Online Health Consultation Method, System and Storage Medium Based on Intelligent AI

Abstract

The present invention discloses an online health consultation method, system and storage medium based on intelligent AI, including: obtaining the physical condition description data of a user to generate consultation information, and using the consultation information to obtain a number of corresponding question-and-answer pairs according to a preset large language model; selecting relevant basic health data from the user's basic health record, obtaining abnormal features through feature extraction, and selecting the question-and-answer pairs corresponding to the abnormal features to obtain user corpus to supplement the consultation information; using the supplemented consultation information and abnormal features to generate a health condition representation of the user in the large language model, and accessing a relevant domain knowledge graph to generate a health condition representation of the user in the knowledge graph; performing learning representation on the fused representation of the user's health condition through a multi-layer perceptron, and selecting answer data for output. The present invention realizes precise parsing of the user's consultation content, provides more targeted and precise consultation feedback in the aspect of medical health for the user, and timely discovers the user's health problems.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to an online health consultation method, system, and storage medium based on intelligent AI. Background Art

[0002] Online health consultation is a new product of the combination of the Internet and medical care, which covers various forms of health services such as health knowledge teaching, information medical inquiry, health electronic files, disease risk assessment, online consultation for medical treatment, online prescriptions, remote diagnosis, online treatment, and rehabilitation examinations with the network as the carrier and technical tool. Online medical care reveals a new direction for the development of the future health industry and is of great help in solving the contradiction between the uneven distribution of medical staff resources and the gradually increasing health care needs of the public. However, while online health consultation is developing, it also faces many problems, such as a significant information asymmetry problem between users and professional doctors.

[0003] With the continuous development and improvement of AI technology, AI-supported online health consultation will play a more important role in the future. Health consultation based on artificial intelligence technology will be able to understand human language and behavior more deeply, provide more accurate and comprehensive health advice, and provide more comprehensive and efficient health management services for people. Currently, most online health consultations achieved through AI obtain consultation results by users uploading disease information or consultation information. Due to the medical professionalism and expression differences of users, errors such as words not expressing their intended meanings will occur, which will lead to unreliable consultation results. Therefore, how to use means such as large models and knowledge graphs to obtain accurate user expressions, provide personalized and accurate medical services for users, and improve consultation efficiency is an urgent problem to be solved. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes an online health consultation method, system, and storage medium based on intelligent AI, aiming to accurately analyze the content of user consultations and provide more targeted and accurate consultation feedback in the medical and health aspects for users.

[0005] The first aspect of the present invention provides an online health consultation method based on intelligent AI, including:

[0006] Obtain the description data of the user's physical condition and the basic health file, preprocess the description data of the physical condition to generate consultation information, and use the consultation information to obtain a number of corresponding question-and-answer pairs according to a preset large language model;

[0007] Select relevant basic health data from the user's basic health record according to the Q&A pairs, extract features from the relevant basic health data to obtain abnormal features, and further obtain user corpus based on the Q&A pairs corresponding to the abnormal features to supplement the consultation information;

[0008] Generate a health status representation of the user in the large language model using the supplemented consultation information and abnormal features. Additionally, connect to a relevant domain knowledge graph, perform an embedding representation on the supplemented consultation information and abnormal features, and generate a health status representation of the user in the knowledge graph through graph embedding;

[0009] Fuse the obtained health status representations of the user, and use a multi-layer perceptron to learn and represent the fused health status representation of the user, and select answer data for the user's consultation information for output.

[0010] In this solution, preprocess the body condition description data to generate consultation information, specifically:

[0011] Obtain the body condition description data input by the user, perform a two-level encoding on the body condition description data, train a BERT model through a medical and health text dataset, and use the trained BERT model to perform an embedding representation on the body condition description data;

[0012] Obtain the feature vectors of keywords and the feature vectors of sentences in the body condition description data, import the feature vectors into the attention layer, use the attention mechanism to further embed the feature vectors contained in the sentence, update the keyword feature vectors through attention weight assignment, and perform normalization processing on the updated keyword feature vectors to obtain keyword semantic features;

[0013] Continue to perform an attention operation on the updated keyword feature vectors and sentence feature vectors, add different weight information to the keyword feature vectors and sentence feature vectors and then perform normalization processing to obtain sentence semantic features;

[0014] Obtain two-way semantic features from the keyword semantic features and sentence semantic features through a Bi-LSTM network, obtain two-level aggregated features through aggregation and activation function processing, and generate consultation information based on the two-level aggregated features.

[0015] In this solution, use the consultation information to obtain several corresponding Q&A pairs according to a preset large language model, specifically:

[0016] Import the consultation information into the preset large language model, extract relevant knowledge paragraphs according to the consultation information, generate hypothetical Q&A pairs based on the preset prompt template, and use the Q&A pairs to represent the questions and corresponding answers involved in the relevant knowledge paragraphs;

[0017] Cascade the Q&A pairs and relevant knowledge passages into a multi-dimensional vector space, and use similarity comparison in the multi-dimensional vector space to perform redundancy removal on the embedded vectors, and extract the health data indicators corresponding to the Q&A pairs.

[0018] In this solution, relevant basic health data is selected from the user's basic health record according to the Q&A pairs, and feature extraction is performed on the relevant basic health data to obtain abnormal features. Specifically:

[0019] Obtain the health data indicators involved in the Q&A pairs, generate retrieval tags using the health data indicators, perform similarity calculation in the user's basic health record through the retrieval tags, obtain the basic health data that meets the similarity standard under each health data indicator, and extract the basic health data sequence corresponding to each health data indicator;

[0020] Use the improved autoencoder model to construct an abnormal feature extraction model. Through the autoencoder model with sparse and noise reduction functions, perform feature encoding on the normal data of each health data indicator, and add batch normalization processing after the hidden layer of the sparse denoising autoencoder;

[0021] Use batch normalization processing to improve the training speed of the abnormal feature extraction model, obtain the benchmark reconstruction error corresponding to the normal data of each health data indicator, and use the Mahalanobis distance as a metric function to represent the benchmark reconstruction error to obtain an abnormal analysis indicator;

[0022] Input the basic health data sequence corresponding to each health data indicator into the trained abnormal feature extraction model to obtain the real-time reconstruction error, and judge whether the abnormal analysis indicator of each data sequence is greater than the preset distance threshold. If it is greater, it is regarded as an abnormal feature.

[0023] In this solution, use the supplemented consultation information and abnormal features to generate the health status representation of the user in the large language model. Additionally, connect to the relevant domain knowledge graph, perform embedded representation on the supplemented consultation information and abnormal features, and generate the health status representation of the user in the knowledge graph through graph embedding. Specifically:

[0024] Generate supplementary questions for the user according to the Q&A pairs corresponding to the abnormal health data, use the user's interaction corpus to supplement the consultation information, generate a query representation from the supplemented consultation information and abnormal features, and perform vector similarity retrieval in the multi-dimensional vector space where the Q&A pairs and relevant knowledge passages are cascaded;

[0025] Obtain the retrieved knowledge vectors, sort them according to the similarity, import them into the large language model, and aggregate the output of the large language model to obtain the health status representation of the user in the large language model;

[0026] Access the knowledge graph in the field of medical and health, use the supplemented consultation information and abnormal features to perform initial positioning in the field knowledge graph, and obtain the neighbor nodes of the consultation information node and the abnormal feature node according to semantic similarity and abnormal health data similarity;

[0027] Perform attention encoding on the neighbor nodes, assign higher weights to adjacent nodes with higher similarity, use the weights to aggregate and update the vector representations of the consultation information node and the abnormal feature node, and perform graph embedding representation learning according to preset rules using random walk to obtain meta-paths containing the consultation information node and the abnormal feature node, and aggregate all meta-paths to generate the health status representation of the user in the knowledge graph.

[0028] In this solution, a multi-layer perceptron is used to learn and represent the fused representation of the user's health status, and answer data is selected for the user's consultation information and output. Specifically:

[0029] Obtain the question-answer pairs and the health problem and corresponding status evaluation system involved in the meta-path according to the large language model and the field knowledge graph, obtain the health status representation through encoding, and use the multi-layer perceptron to learn the fused representation of the user's health status and the health status representation;

[0030] Combine the activation function operation to obtain the prediction probability of the interactive association between the fused representation of the user's health status and the health status representation, select the health status representation with a prediction probability greater than the preset probability threshold, and use the prediction probability as the weight information;

[0031] Based on the user's basic health data, use the status evaluation system in the selected health status representation to evaluate the severity score of the corresponding health problem, and use the weight information to weight the severity score of the health problem to generate the attention priority of the health problem;

[0032] Output the filtered health problems and the corresponding attention priorities as answer data.

[0033] The second aspect of the present invention provides an online health consultation system based on intelligent AI, which is characterized in that the system includes a consultation information acquisition unit, a health data call unit, a large language model unit, a knowledge graph unit, and a consultation output unit;

[0034] The consultation information acquisition unit is responsible for acquiring the user's physical condition description data and generating the user's consultation information through preprocessing;

[0035] The health data call unit is responsible for reading the user's basic health records within a preset time from different medical data sources;

[0036] The large language model unit is responsible for obtaining the question-and-answer pairs that may be involved in the retrieved consultation information, selecting relevant basic health data using the screened question-and-answer pairs, obtaining abnormal features through feature extraction and analysis, and supplementing the consultation information with user corpus according to the question-and-answer pairs corresponding to the abnormal features, so as to generate the health status representation of the user in the large language model;

[0037] The knowledge graph unit is responsible for embedding and representing the supplemented consultation information and abnormal features using the knowledge graph in the relevant field, and generating the health status representation of the user in the knowledge graph through graph embedding;

[0038] The consultation output unit uses a multi-layer perceptron to learn and represent the fused health status representation of the user, and selects answer data for the user consultation information for output.

[0039] The third aspect of the present invention provides a computer-readable storage medium, which includes a program for an online health consultation method based on intelligent AI. When the program for the online health consultation method based on intelligent AI is executed by a processor, the steps of the online health consultation method based on intelligent AI are implemented.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] The present invention uses a large language model to accurately analyze the user's consultation content and basic health data, discovers health problems and future health trends, and provides more targeted and accurate medical and health consultation feedback for users;

[0042] By using the analysis and mining of the large language model and the relevant knowledge graph to find more comprehensive correlation relationships, and through comprehensively analyzing the results of multiple dimensions, the efficiency of health consultation is greatly improved, and the accuracy of health problem recognition is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments or exemplifications of the present invention, the following will briefly introduce the drawings required for use in the embodiments or exemplifications. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the drawings shown without creative efforts.

[0044] Figure 1 Shows the flowchart of the online health consultation method based on intelligent AI;

[0045] Figure 2 Shows the flowchart of screening abnormal features in the basic health data in the embodiment;

[0046] Figure 3The flowchart shows the answer data for obtaining user consultation information in an embodiment;

[0047] Figure 4 The block diagram shows an online health consultation system based on intelligent AI. Detailed implementation manners

[0048] In order to more clearly understand the above objects, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments may be combined with each other.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0050] Figure 1 The flowchart shows an online health consultation method based on intelligent AI.

[0051] As Figure 1 shown, this embodiment provides an online health consultation method based on intelligent AI, including:

[0052] S102, obtaining the description data of the user's physical condition and the basic health record, preprocessing the description data of the physical condition to generate consultation information, and obtaining a plurality of corresponding question-and-answer pairs according to the consultation information using a preset large language model;

[0053] S104, selecting relevant basic health data from the user's basic health record according to the question-and-answer pairs, extracting features from the relevant basic health data to obtain abnormal features, and further obtaining user corpus based on the abnormal features corresponding to the question-and-answer pairs to supplement the consultation information;

[0054] S106, generating a health status representation of the user in the large language model using the supplemented consultation information and abnormal features, additionally accessing a knowledge graph in the relevant field, performing an embedding representation on the supplemented consultation information and abnormal features, and generating a health status representation of the user in the knowledge graph through graph embedding;

[0055] S108, fusing the obtained user health status representations, learning and representing the fused user health status representation through a multi-layer perceptron, and outputting answer data for the user consultation information.

[0056] It should be noted that the description data of the user's physical condition is obtained, double-level encoded for the description data of the physical condition, the BERT model is trained through a medical and health text dataset, the trained BERT model is used to perform embedding representation on the description data of the physical condition, the maximum pooling operation is used to obtain the feature vectors of the keywords and the feature vectors of the sentences in the description data of the physical condition. Maximum pooling can filter out abnormal information in medical and health texts, compress the dimension of the vectors to speed up the operation, and prevent model overfitting caused by feature embedding. The feature vectors are imported into the attention layer, and the attention mechanism is used to further embed the sentences and the included feature vectors. The keyword feature vectors are updated through attention weight allocation, and the updated keyword feature vectors are normalized to obtain the keyword semantic features.

[0057] The attention operation is continued for the updated keyword feature vectors and sentence feature vectors. Different weight information is added to the keyword feature vectors and sentence feature vectors and then normalized. The normalized output includes deeper sentence-level semantic information embedding. The fully connected connection of the keyword feature vectors and sentence feature vectors is used as the query vector for calculating the attention weights to obtain the sentence semantic features; the keyword semantic features and sentence semantic features pass through the Bi-LSTM network to obtain bidirectional semantic features, and the double-level aggregation features are obtained through aggregation and activation function processing. The consultation information is generated based on the double-level aggregation features.

[0058] It should be noted that the consultation information is imported into a preset large language model such as ChatGPT. Using the language understanding and generation capabilities of the large language model, relevant knowledge passages are extracted according to the consultation information, and hypothetical question-and-answer pairs are generated based on a preset prompt template. The question-and-answer pairs are used to represent the questions and corresponding answers involved in the relevant knowledge passages; for example, when the user takes dizziness as the consultation information, the large language model obtains the relevant knowledge passage of "dizziness", divides the corpus of the relevant knowledge passage, and clarifies that bad living habits, hypoglycemia, anemia, electrolyte disorders, etc. can all cause dizziness, and dizziness corresponds to health problems such as hypertension, myocardial ischemia, heart failure, and hepatic and renal insufficiency. According to the 5W1H (What, Why, Where, When, Who, How) rule, a prompt template is set to generate questions and their answers. The question-and-answer pairs and the relevant knowledge passages are cascaded and embedded into a multi-dimensional vector space, and redundancy removal processing of the embedded vectors is performed using similarity comparison in the multi-dimensional vector space, and the health data indicators corresponding to the question-and-answer pairs are extracted.

[0059] Figure 2 The flowchart showing the screening of abnormal features in the basic health data in the embodiment is shown.

[0060] According to an embodiment of the present invention, relevant basic health data is selected from the user's basic health record according to the question-and-answer pair, and feature extraction is performed on the relevant basic health data to obtain abnormal features. Specifically:

[0061] S202, obtain the health data indicators involved in the question-and-answer pair, generate a retrieval tag using the health data indicators, perform a similarity calculation in the user's basic health record through the retrieval tag, obtain the basic health data that meets the similarity standard under each health data indicator, and extract the basic health data sequence corresponding to each health data indicator;

[0062] S204, use the improved autoencoder model to construct an abnormal feature extraction model, perform feature encoding on the normal data of each health data indicator through the autoencoder model introducing sparse and noise reduction functions, and add batch normalization processing after the hidden layer of the sparse denoising autoencoder;

[0063] S206, use batch normalization processing to improve the training speed of the abnormal feature extraction model, obtain the benchmark reconstruction error corresponding to the normal data of each health data indicator, and represent the benchmark reconstruction error through the Mahalanobis distance as a metric function to obtain an abnormal analysis indicator;

[0064] S208, input the basic health data sequence corresponding to each health data indicator into the trained abnormal feature extraction model, obtain the real-time reconstruction error, and determine whether the abnormal analysis indicator of each data sequence is greater than a preset distance threshold. If it is greater, it is regarded as an abnormal feature.

[0065] It should be noted that the user's basic health record is constructed by obtaining the user's health examination data within a preset time or the examination data uploaded by the user. An abnormal feature extraction model is constructed using the autoencoder model improved by adding noise reduction and sparsity. Noise reduction adds noise to the input data and uses the autoencoder for encoding and decoding training, enabling the model to learn the ability to remove noise, reducing the error of the model to increase the generalization ability of the model. And the sparse function is introduced into the autoencoder to limit the nodes of some hidden layers. Only a limited number of nodes in the hidden layer are activated, and its sparse penalty factor is expressed as: , represents the KL divergence, which describes the difference between two distributions, represents the sparse parameter, represents the hidden layer 's average activation number, represents the total number of hidden layers. After adding the sparse constraint, the ability of the model to represent data is improved, preventing the hidden layer from learning meaningless features. In addition, the batch normalization algorithm is introduced to improve the convergence speed and learning efficiency of the model, making the calculation of the hidden layer more in line with the sensitivity constraints of the non-linear activation function.

[0066] It should be noted that supplementary questions for users are generated based on the Q&A pairs corresponding to abnormal health data, and the interactive corpus of users is used to supplement the consultation information. The supplemented consultation information and abnormal features are used to generate a query representation, and vector similarity retrieval is performed in the multi-dimensional vector space where the Q&A pairs and relevant knowledge paragraphs are concatenated. The retrieved knowledge vectors are sorted according to the similarity, imported into the large language model, and the outputs of the large language model are aggregated. The outputs include the possible physical health problems of the user, the severity of the current health problems, the symptom analysis of the health problems, etc., and the health status representation of the user in the large language model is obtained.

[0067] Connect to the knowledge graph in the field of medical and health, use the supplemented consultation information and abnormal features to perform initial positioning in the field knowledge graph, and obtain the neighbor nodes of the consultation information node and the abnormal feature node according to the semantic similarity and the similarity of abnormal health data; perform attention encoding on the neighbor nodes, and assign higher weights to the adjacent nodes with higher similarity. The vector representations of the consultation information node and the abnormal feature node are aggregated and updated using the weights, and graph embedding representation learning is performed according to preset rules such as path length and answer node type using random walk strategies such as DeepWalk and metapath2vec to obtain the meta-path containing the consultation information node and the abnormal feature node, and all meta-paths are aggregated to generate the health status representation of the user in the knowledge graph. Preferably, when obtaining the health status representation of the user in the knowledge graph, adversarial training is introduced, a discriminator module is added, and the meta-path obtained by random walk is imported into the discriminator module; in the discriminator module, the LSTM network is used to encode the random walk sequence corresponding to the meta-path to obtain the path vector representation and score calculation, and its path score is obtained. , , represents the relationship between nodes in the random walk sequence, represents the learnable parameter, represents the meta-path. The deviation between the path score and the expected score is obtained, and the meta-path with a deviation greater than the preset deviation standard is punished according to the deviation, and the deviation between the path score and the expected score is minimized to guide the stacked walk for training, so as to prompt the domain knowledge graph to predict the answer in the correct path direction.

[0068] The two health status representations of the user are fused through an adaptive information fusion function to determine the health status representation of the user in the large language model and the health status representation of the user in the knowledge graph ratio information , , represents the learnable weight, and the ratio information is used to obtain the fused representation of the user's health status , 。

[0069] Figure 3 The flowchart shows the process of obtaining the answer data for the user's consultation information in an embodiment.

[0070] According to an embodiment of the present invention, a multi-layer perceptron is used to learn and represent the fused representation of the user's health status, and answer data is selected and output for the user's consultation information. Specifically:

[0071] S302, obtain the question-and-answer pairs and the health problems and corresponding status evaluation systems involved in the meta-path according to the large language model and the domain knowledge graph, obtain the health status representation through encoding, and use a multi-layer perceptron to learn the fused representation of the user's health status and the health status representation;

[0072] S304, combine the activation function operation to obtain the prediction probability of the existence of an interaction correlation between the fused representation of the user's health status and the health status representation, select the health status representation with a prediction probability greater than the preset probability threshold, and use the prediction probability as the weight information;

[0073] S306, evaluate the severity score of the corresponding health problem based on the user's basic health data using the status evaluation system in the selected health status representation, and weight the severity score of the health problem using the weight information to generate the attention priority of the health problem;

[0074] S308, output the screened health problems and the corresponding attention priorities as answer data.

[0075] It should be noted that a multi-layer perceptron is used to learn the fused representation of the user's health status and the health status representation ; combine the activation function operation to obtain the prediction probability of the existence of an interaction correlation between the fused representation of the user's health status and the health status representation , which is expressed as: , represents the activation function. Obtain the health problems and the corresponding status evaluation systems involved in the question-and-answer pairs and the meta-path, such as the dizziness problem and the evaluation system for dizziness caused by different reasons, such as the duration of bad living habits such as staying up late, and the different severity levels of dizziness caused. Quantify the severity according to the evaluation system constructed based on historical health problem instances, and generate the attention priority in combination with the prediction probability of the health problem, so as to arouse the user's attention to their own physical health status.

[0076] Figure 4 The block diagram shows an online health consultation system based on intelligent AI.

[0077] In a second aspect of the present invention, there is provided an online health consultation system 4 based on intelligent AI, characterized in that the system includes a consultation information acquisition unit 401, a health data call unit 402, a large language model unit 403, a knowledge graph unit 404, and a consultation output unit 405;

[0078] The consultation information acquisition unit is responsible for acquiring the physical condition description data of the user and generating user consultation information through preprocessing;

[0079] The health data call unit is responsible for reading the basic health records of the user within a preset time from different medical data sources;

[0080] The large language model unit is responsible for obtaining the question-and-answer pairs that may be involved in the retrieved consultation information, selecting relevant basic health data using the screened question-and-answer pairs, obtaining abnormal features through feature extraction and analysis, and supplementing the consultation information with the user corpus according to the abnormal feature corresponding question-and-answer pairs to generate the health status representation of the user in the large language model;

[0081] The knowledge graph unit is responsible for embedding and representing the supplemented consultation information and abnormal features using the knowledge graph in the relevant field, and generating the health status representation of the user in the knowledge graph through graph embedding;

[0082] The consultation output unit uses a multi-layer perceptron to learn and represent the fused health status representation of the user, and selects answer data for the user consultation information for output.

[0083] In a third aspect of the present invention, there is provided a computer-readable storage medium, which includes an online health consultation method program based on intelligent AI. When the online health consultation method program based on intelligent AI is executed by a processor, the steps of the online health consultation method based on intelligent AI are implemented.

[0084] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed with each other can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical, or other forms.

[0085] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; and some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0086] In addition, in each embodiment of the present invention, each functional unit may be entirely integrated into one processing unit, or each unit may be separately a unit by itself, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of a combination of hardware and software functional units.

[0087] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks and other various media that can store program codes.

[0088] Alternatively, if the above-mentioned integrated units of the present invention are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks or optical disks and other various media that can store program codes.

[0089] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, and all of them should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. An online health consultation method based on intelligent AI, characterized in that: The following steps are involved: Obtaining the user's physical condition description data and basic health records, preprocessing the physical condition description data to generate consulting information, and using the consulting information to obtain a number of corresponding question-answer pairs according to a preset large language model; Selecting relevant basic health data from the user's basic health record according to the question and answer pair, performing feature extraction on the relevant basic health data to obtain abnormal features, and further obtaining user corpus corresponding to the question and answer pair based on the abnormal features to supplement the consultation information; The supplemented consultation information and abnormal features are used to generate the user's health status representation in the large language model. In addition, the supplemented consultation information and abnormal features are connected to the knowledge graph in the relevant field to embed the supplemented consultation information and abnormal features, and the user's health status representation in the knowledge graph is generated through graph embedding; The acquired representations of the user's health status are fused, and the fused representations of the user's health status are learned and represented by a multi-layer perceptron, and the answer data for the user's consultation information is selected and output; The supplemented consultation information and abnormal features are used to generate the user's health status representation in the large language model. In addition, the relevant domain knowledge graph is accessed to embed the supplemented consultation information and abnormal features, and the user's health status representation in the knowledge graph is generated through graph embedding. Specifically: Generate supplementary questions for users based on the question-answer pairs corresponding to abnormal health data, use the user's interactive corpus to supplement the consultation information, generate query representations based on the supplemented consultation information and abnormal features, and perform vector similarity retrieval in the multi-dimensional vector space where the question-answer pairs are cascaded with relevant knowledge segments; The retrieved knowledge vectors are sorted according to similarity, imported into the large language model, and the output of the large language model is aggregated to obtain the user's health status representation in the large language model; Access the knowledge graph in the medical and health-related fields, use the supplemented consulting information and abnormal features to perform initial positioning in the knowledge graph, and obtain neighbor nodes of the consulting information node and the abnormal feature node based on semantic similarity and abnormal health data similarity; Attention encoding is performed on the neighbor nodes, and higher weights are assigned to neighboring nodes with higher similarity. The vector representations of the consulting information nodes and abnormal feature nodes are updated using weight aggregation. Random walks are used to perform graph embedding representation learning according to preset rules to obtain meta-paths containing consulting information nodes and abnormal feature nodes. All meta-paths are aggregated to generate a representation of the user's health status in the knowledge graph.

2. According to claim 1, an online health consultation method based on intelligent AI is characterized in that: The physical condition description data is preprocessed to generate consulting information, specifically: Obtaining physical condition description data input by a user, performing two-level encoding on the physical condition description data, training a BERT model using a medical health text dataset, and using the trained BERT model to embed the physical condition description data; Obtain the feature vectors of keywords and sentences in the physical condition description data, import the feature vectors into the attention layer, use the attention mechanism to further embed the sentences and the feature vectors contained therein, update the keyword feature vectors through attention weight allocation, and normalize the updated keyword feature vectors to obtain the keyword semantic features; Continue to perform attention operation on the updated keyword feature vector and sentence feature vector, add different weight information to the keyword feature vector and sentence feature vector, and then perform normalization processing to obtain sentence semantic features; The keyword semantic features and sentence semantic features are passed through a Bi-LSTM network to obtain bidirectional semantic features, and two-level aggregation features are obtained after aggregation and activation function processing, and consulting information is generated based on the two-level aggregation features.

3. According to claim 1, an online health consultation method based on intelligent AI is characterized in that: The consultation information is used to obtain a number of corresponding question-answer pairs according to a preset large language model, specifically: Importing the consulting information into a preset large language model, extracting relevant knowledge paragraphs according to the consulting information, generating hypothetical question-answer pairs for the relevant knowledge paragraphs based on a preset prompt template, and using the question-answer pairs to represent the questions involved in the relevant knowledge paragraphs and the corresponding answers; The question-answer pair and the relevant knowledge text are cascaded and embedded into a multidimensional vector space, and similarity comparison is used in the multidimensional vector space to perform redundancy removal on the embedded vectors, and extract health data indicators corresponding to the question-answer pair.

4. The online health consultation method based on intelligent AI according to claim 1, characterized in that: According to the question and answer pair, relevant basic health data is selected from the basic health file of the user, and features are extracted from the relevant basic health data to obtain abnormal features, specifically: Obtain the health data indicators involved in the question-answer pair, generate search tags using the health data indicators, perform similarity calculation in the user's basic health records using the search tags, obtain basic health data that meets the similarity standard under each health data indicator, and extract the basic health data sequence corresponding to each health data indicator; The improved autoencoder model is used to build an abnormal feature extraction model. The normal data of each health data indicator is feature encoded by introducing an autoencoder model with sparse and denoising functions, and batch normalization is added after the hidden layer of the sparse denoising autoencoder. Batch normalization is used to improve the training speed of the abnormal feature extraction model, and the baseline reconstruction error corresponding to the normal data of each health data indicator is obtained. The Mahalanobis distance is used as a metric function to represent the baseline reconstruction error to obtain the abnormal analysis indicator; The basic health data sequence corresponding to each health data indicator is input into the trained abnormal feature extraction model to obtain the real-time reconstruction error, and determine whether the abnormal analysis indicator of each data sequence is greater than the preset distance threshold. If so, it is regarded as an abnormal feature.

5. The online health consultation method based on intelligent AI according to claim 1, characterized in that: The multi-layer perceptron is used to learn and represent the fusion representation of the user's health status, and the answer data is selected for the user's consultation information and output, specifically: Based on the large language model and domain knowledge graph, the health issues and corresponding condition evaluation system involved in the question-answer pairs and meta-paths are obtained, the health status representation is obtained after encoding, and the user's health status fusion representation and health status representation are learned using a multi-layer perceptron; Combined with the activation function operation, the predicted probability of the interactive correlation between the fusion representation of the user's health status and the health status representation is obtained, the health status representation with a predicted probability greater than a preset probability threshold is selected, and the predicted probability is used as weight information; Based on the basic health data of the user, the severity score of the corresponding health problem is evaluated by using the status evaluation system in the selected health status representation, and the severity score of the health problem is weighted by using the weight information to generate the priority of the health problem; The health issues obtained through screening and their corresponding priorities are output as answer data.

6. An online health consultation system based on intelligent AI, characterized in that: Implementing the online health consultation method based on intelligent AI as described in any one of claims 1 to 5, comprising a consultation information acquisition unit, a health data calling unit, a large language model unit, a knowledge graph unit and a consultation output unit; The consulting information acquisition unit is responsible for acquiring the user's physical condition description data and generating user consulting information after preprocessing; The health data calling unit is responsible for reading the user's basic health records within a preset time from different medical data sources; The large language model unit is responsible for obtaining question-answer pairs that may be involved in the search for consulting information, selecting relevant basic health data using the screened question-answer pairs, obtaining abnormal features using feature extraction analysis, obtaining user corpus based on the abnormal features corresponding to the question-answer pairs to supplement the consulting information, and generating a representation of the user's health status in the large language model; The knowledge graph unit is responsible for embedding the supplemented consulting information and abnormal features using the knowledge graph of the relevant field, and generating the health status representation of the user in the knowledge graph through graph embedding; The consultation output unit uses a multi-layer perceptron to learn and represent the fused representation of the user's health status, and selects answer data for the user's consultation information for output.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes an online health consultation method program based on intelligent AI. When the online health consultation method program based on intelligent AI is executed by a processor, the steps of the online health consultation method based on intelligent AI as described in any one of claims 1 to 5 are implemented.

Images