object, a gated attention mechanism is used to determine the personalized attention energy score of the i-th herb to the j-th herb in the traditional Chinese medicine prescription sequence. The calculation formula is as follows:

[0089] in, and The first Weihedi The query vector and key vector of medicinal herbs. The dimension of the key vector; This refers to the personalized context vector of the target object; and The first Weihedi Semantic aggregation vectors of medicinal herbs; , For global personalized bias weight vector and drug pair gating weight vector; Use the Sigmoid activation function; As an indicator function, when medicinal materials In the aforementioned multimodal knowledge graph, it refers to medicinal materials. The function value is 1 when the i-th medicinal herb is a neighbor, and 0 otherwise; N(i) represents the neighborhood set of all entities (mainly other medicinal herbs) that are directly connected to the i-th medicinal herb (i.e., there is an edge between them) in the multimodal knowledge graph; T is the transpose operation on a vector or matrix, and is the transpose operator;

[0090] All personalized attention energy scores Normalized to attention weights using the Softmax function The calculation formula is as follows:

[0091] From this, we obtain It is a value between 0 and 1, representing the overall evaluation. When the risk of the herbal medicine is mentioned, the first The proportion that each herb should account for; The higher the value, the better the evaluation. At that time, the model believes The more critical the information provided; where N represents the length of the Chinese herbal medicine prescription sequence, indicating that the prescription sequence contains N herbs; t is the summation index variable, which iterates through all herbs from the first to the Nth herb in the Chinese herbal medicine prescription sequence;

[0092] The attention weights corresponding to each herb are compared with a preset weight threshold. When the comparison result shows that there is a pair of herbs whose attention weights are equal to or greater than the preset weight threshold, the herb pairs whose attention weights are equal to or greater than the preset weight threshold are identified as risky herb pairs, and the herbal prescription sequence is identified as having a risk. When the comparison result shows that there is no pair of herbs whose attention weights are equal to or greater than the preset weight threshold, the herbal prescription sequence is identified as having no risk, and a no-risk prompt is output.

[0093] The initial vector sequence of the prescription sequence is constructed based on the semantic aggregation vector corresponding to each herb in the prescription sequence. This can be implemented through the following steps: Step 1: Standardize the input prescription text (e.g., unify herb aliases, remove dosage units) to accurately identify each standard herb entity in the sequence. Step 2: Query the pre-generated "herb-semantic aggregation vector mapping table" to retrieve the corresponding semantic aggregation vector for each identified herb. Step 3: Arrange all retrieved semantic aggregation vectors sequentially according to the original herb order in the prescription, assembling them into an ordered vector sequence. For example, for the prescription [Astragalus membranaceus, Angelica sinensis, Ligusticum chuanxiong], the assembled initial vector sequence is [h...]. _黄芪 , h _当归 , h _川芎 ].

[0094] The gating attention mechanism described in this invention involves the following key parameters: query vector Key vector Global personalized bias weight vector and drug-gated weight vector The definitions, value ranges, and methods for obtaining these parameters are as follows:

[0095] 1. Query Vector With key vectors;

[0096] Query vector , is the in the sequence The vector representation of each medicinal herb is used to query or retrieve information about other medicinal herbs in a sequence. Key vector , is the in the sequence The vector representation of each medicinal herb is used to respond to queries, representing its matching features. The two are connected via a dot product operation. A scalar, called the basic attention energy, is calculated, which characterizes the effect of medicinal materials on individual patient (i.e., target) attention without considering the individual factors of the patient. with medicinal materials The strength of the potential, general correlation between them.

[0097] Query vector With key vector It is itself A real vector of dimension, i.e. The dot product of the two Let be a real number whose range is In practical models, the numerical range is constrained by the vector initialization distribution and the model training process.

[0098] Query vector With key vector The parameters are not stored independently, but rather are semantically aggregated vectors of medicinal materials through linear transformation. and The calculation is performed in real time. The specific calculation formula is as follows:

[0099] in, and It is the trainable parameter matrix in the model. semantic aggregation vector The dimension of the matrix. and The acquisition of model parameters follows the standard training process: at the beginning of training, random initial values ​​are assigned using conventional initialization methods (such as Xavier initialization); subsequently, on the training dataset, the gradient is calculated using the loss function of the aforementioned risk quantification assessment task (such as cross-entropy loss) as the optimization objective, through backpropagation, and its value is iteratively updated using an optimizer (such as the Adam optimizer) until the model converges; after training is completed, and The value is fixed and stored.

[0100] 2. Global Personalized Bias Weight Vector Drug-gated weight vector This is a global personalized bias weight vector, whose function is to incorporate the personalized context vector representing the patient's overall state. Through linear transformation This is mapped to a scalar. This scalar represents the general and fundamental risk level shift that the current specific patient's physiological and pathological basis (such as liver and kidney function, and TCM syndrome) has on the overall prescription risk assessment.

[0101] Let be the drug pair gating weight vector, whose function is to gating the drug pair by the first... Weihedi Feature vector formed by concatenating semantic aggregation vectors of medicinal herbs Through linear transformation and the Sigmoid activation function, it is mapped to a gate value between 0 and 1. This gating value acts as a dynamic modulator, determining the extent to which the aforementioned globally personalized bias scalar is applied to the specific drug pair (i, j) currently being evaluated, thereby achieving differentiated risk modulation for different drug combinations.

[0102] and It is a real number vector, and its dimensions are the same as those of the personalized context vector. and concatenated vectors The dimensions match. Global personalized bias scalar. It is a real number. Drug-to-gated value Because it passes through the Sigmoid function When activated, its output is strictly limited to the open interval (0, 1).

[0103] and These are independent, directly trainable parameter vectors in the model. Their acquisition process is similar to... , Parameters are processed synchronously: Initially, they are randomly initialized during model training; during training, the loss function of the aforementioned risk quantification assessment task is used as the optimization objective, and the loss function is calculated relative to the target value using the backpropagation algorithm. and The gradient is calculated, and the value is iteratively optimized using an optimizer; when the model training is complete, and The optimal state is achieved by numerically encoding the rule-based knowledge learned from the training data about "how to modulate attention based on the individual patient (i.e., the target) and the characteristics of the drug pair," which is then stored as part of the system.

[0104] By introducing trainable parameters and By implementing personalized bias and gating mechanisms, the attention model of this invention possesses dynamic adaptability, enabling it to adapt to individualized patient information (through...). ) and specific combination information of drug pairs (via This mechanism refines the general correlations between drugs, enabling the system's risk assessment to evolve from a general model based on population statistics to a highly personalized and precise analysis. It provides the technological guarantee for achieving both intelligence and precision.

[0105] The beneficial effects of the above technical solution are as follows: By combining patient individual information with a knowledge graph-based gating attention mechanism, this invention can quantify and evaluate the interactions between various medicinal materials in a prescription in the form of specific numerical values, thereby achieving in-depth analysis of the complex compatibility structure within the compound prescription. Based on this, the system can automatically identify and prompt combinations of medicinal materials in the prescription that require special attention according to preset weight thresholds, helping users quickly focus on high-risk aspects from a large number of possible interactions, achieving automatic screening of key information and risk warning. Furthermore, the use of gating mechanisms and domain rule indicator functions makes the model's attention logic closer to the professional understanding of traditional Chinese medicine compatibility, reducing the uninterpretability of traditional deep learning models and enhancing the credibility and acceptability of the analysis results in a professional context.

[0106] In one embodiment, when the risk quantification assessment result indicates that the traditional Chinese medicine prescription sequence has a risk, generating and outputting adjustment suggestions for the traditional Chinese medicine prescription sequence based on the multimodal knowledge graph and the personalized context vector includes:

[0107] When the results of the risk quantification assessment indicate that the traditional Chinese medicine prescription sequence carries a risk,

[0108] Obtain the aforementioned risky medicinal material pairs;

[0109] In the multimodal knowledge graph, candidate herbs with the same or similar therapeutic effects as the target herbs in the risky herb pair are sought, and there are no contraindications between them; based on the syndrome and physiological state of the target object represented by the personalized context vector, the most suitable herb is selected from the candidate herbs as a replacement suggestion; and / or, when the interaction of the risky herb pair is dose-related, a dose modification suggestion for at least one herb in the risky herb pair is generated based on the physiological indicators of the target object in the personalized context vector.

[0110] The beneficial effects of the above technical solution are as follows: the system can not only identify potentially highly correlated medicinal material pairs, but also automatically associate and retrieve relevant alternative solutions (such as medicinal material information or dosage reference ranges) based on the knowledge graph, expanding the single analysis function into a continuous information service chain, realizing the functional extension from risk identification to information association services. This technical solution drives the knowledge graph to transform from a passive database query to an intelligent engine that actively provides associated information. It can automatically reason and retrieve based on the input risk node (medicinal material) along a predefined relationship path (such as similar therapeutic effects, no contraindications), thus improving the proactive service and reasoning capabilities of the knowledge graph.

[0111] In one embodiment, the prescription optimization module is further configured to, when outputting the adjustment suggestion, also associate the adjustment suggestion with the reasoning basis from the multimodal knowledge graph.

[0112] Specifically, when the prescription optimization module outputs adjustment suggestions, it simultaneously extracts and correlates the reasoning basis for the suggestions from the multimodal knowledge graph. For example, for suggestions on replacing medicinal materials, the system outputs the efficacy similarity comparison path and incompatibility verification results based on the graph; for suggestions on dosage adjustment, it outputs the correlation analysis between the safe dosage range, metabolic characteristics, and patient physiological indicators of the medicinal material recorded in the graph. These bases are presented along with the suggestions in the form of structured text or visual graph fragments, making the source of the suggestions clear and traceable, and enhancing the interpretability and clinical reference value of the system.

[0113] The beneficial effects of the above technical solution are as follows: by technically linking the conclusions (adjustment suggestions) with the source data (relationships and attributes in the knowledge graph), it ensures that every piece of information output by the system is verifiable, enhancing the integrity and traceability of the system's output information; at the same time, it provides users with an integrated information package of conclusions and supporting evidence, avoiding the need for users to manually cross-verify between the system and multiple knowledge bases, significantly reducing the cost of information acquisition and verification, and optimizing the human-computer interaction experience and information delivery efficiency.

[0114] In one embodiment, the TCM intelligent diagnosis and treatment decision-making system may further include:

[0115] The feedback optimization module is used to perform the following operations:

[0116] Receive efficacy feedback data from clinical practice, wherein the efficacy feedback data includes at least a sequence of traditional Chinese medicine prescriptions, a corresponding patient-specific context vector, and evaluation labels characterizing the clinical efficacy of the prescription;

[0117] Based on the evaluation labels, prescriptions labeled as valid are used as positive samples, and prescriptions labeled as invalid or causing adverse reactions are used as negative samples, thus forming a supervised learning training set.

[0118] A system overall loss function is constructed, the objective of which is to minimize the risk assessment error of the sequence deep learning model for positive samples, while maximizing the risk assessment error for negative samples; the overall loss function is a binary cross-entropy loss function.

[0119] Based on the supervised learning training set and the overall loss function, gradient descent optimization is performed using the backpropagation algorithm, simultaneously updating the following two types of parameters:

[0120] All trainable parameters in the sequence deep learning model, including the linear transformation matrix of the query, key, and value vectors involved in the gating attention mechanism, and the personalized bias weight vector. and drug-gated weight vector ;

[0121] The embedding vectors and relation representation matrices of entities in the multimodal knowledge graph, as well as the trainable parameters in the semantic aggregation vector generation process, including the importance score vector. Characteristic transformation matrix and transform bias vector .

[0122] The beneficial effects of the above technical solution are as follows: it enables continuous self-optimization and iteration of system performance, allowing the system's data processing model and knowledge representation to continuously evolve with the accumulation of new data, effectively alleviating the problem of model performance degradation caused by changes in data distribution or knowledge updates. It forms a complete technical closed loop of data input, analysis output, practical feedback, and model updates, enabling the system to dynamically optimize itself using data from real-world application environments, thus improving its stability and practicality in real-world complex scenarios.

[0123] Furthermore, in the above technical solution, the vectorization process of drug metabolism-related genotype information involves converting discrete, categorized genotype data into a fixed-length numerical vector. The specific process can be implemented as follows:

[0124] A set of gene loci highly associated with the metabolism, transport, and effects of traditional Chinese medicine is predefined, denoted as:

[0125] in, This indicates the total number of gene loci selected. It represents a specific gene locus (e.g., CYP2C9*3 or VKORC1 rs9923231).

[0126] For sets Each gene locus All possible genotypes constitute a finite set, denoted as:

[0127] in, Indicates site The number of genotypes.

[0128] Suppose the target object is at the location The actual test result is the genotype. (Right now Using a one-hot encoding method, the classification result is mapped to a... 2D binary vector :

[0129] The elements of this vector are defined as follows:

[0130] That is, only the actual genotype corresponds to the vector. index position The value is 1.

[0131] All other positions are 0.

[0132] All One-hot encoding vector of each gene locus The genotype vector is generated by splicing the data according to a predefined site order. :

[0133] in, This represents the vector concatenation function. The final result is... The Wei

[0134] Spend The sum of the number of genotypes at all loci:

[0135] In one embodiment, the aforementioned intelligent TCM diagnosis and treatment decision-making system may further include:

[0136] The governance principles and methods generation module is used to execute the following process:

[0137] Step T1: Extract key syndrome attributes from the TCM syndrome vector in the personalized context vector of the target object, and generate a preliminary set of treatment principles and methods for the syndrome based on the TCM treatment principles and methods knowledge base.

[0138] Step T2: Combine the physiological indicator vector and genotype vector of the target object to screen the preliminary set of treatment principles and methods, and exclude treatment options that are not compatible with the current physiological state or metabolic characteristics of the target object, so as to obtain the screened set of treatment principles and methods.

[0139] Step T3: The set of selected treatment principles and methods is used as a constraint and fed back to the prescription optimization module to guide the generation of medicinal material replacement suggestions, ensuring that the adjusted prescription conforms to the TCM prescription principle of unified prescription based on principles.

[0140] The aforementioned technical solution incorporates the "treatment method" aspect of the TCM framework of "theory, method, prescription, and medicine" into the system workflow, achieving structured reasoning from syndrome differentiation to treatment principles and prescription selection. By introducing treatment principles and methods as theoretical constraints for prescription optimization, the system can not only consider compatibility contraindications and efficacy similarity when recommending alternative medicinal materials, but also ensure that the prescription conforms to the principles of TCM treatment, thereby enhancing the rationality of prescription adjustments and consistency with TCM theory.

[0141] In step T1, generating a preliminary set of treatment principles and methods for the syndrome based on the TCM treatment principles and methods knowledge base can be implemented as follows:

[0142] Step T11: Construct a structured knowledge graph sub-module of treatment principles and methods, wherein the nodes include at least: syndrome, treatment principle, treatment method, and the applicable physiological indicator range attribute associated with the treatment method node, as well as the contraindication relationship attribute with drug metabolism genotype;

[0143] Step T12: Calculate the similarity between the syndrome vector of the target object and the vector representation of all syndrome nodes in the knowledge graph submodule of the treatment principles and methods, and match one or more standard syndromes with the highest similarity.

[0144] Step T13: In the knowledge graph submodule, search and reason along the path of "syndrome-related treatment principle-related treatment method" to obtain all treatment principles and treatment methods associated with the matched standard syndrome, thus forming the preliminary set of treatment principles and treatment methods.

[0145] Step T14: For each treatment method in the preliminary treatment principle and treatment method set, extract the applicable physiological indicator range from its node attributes;

[0146] In step T2, the preliminary set of treatment principles and methods is screened to exclude treatment options that are incompatible with the current physiological state or metabolic characteristics of the target object, including:

[0147] Step T21: For each treatment method option in the preliminary treatment principle and treatment method set, obtain its applicable physiological indicator range predefined in the treatment principle and treatment method knowledge graph submodule;

[0148] Step T22: Compare the physiological indicator vector of the target object with the applicable physiological indicator range of each treatment method. If an indicator exceeds the range, it is determined to be incompatible. Count the number of indicators exceeding the range for each treatment method as a measure of its incompatibility.

[0149] Step T23: Based on the drug metabolism-related genotype vector of the target object, query the predefined gene therapy contraindication mapping table and mark the therapy options that are contraindicated with the genotype;

[0150] Step T24: Screening and sorting based on the incompatibility degree measurement and gene taboo markers: First, exclude all treatment options marked as gene taboos; then, sort the remaining treatments in ascending order of their "number of out-of-range indicators," with fewer indicators indicating higher compatibility; finally, select treatments with zero "number of out-of-range indicators," or treatments ranked in the top preset positions, to form the screened set of treatment methods.

[0151] The aforementioned technical solution provides a clear, stable, and easy-to-implement screening mechanism without introducing additional trainable parameters, thus reducing system complexity and the risk of overfitting. By directly comparing physiological indicators with a pre-defined applicable range and using the intuitive metric of "the number of indicators exceeding the range," it achieves an efficient and interpretable assessment of the compatibility between treatment methods and the patient's physiological state. Combined with gene contraindication rules, it ensures the biosafety basis of treatment recommendations. This method is clearly defined, computationally efficient, and entirely based on knowledge and safety principles within the field of Traditional Chinese Medicine, enhancing the stability and clinical interpretability of the system output. This provides a solid logical basis and operability for the individualized screening process in "syndrome differentiation and treatment."

[0152] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. This disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A TCM intelligent diagnosis and treatment decision-making system integrating knowledge graphs and sequence deep learning, characterized in that, include: The knowledge graph construction module is used to extract entities and relationships between entities in the field of traditional Chinese medicine from structured and unstructured data sources, construct a multimodal knowledge graph containing entities and relationships between entities, and perform embedding representation learning on the multimodal knowledge graph to generate semantic aggregation vectors for each medicinal material; wherein, the entities include any one or more of the following: medicinal materials, chemical components, biological targets, pathways, diseases, and TCM syndromes. The data processing module is used to acquire multi-source heterogeneous information of the target object, vectorize the multi-source heterogeneous information separately and then fuse them to generate a personalized context vector of the target object. The multi-source heterogeneous information includes one or more of the following: information from the four diagnostic methods, drug metabolism-related genotype information, and clinical physiological indicators. The intelligent assessment module is used to perform risk quantification assessment on the input traditional Chinese medicine prescription sequence based on the semantic aggregation vector and the personalized context vector using a sequence deep learning model based on a gated attention mechanism. The prescription optimization module is used to generate and output adjustment suggestions for the traditional Chinese medicine prescription sequence based on the multimodal knowledge graph and the personalized context vector when the result of the risk quantification assessment indicates that the traditional Chinese medicine prescription sequence has risks. The step of using a gated attention mechanism to calculate the importance weight of each entity in the set of associated entities of the medicinal material for representing the semantics of the medicinal material, and then multiplying the embedding vector of each associated entity with the calculated importance weight and performing a weighted sum to generate the semantic aggregation vector of the medicinal material includes: Generate the semantic aggregation vector of the i-th medicinal material according to the following formula: in, in, Indicates the i-th medicinal material semantic aggregation vector; This indicates the relationship between the i-th medicinal herb and the multimodal knowledge graph. The set of all directly connected entities; Represents a set One of the entities; Representing entities Vector representation of; Representing entities For characterizing medicinal materials Its importance in the current task; Indicates medicinal materials The initial vector; This represents a pre-trained importance score vector; This represents the pre-trained feature transformation matrix; This represents the pre-trained transformation bias vector; Indicates the vector concatenation symbol; Represents the hyperbolic tangent activation function; The step of vectorizing and fusing the multi-source heterogeneous information to generate a personalized context vector for the target object includes: The four diagnostic methods information of the target object are mapped into TCM syndrome vectors through unique thermal coding or syndrome embedding models. ; The drug metabolism-related genotype information of the target object is mapped into a genotype vector through one-hot encoding. ; The clinical physiological indicators of the target subjects are mapped into physiological indicator vectors through standardization processing. ; The TCM syndrome vector, the genotype vector, and the physiological indicator vector are concatenated to generate a personalized context vector for the target object. The calculation formula is as follows: in, This represents a vector concatenation function.

2. The TCM intelligent diagnosis and treatment decision-making system according to claim 1, characterized in that, The knowledge graph construction module performs embedding representation learning on the multimodal knowledge graph to generate a semantic aggregation vector for each medicinal herb, including: The importance weight of each entity in the set of associated entities of a medicinal material is calculated to represent the semantics of the medicinal material using a gated attention mechanism. The embedding vector of each associated entity is multiplied by the calculated importance weight and then weighted and summed to generate the semantic aggregation vector of the medicinal material.

3. The TCM intelligent diagnosis and treatment decision-making system according to claim 1, characterized in that, The step of performing risk quantification assessment on the input traditional Chinese medicine prescription sequence using a sequence deep learning model based on a gated attention mechanism, based on the semantic aggregation vector and the personalized context vector, includes: Obtain the semantic aggregation vector corresponding to each herb in the input Chinese herbal prescription sequence; Based on the semantic aggregation vector corresponding to each herb in the Chinese medicine prescription sequence, an initial vector sequence of the prescription sequence is constructed. Based on the initial vector sequence and the personalized context vector of the target object, a gated attention mechanism is used to determine the personalized attention energy score of the i-th herb to the j-th herb in the traditional Chinese medicine prescription sequence. The calculation formula is as follows: in, and The first Weihedi The query vector and key vector of medicinal herbs. The dimension of the key vector; This refers to the personalized context vector of the target object; and The first Weihedi Semantic aggregation vectors of medicinal herbs; and For global personalized bias weight vector and drug pair gating weight vector; Use the Sigmoid activation function; As an indicator function, when medicinal materials In the aforementioned multimodal knowledge graph, it refers to medicinal materials. The function value is 1 when the neighbor is selected, and 0 otherwise. All personalized attention energy scores Normalized to attention weights using the Softmax function ; The attention weight of each other herb in the prescription sequence for each herb pair is compared with a preset weight threshold. When the comparison result shows that the attention weight of a herb pair is equal to or greater than the preset weight threshold, all herb pairs with attention weights equal to or greater than the preset weight threshold are identified as risky herb pairs, and the prescription sequence is identified as having a risk. When the comparison result shows that no herb pairs have attention weights equal to or greater than the preset weight threshold, the prescription sequence is identified as having no risk, and a no-risk message is output.

4. The TCM intelligent diagnosis and treatment decision-making system according to claim 1, characterized in that, When the risk quantification assessment result indicates that the traditional Chinese medicine prescription sequence has a risk, adjustment suggestions for the traditional Chinese medicine prescription sequence are generated and output based on the multimodal knowledge graph and the personalized context vector, including: When the results of the risk quantification assessment indicate that the traditional Chinese medicine prescription sequence carries a risk, Obtain the aforementioned risky medicinal material pairs; In the multimodal knowledge graph, candidate herbs with the same or similar therapeutic effects as the target herbs in the risky herb pair are sought, and there are no contraindications between them; based on the syndrome and physiological state of the target object represented by the personalized context vector, the most suitable herb is selected from the candidate herbs as a replacement suggestion; and / or, when the interaction of the risky herb pair is dose-related, a dose modification suggestion for at least one herb in the risky herb pair is generated based on the physiological indicators of the target object in the personalized context vector.

5. The TCM intelligent diagnosis and treatment decision-making system according to claim 1, characterized in that, The prescription optimization module is also used to associate the adjustment suggestion with the reasoning basis from the multimodal knowledge graph when outputting the adjustment suggestion.

6. The TCM intelligent diagnosis and treatment decision-making system according to claim 1, characterized in that, Also includes: The feedback optimization module is used to perform the following operations: Receive efficacy feedback data from clinical practice, wherein the efficacy feedback data includes at least a sequence of traditional Chinese medicine prescriptions, a corresponding patient-specific context vector, and evaluation labels characterizing the clinical efficacy of the prescription; Based on the evaluation labels, prescriptions labeled as valid are used as positive samples, and prescriptions labeled as invalid or causing adverse reactions are used as negative samples, thus forming a supervised learning training set. A system overall loss function is constructed, the objective of which is to minimize the risk assessment error of the sequence deep learning model for positive samples, while maximizing the risk assessment error for negative samples; the overall loss function is a binary cross-entropy loss function. Based on the supervised learning training set and the overall loss function, gradient descent optimization is performed using the backpropagation algorithm, simultaneously updating the following two types of parameters: All trainable parameters in the sequence deep learning model; The embedding vectors and relation representation matrices of entities in the multimodal knowledge graph, as well as the trainable parameters in the semantic aggregation vector generation process.

7. The TCM intelligent diagnosis and treatment decision-making system according to claim 1, characterized in that, Also includes: The governance principles and methods generation module is used to execute the following process: Step T1: Extract key syndrome attributes from the TCM syndrome vector in the personalized context vector of the target object, and generate a preliminary set of treatment principles and methods for the syndrome based on the TCM treatment principles and methods knowledge base. Step T2: Combine the physiological indicator vector and genotype vector of the target object to screen the preliminary set of treatment principles and methods, and exclude treatment options that are not compatible with the current physiological state or metabolic characteristics of the target object, so as to obtain the screened set of treatment principles and methods. Step T3: The set of selected treatment principles and methods is used as a constraint and fed back to the prescription optimization module to guide the generation of medicinal material replacement suggestions, ensuring that the adjusted prescription conforms to the TCM prescription principle of unified prescription based on principles.

Images