Medical intelligent interaction system and intelligent interaction method

Abstract

The invention belongs to the technical field of artificial intelligence and machine question answering, and particularly discloses a medical intelligent interaction system and an intelligent interaction method. Comprising an AIGC intelligent interaction module, a medical knowledge verification module, a user interaction module and an AI knowledge base. The AIGC intelligent interaction module is used for identifying a user consultation intention and generating corresponding candidate answers; the medical knowledge verification module is used for carrying out speciality, accuracy and compliance verification on candidate answers; the user interaction module is used for receiving multi-modal user input and displaying a final answer result; and the AI knowledge base is used for providing medical knowledge resource support. According to the method, workflow arrangement, multi-model cooperation and automation are carried out through the AIGC application platform, and seamless connection of the whole process from question input by a user to final result output is realized.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and machine question answering technology, and more specifically, relates to a medical intelligent interaction system and intelligent interaction method. Background Technology

[0002] Existing medical question-and-answer systems generally suffer from problems such as lagging knowledge updates, insufficient naturalness of interaction, and difficulty in ensuring the professionalism and security of answers. Furthermore, they are significantly lacking in core features such as cross-platform deployment, high-concurrency processing, and strong scalability. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this application is to provide a medical intelligent interaction system and intelligent interaction method, which aims to solve the problems of lagging knowledge updates, insufficient naturalness of interaction, difficulty in ensuring professionalism and security, and poor deployment and scalability of existing medical question-and-answer systems.

[0004] To achieve the above objectives, in a first aspect, this application provides a medical intelligent interaction system, comprising: an AIGC intelligent interaction module, a medical knowledge verification module, a user interaction module, and an AI knowledge base; wherein, the AIGC intelligent interaction module is used to identify the user's consultation intent and generate corresponding candidate answers; the medical knowledge verification module is used to verify the professionalism, accuracy, and compliance of the candidate answers; the user interaction module is used to receive multimodal user input and display the final answer result; and the AI ​​knowledge base is used to provide medical knowledge resource support.

[0005] Preferably, the medical knowledge verification module includes: a medical knowledge matching unit and a compliance check unit; wherein, the medical knowledge matching unit is used to verify the accuracy of the medical knowledge in the candidate answers; the compliance check unit is used to detect the medical compliance of the candidate answers; the compliance check unit includes: a prohibited content detection unit and a risk warning embedding unit; wherein, the prohibited content detection unit is used to screen for illegal content in the candidate answers, and trigger manual review when illegal content is found; the risk warning embedding unit is used to embed standardized risk-related warnings into the candidate answers.

[0006] Preferably, the AIGC intelligent interaction module includes: an intent recognition unit and a response generation unit; wherein, the intent recognition unit is used to parse the core needs of the user's inquiry; and the response generation unit is used to generate candidate answers based on the AI ​​knowledge base.

[0007] Preferably, the user interaction module includes a multimodal input unit and a result display unit; wherein, the multimodal input unit is used to receive multi-form consultation content from users; and the result display unit is used to present the final, verified answer.

[0008] Preferably, the medical intelligent interaction system further includes: an answer optimization unit, used to process the verified candidate answers into a structured and simplified language, and then display them to the user through the user interaction module.

[0009] To achieve the above objectives, in a second aspect, this application provides a medical intelligent interaction method, comprising the following steps: (1) Users submit basic medical consultation requests in text / voice form through multi-terminal interactive entry points; (2) If it is voice input, convert it to text through the speech recognition API and then proceed to (3); if it is text input, proceed directly to (3). (3) The .NET Core API gateway receives the request, completes the permission verification and rate limiting judgment, and routes it to the AIGC intelligent interaction module; (4) The AIGC intelligent interaction module performs word segmentation, entity recognition and semantic analysis on the question, extracts the core entities in the medical field, determines the user's consultation intent, and the AIGC big language model combines the extracted entities and consultation intent to generate candidate answers. When the candidate answer information is incomplete, it automatically generates follow-up questions to guide the user to supplement. (5) Perform semantic matching between the generated candidate answers and the AI ​​knowledge base to correct knowledge biases; detect prohibited / high-risk content and trigger manual review or embed risk warnings; (6) After the verified answers are processed in a structured and simplified manner, they are displayed to the user through the user interaction module; (7) Record interaction data and drive model fine-tuning and knowledge base updates through incremental learning to achieve iterative optimization of the system.

[0010] Preferably, in step (4), the Agent autonomously plans the dialogue process with a goal-oriented approach, proactively guiding the user to supplement key information, specifically including: (1) Initialization phase: First, the Agent parses the core objective associated with the current dialogue and, based on the preset objective-information mapping library, decomposes the core required information and optional supplementary information required to complete the objective. Then, the key information is sorted by weight algorithm to guide the user to fill in the high-priority required information first. (2) Real-time interaction stage: The Agent connects to the output of the intent classifier, combines NLP semantic parsing technology to extract the effective information in the user's current input, compares it with the list of key information required by the target, calculates the current information gap through the information completeness scoring model, and records the user's dialogue style; (3) Agent decision-making stage: dynamic planning of dialogue path.

[0011] Preferably, the Agent decision-making phase implements dynamic planning of the dialogue path in the following ways: (3.1) Dialogue tree generation: The root node is “Complete all required information”, and each missing key information is a child node. The initial dialogue branch is generated according to the information priority. At the same time, the branch jump rules are preset, and the system displays and receives the text input by the patient’s voice in real time. (3.2) Guiding strategy decision: Based on the information gap type and user dialogue style, the Agent autonomously selects the optimal guiding method from the strategy library. The strategy library contains three autonomous optimal guiding methods: direct inquiry strategy, scenario-related guiding strategy, and step-by-step decomposition guiding strategy. (3.3) Dynamic path adjustment: If the user does not directly reply to the current guidance question, the Agent uses the intent drift detection algorithm to determine whether the user's new input is related to the core objective. If it is related, the Agent answers the user's question first and then returns to the original guidance path. If it is not related, the Agent responds politely first and then uses a lightweight intent recall script to bring the user back to the core objective.

[0012] Preferably, in step (7), incremental learning drives model fine-tuning and knowledge base updates, as follows: (7.1) Construct an interactive data collection system specifically for medical scenarios, covering the entire data chain from "user side to agent side to knowledge base side". Perform targeted data cleaning and structured annotation on the collected raw data for medical scenarios to ensure data quality. (7.2) Construct a multi-dimensional value scoring model and select incremental data that is of high value for model fine-tuning and knowledge base updates; (7.3) Achieve lightweight iteration of the model based on high-value incremental datasets; (7.4) Relying on high-value incremental data, realize the hierarchical dynamic updating of the medical knowledge base to ensure the timeliness and accuracy of knowledge. The specific steps are as follows: (7.5) Construct a multi-dimensional evaluation index system covering three dimensions: model performance, knowledge base quality, and user experience. If the evaluation index meets the preset target, the new version of the model and knowledge base will be launched online, and the next round of interactive data collection will be started. If the target is not met, the reasons will be analyzed, the data screening threshold or model fine-tuning strategy will be adjusted, and the incremental learning process will be re-executed.

[0013] Preferably, step (7.3) includes: (7.3.1) For the network structure of the medical semantic matching model, intent classification model and deviation recognition model, freeze the general semantic encoding layer at the bottom of the model and retain the general medical semantic features that it has learned; only unfreeze the top medical task adaptation layer and the newly added incremental feature fusion layer to ensure that only the parameters related to the new data are updated. (7.3.2) Through the incremental feature fusion layer, the exclusive features of the high-value incremental dataset Dinc are extracted and weighted and fused with the original features of the model; mini-batch gradient descent is adopted, combined with the elastic weight integration algorithm, and weight penalty is applied to the key parameters of the original model during the training process. That is, the new data training causes the model to lose the original medical knowledge. The model is trained until the training termination condition is met, the training is stopped, and the iterative incremental model Minc is generated. (7.3.3) Compare the performance of the incremental model Minc and the original model Mold on the core medical indicators. If Minc performs better than Mold, the incremental model parameters are integrated with the original model parameters through the model version fusion algorithm to generate a new version model Mnew, and the old model is archived. If the performance does not meet the standard, the incremental data is discarded and higher-value data is selected for iteration.

[0014] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: First, in response to the problem of lagging knowledge updates in existing medical question-and-answer systems, this application uses the AIGC application platform to orchestrate workflows, facilitate multi-model collaboration, and automate the process, achieving seamless integration from user input of questions to final output results.

[0015] Secondly, to address the issue of insufficient naturalness in the interaction of existing medical question-and-answer systems, this application uses an agent to autonomously plan the dialogue process in a goal-oriented manner, proactively guiding users to supplement key information rather than responding passively. At the same time, it avoids deviating from the core medical objectives. The prompts adopt the design principle of "accurate and warm," which not only ensures the accuracy and traceability of medical advice through evidence-based anchoring, but also incorporates emotional computing factors to alleviate user anxiety with empathetic language.

[0016] Third, existing medical Q&A systems have issues with ensuring professionalism and security. This application addresses these issues by directly accessing professionally verified content from medical professionals through the medical institution's management backend. This avoids knowledge biases caused by scattered online information and non-authoritative literature, ensuring the professionalism of Q&A from the source. Furthermore, a dynamic update mechanism is established to promptly synchronize the latest clinical guidelines and drug instruction manual revisions, preventing security risks caused by outdated knowledge and ensuring that Q&A suggestions are compliant and controllable.

[0017] Fourth, addressing the issues of poor deployment scalability in existing medical question-and-answer systems, this application leverages native optimizations of .NET 8 (such as AOT compilation, GC improvements, and JIT performance enhancements) and utilizes Android Studio offline packaging technology to reduce package size and enable more custom functionality development. Thanks to .NET 8's extreme performance optimization, default TLS 1.3 enhanced security, and more comprehensive OpenAPI 3.0 support, it adapts to the development of multi-turn dialogue interfaces for medical question-and-answer agents, is compatible with Windows / Linux / macOS platforms, meets the needs of heterogeneous IT environments in hospitals, and natively supports containerized deployment and Kubernetes orchestration, facilitating distributed deployment of the medical knowledge base. These core features solidify the system's underlying capabilities, improve core user experience, adapt to the characteristics of medical scenarios, and reduce operational costs. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the structure of a medical intelligent interactive system provided in an embodiment of this application.

[0019] Figure 2 This is a flowchart of a medical intelligent interaction method provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0022] In this application, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0023] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0024] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0025] The embodiments of this application are described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, this application provides a medical intelligent interaction system, which includes: an AIGC intelligent interaction module, a medical knowledge verification module, a user interaction module, and an AI knowledge base; the AIGC intelligent interaction module is used to identify the user's consultation intent and generate corresponding candidate answers; the medical knowledge verification module is used to verify the professionalism, accuracy, and compliance of the candidate answers; the user interaction module is used to receive multimodal user input and display the final answer result; the AI ​​knowledge base is used to provide medical knowledge resource support.

[0027] Preferably, the medical knowledge verification module includes: a medical knowledge matching unit and a compliance check unit; wherein, the medical knowledge matching unit is used to verify the accuracy of the medical knowledge in the candidate answers; and the compliance check unit is used to detect the medical compliance of the candidate answers.

[0028] Preferably, the compliance check unit includes: a prohibited content detection unit and a risk warning embedding unit; wherein, the prohibited content detection unit is used to screen for illegal content in candidate answers and trigger manual review when illegal content is found; the risk warning embedding unit is used to embed standardized risk-related warnings into candidate answers.

[0029] Preferably, the AIGC intelligent interaction module includes: an intent recognition unit and a response generation unit; wherein, the intent recognition unit is used to parse the core needs of the user's inquiry; and the response generation unit is used to generate candidate answers based on the AI ​​knowledge base.

[0030] Preferably, the user interaction module includes a multimodal input unit and a result display unit; wherein, the multimodal input unit is used to receive various forms of user inquiries; and the result display unit is used to present the verified final answer. This application provides multi-terminal user interaction entry points, supports multiple inquiry methods such as text and voice, receives user input, and displays the final answer.

[0031] The medical intelligent interaction system also includes an answer optimization unit, which is used to process the verified candidate answers into a structured and simplified language before displaying them to the user through the user interaction module.

[0032] In summary, the medical intelligent interaction system proposed in this application is a highly intelligent medical assistance platform. It integrates massive amounts of medical knowledge, clinical data, and cutting-edge generative AI algorithms, enabling in-depth analysis of multi-dimensional information such as patient medical history, symptoms, and examination results to quickly generate personalized diagnostic suggestions and treatment plans. Specifically, this application empowers the system's core through an AI knowledge base. This core provides three core values—efficient knowledge management, intelligent question-and-answer interaction, and business decision support—through a complete "knowledge accumulation-intelligent retrieval-precise output" chain. It offers a vast amount of knowledge resources, including medical literature, clinical guidelines, and expert experience, covering all medical specialties such as internal medicine, surgery, gynecology, and pediatrics. By leveraging cutting-edge generative AI algorithms, it addresses the issues of lagging knowledge updates, insufficient naturalness of interaction, and difficulties in ensuring professionalism and security.

[0033] This system is built on the .NET8 technology platform. Its core value lies in providing high-performance, high-reliability, easy-to-expand, and easy-to-maintain technical support for business scenarios by relying on the native capabilities of .NET8 and enterprise-level engineering practices, covering the entire link from the underlying infrastructure to the upper-level business empowerment. Specifically, this application is developed based on the native optimization of .NET8 (such as AOT compilation, GC improvement, and JIT performance improvement): (1) AOT (Ahead-of-Time) compilation can directly compile C# code into machine code, avoiding runtime JIT compilation overhead. The core module is pre-compiled into native code. The compiled deployment package does not need to depend on the .NET runtime (Self-Contained) and can run directly on low-configuration servers / edge devices (such as domestic ARM architecture servers) in primary medical institutions. The deployment package size is reduced by 40%-60% compared with the traditional mode (such as the core service package). (1) Reduced from 200MB to 80MB); (2) For memory-intensive scenarios such as "high-concurrency consultation, large text semantic processing, and knowledge base vector storage" in medical question-answering systems, several core improvements were introduced to GC (garbage collection): such as generational GC optimization, large object heap (LOH) defragmentation, and asynchronous GC support; (3) For dynamic computing scenarios such as "dynamic semantic matching, Agent dialogue path planning, and incremental learning model fine-tuning" in medical question-answering systems, the JIT compiler (RyuJIT) has made significant improvements in loop optimization, vectorization, and generic code optimization, and the execution efficiency of hot code has been improved by about 20%.

[0034] The system backend is developed based on the .NET8 framework. In response to the problems of poor deployment scalability of the medical Q&A system, the technical effects of this application are verified as follows: (1) Deployment scalability: Before optimization, it only supports deployment on a single platform and cannot be expanded; After implementation, it is compatible with all platforms and K8s supports elastic expansion of 1-10 replicas; (2) Package size: Before optimization, the Android package is 500MB and the deployment package is 200MB; After implementation, the Android package is 150MB and the deployment package is 80MB, a reduction of 60%; (3) Response performance: Before optimization, the single request takes 50ms and the QPS is 500; After implementation, the single request takes 10ms and the QPS is 2000+, an improvement of 300%; (4) Security: Before optimization, TLS1.2 is used, which poses a risk of data leakage; After implementation, TLS1.3 is used by default, and there is zero leakage of medical data transmission; (5) Operation and maintenance cost: Before optimization, manual deployment takes 1 hour to recover from a fault; After implementation, automated deployment is used, and fault self-healing is completed within 5 minutes.

[0035] like Figure 2 As shown, this application provides a medical intelligent interaction method, including the following steps: (1) Users initiate consultations through multiple interactive entry points: Users submit basic medical consultation requests in text / voice form through Web, APP, mini-program and other channels.

[0036] (2) Input processing unit performs input processing: If it is voice input, it is converted into text through the speech recognition API and then enters the subsequent processing; if it is text input, the text input directly enters the subsequent process.

[0037] (3) .NET Core API Gateway performs request routing and verification: After receiving the request, it completes permission verification and rate limiting judgment, and routes the request to the AIGC intelligent interaction module. In this application, permission verification is implemented based on Middleware + policy authorization (IAuthorizationPolicy), combined with JWT / OAuth2.0 to complete authentication, and AIGC-specific permission verification is implemented through custom IAuthorizationRequirement; rate limiting judgment refers to integrating AspNetCoreRateLimit (basic frequency rate limiting), Sentinel.NET (distributed rate limiting / circuit breaking), Polly (circuit breaking degradation), and combining Redis to store rate limiting counts (adapting to distributed deployment); logging and monitoring means that all requests that fail permission verification or trigger rate limiting must record complete logs (caller, time, reason), and connect to a monitoring platform (such as Prometheus + Grafana) to monitor the permission violation and rate limiting trigger frequency of the AIGC interface in real time.

[0038] (4) AIGC intelligent interaction module generates user intent and answer: Natural language processing unit performs word segmentation, entity recognition and semantic analysis on the question, and extracts core entities such as symptoms and drugs; intent classifier determines consultation intent (such as drug consultation); calls AIGC large language model to generate candidate answers, and automatically generates follow-up questions to guide users to supplement if the information is incomplete.

[0039] (5) The medical knowledge verification module performs medical knowledge verification: it performs semantic matching between the generated candidate answers and the AI ​​knowledge base to correct knowledge deviations; it detects prohibited / high-risk content and triggers manual review or embeds risk warnings.

[0040] (6) The answer optimization unit optimizes and provides feedback on the answers: After the verified answers are structured and the language is simplified, they are displayed to the user through the user interaction module.

[0041] (7) Iterative optimization: Record interaction data and drive model fine-tuning and knowledge base updates through incremental learning to improve system performance.

[0042] Preferably, in step (4), the intent classifier determines the consultation intent through a rule engine and a machine learning model, specifically as follows: (1) Pre-filtering rules: Before the data enters the model, data that does not meet the requirements or is of low quality is filtered out by preset rules to reduce the interference of noisy data on model training, improve data quality, and thus improve the performance of the model; (2) Model training and evaluation: The machine learning model is trained using cleaned and preprocessed data, and the performance of the model is measured by evaluation metrics (such as accuracy, recall, F1 score, etc.) so as to find the most suitable model parameters for the current data and task through training and evaluation, and ensure that the model can accurately process new data; (3) Algorithm classification fallback: When the model cannot accurately classify or process certain data, use backup algorithms or rules to process it, so as to improve the robustness of the system and ensure that the system can still provide reasonable output even if the main model fails; (4) Data preprocessing: Clean, transform and standardize the original data to make it suitable for model training, so as to improve the consistency and quality of the data, reduce the complexity of model training, and improve the generalization ability of the model; (5) Manual annotation iteration: supplement and optimize the training data of the model by manually annotating the data, especially when the model performance is insufficient or new situations are encountered, so as to improve the accuracy and adaptability of the model through manual intervention, so that it can better handle complex or edge situations.

[0043] Preferably, in step (4), the Agent autonomously plans the dialogue process with a goal-oriented approach, proactively guiding the user to supplement key information. This algorithm is the first of its kind proposed in this application. Specifically, it includes: (1) Initialization Phase: Goal decomposition and key information definition, goal modeling. Specifically, firstly, the Agent parses the core goal associated with the current dialogue and, based on a pre-set goal-information mapping library, decomposes the core mandatory information and optional supplementary information required to complete the goal. Then, a weighted algorithm is used to sort the key information, guiding the user to prioritize the supplementation of high-priority mandatory information, such as first supplementing "whether married" and then "whether having children". The weights can be determined based on the impact of missing information on goal achievement and the user's input cost.

[0044] (2) Real-time interaction stage: dialogue state perception and information gap identification. Specifically, the Agent connects to the output of the intent classifier and combines NLP semantic parsing technology (such as dependency parsing and entity recognition) to extract the effective information in the user's current input. It compares the information with the list of key information required by the target and calculates the current information gap through the information completeness scoring model. The formula is: Information completeness = (number of required information obtained / total number of required information) * 100% to quantify the information gap. At the same time, it marks the type of missing information (such as parameter missing, permission missing, scenario missing) and records the user's dialogue style (such as concise, detailed, and questioning) to prepare for the subsequent guidance script adaptation.

[0045] (3) Agent Decision-Making Stage: Dynamic Planning of Dialogue Path. Specifically, this application proposes for the first time a multi-branch dynamic dialogue tree algorithm, allowing the Agent to autonomously plan the guidance path. The core logic is as follows: (3.1) Dialogue tree generation: The root node is “Complete all required information”, and each missing key information is a child node. The initial dialogue branch is generated according to the information priority. At the same time, the branch jump rules are preset. For example, if the user mentions “voice input”, the system will jump to the “voice guidance” branch. The system displays and receives the text input by the patient’s voice in real time, instead of requiring the patient to click.

[0046] (3.2) Guiding Strategy Decision: Based on the type of information gap and the user's dialogue style, the Agent autonomously selects the optimal guidance method from the strategy library. The strategy library contains three core strategies: direct inquiry strategy, scenario-related guidance strategy, and step-by-step decomposition guidance strategy. Among them, the direct inquiry strategy is suitable for scenarios where the information gap is clear and the user expresses themselves directly, for example: "Please click / tell me the part of your body that you feel uncomfortable"; the scenario-related guidance strategy is suitable for scenarios where the user's definition of information is unclear, for example: when an elderly patient searches for the location of the internal medicine department, the system not only displays the route, but also repeatedly prompts with voice "There is an accessible elevator 5 meters ahead"; the step-by-step decomposition guidance strategy is suitable for scenarios with complex information supplementation, for example: "Is your headache relieved in a quiet environment: ① Slightly relieved; ② More severe, do you need emergency help from professional medical staff? The background has synchronized the generated medical information to the attending physician."

[0047] (3.3) Dynamic path adjustment: If the user does not directly reply to the current guiding question (such as changing the topic), the Agent uses the intent drift detection algorithm to determine whether the user's new input is related to the core objective. If it is related (such as the user asking "Why do I need to upload personal information"), the Agent answers the user's question first and then returns to the original guiding path. If it is not related (such as the user asking "How is the weather today"), the Agent responds politely first and then uses a lightweight intent recall script to bring the core objective back, for example: "The current temperature is 20℃, but in order to help you complete your medical visit as soon as possible, please provide some personal information so that the doctor in the background can make a diagnosis."

[0048] Preferably, in step (5), the medical knowledge matching unit uses a unit model deep learning algorithm to perform semantic matching between candidate answers and the AI ​​knowledge base, correcting knowledge biases. This algorithm is proposed for the first time in this application, and is detailed below: (5.1) Medical knowledge base preprocessing and semantic modeling; (5.2) Semantic encoding and preliminary matching of candidate answers: The semantic similarity between CandidateEmbedding and KnowledgeEmbedding is calculated using a weighted cosine similarity algorithm.

[0049] in: For the first The weights of individual medical entities are assigned, for example, based on the importance of terms, the weight of disease name is determined to be 0.8, the weight of symptom is 0.6, and the weight of medication is 0.7. The first candidate answer Vector values ​​of each entity; For the first in the knowledge base The vector values ​​of each entity. The number of entities is used. Knowledge base entries with a similarity ≥ a preset threshold (e.g., 85%) are selected as the candidate matching set. (5.3) Knowledge bias identification and hierarchical correction; (5.4) Matching effect evaluation and model iteration.

[0050] Preferably, the specific algorithm for driving model fine-tuning and knowledge base updating through incremental learning in step (7) is proposed for the first time in this application, and is as follows: (7.1) End-to-end acquisition and structured storage of interactive data. Specifically, this includes: (7.1.1) Construct a dedicated interactive data collection system for medical scenarios, covering the entire data chain from the user side to the agent side to the knowledge base side. Specifically, this includes: User interaction data, such as user consultation text, supplementary key information (e.g., symptom duration, past medical history), satisfaction rating of responses (1-5 points), manual error correction feedback (e.g., "This medication recommendation does not conform to the guidelines"), and frequently asked follow-up questions; Agent interaction process data, such as intent classification results, dialogue path planning records, candidate answer generation logs, knowledge matching deviation types (terminology / logic / timeliness), and correction strategy execution records; System-side evaluation data, such as knowledge matching similarity, deviation identification accuracy, user interaction rounds, and response time, etc.

[0051] (7.1.2) Perform targeted data cleaning and structured annotation on the collected raw data for medical scenarios to ensure data quality. Specifically, this includes: data noise reduction to filter invalid data, such as meaningless chatter and repeated questions; correcting typos and colloquial expressions, such as standardizing "heartache" to "chest pain"; labeling the data with multiple layers based on medical annotation standards, including intent labels (such as "disease diagnosis consultation" and "medication guidance consultation"), deviation labels (such as "terminology deviation - secondary hypertension" and "logic deviation - medication dosage error"), and feedback labels (such as "satisfied" and "medication advice needs to be corrected"); and using a dual-database architecture of "time-series database + vector database" for structured storage. The time-series database stores interaction process logs (supporting traceability by time dimension), and the vector database stores text semantic vectors (supporting quick retrieval of similar cases), and establishes an association index between data, model, and knowledge base.

[0052] (7.2) Construct a multi-dimensional value scoring model to select incremental data with high value for model fine-tuning and knowledge base updates, avoiding interference from invalid data. This application proposes for the first time a value-stratification algorithm for medical interaction data, the specific steps of which are as follows: (7.2.1) Data value quantification and scoring, the formula is as follows:

[0053] Wherein, V represents the value score calculated for a single piece of interactive data. V1 represents the bias correlation, ranging from 0 to 1, which determines whether the data contains knowledge bias cases (such as terminology errors or logical contradictions). If there is a bias, the score approaches 1. V2 represents user attention, ranging from 0 to 1, which is determined based on the user satisfaction rating and the frequency of follow-up questions. If the user satisfaction is low and the follow-up questions are frequent, the score approaches 1. V3 represents knowledge scarcity, ranging from 0 to 1, which determines whether the medical scenario corresponding to the data is a weak area covered by the knowledge base (such as rare disease consultation). The higher the scarcity, the higher the score. V4 represents the timeliness weight, ranging from 0 to 1, which determines whether the data involves the latest clinical guideline updates. If it does, the score approaches 1. The weighting coefficient is dynamically adjusted based on the experience of medical experts and the needs of model iteration. The weight of V1 can be increased in the early stages of iteration to enhance the ability to correct biases.

[0054] (7.2.2) High-value data screening and clustering. A value score threshold (e.g., 0.7) is set to screen out the high-value incremental dataset Dinc. Using an improved K-means clustering algorithm, Dinc is clustered according to "disease type + deviation type + interaction scenario," generating several incremental data subsets, such as a diabetes medication deviation subset and a rare disease consultation subset, providing data support for subsequent scenario-based fine-tuning. The improved K-means clustering algorithm introduces semantic distance for medical terms. Traditional K-means centroids are based on the sample mean, which cannot adapt to the semantic characteristics of medical terms; therefore, it is improved to use a medical semantic weighted centroid.

[0055] in, For the first The updated centroids of each cluster are semantically weighted centroids, not the conventional mean. For the first The set of all samples contained in a cluster For the first The initial mean centroids of each cluster are determined by using random samples in the first iteration and the centroids updated in the previous iteration thereafter. This represents the similarity weight of a sample to the centroid. The smaller the distance, the greater the weight, in order to strengthen the influence of the core sample on the centroid.

[0056] (7.3) Incremental learning-driven lightweight model fine-tuning. This application proposes for the first time a medical knowledge-enhanced incremental fine-tuning algorithm, based on a high-value incremental dataset, to achieve lightweight model iteration and avoid the high cost of full retraining. The specific process is as follows: (7.3.1) Model layer freezing and parameter isolation. For the network structure of the medical semantic matching model, intent classification model and deviation recognition model, this application adopts a layered parameter freezing strategy. The bottom general semantic encoding layer of the model (such as the bottom 10 layers of MedBERT) is frozen to retain the general medical semantic features that it has learned; only the top medical task adaptation layer (such as the terminology attention layer and deviation classification output layer) and the newly added incremental feature fusion layer are unfrozen to ensure that only the parameters related to the new data are updated, thereby reducing the consumption of computing resources.

[0057] (7.3.2) Incremental Feature Fusion and Model Training. Specifically, this includes: Incremental feature extraction: Through the incremental feature fusion layer, extract the exclusive features of the high-value incremental dataset Dinc (such as novel drug use bias features and rare disease symptom association features), and perform weighted fusion with the original features of the model; Lightweight training strategy: Use Mini-Batch Gradient Descent (Mini-BatchSGD) combined with the Elastic Weight Integration (EWC) algorithm to apply weight penalties to the key parameters of the original model during training to avoid "catastrophic forgetting," that is, the model loses its original medical knowledge due to training on new data; Training termination condition: When the bias identification accuracy on the validation set increases by ≥1% and the user satisfaction score increases by ≥0.5 points, stop training and generate the iterative incremental model Minc.

[0058] (7.3.3) Model Performance Evaluation and Version Fusion. This application compares the performance of the incremental model Minc and the original model Mold on core medical indicators, mainly in terms of deviation identification accuracy, knowledge matching accuracy, and user interaction rounds. If Minc outperforms Mold, the incremental model parameters are integrated with the original model parameters using a model version fusion algorithm to generate a new version model Mnew, and the old model is archived; if the performance is not satisfactory, the incremental data for that round is discarded, and higher-value data is selected for iteration.

[0059] (7.4) Dynamic updates of the knowledge base driven by incremental data. This application relies on high-value incremental data to achieve hierarchical dynamic updates of the medical knowledge base, ensuring the timeliness and accuracy of the knowledge. The specific steps are as follows: (7.4.1) Knowledge base update triggering mechanism. Specifically, three types of update triggering conditions are set, and the knowledge base update will be initiated when any one of the triggering conditions is met: ① Data triggering conditions, such as the incremental data volume of "2024 version of hypertension guidelines medication update" reaching a preset threshold (e.g., 100 entries); ② Performance triggering conditions, such as the model's bias identification accuracy in a certain medical scenario (e.g., diabetes diagnosis and treatment) decreasing by ≥2% for 3 consecutive rounds; ③ Manual triggering conditions, such as medical experts submitting authoritative materials such as the latest clinical guidelines and rare disease diagnosis and treatment standards.

[0060] (7.4.2) Knowledge Base Layered Update Strategy. This application is based on a three-layer architecture of the medical knowledge base (terminology graph layer - scenario knowledge layer - time-sensitive evidence layer), and performs differentiated updates: When the knowledge base layer is at the terminology graph layer, the update content is: adding new medical terms and supplementing terminology relationships. Specifically, this involves: extracting uncollected terms (such as new disease names and new drug names) from incremental data, incorporating them into the terminology graph based on medical expert review, and supplementing synonyms and hyponyms / hypernyms. For example, "GLP-1 receptor agonist" is associated with "semaglutide". When the knowledge base layer is at the scenario knowledge layer, the update content is: correcting erroneous knowledge and supplementing scarce scenario knowledge. Specifically, this involves: correcting erroneous knowledge and supplementing scarce scenario knowledge. The process involves: replacing erroneous content in the knowledge base for logically flawed cases (such as "medication errors in diabetic nephropathy") in incremental data; supplementing corresponding diagnosis and treatment procedures and symptom identification points for scarce scenario data (such as rare disease consultations); and updating the knowledge base at the time-sensitive evidence level by updating the knowledge evidence level and labeling it with a time-sensitive tag. Specifically, the process involves updating the evidence level tag of the latest guideline content (such as the 2024 edition of the "Guidelines for the Prevention and Treatment of Type 2 Diabetes in China") in the incremental data and labeling it with the time-sensitive tag "2024 Guideline Update," while eliminating outdated knowledge (such as the 2018 edition of the old medication regimen).

[0061] (7.4.3) Consistency verification after update. After the knowledge base is updated, this application uses a medical knowledge consistency verification algorithm to verify whether there is a logical contradiction between the updated content and the existing knowledge (such as whether the contraindications of the new drug conflict with other drugs). After the verification is passed, the semantic embedding vector of the knowledge base is updated synchronously to ensure the accuracy of subsequent knowledge matching.

[0062] (7.5) Evaluation of iterative effects and closed-loop iteration.

[0063] (7.5.1) Multi-dimensional evaluation index system. This application constructs an iterative effect evaluation index specifically for medical scenarios, covering three dimensions: model performance, knowledge base quality, and user experience. Among them, model performance indicators include the improvement rate of deviation identification accuracy, the improvement rate of knowledge matching accuracy, the F1 score of intent classification, and the change rate of model inference latency; knowledge base quality indicators include the knowledge timeliness coverage rate (the proportion of the latest guideline content), the terminology standardization rate, and the knowledge logic consistency pass rate; user experience indicators include the improvement rate of user satisfaction rating, the decrease rate of average interaction rounds, and the decrease rate of manual error correction feedback.

[0064] (7.5.2) Closed-loop iterative optimization. If the evaluation indicators reach the preset target, such as the deviation identification accuracy rate is improved by ≥3%, the new version of the model and knowledge base will be launched online, and the next round of interactive data collection will be started. If the target is not met, the reasons will be analyzed, such as insufficient value of incremental data or unreasonable fine-tuning parameter settings. The data screening threshold or model fine-tuning strategy will be adjusted, and the incremental learning process will be re-executed.

[0065] Example Scenario: A user consults a medical intelligent interactive system about "a cold with cough and phlegm that has lasted for 2 days, what medicine should I take?"

[0066] User operation: Open the medical intelligent interactive system, click on the human body part diagram, and the background will initiate a consultation on the selected part, which will last for ten rounds of question and answer. You can also take pictures of relevant examination and test reports and upload them for clear pathological analysis.

[0067] Request processing: The API gateway receives the request, verifies it, and then routes it to the AIGC intelligent interaction module.

[0068] Semantic parsing: The natural language processing unit segments the question into words (cold / cough / accompanied by / sputum / lasting / 2 days / should / eat / what / medicine), entity recognition extracts "cold", "cough", "sputum", "2 days" and "medicine", and the intent classifier determines it as "medicine consultation".

[0069] Candidate answer generation: The AIGC large language model calling unit combines the extracted entities and intents, and calls Tongyi Qianwen 3.0 to generate candidate answers: "For colds and coughs with phlegm that last for 2 days, you can choose medications such as compound cold medicine granules and ambroxol oral solution. Compound cold medicine granules can relieve symptoms such as cough and fever caused by colds, and ambroxol oral solution can dilute phlegm and promote phlegm discharge."

[0070] Medical knowledge verification: 1) Knowledge matching: Match candidate answers with a structured knowledge base with a similarity of 92% (above the threshold of 85%), no correction is required; 2) Prohibited Detection: No illegal content was detected; 3) Risk warning embedded: Automatically add "This answer is for basic health consultation only and does not constitute medical advice. Please read the instructions or follow your doctor's advice before using the medication. If symptoms worsen, please seek medical attention promptly."

[0071] Answer Optimization: The structured optimization is as follows: "Core Conclusion: For colds with cough and phlegm (lasting 2 days), Compound Cold Relief Granules and Ambroxol Oral Solution can be taken under the guidance of a doctor; Key Basis: Compound Cold Relief Granules relieve cold-related cough symptoms, and Ambroxol Oral Solution promotes phlegm expectoration; Supplementary Suggestions: Read the instructions before use to avoid allergy risks, drink plenty of water, and eat a light diet; Risk Warning: This answer is only a basic health consultation and does not constitute medical advice. Please read the instructions or follow your doctor's advice before use. If symptoms worsen, please seek medical attention promptly." The language has been simplified and no complex terminology has been removed.

[0072] Results Display: The optimized response is displayed to the user in a professional medical record format report. At the same time, the system records the interaction data for subsequent model optimization.

[0073] It is understood that the detailed functional implementation of each of the above units / modules can be found in the description in the aforementioned method embodiments, and will not be repeated here.

[0074] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0075] Based on the methods in the above embodiments, this application provides an electronic device that may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor may invoke logical instructions stored in the memory to execute the methods in the above embodiments.

[0076] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0077] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0078] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0079] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0080] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0081] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0082] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0083] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A medical intelligent interactive system, characterized in that, include: The system comprises an AIGC intelligent interaction module, a medical knowledge verification module, a user interaction module, and an AI knowledge base; wherein, the AIGC intelligent interaction module is used to identify the user's consultation intent and generate corresponding candidate answers; the medical knowledge verification module is used to verify the professionalism, accuracy, and compliance of the candidate answers. The user interaction module is used to receive multimodal user input and display the final answer result; the AI ​​knowledge base is used to provide medical knowledge resource support.

2. The system as described in claim 1, characterized in that, The medical knowledge verification module includes a medical knowledge matching unit and a compliance check unit; wherein, the medical knowledge matching unit is used to verify the accuracy of the medical knowledge in the candidate answers; the compliance check unit is used to detect the medical compliance of the candidate answers; the compliance check unit includes a prohibited content detection unit and a risk warning embedding unit; wherein, the prohibited content detection unit is used to screen for illegal content in the candidate answers, and triggers manual review when illegal content is found; the risk warning embedding unit is used to embed standardized risk-related warnings into the candidate answers.

3. The system as described in claim 1, characterized in that, The AIGC intelligent interaction module includes an intent recognition unit and a response generation unit; wherein, the intent recognition unit is used to parse the core needs of the user's inquiry; and the response generation unit is used to generate candidate answers based on the AI ​​knowledge base.

4. The system as described in claim 1, characterized in that, The user interaction module includes a multimodal input unit and a result display unit; wherein, the multimodal input unit is used to receive multi-form consultation content from users; and the result display unit is used to present the final, verified answer.

5. The system as described in claim 1, characterized in that, The medical intelligent interaction system also includes an answer optimization unit, which is used to process the verified candidate answers into a structured and simplified language before displaying them to the user through the user interaction module.

6. A medical intelligent interaction method, characterized in that, Includes the following steps: (1) Users submit basic medical consultation requests in text / voice form through multi-terminal interactive entry points; (2) If it is voice input, convert it to text through the speech recognition API and then proceed to (3); if it is text input, proceed directly to (3). (3) The API gateway receives the request, completes the permission verification and rate limiting judgment, and routes it to the AIGC intelligent interaction module; (4) The AIGC intelligent interaction module performs word segmentation, entity recognition and semantic analysis on the question, extracts the core entities in the medical field, determines the user's consultation intent, and the AIGC big language model combines the extracted entities and consultation intent to generate candidate answers. When the candidate answer information is incomplete, it automatically generates follow-up questions to guide the user to supplement. (5) Perform semantic matching between the generated candidate answers and the AI ​​knowledge base to correct knowledge biases; Detect prohibited / high-risk content, triggering manual review or embedding risk warnings; (6) After the verified answers are processed in a structured and simplified manner, they are displayed to the user through the user interaction module; (7) Record interaction data and drive model fine-tuning and knowledge base updates through incremental learning to achieve iterative optimization of the system.

7. The method as described in claim 6, characterized in that, In step (4), the Agent autonomously plans the dialogue process with a goal-oriented approach, proactively guiding the user to supplement key information, specifically including: (1) Initialization phase: First, the Agent parses the core objective associated with the current dialogue and, based on the preset objective-information mapping library, decomposes the core required information and optional supplementary information required to complete the objective. Then, the key information is sorted by weight algorithm to guide the user to fill in the high-priority required information first. (2) Real-time interaction stage: The Agent connects to the output of the intent classifier, combines NLP semantic parsing technology to extract the effective information in the user's current input, compares it with the list of key information required by the target, calculates the current information gap through the information completeness scoring model, and records the user's dialogue style; (3) Agent decision-making stage: dynamic planning of dialogue path.

8. The method as described in claim 7, characterized in that, The Agent decision-making phase achieves dynamic planning of the dialogue path in the following ways: (3.1) Dialogue tree generation: With "Complete all required information" as the root node and each missing key information as a child node, the initial dialogue branch is generated according to the information priority; Simultaneously, the system presets branch jump rules and displays and receives text input by the patient's voice in real time; (3.2) Guiding strategy decision: Based on the information gap type and user dialogue style, the Agent autonomously selects the optimal guiding method from the strategy library. The strategy library contains three autonomous optimal guiding methods: direct inquiry strategy, scenario-related guiding strategy, and step-by-step decomposition guiding strategy. (3.3) Dynamic path adjustment: If the user does not directly reply to the current guidance question, the Agent uses the intent drift detection algorithm to determine whether the user's new input is related to the core objective. If it is related, the Agent answers the user's question first and then returns to the original guidance path. If it's irrelevant, respond politely first, then use a lightweight, targeted recall message to bring the core objective back.

9. The method as described in claim 6, characterized in that, In step (7), incremental learning drives model fine-tuning and knowledge base updates, as follows: (7.1) Construct an interactive data collection system specifically for medical scenarios, covering the entire data chain from "user side to agent side to knowledge base side". Perform targeted data cleaning and structured annotation on the collected raw data for medical scenarios to ensure data quality. (7.2) Construct a multi-dimensional value scoring model and select incremental data that is of high value for model fine-tuning and knowledge base updates; (7.3) Achieve lightweight iteration of the model based on high-value incremental datasets; (7.4) Relying on high-value incremental data, realize the hierarchical dynamic updating of the medical knowledge base to ensure the timeliness and accuracy of knowledge. The specific steps are as follows: (7.5) Construct a multi-dimensional evaluation index system covering three dimensions: model performance, knowledge base quality, and user experience. If the evaluation index meets the preset target, the new version of the model and knowledge base will be launched online, and the next round of interactive data collection will be started. If the target is not met, the reasons will be analyzed, the data screening threshold or model fine-tuning strategy will be adjusted, and the incremental learning process will be re-executed.

10. The method as described in claim 9, characterized in that, Step (7.3) includes: (7.3.1) For the network structure of the medical semantic matching model, intent classification model and deviation recognition model, freeze the general semantic encoding layer at the bottom of the model and retain the general medical semantic features that it has learned; only unfreeze the top medical task adaptation layer and the newly added incremental feature fusion layer to ensure that only the parameters related to the new data are updated. (7.3.2) Through the incremental feature fusion layer, the exclusive features of the high-value incremental dataset Dinc are extracted and weighted and fused with the original features of the model; mini-batch gradient descent is adopted, combined with the elastic weight integration algorithm, and weight penalty is applied to the key parameters of the original model during the training process. That is, the new data training causes the model to lose the original medical knowledge. The model is trained until the training termination condition is met, the training is stopped, and the iterative incremental model Minc is generated. (7.3.3) Compare the performance of the incremental model Minc and the original model Mold on the core medical indicators. If Minc performs better than Mold, the incremental model parameters are integrated with the original model parameters through the model version fusion algorithm to generate a new version model Mnew, and the old model is archived. If the performance does not meet the standard, the incremental data is discarded and higher-value data is selected for iteration.

Images