Remote health consultation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence health consultation, and discloses a remote health consultation method and system based on artificial intelligence. The method comprises the following steps: receiving a health consultation request submitted by a user through terminal equipment, wherein the request comprises basic health information of the user and a specific consultation question; the basic health information is preprocessed to generate a standardized health data record, and the problems that existing health information is not uniform in format and difficult to efficiently analyze are solved; the standardized data are input into the trained health assessment model, and comprehensive and objective analysis of the health condition of the user is completed; retrieving associated medical knowledge nodes from a medical knowledge graph according to a model analysis result, and providing a professional medical basis for a consultation scheme; and in combination with the medical knowledge node and the user consultation question, constructing a personalized health consultation scheme and sending the personalized health consultation scheme to the user terminal equipment. The method can quickly respond to the consultation demand of the user, guarantee the scientificity and pertinence of the consultation scheme, and effectively improve the remote health consultation experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence health consultation technology, specifically to a remote health consultation method and system based on artificial intelligence. Background Technology

[0002] With the accelerating aging of the population and the continuous improvement of people's health awareness, the demand for convenient and efficient health consultation services is growing. Traditional health consultation models rely heavily on offline medical institutions, and users often need to make appointments in advance and wait in line at designated locations. This not only consumes a lot of time and transportation costs, but is also significantly limited by geographical location. Especially in remote areas or where medical resources are scarce, users find it difficult to quickly obtain professional health guidance.

[0003] In recent years, remote health consultation services have gradually emerged, breaking down geographical barriers to some extent, allowing users to communicate with medical professionals through terminal devices. However, current remote health consultation services still have many limitations. Faced with a large number of user inquiries, the number of medical professionals is limited, making it difficult to provide immediate responses, resulting in a poor user consultation experience. Some urgent health issues may be delayed due to excessively long waiting times. Different users have individual differences in their health conditions. Existing remote consultation services mostly adopt a generic response model, lacking in-depth analysis and integration of users' basic health information, and are unable to provide personalized health plans based on specific user circumstances, resulting in insufficient targeting and practicality of the consultation plans.

[0004] During health consultations, the retrieval and application of medical knowledge rely on the personal experience of medical personnel, which may lead to incomplete or inaccurate knowledge retrieval, thus affecting the scientific validity of health advice. Furthermore, the diverse forms of health information submitted by users, including textual descriptions and numerical data, lack unified preprocessing standards. This makes it difficult to effectively transform health information into standardized data suitable for analysis, further hindering the improvement of the quality and efficiency of health consultation services. These problems make existing remote health consultation services insufficient to meet users' needs for efficient, personalized, and scientific health guidance, necessitating a new technological solution to address these shortcomings. Summary of the Invention

[0005] The purpose of this invention is to provide a remote health consultation method and system based on artificial intelligence to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a remote health consultation method based on artificial intelligence, the method comprising: Receive health consultation requests submitted by users through terminal devices, the health consultation requests including the user's basic health information and consultation questions; The basic health information is preprocessed to generate standardized health data records; The standardized health data records are input into the trained health assessment model for analysis. Based on the analysis results of the health assessment model, relevant medical knowledge nodes are retrieved from the medical knowledge graph; Based on the aforementioned medical knowledge nodes and the aforementioned consultation questions, a personalized health consultation plan is constructed. The personalized health consultation plan will be sent to the user's terminal device; The medical knowledge nodes include disease nodes, treatment plan nodes, and medication recommendation nodes.

[0007] Preferably, the preprocessing of the basic health information includes: Identify unstructured text data in the basic health information; Extract key health features from the unstructured text data; The key health characteristics were matched with a standard medical terminology database; Generate standardized health data records that include structured health characteristics.

[0008] Preferably, the training process of the health assessment model includes: Collect historical health consultation case data, which includes user health information and corresponding medical diagnosis results; Feature engineering was performed on the case data to extract multidimensional health features; Build a deep neural network model and set multiple hidden layers; The deep neural network model is iteratively trained using the multidimensional health features; Verify the prediction accuracy of the trained deep neural network model on the test dataset.

[0009] Preferably, the step of retrieving relevant medical knowledge nodes from the medical knowledge graph includes: Analyze the results output by the health assessment model to determine key diagnostic indicators; Search the medical knowledge graph for disease nodes associated with the key diagnostic indicators; Obtain the treatment plan node and medication suggestion node corresponding to the disease node; Calculate the matching degree between each medical knowledge node and the user's health data.

[0010] Preferably, the construction of a personalized health consultation plan includes: Based on the matching degree of the medical knowledge nodes, select the medical knowledge nodes with the highest priority. Based on the user's personal health record, adjust the treatment plan parameters in the medical knowledge node; Generate a consultation protocol framework that includes diagnostic recommendations, treatment plans, and health guidance; The aforementioned consultation solution framework is processed using natural language generation to form readable text.

[0011] Preferably, the method further includes: Receive user feedback on the personalized health consultation plan; Analyze the user satisfaction data in the feedback information; The parameters of the health assessment model are updated based on the user satisfaction data; Store the updated parameters of the health assessment model in the model version repository.

[0012] Preferably, the analysis of user satisfaction data in the feedback information includes: Extract the text evaluation content from the feedback information; Sentiment analysis is performed on the text evaluation content to calculate the sentiment tendency value; Identify key opinion points in the text evaluation content; The sentiment scores and key opinion points are mapped to dimensions for improving the health assessment model.

[0013] Preferably, the parameters for updating the health assessment model include: The range of model parameters that need to be adjusted is determined based on the dimensions for improvement of the health assessment model. Calculate the parameter adjustment amount, which is proportional to the sentiment tendency value; Parameters are fine-tuned while ensuring the overall stability of the health assessment model. Record the version information and adjustment effects of each parameter update.

[0014] Preferably, the method further includes: Regularly retrieve the latest medical research findings from medical databases; Transform the aforementioned medical research findings into new medical knowledge nodes; Update the node relationships in the medical knowledge graph; The health assessment model is incrementally trained to adapt to new medical knowledge.

[0015] Preferably, the present invention also includes an artificial intelligence-based remote health consultation system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the artificial intelligence-based remote health consultation method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By receiving health consultation requests submitted by user terminal devices, the system eliminates the need for users to wait for offline appointments or manual responses from medical personnel. The automated processing workflow can quickly initiate the health consultation service process, shortening the time from submitting a request to receiving a consultation plan and improving the immediacy of user consultations. This is especially beneficial for users who do not require urgent health advice but need timely health information, significantly improving their consultation experience and allowing them to receive prompt feedback when they need health guidance.

[0017] This method preprocesses users' basic health information and generates standardized health data records, addressing the problem of diverse and difficult-to-integrate health information in existing systems. Health information submitted by different users may exhibit inconsistent formats and non-standard expressions. The standardized data records generated after preprocessing eliminate interference from these format differences, making the health information easier for subsequent health assessment models to identify and analyze. This ensures the validity and usability of the health information, providing a solid data foundation for accurate analysis of users' health status and avoiding analytical biases caused by non-standard information.

[0018] Analyzing standardized health data records by inputting them into a trained health assessment model fully leverages the advantages of artificial intelligence technology to achieve a deep understanding of users' health conditions. The health assessment model, trained on a large amount of medical data, can uncover potential health characteristics from standardized data. Compared to traditional analysis methods that rely on the personal experience of medical personnel, it can more comprehensively and objectively assess users' health conditions, reducing oversights that may arise from human judgment. This makes the analysis results more scientific and reliable, providing a strong basis for subsequently developing appropriate health consultation plans.

[0019] By retrieving relevant medical knowledge nodes from a medical knowledge graph based on the analysis results of a health assessment model, the professionalism and comprehensiveness of health consultation plans can be ensured. The medical knowledge graph integrates a vast amount of standardized medical knowledge. By targeting knowledge nodes that match the user's health condition, it avoids the randomness of medical knowledge retrieval, ensuring that the acquired medical knowledge is highly relevant to the user's specific health problems. This reduces biases in consultation advice caused by incomplete or inaccurate knowledge retrieval, providing a solid foundation of medical knowledge for the development of health consultation plans and enhancing their scientific rigor and credibility.

[0020] Personalized health consultation plans built upon medical knowledge nodes and consultation questions can meet the individual health needs of different users. This method fully considers users' basic health information and specific consultation questions, abandoning the traditional generic response model and tailoring consultation plans for each user. This makes the plan content more closely aligned with the user's actual health condition and needs. For example, for users with chronic diseases, the plan can provide targeted dietary and exercise suggestions based on their medical history and current health indicators; for general health consultation users, personalized preventive and health care guidance can be provided based on their consultation questions and health data, allowing each user to receive health advice tailored to their own situation and enhancing the practicality and relevance of the consultation plan.

[0021] Sending personalized health consultation plans to users' devices meets their need for convenient access to health information. Users don't need to travel to a specific location; they can view the consultation plan simply through their own devices, without time or geographical limitations, further enhancing the convenience of health consultation services. At the same time, the personalized plans are presented to users in a clear format, making them easy to understand and implement. This helps users better adjust their health management methods according to the plan, promoting improved health levels and driving the development of remote health consultation services towards greater efficiency, accuracy, and personalization. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based remote health consultation method described in this invention.

[0023] Figure 2 Flowchart for training a health assessment model;

[0024] Figure 3 A flowchart for developing personalized health consultation plans. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, 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.

[0026] Please see Figure 1This invention provides an artificial intelligence-based remote health consultation method. The method includes: a user submitting a health consultation request through their terminal device, such as a smartphone, tablet, or personal computer. This request includes basic health information actively filled in or selected by the user, such as age, gender, height, weight, past medical history, current symptom description, and specific consultation questions. Upon receiving this request, the system first preprocesses the potentially unstructured or semi-structured basic health information to eliminate data noise, standardize data format, and generate a standardized health data record. This record exists in the form of structured data fields, facilitating processing by a computer model. This standardized health data record is then used as input to a pre-trained health assessment model. This model is typically a complex machine learning model capable of deep analysis of the input health data and outputting analysis results including potential health risk indicators and symptom correlation scores. The system then accesses a pre-constructed medical knowledge graph containing numerous medical concepts and their relationships based on the analysis results output by the health assessment model. Within this knowledge graph, it retrieves medical knowledge nodes highly relevant to the current analysis results, such as disease nodes, treatment plan nodes, or medication recommendation nodes. After the search is completed, the system will integrate the retrieved medical knowledge nodes with the user's initial consultation question, and use certain logical rules and algorithms to construct a personalized health consultation plan for the current user. This plan may include health assessment conclusions, lifestyle suggestions, and medical guidance. The completed personalized health consultation plan will be sent back to the user's terminal device via the network, presented to the user in the form of text, images, or voice, completing a full remote health consultation process.

[0027] Example 1: In health consultation requests submitted by users through the terminal interface, basic health information typically consists of two parts: one part is structured data fields, such as the age range selected by the user in the drop-down menu, the gender determined by the radio button, and the height and weight values ​​entered in the input box; the other part is unstructured text data, which mainly manifests as detailed statements entered by the user in free text areas such as "symptom description" or "consultation questions." This type of text is usually written in natural language and includes subjective descriptions of the user's discomfort, medical history, and lifestyle habits, with a high degree of arbitrariness and diversity in expression. Identifying unstructured text data is the first step in preprocessing. The system needs to accurately separate these free text fields from the entire received request data packet, as these fields are the main source of information for subsequent in-depth analysis. Extracting key health features from unstructured text data is a sophisticated process that relies on natural language processing technology. The system will call the built-in text analysis module to perform multi-level processing on the identified text segments. First, word segmentation is performed to divide continuous sentences into independent lexical units. Then, part-of-speech tagging and syntactic analysis are conducted to understand the text's grammatical structure. Based on this, named entity recognition (NER) technology is used to identify and locate healthcare-related entities in the text. These entities typically include body parts, clinical symptoms, disease names, drug names, examination items, time descriptions, and adverbs of degree. For example, from the description "I've had a persistent, severe headache for the past three days, especially a throbbing pain in my temples in the afternoon, and I occasionally feel nauseous but haven't vomited," the NER model needs to accurately extract a series of key health features such as "three days," "severe," "headache," "afternoon," "temples," "throbbing pain," "nausea," and "vomiting." This process not only requires identifying the entities themselves but also capturing the modifying relationships and temporal dynamics between entities as much as possible.

[0028] Matching extracted key health features with a standard medical terminology database is a core step in achieving information standardization. Due to the colloquial and regional variations in user descriptions, the same medical concept may have multiple expressions; for example, "headache" and "spike headache" refer to the same concept, and "diarrhea" and "diarrhea" have the same meaning. The system maintains a standard medical terminology database built upon international standard medical terminology systems such as SNOMEDCT and ICD. Each term in this database has a unique identifier and a clear definition. The matching process involves extracting potentially non-standard feature words from the text and mapping them to the most appropriate standard terms in the database using methods such as semantic similarity calculation, synonym mapping, and contextual analysis. For example, the user description "headache" is mapped to the standard term "spike headache"; "throbbing pain in the temples" is further precisely mapped to "pulsating headache" and associated with its location attribute as "temporal region"; "nausea" is mapped to "nausea"; and "vomiting," although not actually occurring, is mentioned as a related symptom. This mapping process greatly eliminates linguistic ambiguity, transforming vague, subjective descriptions into precise medical concepts that can be understood and reasoned about by computers. Generating standardized health data records containing structured health features is the final output of the preprocessing workflow. The information processed in the aforementioned steps is no longer disorganized text but a well-organized, structured data object. This data record typically uses a machine-readable format such as JSON or XML, containing all standardized fields extracted from the user's original information, for example: {"symptoms":[{"term":"headache","character":"pulsation","location":"temporal","duration":"3 days","severity":"severe","timing":"afternoon"}],"associated_symptoms":[{"term":"nausea","occurrence":"occurrence"}],...}. Simultaneously, information obtained from the structured portion submitted by the user, such as age and gender, is also integrated into this record, forming a complete snapshot of health data regarding the current consultation event. This standardized record provides high-quality, low-noise input for health assessment models, allowing them to focus on pattern recognition and risk assessment at the medical level, without having to deal with uncertainties at the language understanding level.

[0029] The entire preprocessing workflow relies on a well-designed software pipeline consisting of multiple sequentially executed microservices or function modules. The text recognition module receives the original request and separates text fields; the feature extraction module loads the NLP model for entity recognition; the terminology matching module accesses the terminology database to complete standardized mapping; and finally, the data assembly module generates the final structured record. Each module in the pipeline has logging and error handling capabilities. If an anomaly occurs in any processing step, such as failing to find a corresponding item in the terminology database, the system will mark the feature as "unmapped" and log it for subsequent manual review and database expansion, rather than causing the entire process to be interrupted. This design ensures the system's robustness. The effectiveness of preprocessing largely depends on the accuracy of the natural language processing model used and the coverage and update frequency of the standard medical terminology database. Therefore, maintaining an up-to-date and comprehensive terminology database and regularly updating the NLP model are fundamental to ensuring the efficiency of the preprocessing stage. The output of the preprocessing stage, namely the standardized health data record, needs a data structure design that fully considers the input requirements of downstream health assessment models. The naming, value range, and encoding methods of data fields must be strictly consistent with the data format used during model training. Any deviation may cause the model to fail to correctly interpret the input information and produce incorrect output. Therefore, during system development and iteration, data engineers and algorithm engineers need to collaborate closely to define and maintain this standardized data model. From the user's raw text input to the generation of standardized records, the entire preprocessing process is completed at high speed without the user's awareness. Its ultimate goal is to deliver "clean," "organized," and "semantically clear" data, laying a solid information foundation for AI-driven health assessment.

[0030] Example 2: See Figure 2The training process of the health assessment model is built upon a large amount of historical data. Its purpose is to enable the model to learn complex patterns from past health consultation cases, thereby acquiring the ability to scientifically assess newly input user health data. The training process begins with the collection of historical health consultation case data. This data typically comes from anonymized records of partner medical institutions, authorized anonymized health platform data, or other sources that comply with ethical and legal regulations. A qualified case data unit needs to contain a relatively complete fragment of user health information and a corresponding diagnostic result or authoritative assessment conclusion confirmed by medical experts. Health information may cover multiple dimensions such as demographic data, chief complaint, present illness history, past medical history, physical examination indicators, laboratory test results, and imaging report summaries. The corresponding diagnostic result is the target label that the model aims to approximate. The quality and quantity of data directly determine the upper limit of model performance. Therefore, significant effort needs to be invested in planning during the data collection phase to ensure that the data covers as many disease types, population characteristics, and health states as possible, while establishing a strict data anonymization process to protect user privacy.

[0031] Feature engineering of the collected raw case data serves as a bridge between the raw data and the machine learning model. This step aims to transform the chaotic raw information into numerical features that the model can effectively learn. Feature engineering processes structured numerical fields (such as age and blood pressure) and semi-structured or unstructured text descriptions (such as the chief complaint). For numerical features, normalization or standardization is typically required to scale them to a uniform scale, for example, mapping age from 0-100 years to 0-1, to avoid disproportionately impacting the model due to the large numerical range of some features. For categorical features, such as gender and blood type, one-hot encoding is used to convert them into binary vector representations. For text descriptions, natural language processing techniques, such as bag-of-words models, TF-IDF, or more advanced word embedding methods, are used to convert the text into fixed-length numerical vectors. More advanced feature engineering may involve feature construction, i.e., creating new features based on domain knowledge, such as calculating body mass index based on height and weight, combining a series of related symptoms into a new comprehensive index, or extracting trend features from time series data. Ultimately, each historical case is transformed into a multidimensional health feature vector, which comprehensively represents the health status of that case.

[0032] Building a deep neural network model is the core of learning complex nonlinear relationships. Deep neural networks consist of multiple interconnected artificial neurons, and their multi-layered structure allows for multi-level feature abstraction. The number of neurons in the input layer is equal to the dimension of the aforementioned multi-dimensional health feature vector, and it is responsible for receiving data. The design of the output layer depends on the nature of the prediction task. If it is a multi-class classification problem (e.g., determining which disease it belongs to), the number of neurons in the output layer equals the number of classes, and a Softmax activation function is used. If it is a regression problem (e.g., predicting the future value of a certain indicator), there may be only one output neuron using a linear activation function. Between the input and output layers are multiple hidden layers. The number of hidden layers and the number of neurons in each layer are hyperparameters that need careful design. Too few layers may lead to insufficient model learning ability and an inability to capture complex patterns; too many layers may cause overfitting and a surge in computational costs. Each layer typically consists of a linear transformation and a nonlinear activation function. Commonly used activation functions include ReLU, Sigmoid, and Tanh, which introduce nonlinearity into the model. The initialization strategy for model parameters and regularization methods to prevent overfitting (such as Dropout and L2 regularization) also need to be determined during the model building phase.

[0033] The core process of model learning is iteratively training the constructed deep neural network using the processed multidimensional health feature vectors. At the start of training, all parameters of the model (weights and biases) are typically initialized to random small values. In each iteration, a batch of training data is fed into the network, and forward propagation is performed to obtain the predicted output. The predicted output is then compared with the true diagnostic labels, and a loss function is used to quantify the difference between the current prediction and the true value. Cross-entropy loss is commonly used for classification tasks, while mean squared error loss is commonly used for regression tasks. Next, the gradient of the loss function with respect to each parameter in the model is calculated using the backpropagation algorithm. The gradient indicates the direction and magnitude in which each parameter should be adjusted to reduce the loss. Finally, an optimization algorithm, such as stochastic gradient descent and its variants (e.g., Adam, RMSprop), is used to update the parameters in the network based on the calculated gradient. This process of forward propagation, loss calculation, backpropagation, and parameter update is repeated multiple times on the entire training dataset; each round is called an epoch. The learning rate is a key hyperparameter of the optimizer, controlling the step size of parameter updates, and needs to be set carefully.

[0034] Validating the predictive performance of a trained deep neural network model on an independent test dataset is essential for evaluating its generalization ability. The test dataset must be completely independent of the training and validation datasets; that is, the cases in the test dataset must not have been involved in the model's training or hyperparameter tuning process in any way. The test dataset is input into the trained model, which outputs a prediction for each test case. These predictions are compared with the true labels of the test cases to calculate a series of objective performance metrics. For classification tasks, commonly used metrics include accuracy, precision, recall, F1 score, and these metrics under macro-average or weighted averages. More detailed analysis will examine the confusion matrix to understand which categories the model is prone to confusion with. For regression tasks, commonly used metrics include mean absolute error, root mean square error, and coefficient of determination. These performance metrics on the test set can relatively objectively reflect the model's performance when faced with new and unseen data, and are important criteria for judging whether the model meets practical standards and can be deployed in real-world environments. If the model performs poorly on the test set, it may indicate overfitting, underfitting, or a mismatch in data distribution. In such cases, it's necessary to backtrack and check data quality, feature engineering, or model structure, and then make further adjustments and retrain. The entire training, validation, and testing process constitutes a complete iterative cycle until the model performance meets the preset deployment requirements.

[0035] Example 3: See Figure 3Generating personalized consultation plans based on the analysis results of a health assessment model is a progressive process that deeply integrates machine inference with structured medical knowledge. This process begins with a deep analysis of the model's output. The health assessment model typically outputs a multi-dimensional vector or structured object, which may include probabilistic judgments of potential health problems, scores for the degree of abnormality of a series of key physiological indicators, and the strength of association between different symptom combinations and known disease patterns. The focus of the analysis is to identify a few of the most indicative key diagnostic indicators from these complex outputs. For example, the model might output that the diagnostic hypothesis of "migraine" has the highest confidence, while marking "duration" and "pain nature" as high-influence features. These refined indicators form the coordinates for navigating the vast ocean of medical knowledge. After identifying the key diagnostic indicators, the system begins to traverse and retrieve data from a pre-constructed medical knowledge graph. The medical knowledge graph can be viewed as a vast semantic network, where nodes represent medical concept entities, and edges represent the relationships between entities. The retrieval process typically begins with key diagnostic indicators. For example, using "migraine" as the central node, the system first locates the corresponding disease node in the atlas. Then, based on pre-defined inference rules and path strategies, the system traverses multiple relational dimensions directly or indirectly connected to the central node. These relations may include "has symptoms," "may have complications," "requires examination," "preferred drug treatment," "non-drug therapy," "contraindicated drugs," and "lifestyle recommendations." Through this atlas traversal, the system can obtain a set of medical knowledge nodes with varying degrees of correlation, centered on key diagnostic indicators. For example, nodes such as "ibuprofen," "triptans," "electroencephalogram," and "relaxation training," along with their attribute information, will be retrieved simultaneously.

[0036] To filter the most relevant content from a large number of retrieved nodes, the system needs to calculate the matching degree between each medical knowledge node and the user's complete health data background. This calculation considers not only the correlation strength between the node and the core diagnosis, but also the user's personalized attributes. The matching degree calculation can be quantified using a function, for example:

[0037]

[0038] in: Represents specific medical knowledge nodes With users Overall matching degree. Represents a node With core diagnostics The standard correlation strength between them, which is usually defined by medical experts or derived from authoritative literature when constructing a knowledge graph. Represents a node With user profile The compatibility score is based on personal profiles that include constraints such as age, gender, allergy history, comorbidities, and liver and kidney function. and These are weighting coefficients used to balance the relative importance of disease association strength and personalized constraints in the final match. For example, a certain drug may be very effective for the core diagnosis ( If the value is high, but the user profile shows that they have contraindications to the drug (such as severe liver disease), then their The value will be extremely low, resulting in a low final match rate. The score is not high enough, so it is downgraded during the screening process.

[0039] After the matching degree calculation is completed, the construction of personalized health consultation plans enters the substantive stage. The system first sorts all relevant medical knowledge nodes in descending order based on the calculated matching degree, selecting the highest-priority subset of nodes as the candidate content library for consultation plans. Next comes the personalized fine-tuning of the plan. The system deeply integrates with the user's personal health record, adaptively adjusting the specific parameters in the selected treatment plan nodes. For example, for a medication node, the system calculates a suitable initial dosage range for the user based on their age, weight, and kidney function indicators, referring to the algorithm in the drug's instructions; for an exercise suggestion node, it considers the user's age and physical condition, recommending exercise intensity and frequency. This adjustment transforms the plan from general medical advice into highly individualized guidance. Based on the determined content and parameters, the system generates a structured consultation plan framework. This framework typically follows a certain clinical logic, including modules such as "Assessment Conclusion," "Diagnostic Basis Analysis," "Treatment Plan," "Medication Guidance," "Lifestyle Intervention," "Follow-up Recommendations," and "Risk Warnings." The content of each module is filled with the previously selected and adjusted relevant medical knowledge node information. For example, the "Treatment Plan" module integrates information on drug treatment, physical therapy, and other related nodes, while the "Lifestyle Intervention" module integrates suggestions on diet, exercise, stress management, and other related nodes.

[0040] The final step is to transform this structured framework into user-friendly natural language text. This process is handled by the Natural Language Generation (NLG) module, which uses predefined templates, grammatical rules, and contextualization techniques to organize discrete data points and logical relationships into coherent, fluent paragraph-style text. For example, it transforms the structured data "Drug: Ibuprofen; Dosage: 200 mg; Frequency: Oral if necessary, no more than 3 times daily" into "When a headache occurs, you may consider taking ibuprofen orally to relieve pain, at a dose of 200 mg each time. Please note that if the pain persists, the frequency of administration should not exceed three times daily." NLG technology also ensures that the generated text has appropriate tone and clear expression, and adjusts the use of terminology according to the user's education level, thus producing a truly personalized and highly readable health consultation plan. The entire process starts with machine-readable data, and through multiple steps of reasoning, matching, adjustment, and language generation, ultimately delivers a humanized and professional health consultation report to the user.

[0041] Example 4: After pushing a personalized health consultation plan to the user's terminal device, the system will proactively open a feedback window, inviting the user to evaluate the received plan. Feedback channels are designed to be as convenient and diverse as possible. For example, a five-star rating slider is placed at the bottom of the plan display page, next to which is an optional text input box. The label might be "Your opinion is very important to us" or "Please share your specific feelings." Additionally, the system may send a brief follow-up link via push notification or SMS 24 hours after the plan is sent, inquiring whether the user's symptoms have improved or whether the plan is clear and feasible. Users can submit structured rating data and unstructured text comments through these channels. Received feedback information is categorized and stored by the system. Structured data, such as five-star ratings (e.g., 4 stars), is directly recorded and associated with the corresponding consultation session ID. Unstructured textual reviews require more in-depth analysis. For example, a user might write: "The plan suggests I do more aerobic exercise to relieve anxiety, which is the right direction. However, the types of exercise mentioned are not suitable for someone like me who has injured their knee. I hope to get more specific suggestions that put less stress on my knee joint. Also, the part about diet is rather general." This text contains the user's positive attitude ("the direction is right") and specific suggestions for improvement.

[0042] Analyzing user satisfaction data in feedback begins with extracting the text evaluation content. The system retrieves the raw text comments submitted by users from the database and performs basic cleaning and preprocessing, such as removing irrelevant symbols and correcting obvious spelling errors. Next, sentiment analysis is performed on the cleaned text. This process uses a trained sentiment analysis model that understands the overall sentiment tendency of the text and quantifies it as a numerical value, such as a continuous value from -1 (extremely negative) to +1 (extremely positive). In the example above, the model might analyze that the text contains both positive and constructive criticisms, ultimately giving a slightly positive sentiment value, such as +0.3. This numerical value alone cannot fully reflect the details of the issue, so the next step is to identify the key opinion points in the text evaluation content. By using text mining techniques, such as rule-based or machine learning-based keyword / key phrase extraction and opinion mining, the system needs to find the specific objects the user is discussing (i.e., the evaluation entities) and their corresponding opinions within the paragraphs. In the above comments, key points of opinion may be identified as follows: Regarding "exercise method", users expressed a negative opinion that it was "inappropriate" and provided the constraint that "the knee has been injured"; Regarding "dietary advice", users expressed a negative opinion that it was "too general".

[0043] Mapping sentiment scores and key opinions to model improvement dimensions is a crucial step in connecting user subjective feedback with system technical optimization. This step requires a pre-defined mapping rule base. The rule base defines how to transform user opinions described in natural language into specific, actionable model or knowledge base improvement items. For example, the rule base might stipulate that when a comment mentions "exercise method" and relates it to specific physical conditions such as "knee injury," the opinion should be mapped to the "personalized exercise advice" dimension; when it mentions "dietary advice" and relates it to "general," it should be mapped to the "detailedness of dietary advice" dimension. The positive or negative value and magnitude of the sentiment score indicate whether the dimension needs to be strengthened or maintained, and the urgency of improvement. A negative sentiment score usually means that the corresponding dimension is deficient and needs priority improvement. Through this mapping, vague textual feedback is transformed into clear improvement signals pointing to specific functional modules. See Table 1, which shows a simplified mapping of feedback analysis results.

[0044] Table 1: Mapping of User Feedback Analysis Results

[0045] After mapping is complete, the system does not immediately modify the running model. Instead, it stores the analysis results, including the mapped improvement dimensions, urgency levels, and original feedback snippets, in a dedicated feedback analysis database. Once this data accumulates to a certain scale, such as weekly or monthly, it is reviewed in batches by the system administrator or algorithm engineer. The review process focuses on frequently occurring improvement dimensions and high-urgency issues. For example, if the "personalized exercise recommendations" dimension is mentioned multiple times in a short period and is often associated with specific constraints such as "knee injuries" or "lumbar spine problems," this provides a clear signal that the nodes in the knowledge graph related to exercise therapy need to be expanded to include alternative exercise programs for different physical limitations. Similarly, if "detailed dietary recommendations" is frequently mentioned, it may be necessary to enhance the template library of the natural language generation module in this domain, providing more specific recipe examples or food selection lists.

[0046] This feedback-based optimization mechanism enables the system to transcend static, one-off model deployments and evolve into an intelligent agent capable of continuous learning from real-world interactions. Every user rating and comment, whether positive or negative, becomes crucial fuel driving system evolution. Positive feedback helps confirm the correctness of the current path, while constructive criticism accurately reveals system weaknesses and potential user needs, pointing the way for subsequent iterations and upgrades. By systematically transforming unstructured text feedback into structured improvement instructions, the system establishes an effective closed loop, ensuring that its health consultation services increasingly align with users' true expectations and individual circumstances.

[0047] Example 5: The model parameter update operation is based on the user feedback analysis completed in Example 4. After the system determines the model improvement dimensions that need adjustment through analysis, such as the "personalized exercise suggestions" dimension being marked as high urgency, the next step is to determine the range of model parameters that need adjustment. The health assessment model, as a deep neural network, contains millions or even billions of parameters distributed across different layers of neuron connections. The system needs to analyze which network layer or parameter groups have the greatest impact on the "personalized exercise suggestions" output feature. For example, analysis may reveal that the neuron parameters in the final output layer responsible for determining "whether exercise suggestions are needed," and the parameters in a hidden layer adjacent to the output layer that handle the "user's exercise contraindications" feature, are the most closely related to this improvement dimension. This process of determining the parameter range relies on the interpretability analysis of the model's internal connection weights and activation patterns to ensure the accuracy of the adjustments and avoid unnecessary interference with other unrelated functions of the model. After defining the target parameter range, the system calculates the specific adjustment amount for each parameter to be adjusted. The calculation of adjustments is not arbitrary but follows certain rules. The core principle is that the magnitude of the adjustment is directly proportional to the magnitude of the user feedback sentiment value that triggered the adjustment. Suppose the system receives negative feedback on exercise suggestions, with a calculated sentiment value of -0.8 (range -1 to +1), indicating a high level of user dissatisfaction. The system sets a base adjustment rate and then uses the absolute value of the sentiment value (0.8) as a multiplier for the adjustment intensity. For weighted parameters that need to be downgraded (i.e., those that lead to inappropriate general exercise suggestions), the adjustment is multiplied by 0.8 from the base adjustment rate, resulting in a relatively larger adjustment. For weighted parameters that need to be upgraded (i.e., those that should have triggered the system to consider the user's knee injury history but failed to do so sufficiently), the adjustment is also increased proportionally. This proportional relationship ensures that strong negative feedback triggers more significant model corrections, while weak or positive feedback only causes minor optimizations or maintains stability.

[0048] The parameter tuning process must be fine-tuned while ensuring the overall stability of the health assessment model; this is a crucial constraint. Making large-scale modifications to a large number of parameters at once can lead to unpredictable and drastic changes in model behavior, or even cause the model to "forget" previously learned knowledge. Therefore, in practice, a conservative optimization strategy is usually adopted, such as using a very small learning rate for a single or a few gradient descent steps. A common practice is to select a small batch of recent high-weight feedback cases (including user query data, original model output, and feedback results) relevant to the improvement dimension from the feedback analysis database, forming a mini-training batch. Then, for the previously determined target parameter range, a forward and backward propagation is performed on this mini-batch with a learning rate much lower than the initial training rate, to calculate the gradient and update the parameters. This "fine-tuning" method is like meticulously sculpting the model, modifying only the parts that need adjustment while preserving the model's overall knowledge architecture and generalization ability to the greatest extent. Each parameter update operation is recorded in detail by the system, forming a complete version log. The recorded information includes, but is not limited to: the timestamp of this update, the set of main feedback case IDs that triggered the update, a description of the range of parameters adjusted, the learning rate used, and a comparison of the model's performance changes on a fixed validation dataset before and after the adjustment (e.g., minor changes in accuracy and recall). These records are stored in a model repository, which not only saves the parameter files of the currently active model but also maintains a clear model evolution history. This version control mechanism provides important traceability and rollback capabilities. If, after an update, monitoring reveals an unexpected decline in model performance in a certain aspect, operations personnel can quickly locate the specific update version and choose to revert to a previous stable state, ensuring the reliability of online services.

[0049] In addition to passive optimization based on user feedback, the system also employs a mechanism for regularly acquiring new knowledge from external sources and proactively updating itself. The system periodically retrieves the latest medical research findings from pre-defined authoritative medical databases. These data sources may include API interfaces of academic literature databases such as PubMed, official websites of authoritative medical journals, or the latest clinical guidelines published by the Centers for Disease Control and Prevention. The retrieved content typically includes structured metadata (such as title, author, abstract, and keywords) and full-text text. The system utilizes natural language processing techniques, particularly information extraction techniques, to identify new medical concepts, new treatment methods, new drug associations, or revisions to existing medical knowledge from these new documents. For example, a newly published randomized controlled trial might demonstrate that a known nonsteroidal anti-inflammatory drug (NSAID) is more effective than previous first-line drugs in treating a specific type of migraine. The identified new knowledge needs to be transformed into a form that can be accommodated by the medical knowledge graph. This process involves creating new medical knowledge nodes or updating the attributes of existing nodes. For example, for the aforementioned new research, the system might need to create a new "treatment effectiveness" node, linked to specific drug and disease nodes, and containing new evidence levels and efficacy data. Simultaneously, it may be necessary to lower the evidence level of existing preferred drug nodes. Updating node relationships in a medical knowledge graph is a rigorous process, typically requiring a semi-automated pipeline, including new knowledge extraction, conflict detection with the existing graph, rule-based or statistical fusion decisions, and finally, automatic or manual review by the system to perform the update operation, ensuring the logical consistency and accuracy of the knowledge graph.

[0050] After the medical knowledge graph is updated, the health assessment model also needs corresponding incremental training to enable it to understand and apply this new knowledge. Incremental training does not involve training from scratch with entirely new data, which is costly and inefficient. Instead, it simulates and generates a new batch of training cases from clinical studies corresponding to the latest medical literature, or directly introduces a batch of recently occurring, anonymized real cases that meet the new guideline standards. Then, this new data containing new knowledge, combined with some historical data, is used to perform an incremental training round on the existing model. During training, a small learning rate is typically used, and the parameters of the last few layers of the model may be unfrozen to allow it to adapt to new knowledge quickly while maintaining the model's ability to extract general features from its underlying layers. Through this continuous, incremental learning, the health assessment model can keep pace with medical advancements, ensuring that its health assessments and recommendations are always based on the most reliable evidence available. The entire implementation described in Example 5 constitutes the core maintenance and evolution mechanism of an intelligent health consultation system that can continuously learn and dynamically evolve from internal feedback and external progress.

[0051] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A remote health consultation method based on artificial intelligence, characterized in that, Includes the following steps: Receive health consultation requests submitted by users through terminal devices, the health consultation requests including the user's basic health information and consultation questions; The basic health information is preprocessed to generate standardized health data records; The standardized health data records are input into the trained health assessment model for analysis. Based on the analysis results of the health assessment model, relevant medical knowledge nodes are retrieved from the medical knowledge graph; Based on the aforementioned medical knowledge nodes and the aforementioned consultation questions, a personalized health consultation plan is constructed. The personalized health consultation plan will be sent to the user's terminal device; The medical knowledge nodes include disease nodes, treatment plan nodes, and medication recommendation nodes; The process of retrieving relevant medical knowledge nodes from a medical knowledge graph includes: Analyze the results output by the health assessment model to determine key diagnostic indicators; Search the medical knowledge graph for disease nodes associated with the key diagnostic indicators; Obtain the treatment plan node and medication suggestion node corresponding to the disease node; Calculate the matching degree between each medical knowledge node and the user's health data.

2. The remote health consultation method based on artificial intelligence according to claim 1, characterized in that, The preprocessing of the basic health information includes: Identify unstructured text data in the basic health information; Extract key health features from the unstructured text data; The key health characteristics were matched with a standard medical terminology database; Generate standardized health data records that include structured health characteristics.

3. The remote health consultation method based on artificial intelligence according to claim 2, characterized in that, The training process of the health assessment model includes: Collect historical health consultation case data, which includes user health information and corresponding medical diagnosis results; Feature engineering was performed on the case data to extract multidimensional health features; Build a deep neural network model and set multiple hidden layers; The deep neural network model is iteratively trained using the multidimensional health features; Verify the prediction accuracy of the trained deep neural network model on the test dataset.

4. The remote health consultation method based on artificial intelligence according to claim 1, characterized in that, The construction of personalized health consultation plans includes: Based on the matching degree of the medical knowledge nodes, select the medical knowledge nodes with the highest priority. Based on the user's personal health record, adjust the treatment plan parameters in the medical knowledge node; Generate a consultation protocol framework that includes diagnostic recommendations, treatment plans, and health guidance; The aforementioned consultation solution framework is processed using natural language generation to form readable text.

5. The remote health consultation method based on artificial intelligence according to claim 4, characterized in that, Also includes: Receive user feedback on the personalized health consultation plan; Analyze the user satisfaction data in the feedback information; The parameters of the health assessment model are updated based on the user satisfaction data; Store the updated parameters of the health assessment model in the model version repository.

6. The remote health consultation method based on artificial intelligence according to claim 5, characterized in that, The analysis of user satisfaction data in the feedback information includes: Extract the text evaluation content from the feedback information; Sentiment analysis is performed on the text evaluation content to calculate the sentiment tendency value; Identify key opinion points in the text evaluation content; The sentiment scores and key opinion points are mapped to dimensions for improving the health assessment model.

7. The remote health consultation method based on artificial intelligence according to claim 6, characterized in that, The parameters for updating the health assessment model include: The range of model parameters that need to be adjusted is determined based on the dimensions for improvement of the health assessment model. Calculate the parameter adjustment amount, which is proportional to the sentiment tendency value; Parameters are fine-tuned while ensuring the overall stability of the health assessment model. Record the version information and adjustment effects of each parameter update.

8. The remote health consultation method based on artificial intelligence according to claim 7, characterized in that, Also includes: Regularly retrieve the latest medical research findings from medical databases; Transform the aforementioned medical research findings into new medical knowledge nodes; Update the node relationships in the medical knowledge graph; The health assessment model is incrementally trained to adapt to new medical knowledge.

9. A remote health consultation system based on artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based remote health consultation method according to any one of claims 1 to 8.

Images