Automatic explainable disease automatic diagnosis device based on knowledge graph enhancement

Abstract

The present disclosure relates to an explainable disease automatic diagnosis device based on knowledge graph enhancement, which comprises: a data acquisition module for acquiring the symptoms of a patient and the supplementary description of each symptom by a doctor; a data enhancement module for calculating the enhanced representation of each symptom in combination with the supplementary description; a path inference module for performing path inference in a knowledge graph based on the enhanced representation to obtain a path set corresponding to each symptom; wherein the entities in the knowledge graph include diseases, symptoms, operations, drugs, descriptions, related diseases and related symptoms, the relationships in the knowledge graph include being related to, having operation history and disease performance, and the paths in the path set are paths with disease entities as the terminal; a symptom inference module for calculating the disease occurrence probability and / or reward corresponding to all path terminals to obtain the diagnosis result of the symptom. The present application realizes the two purposes of disease diagnosis and cause explanation.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and more specifically to an automated, interpretable disease diagnosis device based on knowledge graph enhancement. Background Technology

[0002] Traditional disease diagnosis relies heavily on doctors' experience, and diagnostic results are influenced by environmental and human factors. Furthermore, the uneven distribution of medical resources in my country leaves many rural residents without access to sufficient and high-quality healthcare. To address this issue, the appropriate application of machine learning technology to disease diagnosis is essential. AI-based automated disease diagnosis methods can help patients detect diseases early and effectively assist doctors in diagnosis, reducing the probability of misdiagnosis and missed diagnosis.

[0003] For the problem of automatic disease diagnosis, there are two main approaches: one is to treat it as a disease classification task. A diagnostic model is trained, taking the patient's symptoms and other relevant information as input, calculating the probability of various diseases, and finally selecting the disease with the highest probability as the diagnosis result. Based on this idea, CNNs, RNNs, and other advanced technologies are currently used.

[0004] Various methods based on structures such as Transformer have been proposed. For example, the literature (Narges Razavian. "Deep ehr: Chronic disease prediction using medical notes." Machine Learning for Healthcare Conference. PMLR)

[0005] The authors in 2018 compared various methods, including CNN, BiLSTM, and a combination of CNN and LSTM, using these models to extract features from the input information, then inputting them into a linear layer for classification, ultimately achieving diagnosis. Since the process of disease diagnosis in reality can be viewed as a decision-making process based on multiple pieces of information, another approach is to transform the diagnosis into a Markov decision process. This approach is mostly based on reinforcement learning, training an intelligent agent to solve the Markov decision problem (Liu, Siqi, Kee Yuan Ngiam, and Mengling Feng. "Deepreinforcement learning for clinical decision support: a brief survey." arXivpreprint arXiv:1907.09475 (2019).). The literature (Wei, Zhongyu, et al. "Task-oriented dialogue system for automatic diagnosis." Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2018.) proposed a framework for an automatic diagnostic dialogue system based on reinforcement learning, which achieves disease diagnosis by analyzing dialogue with patients. Xu et al. (Xu, Lin, et al. "End-to-end knowledge-routed relational dialogue system for automatic diagnosis." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. No. 01. 2019.) improved upon this framework and proposed a method called KR-DQN, which can seamlessly integrate rich prior medical knowledge into the diagnostic process. Their training data was collected from the internet and simplified to a certain extent.

[0006] However, existing methods have the following problems:

[0007] 1. In methods that treat automatic disease diagnosis as a classification problem, a large amount of high-quality training data is required. Even for the same disease, changes in its description can affect the classification results. Furthermore, the accuracy of classification-based methods decreases significantly as the number of disease categories increases. Most importantly, due to the "black box" nature of deep learning models, this invention struggles to explain the behavior of the classification model, making it difficult to provide a reasonable explanation for the final diagnostic results. This undermines confidence in the model and severely limits the practical application of classification-based automatic diagnostic models in the medical field.

[0008] 2. Another type of dialogue-based automated diagnosis method also has many problems. In existing dialogue-based automated diagnosis methods, the model extracts multiple symptom data pairs through multiple rounds of interaction with the patient, ultimately completing the disease diagnosis. The model always tends to obtain more information, thus generating a large number of meaningless dialogue rounds during the diagnosis process, reducing the user experience. At the same time, when the amount of symptom data to be extracted in the dataset is small, its diagnostic accuracy drops significantly. In the dataset used in one example, most diseases present with 1 to 3 symptoms, such as... Figure 1 As shown, approximately 75% of diseases present with a maximum of three symptoms. In such cases, dialogue-based automated diagnostic methods are not very effective.

[0009] 3. Existing methods still rely solely on labeled data for training, incorporating little or no prior medical knowledge. Integrating specialized medical knowledge into the automated disease diagnosis process would significantly improve accuracy. Furthermore, the quality of automatically generated dialogues in dialogue-based diagnostic models limits subsequent disease diagnosis tasks. The dialogue datasets used in existing methods are either machine-generated or heavily simplified, making it difficult to generalize models to real-world tasks and severely restricting the application of these methods. Summary of the Invention

[0010] To address the aforementioned issues, this invention discloses an automatic and interpretable disease diagnosis device based on knowledge graph enhancement. The problem is defined as a Markov decision problem on a medical knowledge graph. The aim is to construct an automatic diagnosis model based on reinforcement learning, inputting the patient's symptoms and other supplementary information, integrating prior knowledge from the medical knowledge graph, and finally diagnosing the disease and returning an interpretable decision path, thereby achieving the dual objectives of disease diagnosis and cause explanation.

[0011] The technical content of this invention includes:

[0012] An automated diagnostic device for interpretable diseases based on knowledge graph enhancement, the device comprising:

[0013] The data acquisition module is used to acquire the patient's symptoms and the doctor's supplementary explanations for each symptom;

[0014] The data augmentation module is used to calculate the augmented representation of each symptom in conjunction with the supplementary description;

[0015] The path inference module is used to perform path inference based on the enhanced representation in the knowledge graph to obtain a set of paths corresponding to each symptom; wherein, the entities in the knowledge graph include: disease, symptom, surgery, drug, description, related disease and related symptom, the relationships in the knowledge graph include: related to, surgical history and disease manifestation, and the paths in the path set are paths with disease entities as the endpoint;

[0016] The symptom inference module is used to calculate the probability of disease occurrence and / or reward corresponding to all path endpoints in order to obtain the diagnostic result of the symptom.

[0017] Furthermore, the calculation of the enhancement representation of each symptom, in conjunction with the aforementioned supplementary explanation, includes:

[0018] Using a Word2Vec model pre-trained with an embedding layer or a medical dictionary, the vector representation V of symptom m is computed respectively. m Supplementary explanation t corresponding to the symptom m i vector representation

[0019] Based on the vector representation V m and the vector representation The enhancement of the symptom m is represented by V′. m .

[0020] Further, the path inference process performed on the enhanced representation in the knowledge graph to obtain a set of paths corresponding to each symptom includes:

[0021] In the knowledge graph, find the symptom entity e corresponding to the symptom. m And the enhancement is represented as the symptom entity e m Attributes;

[0022] Based on the current path endpoint e t The historical information h_t of the generated path and the symptom entity e m The attribute is used to calculate the state s of the current path endpoint. t Where t represents the number of inference steps;

[0023] The state s is generated based on the reinforcement learning agent. t Full motion space A t The reinforcement learning agent comprises: a two-layer fully connected neural network, a numerical network, and a policy network.

[0024] For the full motion space A t Pruning is performed to obtain the pruning motion space.

[0025] For the space of pruning action The actions in the video are scored using cosine similarity f. p Select the best action and execute it to obtain node e. t+1 And add the node to the existing path.

[0026] After reaching the maximum number of steps T, based on the pruning action space The set of paths corresponding to the symptoms is obtained.

[0027] Furthermore, the steps of training the reinforcement learning agent include:

[0028] Construct the training set;

[0029] The state s is learned using the aforementioned two-layer fully connected neural network. t The representation of x t ;

[0030] Based on the representation x t The policy network calculates the next action a. t ;

[0031] Using a numerical network based on the representation x t Let's learn about the numerical value v;

[0032] The policy network optimizes its parameters based on the value v;

[0033] The numerical network updates its parameters based on the training data.

[0034] Furthermore, the entire action space A t Pruning is performed to obtain the pruning motion space. include: Where r is the relation, e is the entity, m is the symptom, and A is the relation. t For the full motion space, f p The cosine similarity score is given between the symptom entity m and the current entity.

[0035] Furthermore, the cosine similarity score Among them, V m The vector representation of the symptoms, V′ m The enhancement of the symptoms is indicated by r, which represents the relationship, and e represents the relationship. k Let D represent the set of disease entities.

[0036] Furthermore, the knowledge graph is constructed based on the TransE learning method.

[0037] Further, the probability of disease occurrence and / or reward corresponding to all path endpoints are calculated to obtain a diagnostic result for the symptom, including:

[0038] Based on actual needs, one or more of the disease occurrence probability or reward can be selected as diagnostic criteria.

[0039] If the probability of disease occurrence is selected, the number of times the disease occurs corresponding to all path endpoints is counted, and the diagnosis result of the symptom is obtained based on the probability of occurrence of each disease.

[0040] If the reward is selected, the enhanced representations of each symptom are fused to obtain the fused symptom embedding. And by calculating the reward at the end of all paths. To obtain a diagnostic result for the described symptoms; wherein, For the path endpoint e t The representation of f r V′ m and The dot product;

[0041] If you choose disease occurrence probability and reward, then combine the occurrence probability of each disease with the reward R. t The diagnosis of the symptoms was obtained.

[0042] Compared with the prior art, the positive effects of the present invention are as follows:

[0043] 1. In this paper, the automatic disease diagnosis task is defined as a decision problem on a disease knowledge graph, and intelligent agents are used to explicitly reason about knowledge to make decisions.

[0044] 2. This invention proposes a path reasoning method based on reinforcement learning, which features soft reward and action pruning strategies for disease diagnosis.

[0045] 3. In order to incorporate more information into the knowledge graph, this invention proposes an enhanced embedding method.

[0046] 4. Experimental results show that the model of this invention outperforms all baseline models in terms of accuracy and interoperability. A comparison of the performance of the proposed model method with traditional methods is presented. Experimental verification uses the same validation data, and parameter tuning is performed on multiple algorithm models to obtain the best performance under the current model. Attached Figure Description

[0047] Figure 1 A statistical chart of the number of symptoms for each disease.

[0048] Figure 2Overall structure of the model.

[0049] Figure 3 A schematic diagram of a reinforcement learning module.

[0050] Figure 4 Block diagram of an automated, interpretable disease diagnosis device based on knowledge graph enhancement.

[0051] Figure 5 Top-n accuracy of each model.

[0052] Figure 6 Experimental results of different enhancement embedding methods. Detailed Implementation

[0053] The technical solutions of the present invention will now be clearly and completely described in conjunction with the embodiments and accompanying drawings. It should be understood that the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] I. Model Training.

[0055] The training process of this invention includes data analysis and processing, knowledge graph representation learning, reinforcement learning module construction and training, and disease reasoning. The dataset for this method comes from real electronic medical records. After preprocessing and feature extraction, triples containing various entities such as disease, symptoms, drugs, and surgeries, and the relationships between these entities are obtained. Negative sampling is used to construct negative examples of the triples, and the TransE method is used to learn the representations of entities and relationships. A reinforcement learning module is then constructed to train the decision-making agent. The agent starts from multiple symptoms and ultimately returns multiple paths from symptoms to diseases. In the disease reasoning part, a scoring function is designed to evaluate the multiple returned results and select the best result as the final diagnosis. The final structure of the model is as follows: Figure 2 As shown, structured data is extracted from EMR to construct a disease knowledge graph. Unstructured text data is extracted to embed into the KG. The model uses an Actor-Critic-based reinforcement learning approach for path inference, ultimately returning the path from symptoms to disease.

[0056] 1: Data Analysis and Processing

[0057] This method first performs data processing and analysis. After reviewing relevant literature and consulting with domain experts, the entity types required for the task were defined, including "disease," "symptom," "surgery," "medication," "description," "related diseases," and "related symptoms." Relationships were defined for these entities, such as "related to," "surgical history," and "disease manifestation." For text data, this invention uses Jieba segmentation and stop word removal techniques for initial cleaning, and the TF-IDF algorithm is used to extract feature words from the patient's chief complaint. In addition to entity and relation information, the dataset also includes supplementary descriptions of the patient's symptoms provided by the doctor during diagnosis, referred to as "attributes." During the data processing stage, this attribute information is saved in a shorter text format.

[0058] 2: Knowledge graph-enhanced representation learning.

[0059] The purpose of this method is to build a knowledge graph using existing entities and relations, and to learn representations for these entities and relations. The simplest approach is to label each entity using one-hot vectors, but this results in an extremely sparse vector space, which is detrimental to subsequent computation. TransE is a simple and effective representation learning method. For an event triple (head entity h, relation r, tail entity t), this invention first uses an embedding layer to obtain the vector representation of the entity and defines several vectors to represent the relations defined above. The training objective of this invention is to ensure that the embedded vectors satisfy the condition: h + r ≈ t. In the implementation process, cosine similarity is used to evaluate the similarity between h + r and t. Most entities in a knowledge graph are unlabeled; from a computational perspective, it is impossible to label all pairs as negative. Therefore, negative sampling techniques are used to ensure accuracy while also accelerating model training. Simply put, given a head entity h, negative sampling techniques randomly select combinations of entities and relations that have not appeared in the triple records from a large number of relations and entities, using these as negative examples for training. When the model converges, in positive examples, the cosine similarity between h+r and t is as large as possible, while in negative examples the opposite is true.

[0060] Building upon TransE, this method aims to incorporate more information into the knowledge graph to improve the performance of downstream tasks. Therefore, an augmented representation learning method is proposed, which transforms the "attributes" stored during data processing into vectors and embeds them into entity representations. Two methods for obtaining attribute vectors were explored: directly obtaining vectors using an embedding layer, and obtaining vectors using a Word2Vec model pre-trained with a medical dictionary. The attribute vector is denoted as V. t The symptom entity vector is denoted as V. m Then the augmentation vector representation V′ m The calculation formula is as follows:

[0061] 3: Construction and training of reinforcement learning modules.

[0062] After completing data analysis and knowledge graph representation learning, the next goal is to implement disease diagnosis using reinforcement learning methods. The reinforcement learning module uses an Actor-Critic implementation. This module uses a two-layer fully connected neural network to learn the state s. t The representation of x t A policy network π(A) t A numerical network is based on the representation x t Let's learn the numerical value v. The structure of the reinforcement learning module is defined as follows:

[0063] x t =σ(σ(s) t W1+b1)W2+b2)

[0064] π(A t ) = softmax((x t W p +b P )ΘA t )

[0065] v t =x t W v +b v

[0066] Here, σ is a nonlinear activation function, and the present invention uses a corrected linear unit (ReLU), Θ represents the Hadamard product, {W1, W2, W... p W v ,b1,b2,b p ,b v} represents the parameters and biases in a neural network. Its network structure is as follows: Figure 3 As shown.

[0067] When the knowledge graph is large, each node may be connected to a large number of nodes, resulting in a large number of candidate options during the transition steps. To alleviate the accuracy degradation caused by too many candidate options, this method proposes an efficient pruning strategy to control the number of candidate options. The pruning strategy can be described as follows:

[0068]

[0069] Where r is the relation, e is the entity, m is the symptom, and A is the relation. t For the full motion space, f p The cosine similarity score between symptom entity m and the current entity is calculated using the following formula:

[0070]

[0071] Where m′ is the enhanced representation of the symptom entity, D is the disease set, and e k This refers to the k-th entity in the path.

[0072] A soft reward strategy was also designed to encourage the intelligent agent to fully utilize the heterogeneity of the knowledge graph to explore more possible nodes. The soft reward strategy is described as follows:

[0073]

[0074] Where f r (m′,e t Calculate m′ and e t The dot product.

[0075] 4: Disease Reasoning

[0076] Given a number of symptoms (m1,…,m) n The goal of disease reasoning is to find the most likely set of diseases {d} and return a reasoning path. A simple approach is to select the n highest-scoring paths generated by the policy network π as the result, but this method cannot guarantee path diversity because the policy network π tends to select the same high-scoring paths. Therefore, this invention proposes a search method guided by action probability and reward to explore candidate paths and possible diseases for each symptom. The disease reasoning algorithm can be described as follows:

[0077] 1) Given symptom m, maximum number of steps T, policy network And the predefined sample size {K1,K2,…,K} for each step T}

[0078] 2) In the initial state, there is only the starting point m on the path, the probability set Q is initialized to {1}, and the reward R is initialized to 0.

[0079] 3) Obtain the current node's state s t The full action space is generated based on the reinforcement learning agent, and then a pruning strategy is used to obtain the pruned action space. The size of the action after pruning is limited to K. t .

[0080] 4) Add the current node to the path. For each action... Calculate the probability q and reward r, and store the results in Q and R, respectively. The probability calculation is based on the frequency of the action, while the reward calculation follows the soft reward strategy mentioned above.

[0081] 5) Repeat steps 3) and 4) T times to obtain multiple paths of length T. Save the paths that terminate at the disease. The path information includes the corresponding probability and score.

[0082] The disease inference algorithm takes a given symptom m, a policy network π, a maximum number of hops T, and a predefined sample size {K1, K2, ..., K} for each step. T Given the disease set {d} as input, return the corresponding disease set {d} and T-hop path, and calculate the probability Q of each path. T Score S T The disease reasoning algorithm takes a single symptom as input and returns an interpretable path and diagnosis. For cases with multiple symptoms, the algorithm inputs each symptom individually and then selects the best result from all returned results based on the fused symptom.

[0083] II. Automated Disease Diagnosis Test.

[0084] The present invention provides an interpretable automated diagnostic device for diseases, such as Figure 4 As shown, it includes a data acquisition module 410, a data enhancement module 420, a path inference module 430, and a symptom inference module 440.

[0085] The data acquisition module 410 is used to acquire the patient's symptoms and the doctor's supplementary explanations for each symptom;

[0086] Data augmentation module 420 is used to calculate the augmented representation of each symptom in conjunction with the aforementioned supplementary description;

[0087] The path inference module 430 is used to perform path inference based on the enhanced representation in the knowledge graph to obtain a set of paths corresponding to each symptom; wherein, the entities in the knowledge graph include: disease, symptom, surgery, drug, description, related disease and related symptom, the relations in the knowledge graph include: related to, surgical history and disease manifestation, and the paths in the path set are paths with disease entities as the endpoint;

[0088] The symptom inference module 440 is used to calculate the probability of disease occurrence and / or reward corresponding to all path endpoints in order to obtain the diagnostic result of the symptom.

[0089] Based on the above-described automatic disease diagnosis device 400, a complete reasoning process of the present invention can be interpreted as: inputting a set of symptoms (m1,...,m...). n ) and text labels (t1,…,t i The text tags are embedded into multiple symptom embeddings using an enhanced embedding method to obtain an enhanced symptom embedding (m′1,…,m′). n The disease reasoning algorithm is used for path inference, obtaining multiple sets of paths (p0,...,p) starting from each symptom.n Simultaneously, computational fusion symptom embeddings In step 4) of disease reasoning, m is used f Instead of m, calculate the reward and probability. Finally, depending on the requirements, you can choose the result with the highest reward or the result with the highest probability as the final return result.

[0090] In summary, this invention utilizes real electronic medical records to construct a medical knowledge graph, and performs disease querying and reasoning based on this graph. This introduces prior knowledge into diagnostic methods, thereby improving the accuracy of results. The model designed and implemented in this method uses reinforcement learning to achieve path reasoning on the graph. The returned results include both the diagnostic results and the complete reasoning path. Compared to existing methods that only return diagnostic results, the results of this method have stronger interpretability, and each node and edge on the path can provide more auxiliary information for doctors and patients. This research addresses the specific problem scenario by transforming the traditional disease classification problem into a path inference problem based on a medical knowledge graph, thus achieving better automatic disease diagnosis results than traditional algorithms.

[0091] The proposed model method is compared with traditional methods. Experiments were conducted using the same validation data, and parameters were fine-tuned for various algorithm models to obtain the best performance under the current model.

[0092] 1. Evaluation Indicators

[0093] This study uses evaluation metrics similar to those used in classification analysis. Common metrics in classification tasks include accuracy, precision, recall, and F1-score. Furthermore, this invention statistically analyzes the accuracy of the top-n results, where n ranges from 1 to 5.

[0094] Specifically, this invention defines True Positive, True Negative, False Positive, and False Negative as follows (Table 1), making the evaluation metrics more reasonable and meaningful. The final evaluation metrics—recall, precision, and F1-score—are calculated using the redefined True Positive, True Negative, False Positive, and False Negative metrics. The calculation methods for Precision, Recall, and F1-score are as follows:

[0095]

[0096]

[0097]

[0098]

[0099] name meaning True Positive Positive samples that are correctly predicted True Negative Negative samples that were predicted correctly False positives Positive samples that were predicted incorrectly False Negative Negative samples predicted incorrectly

[0100] Table 1 Definition of Evaluation Indicators

[0101] 2. Comparison of Model Performance

[0102] The experiment compared different automatic disease diagnosis algorithms with the disease diagnosis model proposed in this invention on a test set, and also compared the impact of different augmentation embedding methods on the final prediction accuracy. This invention selected several state-of-the-art models for comparison, namely the Diaformer model, the BED model, and the GAMP model. In addition, this invention also compared with traditional machine learning classification methods such as XGBoost and SVM. The experimental results are shown in Table 2. The model of this invention achieves the highest scores in all metrics without using augmentation embedding methods. This invention also provides the top-n accuracy of different models, as shown in the table below. Figure 5 As shown, this method achieves optimal results when n ranges from 1 to 5. The results demonstrate the correctness of the approach of transforming the automatic disease diagnosis task into a path inference task on a knowledge graph, and provide a foundation for subsequent enhanced embedding experiments.

[0103]

[0104] Table 1 Comparison of Model Results

[0105] 3. Enhance the results of the embedding experiment

[0106] Real-world disease diagnosis is a complex process, requiring doctors to consider a vast amount of information to make a diagnosis. Furthermore, a significant amount of unused information exists in real electronic medical records. The goal of augmented embedding methods is to integrate this information into knowledge graph representations and then utilize it appropriately in subsequent predictions to improve accuracy. This paper attempts two augmented embedding approaches: one uses an embedding layer to obtain attribute embedding vectors, and the other uses a pre-trained word2vec model to obtain embedding vectors. The accuracy of the two methods is as follows: Figure 6 As shown in the figure. Experimental results show that the model performance is significantly improved after using the enhanced embedding method.

[0107] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the scope of the invention. Any person skilled in the art can make modifications and variations to the embodiments without departing from the technical principles and spirit of the present invention. The scope of protection of the present invention should be determined by the claims.

Claims

1. An automated disease diagnosis device based on knowledge graph enhancement, characterized in that, The device includes: The data acquisition module is used to acquire the patient's symptoms and the doctor's supplementary explanations for each symptom; The data augmentation module is used to calculate the augmented representation of each symptom in conjunction with the supplementary description; The path inference module is used to perform path inference based on the enhanced representation in the knowledge graph to obtain a set of paths corresponding to each symptom; wherein, the entities in the knowledge graph include: disease, symptom, surgery, drug, description, related disease and related symptom, the relationships in the knowledge graph include: related to, surgical history and disease manifestation, and the paths in the path set are paths with disease entities as the endpoint; The symptom inference module is used to calculate the probability of disease occurrence and / or reward corresponding to all path endpoints in order to obtain the diagnostic result of the symptom; The step of performing path inference on the enhanced representation in the knowledge graph to obtain a path set corresponding to each symptom includes: Search for the symptom entity corresponding to the symptom in the knowledge graph. and the enhancement is represented as the symptom entity. Attributes; Based on the current path endpoint Historical information of the generated path With the symptom entity The attribute is used to calculate the state of the current path endpoint. ;in, Indicates the number of inference steps; The state is generated based on the reinforcement learning agent. Full motion space The reinforcement learning agent comprises: a two-layer fully connected neural network, a numerical network, and a policy network. For the entire motion space Pruning is performed to obtain the pruning motion space. ; For the space of pruning action The actions in the video are scored using cosine similarity. Select the best action and execute it to obtain the node. And add the node to the existing path; Reaching the maximum number of steps Then, based on the pruning action space The set of paths corresponding to the symptoms is obtained.

2. The apparatus as claimed in claim 1, characterized in that, The calculation of the enhancement representation of each symptom, in conjunction with the aforementioned supplementary explanation, includes: Using Word2Vec models pre-trained with either an embedding layer or a medical dictionary, symptoms were calculated respectively. vector representation and the symptoms Corresponding supplementary explanations vector representation ; Based on the vector representation and the vector representation The symptoms were obtained. Enhanced representation .

3. The apparatus as described in claim 1, characterized in that, The steps for training the reinforcement learning agent include: Construct the training set; The state is learned using the aforementioned two-layer fully connected neural network. The representation of ; Based on the representation The next action that the policy network calculates exists. ; Using a numerical network based on the representation Come to learn numerical ; The policy network is based on the numerical value Perform parameter optimization; The numerical network updates its parameters based on the training data.

4. The apparatus as claimed in claim 1, characterized in that, The full motion space Pruning is performed to obtain the pruning motion space. ,include: ,in, For the relationship, For entities, As symptoms, For the full motion space, symptom entity Cosine similarity score with the current entity.

5. The apparatus as described in claim 4, characterized in that, The cosine similarity score ;in, The vector representation of the symptoms. This indicates an enhancement of the symptoms. Indicates a relationship. Represents an entity, This represents a set of disease entities.

6. The apparatus as claimed in claim 1, characterized in that, The knowledge graph is constructed based on the TransE learning method.

7. The apparatus as claimed in claim 1, characterized in that, Calculate the disease occurrence probability and / or reward corresponding to all path endpoints to obtain a diagnostic result for the symptom, including: Based on actual needs, one or more of the disease occurrence probability or reward can be selected as diagnostic criteria. If the probability of disease occurrence is selected, the number of times the disease occurs corresponding to all path endpoints is counted, and the diagnosis result of the symptom is obtained based on the probability of occurrence of each disease. If the reward is selected, the enhanced representations of each symptom are fused to obtain the fused symptom embedding. And by calculating the reward at the end of all paths. The diagnosis of the symptoms was obtained; among which, End point of the path The expression, express and The dot product; If you choose disease occurrence probability and reward, then the occurrence probability of each disease and the aforementioned reward will be combined. The diagnosis of the symptoms was obtained.

Images