Electronic medical record discharge nodule automatic generation method, system and device based on information enhancement and medium

Abstract

The invention discloses an information enhancement-based automatic generation method for a discharge nodule of an electronic medical record, which comprises the following steps of: respectively encoding information of three entities of diseases, symptoms and examination in the electronic medical record and a relation triple formed by the three entities to obtain an entity information embedded representation and a knowledge graph embedded representation; the method comprises the following steps of: establishing a relationship between entities and the relationship, embedding information contained in the entities and the relationship into an encoder, carrying out information fusion to enhance the capability of the model, and enhancing the generalization capability of model generation by utilizing a replication mechanism of a pointer generation network in consideration that the content of a discharge nodule is mostly from an original text.

Description

Technical Field

[0001] The present invention provides a method, system, device and medium for automatically generating an electronic medical record discharge summary based on information enhancement, and belongs to the technical field of electronic medical record processing. Background Art

[0002] With the increasing attention paid to health issues and the popularity of electronic medical records in the medical field, most hospitals will build electronic medical records for patients. These medical records cover information such as the patient's medical history, examination records, and examination results, and each patient may have more than one medical record. When doctors write discharge summaries for patients, they need to read all the electronic medical records of the patient during hospitalization. The complicated and redundant information in the medical records often takes up a lot of doctors' time. Automatic text summarization technology, also known as automatic abstracting, compresses a large amount of text into short sentences, retains the key information in the text, and solves the problem of redundant and complicated content in the original text. It can effectively reduce the burden of a large amount of information and increase the speed at which users obtain information. Generating discharge summaries through automatic abstracting can help doctors save a lot of manpower and material resources.

[0003] The application scope of automatic summarization is very wide, including news, public opinion analysis, opinion / sentiment summary, scientific paper summary and other fields. In the early stage, automatic summarization mainly used statistical knowledge to generate summaries for texts based on information such as word frequency or word distribution in the text. With the rapid development of artificial intelligence technology, there have been a large number of studies on automatic summarization based on deep learning, but they are mainly concentrated in the fields of economy and news. There are few related works in the medical field, and most of the work in the medical field is concentrated on medical entity relationship extraction, disease risk prediction, medical imaging report generation, auxiliary diagnosis, etc. Although there are some works on summarization research, for example, generating clinical impressions by summarizing radiology reports and generating medical histories by summarizing patients' health records, the research is mostly concentrated in the English medical field, and there are few studies in the Chinese field, especially on the task of generating discharge summaries.

[0004] The research on automatic generation of discharge summaries in electronic medical records has important theoretical value and broad application prospects. In terms of theory, automatic abstract generation is based on the methods and theories of disciplines such as machine learning, data mining, and natural language processing, which can promote the development and improvement of the above-mentioned disciplines. In terms of application, the automatic generation of discharge summaries in electronic medical records can quickly extract key information from a large number of electronic medical records, freeing doctors from heavy document work and focusing more on other clinical tasks for patients; secondly, the automation process reduces the errors that may occur when manually writing discharge summaries, thereby improving the accuracy and reliability of documents; in addition, the automatically generated discharge summaries maintain consistency in format and content, which helps to clearly and accurately transmit medical information between different medical institutions. This fast and accurate information extraction is essential to support clinical decision-making and rapid response in emergency situations. Summary of the invention

[0005] Purpose of the invention: In order to effectively utilize the entity relationship information in the electronic medical record text to improve the quality of text generation, the present invention proposes an automatic generation method for electronic medical record discharge summaries based on information enhancement.

[0006] Technical solution, in order to solve the above technical problems, the present invention proposes a method for automatically generating an electronic medical record discharge summary based on information enhancement, the method comprising the steps of:

[0007] S1. Use the T5 framework to process the input electronic medical record text and obtain the text vector representation h of contextual semantic information t ;

[0008] S2. Use the Bert encoder to process the disease, symptom and examination entity information in the input electronic medical record to obtain the entity vector representation h of the contextual semantic information d 、h s 、h e , and use the graph convolutional network to process the input relationship triples consisting of the three entities of disease, symptom and examination to obtain the knowledge graph embedding representation feature h g ;

[0009] S3. Use multi-granularity information fusion to represent the text vector h t , entity vector representation h d 、h s 、h e And the knowledge graph embedding representation feature h g Fusion is performed to obtain the fused feature representation h a ;

[0010] S4. Use the decoding method based on the pointer generation network to represent the fused feature h aDecode and obtain the final summary result, namely the discharge summary.

[0011] Furthermore, in step S1, the text vector representation h of the contextual semantic information t The calculation process is as follows:

[0012] (1.1) The text is segmented into a series of tokens, where the tokens are words, subwords, and special symbols. Special start and end tokens are added at the beginning and end of the token sequence after segmentation. <s>and the end < / s> token, used to mark the boundary of the sequence;

[0013] (1.2) Add a corresponding position code to each word, which will be input into the first layer of the Transformer model to provide position information;

[0014] (1-3) Each word obtains its corresponding embedding vector representation by searching the embedding matrix;

[0015] (1-4) The positional encoding is added to the embedded vector representation to obtain the input sequence, which is input into the Transformer model encoder module and encoded through a multi-layer encoder structure to generate the corresponding input representation, i.e., h t .

[0016] Furthermore, in step S2, the disease, symptom and examination entity information in the input electronic medical record is processed using the Bert encoder to obtain the entity vector representation h of the contextual semantic information. d 、h s 、h e , the specific method is as follows:

[0017] (2.1) Define the disease sequence set as D = {d1, d2, ..., d n }, where n is the length of the disease entity sequence; the symptom sequence set is defined as S = {s1, s2, ..., s m }, where m is the length of the symptom entity sequence; the inspection sequence set is defined as E = {e1, e2, ..., e k }, where k is the length of the check entity sequence; the original sentence sequence is defined as T = {t1, t2, ..., t l }, where l is the length of the original sentence sequence; obtain entity input representations of diseases, symptoms, and examinations to capture the meaning of the unit, the sentence to which the unit belongs, and the position of the unit in the sequence, the process is as follows;

[0018] ED(p i )=T(p i )+S(p i )+P(l)

[0019] Where p refers to D, S or E, ED(p i ) is the input representation of the i-th unit, T(p i ) is the word embedding, S(p i ) is the segment embedding, P(i) is the position embedding;

[0020] Denote the input of all units as ED(p i ) into a sequence E p , as shown below;

[0021] E p =[ED(p1),ED(p2),...,ED(p n )]

[0022] (2.2) Sequence E p The first encoding layer of the Transformer encoder is represented as

[0023]

[0024] Randomly initialize the learnable parameter matrix W i Q , W i K and W i V , the calculation process of the lth layer in Transformer is as follows;

[0025]

[0026] head i =Attention(Q i , K i , V i )

[0027]

[0028] MultiHead(Q,K,V)=Concat(head1,...,head h )W

[0029] FFN(x)=GeLu(xW1+b1)W2+b2

[0030]

[0031] in, is the input representation of the lth layer, W i Q , W i K and W i V is the learnable parameter matrix, head i is the output of the ith head, h is the number of attention heads, and Q represents the global query matrix. l K represents the global key matrix, which is composed of K i V represents the global value matrix, which is composed of V i Composition, d K is the dimension of the key vector K, and W is the sum of Concat(head1, ..., head h ) is the weight matrix obtained by linear transformation, W1, W2, b1, b2 are the parameters of the feedforward network, is the output representation of the lth layer;

[0032] (2.3) Repeat the above process, and after L layers, the final output is obtained The three entities of disease, symptom and examination are coded respectively, and the process is as follows:

[0033]

[0034]

[0035] Among them, h d is the disease entity hidden vector, h s is the hidden state vector of the symptom entity, h e To check the entity hidden state vector.

[0036] Furthermore, in step S2, the graph convolutional network is used to process the relationship triples consisting of the three entities of disease, symptom and examination in the input electronic medical record to obtain the knowledge graph embedding representation feature h g , the calculation process includes:

[0037] (1) Consider the knowledge graph consisting of three entities, disease, symptom, and examination, and their relationship triples as a graph The nodes represent the three entities of disease, symptom, and examination, and the edges represent the relationship between the three entities. (0) is the node feature matrix, is the node set, E is the edge set, H (0) The dimension is in, is the number of nodes, F in is the dimension of the input feature. The calculation process for each layer of graph convolution is as follows:

[0038]

[0039] Among them, H (l) is the output feature matrix of the lth layer, is the adjacency matrix A of the graph G plus the self-connection identity matrix I, is the diagonal matrix of the degree matrix, W (l) is the weight matrix of the lth layer;

[0040] (2) After two layers of graph convolution and ReLU activation function, the updated feature matrix of each node i is obtained

[0041] (3) Aggregate the node features of graph G into a graph-level vector representation h g , calculated as follows:

[0042]

[0043] Furthermore, in step S3, the multi-granularity information fusion calculation process includes:

[0044] (3.1) Add the disease entity latent vector and the symptom entity latent vector to get a new vector, namely:

[0045]

[0046] (3.2) Construct a fully connected network to process the concatenated vector. First, map the vector to a hidden layer, and then perform a linear transformation on the output of the hidden layer. The process is as follows;

[0047]

[0048] Among them, h (1) is the output of the hidden layer, h (2) is the output after linear transformation, is a parameter;

[0049] (3.3) According to the linear transformation output h (2) Get different weight sizes W f , and calculate the fusion h d ,h s The vector after represents fusion, and the process is as follows;

[0050] W f =sigmoid(h (2) )

[0051] fusion=W f *h d +(1-W f )*h s

[0052] (3.4) h d ,h s The above process of information fusion is defined as Fusion(h d ,h s ), and then fusion with h e Check the entity hidden vector fusion to get the result F after fusing the three entity information fusion ;

[0053] F fusion =Fusion(fusion,h e )

[0054] (3.5) The fused feature representation is used in the same way as the embedding representation feature h of the knowledge graph g The fusion is performed to obtain the sequence feature h with enhanced entity relationship information m , the process is as follows;

[0055] h m =Fusion(F fusion ,h t )

[0056] h m And the text vector representation h after the generative model encoding t Fusion, get the final sequence feature h a , as shown below;

[0057] h a =Fusion(h g ,h m ).

[0058] Furthermore, in step S4, the decoding method calculation process based on the pointer generation network includes:

[0059] (1) Calculate the probability P of each word or phrase appearing in the generated discharge summary gen (w), that is, the sequence feature h of the T5 model with enhanced entity relationship information a The decoding generation result produces a generated probability distribution representation on the target vocabulary, and the initial hidden state of the decoder is set to h a , at decoding time step o, the decoder receives the word embedding of the previous word and has a decoder hidden state h o , the calculation process is as follows:

[0060] P gen (w) = softmax(W3h o +b)

[0061] Among them, W3,b are model parameters;

[0062] (2) Calculate the probability of the model copying each specific word or phrase from the source sequence to the target text when generating the discharge summary, that is, determine the attention distribution α of the T5 model on the source sequence during the decoding process o , and based on this, the probability distribution of the replication channel is generated. The process is as follows:

[0063]

[0064] Among them, w is the source sequence, w i is the lth word in the source sequence, α o，i is the attention weight of the i-th word in the o-th step; α o The attention weight α of the i-th word in step o is o，i composition;

[0065] (3) Based on the decoder hidden state h o And the source sequence context vector c obtained by the attention mechanism o , the gate control unit g is obtained by mapping a single-layer network with the sigmoid function. The calculation process is as follows

[0066] g = sigmoid(W g [h o , c o ]+b g )

[0067] Among them, W g and b g are the weights and bias values ​​of the fully connected layer respectively;

[0068] (4) Use the gating unit g to adjust the balance control between generation and replication in the target sequence. The output logits is a convex combination of the generation distribution and the replication distribution. The calculation process is as follows:

[0069] P(w)=g*P copy (w)+(1-g)*P gen (w).

[0070] In addition, the present invention proposes a computer device for automatically generating electronic medical record discharge summaries based on information enhancement, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods when executing the computer program.

[0071] In addition, the present invention provides a storage medium having a computer program stored thereon, wherein the computer program implements the steps of any one of the methods when executed by a processor.

[0072] In addition, the present invention also proposes an automatic generation system of electronic medical record discharge summary based on information enhancement, which includes the following modules:

[0073] The text vector acquisition module uses the T5 framework to process the input electronic medical record text and obtain the text vector representation h of the contextual semantic information. t ;

[0074] The entity relationship vector representation module uses the Bert encoder to process the disease, symptom and examination entity information in the input electronic medical record to obtain the entity vector representation h of the contextual semantic information. d 、h s 、h e , and use the graph convolutional network to process the input relationship triples consisting of the three entities of disease, symptom and examination to obtain the knowledge graph embedding representation feature h g ;

[0075] The feature fusion module uses multi-granularity information fusion to represent the text vector h t , entity vector representation h d 、h s 、h e And the knowledge graph embedding representation feature h g Fusion is performed to obtain the fused feature representation h a ;

[0076] The summary generation module uses a decoding method based on the pointer generation network to represent the fused feature h a Decode and obtain the final summary result, namely the discharge summary.

[0077] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0078] (1) The basic generative model is combined with key entity information in the electronic medical record, so that the model can more accurately reflect the patient's specific situation when generating summaries, effectively improving the relevance of the summary and the accuracy of the information.

[0079] (2) It not only integrates entity information, but also integrates the relationships between entities, and further improves the performance of the model through knowledge graph embedding technology. The use of knowledge graphs enables the model to effectively understand and apply complex relationships in medical records, such as the relationship between diseases and symptoms. This deep understanding can generate more accurate and comprehensive summaries.

[0080] (3) In addition, the decoding method based on the pointer network allows the key information in the original text to be directly quoted when generating the summary, which is extremely important for ensuring accuracy and professionalism in the medical field, and also enhances the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 A flowchart of the method is shown in FIG. DETAILED DESCRIPTION

[0082] The following experimental and result analysis process is used to verify the automatic generation model of electronic medical record discharge summaries based on information enhancement.

[0083] Example

[0084] 1. Dataset

[0085] The experiment uses the EMRDS dataset, which is an electronic medical record discharge summary dataset with entity and entity relationship annotations. It is constructed from electronic medical records of 17 diseases such as diabetes, cardiovascular, stroke, and lung cancer in a tertiary hospital. The statistical results of the dataset are shown in Table 1.

[0086] Table 1 Statistical results of EMRDS dataset

[0087]

[0088] 2. Hyperparameter Settings

[0089] The experimental environment configured during the experiment, the model preset hyperparameters and other settings are described in detail.

[0090] As shown in the table.

[0091] Table 2 Experimental environment and hyperparameter settings

[0092]

[0093]

[0094] 3. Evaluation Method

[0095] Evaluation metrics provide a quantifiable method to measure the similarity between the generated summary and the reference summary, thereby evaluating the performance of the model. Automatic evaluation provides an objective evaluation method, avoids the influence of subjective bias, and helps ensure the fairness and accuracy of the evaluation results. ROUGE and BLEU are two commonly used evaluation metrics that can effectively quantify the similarity between the summary generated by the system and the reference summary, and are widely used in text summarization tasks.

[0096] ROUGE focuses on the degree of overlap between the generated summary and the reference summary. For example, ROUGE-1 measures the overlap between the two at the level of a single word, ROUGE-2 measures the overlap between the two at the level of two consecutive words, and ROUGE-L measures the similarity between the two based on the longest common subsequence between the two, reflecting the matching of the overall structure of the sentence. BLEU evaluates the quality of the generated summary by comparing the overlap of n-grams.

[0097] 4. Comparison Models

[0098] To fully verify the effect of the model, this method compares two unsupervised models LEAD-3 and TextRank, four generative pre-trained language models T5 model, Pegasus, CPT model and T5-pegasus model, and two large language models ChatGPT and ChatGLM2-6B.

[0099] 5. Results Analysis

[0100] The experimental results are compared with the introduced baseline model and are shown in Table 1, where R-1, R-2 and RL are represented by Rouge-1, Rouge-2 and Rouge-L indicators respectively, and R-avg is the average of these three indicators.

[0101] Table 1 Experimental results of DSGE model framework on EMRDS dataset (%)

[0102]

[0103]

[0104] The results of comparing the unsupervised model and the deep learning model show that TextRank and LEAD-3, as traditional unsupervised learning methods, have limitations and perform poorly in the automatic text summarization task. They cannot effectively understand long medical records, professional terms, and the coherence of the condition, cannot effectively learn the narrative logic unique to the medical record from the context, and cannot accurately determine which information is important for writing the discharge summary. Therefore, the scores on the evaluation indicators are relatively low.

[0105] Compared with unsupervised models, the performance of pre-trained generative models in the task of generating electronic medical record summaries has been significantly improved. Pre-trained generative models can comprehensively use the general rules and special domain knowledge learned from language models to generate summaries that are both accurate and meet medical professional requirements. In addition, the results of the ChatGPT and ChatGLM models are not good, and they do not even exceed the results of the T5-Pegasus model. The possible reason is that they only use the ICL method to utilize large models, while the T5-Pegasus model uses a large amount of relevant training data to fine-tune and optimize the pre-trained model.

[0106] Compared with the pre-trained generative model, this method has improved the performance on the test set when using only entity information. It can combine the basic generative model with the key entity information in the electronic medical record to more accurately reflect the patient's specific situation when generating the summary. The decoding method based on the pointer network allows the key information in the original text to be directly quoted when generating the summary, which is extremely important to ensure the accuracy and professionalism in the medical field, and also enhances the generalization ability of the model. This method performs best among all models when using entity and relationship information at the same time. The model not only integrates entity information, but also integrates the relationship between entities. By using the knowledge graph, the model can effectively understand and use the complex relationships in the medical record, generate more accurate and comprehensive summaries, and lead the scores in the evaluation indicators. It shows the great potential and advantages of combining deep learning and knowledge graph technology in the automatic generation of discharge summaries of electronic medical records.

[0107] 6. Case Analysis

[0108] In order to fully illustrate the effectiveness of this method and increase the transparency and credibility of the research, a case analysis was conducted on a group of summary pairs of the first course of illness and hospitalization, as shown in Table 4. The table shows the generation results of the baseline model T5 and the generation results using only entity information enhancement and entity and relationship information enhancement.

[0109] In the results, the T5 model directly extracts information from the original text and generates a simplified summary. It does not pay attention to key entity elements in the original text, such as "dizziness, numbness in the hands and corners of the mouth, diabetes, and cerebral infarction", but relies more on direct restatement and simplification of the original text content. When using entity information enhancement, not only the above-mentioned key entity information is identified, but also the key time information in the original text, "worsened for 3 days", is retained. This relies on the copying mechanism used by the model to better fit the original text. In addition, when using entity and relationship information enhancement at the same time, not only the entity is focused on, but also the knowledge graph is used to understand the relationship between these entities and the deep knowledge behind them. Therefore, it can further identify the "numbness in the hands and corners of the mouth" caused by the disease, optimize the effect of summary generation, and provide richer and more accurate summary generation results.

[0110] Table 4 Information enhancement case analysis

[0111]

[0112] < / s>

Claims

1. A method for automatically generating an electronic medical record discharge summary based on information enhancement, characterized in that: The method comprises the steps of: S1. Use the T5 framework to process the input electronic medical record text and obtain the text vector representation h of contextual semantic information t ; S2. Use the Bert encoder to process the disease, symptom and examination entity information in the input electronic medical record to obtain the entity vector representation h of the contextual semantic information d 、h s 、h e , and use the graph convolutional network to process the input relationship triples consisting of the three entities of disease, symptom and examination to obtain the knowledge graph embedding representation feature h g ; S3. Use multi-granularity information fusion to represent the text vector h t , entity vector representation h d 、h s 、h e And the knowledge graph embedding representation feature h g Fusion is performed to obtain the fused feature representation h a ; S4. Use the decoding method based on the pointer generation network to represent the fused feature h a Decode and obtain the final summary result, namely the discharge summary.

2. According to the method for automatically generating an electronic medical record discharge summary based on information enhancement according to claim 1, it is characterized in that: In step S1, the text vector representation h of contextual semantic information t The calculation process is as follows: (1.1) The text is segmented into a series of tokens, where the tokens are words, subwords, and special symbols. Special start and end tokens are added at the beginning and end of the token sequence after segmentation. <s> and the end < / s> token, used to mark the boundary of the sequence;< / s> <s> (1.2) Add a corresponding position code to each word, which will be input into the first layer of the Transformer model to provide position information; (1.3) Each word obtains its corresponding embedding vector representation by searching the embedding matrix; (1.4) The positional encoding is added to the embedding vector representation to obtain the input sequence, which is input into the Transformer model encoder module and encoded through a multi-layer encoder structure to generate the corresponding input representation, i.e., h t .

3. The method for automatically generating an electronic medical record discharge summary based on information enhancement according to claim 1, characterized in that: In step S2, the Bert encoder is used to process the disease, symptom and examination entity information in the input electronic medical record to obtain the entity vector representation h of the contextual semantic information. d 、h s 、h e , the specific method is as follows: (2.1) Define the disease sequence set as D = {d1, d2, ..., d n }, where n is the length of the disease entity sequence; the symptom sequence set is defined as S = {s1, s2, ..., s m }, where m is the length of the symptom entity sequence; the inspection sequence set is defined as E = {e1, e2, ..., e k }, where k is the length of the check entity sequence; the original sentence sequence is defined as T = {t1, t2, ..., t l }, where l is the length of the original sentence sequence; obtain entity input representations of diseases, symptoms, and examinations to capture the meaning of the unit, the sentence to which the unit belongs, and the position of the unit in the sequence, the process is as follows; ED(p i )=T(p i )+S(p i )+P(i) Where p refers to D, S or E, ED(p i ) is the input representation of the i-th unit, T(p i ) is the word embedding, S(p i ) is the segment embedding, P(i) is the position embedding; Denote the input of all units as ED(p i ) into a sequence E p , as shown below; AND p =[ED(p1),ED(p2),…,ED(p n )] (2.2) Sequence E p The first encoding layer of the Transformer encoder is represented as Randomly initialize the learnable parameter matrix W i Q , W i K and W i V , the calculation process of the lth layer in Transformer is as follows; head i =Attention(Q l ,K i ,V i ) MultiHead(Q,K,V)=Concat(head1,...,head h )W FFN(x)=GeLu(xW1+b1)W2+b2 in, is the input representation of the lth layer, W i Q , W i K and W i V is the learnable parameter matrix, head i is the output of the ith head, h is the number of attention heads, and Q represents the global query matrix. i K represents the global key matrix, which is composed of K i V represents the global value matrix, which is composed of V i Composition, d K is the dimension of the key vector K, and W is the sum of Concat(head1, ..., head h ) is the weight matrix obtained by linear transformation, W1, W2, b1, b2 are the parameters of the feedforward network, is the output representation of the lth layer; (2.3) Repeat the above process, and after L layers, the final output is obtained The three entities of disease, symptom and examination are coded respectively, and the process is as follows: Among them, h d is the disease entity hidden state vector, h s is the hidden state vector of the symptom entity, h e To check the entity hidden vector.

4. According to claim 1, a method for automatically generating an electronic medical record discharge summary based on information enhancement, characterized in that: In step S2, the graph convolutional network is used to process the relationship triples consisting of the three entities of disease, symptom and examination in the input electronic medical record to obtain the knowledge graph embedding representation feature h g , the calculation process includes: (1) Consider the knowledge graph consisting of three entities, disease, symptom, and examination, and their relationship triples as a graph The nodes represent the three entities of disease, symptom and examination, and the edges represent the relationship between the three entities. H(0) is the node feature matrix. is the node set, E is the edge set, and the dimension of H(0) is in, is the number of nodes, F in is the dimension of the input feature. The calculation process for each layer of graph convolution is as follows: Among them, H(l) is the output feature matrix of the lth layer, is the adjacency matrix A of the graph G plus the self-connection identity matrix I, is the diagonal matrix of the degree matrix, W( l ) is the weight matrix of the lth layer; (2) After two layers of graph convolution and ReLU activation function, the updated feature matrix of each node i is obtained (3) Aggregate the node features of graph G into a graph-level vector representation h g , calculated as follows:

5. The method for automatically generating an electronic medical record discharge summary based on information enhancement according to claim 2, characterized in that: In step S3, the multi-granularity information fusion calculation process includes: (3.1) Add the disease entity latent vector and the symptom entity latent vector to get a new vector, namely: (3.2) Construct a fully connected network to process the concatenated vector. First, map the vector to a hidden layer, and then perform a linear transformation on the output of the hidden layer. The process is as follows; Among them, h(1) is the output of the hidden layer, h(2) is the output after linear transformation, is a parameter; (3.3) According to the linear transformation output h (2) Get different weight sizes W f , and calculate the fusion h d ,h s The vector after represents fusion, and the process is as follows; W f =sigmoid(h (2) ) fusion=W f *h d +(1-W f )*h s (3.4) h d ,h s The above process of information fusion is defined as Fusion(h d ,h s ), and then fusion with h e Check the entity hidden vector fusion to get the result F after fusing the three entity information fusion ; F fusion =Fusion(fusion,h e ) (3.5) The fused feature representation is used in the same way as the embedding representation feature h of the knowledge graph g The fusion is performed to obtain the sequence feature h with enhanced entity relationship information m , the process is as follows; h m =Fusion(F fusion ,h t ) h m And the text vector representation h after the generative model encoding t Fusion, get the final sequence feature h a , as shown below; h a =Fusion(h g ,h m )。 6. According to claim 1, a method for automatically generating an electronic medical record discharge summary based on information enhancement, characterized in that: In step S4, the decoding method calculation process based on the pointer generation network includes: (4.1) Calculate the probability P of each word or phrase appearing in the generated discharge summary gen (w), that is, the sequence feature h of the T5 model with enhanced entity relationship information a The decoding generation result produces a generated probability distribution representation on the target vocabulary, and the initial hidden state of the decoder is set to h a , at decoding time step o, the decoder receives the word embedding of the previous word and has a decoder hidden state h o , the calculation process is as follows: P gen (w)=softmax(W3h o +b) Among them, W3,b are model parameters; (4.2) Calculate the probability of the model copying each specific word or phrase from the source sequence to the target text when generating the discharge summary, that is, determine the attention distribution α of the T5 model on the source sequence during the decoding process o , and based on this, the probability distribution of the replication channel is generated. The process is as follows: Among them, w is the source sequence, w l is the lth word in the source sequence, α o，i is the attention weight of the i-th word in the o-th step; α o The attention weight α of the lth word in the oth step o，i composition; (4.3) Based on the decoder hidden state h o And the source sequence context vector c obtained by the attention mechanism o , the gate control unit g is obtained by mapping a single-layer network with the sigmoid function. The calculation process is as follows g=sigmoid(W g [h o ,c o ]+b g ) Among them, W g and b g are the weights and bias values ​​of the fully connected layer respectively; (4.4) The gated unit g is used to adjust the balance control between generation and replication in the target sequence. The output logits is a convex combination of the generation distribution and the replication distribution. The calculation process is as follows: P(w)=g*P copy (w)+(1-g)*P gen (w)。 7. A computer device for automatically generating an electronic medical record discharge summary based on information enhancement, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. An automatic generation system for electronic medical record discharge summaries based on information enhancement, characterized in that: The system includes the following modules: The text vector acquisition module uses the T5 framework to process the input electronic medical record text and obtain the text vector representation h of the contextual semantic information. t ; The entity relationship vector representation module uses the Bert encoder to process the disease, symptom and examination entity information in the input electronic medical record to obtain the entity vector representation h of the contextual semantic information. d 、h s 、h e , and use the graph convolutional network to process the input relationship triples consisting of the three entities of disease, symptom and examination to obtain the knowledge graph embedding representation feature h g ; The feature fusion module uses multi-granularity information fusion to represent the text vector h t , entity vector representation h d 、h s 、h e And the knowledge graph embedding representation feature h g Fusion is performed to obtain the fused feature representation h a ; The summary generation module uses a decoding method based on the pointer generation network to represent the fused feature h a Decode and obtain the final summary result, namely the discharge summary. < / s>

Images