Intelligent orthodontic aid decision-making method based on large language model and knowledge graph
By constructing orthodontic knowledge graphs and embedding large language models, the problem of insufficient professionalism in orthodontic decision-making in the existing technology is solved, and the efficiency and reliability of orthodontic treatment are improved.
Patent Information
- Application Number
- CN202510094088.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-21
- Publication Date
- 2025-05-16
AI Technical Summary
Existing large language models have insufficient professionalism in orthodontic decision-making, resulting in low reliability.
Build a professional orthodontic knowledge graph and embed it into a large language model, and generate orthodontic decision-making plans through steps such as entity recognition, knowledge entity relationship extraction and intention recognition.
Improves the efficiency of orthodontic treatment and the reliability of large language models in orthodontic treatment, providing more professional and reliable decision support.
Smart Images
Figure CN120012889A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of orthodontic treatment, and more specifically to an intelligent orthodontic auxiliary decision-making method based on a large language model and a knowledge graph. Background Art
[0002] Orthodontic treatment mainly uses various corrective devices to adjust the coordination between facial bones, teeth, and maxillofacial nerves and muscles, that is, to adjust the abnormal relationship between the upper and lower jaws, between the upper and lower teeth, between the teeth and the jaws, and between the nerves and muscles that connect them. The ultimate goal of orthodontic treatment is to achieve balance, stability, and beauty of the oral and maxillofacial system. Traditional orthodontic decision-making schemes rely on the manual judgment of doctors. With the development of large language models, they have been applied to the medical field. Large language models refer to deep learning models trained with a large amount of text data, which can generate natural language text or understand the meaning of language text. Large language models can handle a variety of natural language tasks, such as text classification, question-answering, dialogue, etc., and are an important path to artificial intelligence. However, the existing large language models are not professional enough, and if used for orthodontic decision-making, their reliability is low. Therefore, how to provide an intelligent orthodontic auxiliary decision-making method based on large language models and knowledge graphs is a problem that technicians in this field urgently need to solve. Summary of the invention
[0003] In view of this, the present invention provides an intelligent orthodontic decision-making assistance method based on a large language model and a knowledge graph, and constructs a professional orthodontic knowledge graph to assist the large language model in orthodontic decision-making to improve its professionalism.
[0004] In order to achieve the above object, the present invention provides the following technical solutions:
[0005] An intelligent orthodontic auxiliary decision-making method based on a large language model and a knowledge graph comprises the following steps:
[0006] S1. Build an entity recognition model to obtain orthodontic-related expertise, clinical guidelines and patient orthodontic management record data, and identify knowledge entities in the data based on the entity recognition model;
[0007] S2, extract the entity relationships between knowledge entities in the data and construct an orthodontic knowledge graph;
[0008] S3, embedding the knowledge graph into the Llama model;
[0009] S4. Design the prompt strategy of the Llama model;
[0010] S5. Perform entity recognition on the patient's decision-making data based on the entity recognition model, extract entity relationships, and obtain the patient's ternary information group;
[0011] S6, performing intention recognition on the patient's decision-making data to obtain the patient's intention information;
[0012] S7. The patient's ternary information group and intention information are used as the input of the Llama model, and the orthodontic decision plan is obtained by combining the prompt strategy.
[0013] Optionally, S1 is:
[0014] S11. Build and train an entity recognition model based on Word2Vec-BiLSTM-CNN-CRF;
[0015] S12, obtain orthodontic related professional knowledge, clinical guidelines and patient orthodontic treatment record data, and convert the sentences in the data into feature vectors based on the Word2Vec module of the entity recognition model;
[0016] S13, inputting the feature vector into the entity recognition model, and the BiLSTM module extracts the hidden state based on the feature vector;
[0017] S14, inputting the hidden state into a CNN module of an entity recognition model, and the CNN module extracting data features based on the hidden state;
[0018] S15. Input the data features into the CRF module of the entity recognition model. The CRF module calculates the probability that the data features belong to different labels, and obtains the label sequence with the highest score as the output of the entity recognition model.
[0019] Optionally, S2 specifically includes: constructing predefined relationships between orthodontic knowledge entities, sorting all knowledge entities identified in the patient's decision-making data according to the order of the original sentences, sequentially searching for predefined relationships between all knowledge entities and other knowledge entities, obtaining ternary information groups of knowledge entity relationships, and fusing knowledge entities based on their attributes.
[0020] Optionally, the knowledge entities are fused based on their attributes as follows: the semantic similarity between the feature vectors of each knowledge entity is calculated, a semantic adjacency graph is constructed, knowledge entities whose semantic similarity exceeds a preset threshold are connected, connected components of the semantic adjacency graph are used as a fusion candidate set, and for each connected component, the knowledge entity with the most connections to other knowledge entities is selected as a representative entity, the connection relationships of the remaining knowledge entities are transferred to the representative entity, and other knowledge entities are deleted to obtain a fused knowledge entity.
[0021] Optional, S3 is:
[0022] S31. Divide the ternary information group (h, r, t) of the knowledge entity relationship in the orthodontic knowledge graph into two tuples (h, r h ) and (rt ,t), calculate the eigenvectors of (h,r) and (r,t) and
[0023] S32, the feature vector v h and v t Stacked as a matrix A = [v h ,v t ], perform convolution operation on matrix A to generate feature map V = [v 1 ,v 2 ,...,v d ]:
[0024] v i =f(ω·A i +b)
[0025] In the formula, v i ∈V, f(·) represents the activation function, ω is the convolution kernel, A i is the i-th row of matrix A, and b is the bias parameter;
[0026] S33, set c different convolution kernels to perform convolution operation on matrix A to generate feature map V 1 ,V 2 ,...,V c , concatenate all feature maps to get the feature matrix V, set the weight matrix W, and calculate the features of the i-th dimension:
[0027] e i =f(V i ·W i )+b)
[0028] Where V i is the i-th row of the feature matrix V, W i is the i-th row of the weight matrix W;
[0029] S34. Calculate the scoring function of the triple information group (h, r, t):
[0030] f r (h,t)=concat(e 1 ,e 2 ,...,e d )·w
[0031] In the formula, concat represents the connection function, w is the scoring weight, and the distributed representation of knowledge entities and relations is learned by maximizing the scores of correct triples in the orthodontic knowledge graph.
[0032] Optionally, S6 is specifically:
[0033] Calculate the word vector of the decision-making data as the patient's intention information:
[0034] TI={TI 1 ,TI 2 ,…,TI i ,…,TI k}
[0035]
[0036] Where TI i represents the importance of knowledge entity i, n i,j represents the number of occurrences of knowledge entity i in sentence j, Σ k n k,j represents the total number of knowledge entities in sentence j, |D| is the total number of sentences, |j:t i ∈d j | represents the number of sentences containing knowledge entity i.
[0037] Optionally, S7 specifically includes: combining the patient's triple information and intention information with the prompt strategy, converting the patient's decision data into an input question of the Llama model, and the Llama model searches the professional knowledge base based on the input question to generate an orthodontic decision plan.
[0038] It can be seen from the above technical solution that compared with the prior art, the present invention provides an intelligent orthodontic decision-making assistance method based on a large language model and a knowledge graph, which has the following beneficial effects: the present invention constructs a professional knowledge graph for orthodontics and embeds it into a large language model, and automatically generates an orthodontic decision-making assistance plan based on the large language model for doctors to refer to, thereby improving the efficiency of orthodontic treatment, and improving the reliability of the large language model for orthodontic treatment by embedding the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0040] Figure 1 It is a flow chart of the intelligent orthodontic decision-making assistance method of the present invention;
[0041] Figure 2 This is a flow chart of knowledge entity recognition of the present invention. DETAILED DESCRIPTION
[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0043] The embodiment of the present invention discloses an intelligent orthodontic auxiliary decision-making method based on a large language model and a knowledge graph, such as Figure 1 As shown, the following steps are included:
[0044] S1. Build an entity recognition model to obtain orthodontic-related expertise, clinical guidelines and patient orthodontic management record data, and identify knowledge entities in the data based on the entity recognition model;
[0045] S2, extract the entity relationships between knowledge entities in the data and construct an orthodontic knowledge graph;
[0046] S3, embedding the knowledge graph into the Llama model;
[0047] S4. Design the prompt strategy of the Llama model;
[0048] S5. Perform entity recognition on the patient's decision-making data based on the entity recognition model, extract entity relationships, and obtain the patient's ternary information group;
[0049] S6, performing intention recognition on the patient's decision-making data to obtain the patient's intention information;
[0050] S7. The patient's ternary information group and intention information are used as the input of the Llama model, and the orthodontic decision plan is obtained by combining the prompt strategy.
[0051] Further, such as Figure 2 As shown, S1 is specifically:
[0052] S11. Build and train an entity recognition model based on Word2Vec-BiLSTM-CNN-CRF;
[0053] S12, obtain orthodontic related professional knowledge, clinical guidelines and patient orthodontic treatment record data, and convert the sentences in the data into feature vectors based on the Word2Vec module of the entity recognition model;
[0054] S13, inputting the feature vector into the entity recognition model, and the BiLSTM module extracts the hidden state based on the feature vector;
[0055] S14, inputting the hidden state into a CNN module of an entity recognition model, and the CNN module extracting data features based on the hidden state;
[0056] S15. Input the data features into the CRF module of the entity recognition model. The CRF module calculates the probability that the data features belong to different labels, and obtains the label sequence with the highest score as the output of the entity recognition model.
[0057] Furthermore, in the embodiment of the present invention, BiLSTM runs two LSTM units simultaneously at each time step: one processes data in the forward order of the sequence, and the other processes data in the reverse order, which can more comprehensively capture the features and contextual relationships in the sequence;
[0058] The LSTM network consists of three gate structures, including the forget gate, input gate, and output gate. The forget gate outputs a value between 0 and 1 through the Sigmoid activation function to determine the degree of retention of each state. The input gate mainly controls the input at the current moment and updates the unit state. The output gate generates an output value through the Sigmoid function. Based on the output value and the state value at the current time, the hidden state h at the current moment is obtained. t , specifically:
[0059] f t =σ(W f ·[h t-1 ,x t ]+b f )
[0060] i t =σ(W i ·[h t-1 ,x t ]+b i )
[0061]
[0062] o t =σ(W o ·[h t-1 ,x t ]+b o )
[0063]
[0064] h t =o t *tanh(C t )
[0065] Where W f , W i , W C , W o are the weight matrices of the forget gate, input gate, candidate value, and output gate, respectively, and b f 、b i 、bC 、b o are the bias vectors of the forget gate, input gate, candidate value, and output gate, respectively. σ(·) represents the Sigmoid activation function. t-1 is the hidden state at the previous moment, h t is the hidden state at the current moment, x t is the current input, is the candidate value, i t is the updated value, o t is the output value, c t-1 is the state value at the previous moment, f t is the degree of retention, c t is the current state value.
[0066] The CNN module includes convolutional layer, pooling layer and fully connected layer. The convolutional layer is based on the hidden state h at the current moment. t The features are obtained, the pooling layer compresses the input features and retains the valid information in the data, and the fully connected layer performs adjustment and regularization to obtain the final feature H.
[0067] The CRF module calculates the label output Y=[y 1 ,y 2 ,...,y n ] probability:
[0068]
[0069] In the formula, Ψ i (y i-1 ,y i ,H) is the potential function of the CRF module, i∈n, Respectively represent y i-1 ,y i The real label.
[0070] Furthermore, S2 is specifically as follows: constructing predefined relationships between orthodontic knowledge entities, sorting all knowledge entities identified in the patient's decision-making data according to the order of the original sentences, sequentially searching for predefined relationships between all knowledge entities and other knowledge entities, obtaining ternary information groups of knowledge entity relationships, and fusing knowledge entities based on their attributes.
[0071] Furthermore, in an embodiment of the present invention, the attribute tags of knowledge entities and the predefined relationships between knowledge entities are defined based on orthodontic-related professional materials.
[0072] Furthermore, the knowledge entities are fused based on their attributes as follows: the semantic similarity between the feature vectors of each knowledge entity is calculated, a semantic adjacency graph is constructed, knowledge entities whose semantic similarity exceeds a preset threshold are connected, connected components of the semantic adjacency graph are used as a fusion candidate set, and for each connected component, the knowledge entity with the most connections to other knowledge entities is selected as a representative entity, the connection relationships of the remaining knowledge entities are transferred to the representative entity, and other knowledge entities are deleted to obtain the fused knowledge entity.
[0073] Furthermore, in the embodiment of the present invention, the calculation of semantic similarity is specifically as follows:
[0074]
[0075] S AB Represents the similarity between word vector A and word vector B.
[0076] Furthermore, S3 is specifically:
[0077] S31. Divide the ternary information group (h, r, t) of the knowledge entity relationship in the orthodontic knowledge graph into two tuples (h, r h ) and (r t ,t), calculate the eigenvectors of (h,r) and (r,t) and
[0078] S32, the feature vector v h and v t Stacked as a matrix A = [v h ,v t ], perform convolution operation on matrix A to generate feature map V = [v 1 ,v 2 ,...,v d ]:
[0079] v i =f(ω·A i +b)
[0080] In the formula, v i ∈V, f(·) represents the activation function, ω is the convolution kernel, A i is the i-th row of matrix A, and b is the bias parameter;
[0081] S33, set c different convolution kernels to perform convolution operation on matrix A to generate feature map V 1 ,V 2 ,...,V c , concatenate all feature maps to get the feature matrix V, set the weight matrix W, and calculate the features of the i-th dimension:
[0082] ei =f(V i ·W i +b)
[0083] Where V i is the i-th row of the feature matrix V, W i is the i-th row of the weight matrix W;
[0084] S34. Calculate the scoring function of the triple information group (h, r, t):
[0085] f r (h,t)=concat(e 1 ,e 2 ,...,e d )·w
[0086] In the formula, concat represents the connection function, w is the scoring weight, and the distributed representation of knowledge entities and relations is learned by maximizing the scores of correct triples in the orthodontic knowledge graph.
[0087] Furthermore, the prompt strategy is used to prompt the Llama model's answer and guide the Llama model to generate a specific type of response, for example: The patient's dental condition is (), please generate an orthodontic strategy () based on the following orthodontic-related knowledge ().
[0088] Furthermore, S6 is specifically:
[0089] Calculate the word vector of the decision-making data as the patient's intention information:
[0090] TI={TI 1 ,TI 2 ,…,TI i ,…,TI k}
[0091]
[0092] Where TI i represents the importance of knowledge entity i, n i,j represents the number of occurrences of knowledge entity i in sentence j, ∑ k n k,j represents the total number of knowledge entities in sentence j, |D| is the total number of sentences, |j:t i ∈d j | represents the number of sentences containing knowledge entity i.
[0093] Furthermore, S7 is specifically as follows: the patient's triple information and intention information are combined with the prompt strategy to convert the patient's decision data into an input question of the Llama model. The Llama model searches the professional knowledge base based on the input question to generate an orthodontic decision plan.
[0094] Furthermore, in an embodiment of the present invention,
[0095] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0096] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An intelligent orthodontic decision-making aid method based on a large language model and knowledge graph, characterized in that: The following steps are involved: S1. Build an entity recognition model to obtain orthodontic-related expertise, clinical guidelines and patient orthodontic management record data, and identify knowledge entities in the data based on the entity recognition model; S2, extract the entity relationships between knowledge entities in the data and construct an orthodontic knowledge graph; S3, embedding the knowledge graph into the Llama model; S4. Design the prompt strategy of the Llama model; S5. Perform entity recognition on the patient's decision-making data based on the entity recognition model, extract entity relationships, and obtain the patient's ternary information group; S6, performing intention recognition on the patient's decision-making data to obtain the patient's intention information; S7. The patient's ternary information group and intention information are used as the input of the Llama model, and the orthodontic decision plan is obtained by combining the prompt strategy.
2. According to claim 1, an intelligent orthodontic decision-making aid method based on a large language model and a knowledge graph is characterized in that: S1 is specifically: S11. Build and train an entity recognition model based on Word2Vec-BiLSTM-CNN-CRF; S12, obtain orthodontic related professional knowledge, clinical guidelines and patient orthodontic treatment record data, and convert the sentences in the data into feature vectors based on the Word2Vec module of the entity recognition model; S13, inputting the feature vector into the entity recognition model, and the BiLSTM module extracts the hidden state based on the feature vector; S14, inputting the hidden state into a CNN module of an entity recognition model, and the CNN module extracting data features based on the hidden state; S15. Input the data features into the CRF module of the entity recognition model. The CRF module calculates the probability that the data features belong to different labels, and obtains the label sequence with the highest score as the output of the entity recognition model.
3. According to claim 1, an intelligent orthodontic decision-making aid method based on a large language model and knowledge graph is characterized in that: S2 is specifically as follows: construct predefined relationships between orthodontic knowledge entities, sort all knowledge entities identified in the patient's decision-making data according to the order of the original sentences, search for predefined relationships between all knowledge entities and other knowledge entities in order, obtain the ternary information group of the knowledge entity relationship, and fuse the knowledge entities based on their attributes.
4. The intelligent orthodontic decision-making assistance method based on a large language model and knowledge graph according to claim 3 is characterized in that: The fusion of knowledge entities based on their attributes is specifically as follows: the semantic similarity between the feature vectors of each knowledge entity is calculated, a semantic adjacency graph is constructed, knowledge entities whose semantic similarity exceeds a preset threshold are connected, and the connected components of the semantic adjacency graph are used as a fusion candidate set. For each connected component, the knowledge entity with the most connections with other knowledge entities is selected as the representative entity, the connection relationship of the remaining knowledge entities is transferred to the representative entity, and other knowledge entities are deleted to obtain the fused knowledge entity.
5. The intelligent orthodontic decision-making assistance method based on a large language model and knowledge graph according to claim 1 is characterized in that: S3 is specifically: S31, the ternary information group (h, r， t) is divided into two groups (h, r h ) and (r t , t), calculate the eigenvectors of (h, r) and (r, t) and S32, the feature vector v h and v t Stacked as a matrix A = [v h , v t ], perform convolution operation on matrix A to generate feature map V = [v1, v2, ..., v d ]: v i =f(ω·A i +b) In the formula, v i ∈V, f(·) represents the activation function, ω is the convolution kernel, A i is the i-th row of matrix A, and b is the bias parameter; S33, set c different convolution kernels to perform convolution operation on matrix A to generate feature map V 1 , V 2 , ..., V c , concatenate all feature maps to get the feature matrix V, set the weight matrix W, and calculate the features of the i-th dimension: e i =f(V i ·W i +b) Where V i is the i-th row of the feature matrix V, W i is the i-th row of the weight matrix W; S34, calculate the scoring function of the triple information group (h, r, t): f r (h,t)=concat(e1,e2,...,e d )·w In the formula, concat represents the connection function, w is the scoring weight, and the distributed representation of knowledge entities and relations is learned by maximizing the scores of correct triples in the orthodontic knowledge graph.
6. The intelligent orthodontic decision-making assistance method based on a large language model and knowledge graph according to claim 1 is characterized in that: S6 is specifically: Calculate the word vector of the decision-making data as the patient's intention information: TI = {TI1, TI2, ..., TI i ,…,OF k } Where TI i represents the importance of knowledge entity i, n i，j represents the number of occurrences of knowledge entity i in sentence j, ∑ k n k，j represents the total number of knowledge entities in sentence j, |D| is the total number of sentences, |j:t i ∈d j | represents the number of sentences containing knowledge entity i.
7. The intelligent orthodontic decision-making assistance method based on a large language model and knowledge graph according to claim 1 is characterized in that: S7 is specifically as follows: the patient's triple information and intention information are combined with the prompt strategy to convert the patient's decision data into input questions for the Llama model. The Llama model searches the professional knowledge base based on the input questions and generates an orthodontic decision plan.