Medical service system and method based on multi-granularity semantic analysis and knowledge reasoning engine

Abstract

The invention relates to the technical field of artificial intelligence and medical health, discloses a medical service system and method based on multi-granularity semantic analysis and a knowledge inference engine, and aims to solve the problems of low intelligent inquiry interaction efficiency and insufficient accuracy in the prior art. The system comprises a multi-granularity semantic analysis module and a core mixed knowledge inference engine, the engine selects an optimal question by calculating the information value VoI of each potential question based on a dynamic medical knowledge graph, the VoI is formally defined as the expected reduction amount of diagnosis uncertainty measured by Shannon entropy, and meanwhile, the VoI is defined as the expected reduction amount of diagnosis uncertainty measured by Shannon entropy. And the inference engine also actively instructs the analysis module to dynamically adjust the analysis granularity to form a feedback closed loop of inference guidance analysis. According to the method, a formalized optimal dialogue strategy is used for replacing a traditional heuristic rule, so that the convergence speed and reliability of diagnosis are remarkably improved, and efficient and accurate intelligent medical service is achieved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and medical health technology, and in particular to a clinical decision support system and method for intelligent auxiliary diagnosis using natural language processing and knowledge reasoning. Background Art

[0002] With the deep integration of information technology and the medical industry, intelligent medical consultation or telemedicine systems have become an important direction for improving the accessibility and efficiency of medical services; however, existing technologies still face many challenges in achieving truly intelligent, efficient and reliable doctor-patient interaction.

[0003] At present, early telemedicine systems are mainly based on video conferencing technology, which integrates medical equipment such as electrocardiogram and ultrasound, and connects to hospital information systems, laboratory information systems or image archiving and communication systems to achieve remote consultation. Such systems are essentially auxiliary tools for data transmission and communication. Although they provide connectivity, they do not have any diagnostic reasoning capabilities and cannot actively guide the consultation process.

[0004] Since then, technological developments have introduced medical knowledge graphs for auxiliary diagnosis; this type of method matches the symptoms entered by the user with entities in the knowledge graph and infers possible diseases based on the path relationships in the graph.

[0005] However, the reasoning logic of such methods is relatively rigid; they usually adopt a static query-retrieval mode, which makes it difficult to handle the ambiguity and uncertainty of user input; when a symptom is associated with multiple diseases in the graph, the system lacks an effective mechanism to actively eliminate ambiguity, and its reasoning process is often one-time, rather than a dialogue process that evolves dynamically according to the interaction, which means that the risk of misdiagnosis under different interaction paths is dynamically changing. Therefore, there is an urgent need for a mechanism that can actively monitor and manage interaction uncertainty to reduce potential diagnostic risks and truly meet the user's diagnosis and treatment needs. Therefore, the present invention proposes a medical service system and method based on multi-granularity semantic parsing and knowledge reasoning engine.

[0006] In recent years, the rise of large language models has promoted the application of intelligent agents built on them to complex real-world problems, such as medical consultation; such intelligent agents can conduct more fluent and natural conversations.

[0007] However, when they are applied to the high-risk, high-precision medical field, a series of serious technical defects are exposed. The first problem is the accumulation of hallucinations, that is, the model may lack factual basis when generating content, resulting in untrue or erroneous information, and this error will be continuously amplified in the subsequent reasoning chain, posing a catastrophic risk to medical diagnosis. Secondly, these intelligent agents are weak in calling external tools and adapting to environmental changes, and have poor robustness. In addition, they generally lack the ability to persist in memory of users' long-term behavior and historical tasks, and it is difficult to form a coherent understanding of the case in multiple rounds of dialogue.

[0008] To solve the above problems, the current cutting-edge technology direction is retrieval-enhanced generation, especially solutions combined with knowledge graphs. For example, the MedRAG framework proposed by some researchers combines knowledge graphs with retrieval-enhanced generation models, and uses the structured knowledge in knowledge graphs to constrain and guide the generation process of large language models to improve the accuracy of diagnosis. MedRAG even proposes an active diagnostic questioning mechanism, which aims to obtain more information by asking questions.

[0009] However, after in-depth analysis, it was found that the decision-making logic of this mechanism has fundamental limitations; MedRAG decides which question to ask based on a discriminability score, which is calculated according to the degree centrality of the feature nodes in the knowledge graph; the fewer diseases a feature is connected to, the lower its degree centrality and the higher the discriminability score; this method is essentially a heuristic rule that relies on the static topological structure of the knowledge graph without considering the dynamic diagnostic context for the current specific patient; a question that is highly discriminative under normal circumstances may not be the most necessary question to ask for a specific case in which most possibilities have been ruled out.

[0010] Therefore, existing technologies, even the most advanced knowledge graph-enhanced retrieval-enhanced generation systems, still lack a formally optimized, mathematically-based, and dynamically adaptive dialogue management strategy; their questioning strategies are suboptimal and may lead to unnecessary interaction rounds, thereby reducing the efficiency and accuracy of diagnostic convergence. Summary of the Invention

[0011] The present invention aims to solve the technical problems of low dialogue efficiency, slow diagnostic convergence, and insufficient final diagnostic accuracy in the diagnostic interaction process of existing intelligent medical service systems, because their core reasoning logic relies on unreliable black-box generation models or suboptimal heuristic rules; existing technical solutions generally lack a formalized optimal dialogue strategy that can adaptively adjust according to the dynamically changing dialogue context, which makes it difficult for them to achieve efficient and reliable information acquisition and uncertainty elimination in complex clinical scenarios.

[0012] In order to solve the above technical problems, the present invention provides the following technical solutions: A medical service system and method based on multi-granularity semantic parsing and knowledge reasoning engine, the method provided by the present invention comprises the following basic steps: In steps S1 and S2, the system first receives an initial query input by a user in natural language, and performs preliminary semantic parsing on the query to extract one or more coarse-grained medical entities therefrom.

[0013] Step S3: Based on these extracted medical entities and a dynamic medical knowledge graph, the system initializes or updates the posterior probability distribution of a set of candidate diagnoses in its internal hybrid knowledge reasoning engine.

[0014] Step S4, the core step, is that the system calculates the information value of each potential follow-up question, which is precisely defined as the expected contribution of the question in reducing the uncertainty of the current candidate diagnosis probability distribution.

[0015] Step S5: The system selects the target follow-up question with the highest computational information value according to the principle of maximizing information value.

[0016] Step S6: The system instructs its multi-granularity semantic parsing module to convert the target question in a logical form into a natural language question sentence that conforms to human communication habits.

[0017] Step S7: The system sends the generated question to the user, receives the user's response as new evidence, and loops through the above steps from updating the probability distribution to receiving the response until the entire diagnostic interaction process meets one or more pre-set termination conditions.

[0018] Preferably, in order to implement the above technical solution, the key links are optimized as follows: The information value is calculated based on a formal measurement of the expected reduction in the Shannon entropy of the candidate diagnosis probability distribution; this provides a solid mathematical basis for quantifying the information content of the problem, replacing the fuzzy or static topology-based heuristic rules in the prior art.

[0019] At the same time, the multi-granularity semantic parsing module is specially configured to respond to specific instructions from the hybrid knowledge reasoning engine to perform dynamic, goal-driven decomposition of coarse-grained entities that have been identified in previous user input, thereby extracting finer-grained attributes when needed.

[0020] In addition, the medical knowledge graph in the present invention is dynamic. It is configured to be updated instantly according to the user's real-time response during the current conversation session, incorporating entities and relationships specific to the user session to form a temporary, context-rich conversation graph instance.

[0021] Further preferably, the technical details of the above preferred solution are further elaborated: When calculating the information value, the specific calculation process is as follows: first calculate the total entropy value of the current candidate diagnosis probability distribution, then for each potential question, estimate the probabilities of all possible answers, and calculate the entropy value of the system's new posterior probability distribution after obtaining each answer, and finally calculate the expected posterior entropy after asking the question through weighted average; the difference between the current total entropy and the expected posterior entropy is the information value of the question.

[0022] Among them, the probability that a user will give a specific answer to a question can be estimated based on the large-scale clinical statistical data stored in the dynamic medical knowledge graph; in terms of dynamic decomposition of multi-granularity semantic parsing, a specific implementation example is to further decompose a initially identified coarse-grained entity "headache" into multiple fine-grained attributes such as "pain location", "pain severity", and "pain nature" according to the instructions of the inference engine, and generate targeted follow-up questions around these attributes.

[0023] In order to ensure the natural fluency of the generated questions, the system may further include a large language model, which is specifically used to convert the structured question instructions output by the reasoning engine into final natural language questions; More preferably, in order to ensure the robustness, security and integrity of the present invention in practical applications, the following more detailed limitations are made.

[0024] The termination conditions of the diagnostic interaction are clearly defined as including at least one of the following two items: one is the confidence condition, that is, when the posterior probability of a candidate diagnosis exceeds a predetermined high confidence threshold, such as ninety-five percent; the other is the information gain condition, that is, when the maximum information value of all potential follow-up questions is lower than a predetermined minimum threshold, which indicates that continuing to ask questions can no longer significantly reduce the uncertainty of the diagnosis, and the interaction should also be terminated at this time.

[0025] In order to overcome the inherent "hallucination" risk of large language models, a key design of the present invention is that when using large language models to generate natural language questions, the generation process is strictly constrained by the structured, factual content output from the hybrid knowledge reasoning engine.

[0026] This means that large language models only serve as "polishing" and "translation" tools, and the content they generate is strictly limited to the factual framework determined by the probabilistic logic reasoning module, thereby fundamentally preventing the generation and accumulation of information illusions and ensuring the rigor and safety of medical conversations.

[0027] In addition, at the beginning of the diagnostic process, that is, before receiving any user input, the initial probability distribution of candidate diagnoses in the system can be set a priori based on disease epidemiological data obtained from the dynamic medical knowledge graph, thereby providing a more realistic starting point for the entire Bayesian reasoning process.

[0028] Compared with the prior art, the present invention has the following beneficial effects: The present invention achieves significant and unexpected beneficial effects by introducing a dialogue management strategy based on information value theory and reasoning-driven multi-granularity semantic parsing. First, it maximizes dialogue efficiency. By asking the most informative questions at each step, the present invention avoids redundant and inefficient questioning, thereby converging to the final diagnosis with provably fewer interaction rounds, significantly shortening the consultation time. Second, it enhances diagnostic accuracy. The method systematically and specifically eliminates ambiguity, showing higher robustness in distinguishing diseases with similar symptoms. Through a probabilistic reasoning framework, the system can more reliably handle uncertainty and reduce the risk of misdiagnosis. Third, it is highly dynamic and adaptable. The dialogue strategy of the present invention is not fixed or based on static rules, but is recalculated and optimized at each interaction step based on the currently accumulated evidence. This enables the entire consultation process to be uniquely and individually customized for the specific situation of each user, thereby achieving truly adaptive interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without inventive effort. Figure 1 This is a system architecture diagram of the medical service system in the medical service system and method based on multi-granularity semantic parsing and knowledge reasoning engine of the present invention; Figure 2 This is a specific flow chart of the iterative diagnosis method in the medical service system and method based on multi-granularity semantic parsing and knowledge reasoning engine of the present invention; Figure 3 Schematic diagram of the feedback closed loop concept between the hybrid knowledge reasoning engine and the multi-granularity semantic parsing module in the medical service system and method based on multi-granularity semantic parsing and knowledge reasoning engine of the present invention; DETAILED DESCRIPTION The following will be combined with the drawings and embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work shall fall within the scope of protection of the present invention; Figure 1-3 The present invention is further described.

[0030] A medical service system 100 based on multi-granularity semantic analysis and knowledge reasoning engine, referring to Figure 1 , characterized by comprising: The data interface 110 is responsible for communicating with the user terminal, receiving the user's natural language input, and sending the questions generated by the system to the user.

[0031] The multi-granularity semantic parsing module 120 is innovative in that it fully serves the dynamic needs of the hybrid knowledge reasoning engine 130. This module can use standard technologies in the field of natural language processing, such as named entity recognition and relationship extraction, but its core function is dynamic granularity adjustment. Its working mechanism, in one embodiment, specifically includes: Initial coarse-grained parsing, when the system receives the user's initial query through the data interface 110, an example is "I have a stomachache and a fever", the module 120 will perform standard named entity recognition and extract coarse-grained medical entities, such as the symptom stomachache and the symptom fever.

[0032] Inference-driven fine-grained parsing, these coarse-grained entities are passed to the hybrid knowledge reasoning engine 130; after calculation, the engine may determine that the most valuable information at the moment is the specific location of the stomach pain.

[0033] At this time, the engine sends an instruction to the parsing module 120, which requires decomposing the entity stomachache to obtain its location attribute.

[0034] After receiving the instruction, the natural language generation and parsing module 120 will work with a controlled large language model to generate a question that conforms to natural language habits and has a clear goal. An example is: "In order to better help you, can you point out which specific part of your stomach hurts?" This process shows that the system goes beyond simple slot filling and is capable of more detailed and intelligent conversational interactions.

[0035] The hybrid knowledge reasoning engine 130 is the core innovation of the present invention. It is responsible for the thinking and decision-making of the entire diagnostic process; the engine is mainly composed of a dynamic medical knowledge graph 140 and a probabilistic logic reasoning module 132.

[0036] The dynamic medical knowledge graph 140 is the basis for the system to perform reasoning; its structure can be represented as a graph, which contains an entity set and a relationship set, wherein the entity set includes diseases, symptoms, examinations, drugs, etc., and the relationship set represents the connection between entities, such as whether there are symptoms, what kind of drugs are used for treatment, etc.; the construction of this graph can be based on authoritative medical literature, clinical guidelines, and existing medical ontology libraries; unlike static knowledge bases, the knowledge graph of the present invention is dynamic; the system will create a temporary graph instance for each diagnostic session, and add the current user's specific information as new nodes and edges to this instance.

[0037] The probabilistic logic reasoning module 132 is the key to realizing the creativity of the present invention; it transforms the diagnostic process from being based on heuristic rules to being based on strict probabilistic reasoning and decision theory; the probabilistic logic reasoning module 132 establishes the diagnostic decision-making process on strict probabilistic reasoning and decision theory. Specifically, it introduces the information value theory to guide the dialogue strategy, which is significantly different from the method in the prior art that relies on static heuristic rules.

[0038] In one embodiment, referring to Figure 2 , its detailed working mechanism includes: Status indicates that at any step in the diagnostic process , the system will maintain a set of all possible diagnoses The posterior probability distribution of ;in, Represents all the evidence collected so far; uncertainty quantification, the uncertainty of the current diagnostic status can be accurately measured by the Shannon entropy of the probability distribution, which is calculated as follows: Where, For diagnostic collection The Shannon entropy of For evidence Next diagnosis The higher the entropy value, the greater the uncertainty of the system in the final diagnosis.

[0039] Information value calculation: the system generates a set of potential follow-up questions from the knowledge graph. For each of these questions , there is a set of possible answers ;question The information value of is defined as the expected reduction in entropy after asking the question; its calculation formula is: Where, For the problem the value of information; is the current entropy before the question is asked; The existing evidence Next, the user answers The probability of , which can be estimated from statistical data stored in the knowledge graph or large-scale clinical data; Is receiving a response Afterwards, the new posterior entropy of the system state.

[0040] Action selection: the system calculates the information value of all potential questions and selects the question that maximizes the information value As the next question; the selection formula is: Instruction generation and interaction, refer to Figure 3 , the best question selected It will be passed to the multi-granularity semantic parsing module 120, which generates a natural language form and presents it to the user; this process demonstrates the feedback loop formed between the inference engine 130 and the parsing module 120, that is, reasoning guides parsing, and parsing serves reasoning.

[0041] In order to more specifically illustrate the workflow and advantages of the present invention, a typical clinical scenario is used below for illustration: A user issues an initial query to the system: "I have chest pain"; "chest pain" is a symptom with a wide and dangerous differential diagnosis range, with possible causes including myocardial infarction, gastroesophageal reflux, pulmonary embolism, anxiety, etc.

[0042] In the first step, after receiving the user's initial query "My chest hurts", the system extracts the coarse-grained symptom entity "chest pain"; at this time, the candidate diagnosis probability distribution within the system is relatively diffuse, for example, the probability of myocardial infarction is 0.3, the probability of gastroesophageal reflux is 0.4, and the probability of pulmonary embolism is 0.3; the system then calculates the information value of multiple potential follow-up questions and finds that the information value of asking "nature of pain" is the highest, with a value of 0.25, while the information value of asking "duration" is 0.18; therefore, the system chooses the former and generates the question: "Can you describe the feeling of pain? Is it a tingling, dull pain, or a squeezing feeling?"

[0043] In the second step, the user answers: "It's a squeezing feeling." After receiving this new evidence, the system updates its internal probability distribution. At this point, the probability of confirming myocardial infarction increases significantly to 0.7, while the probability of gastroesophageal reflux decreases to 0.1, and the probability of pulmonary embolism is 0.2. The system's focus shifts accordingly, and it recalculates the information value of all potential questions. The calculation results show that the information value of asking "Does it radiate to the arms" is the highest at this moment, with a value of 0.31, because this is the key question to confirm or exclude the most likely diagnosis at the moment. The information value of asking "Is it related to eating" becomes very low, with a value of 0.05, because the corresponding possibility of gastroesophageal reflux has been greatly reduced. This is in stark contrast to systems that rely on static heuristic rules, which may still ask a general but no longer optimal question at this moment. The system therefore generates the question: "Does this squeezing feeling radiate to your arms, neck, or chin?"

[0044] In the third step, the user answers: "Yes, it radiates to my left arm." After obtaining this strong evidence, the system updates the probability distribution again, and the probability of myocardial infarction soars to 0.92, far exceeding other diagnoses. At this point, the system calculates that the information value of any further questions is lower than the preset termination threshold, indicating that the diagnosis is basically clear and there is little value in continuing to ask questions. The system then triggers the termination condition and gives a final suggestion: "Based on your symptom description, there is a high probability that it is a heart-related problem. Please seek emergency medical assistance immediately."

[0045] This example clearly demonstrates the beneficial effects of efficiency and accuracy brought about by the present invention; each step of the system takes the shortest path to a correct diagnosis, realizing the formalization and optimization management of the diagnostic process; this ability to handle complex uncertainties and perform reliable reasoning is crucial for fields such as medicine that require rigor and verifiability.

[0046] In summary, the advantages of the present invention are: monitoring the multi-disease diagnosis process, realizing intelligent dialogue strategy adjustment, ensuring that the diagnosis in multiple rounds of interaction is in the optimal convergence path, reducing the ambiguity of the diagnosis process, and realizing efficient and safe management of medical services.

[0047] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A medical service method based on multi-granularity semantic parsing and knowledge reasoning engine, characterized by: include: Step S1: receiving the initial natural language query input by the user; Step S2: performing preliminary semantic parsing on the initial natural language query to extract one or more coarse-grained medical entities; Step S3: Initializing or updating a probability distribution of a set of candidate diagnoses in a hybrid knowledge reasoning engine based on the extracted medical entities and a dynamic medical knowledge graph; Step S4: calculating the information value of each of the multiple potential follow-up questions, wherein the information value is used to quantify the expected effect of the question on reducing the uncertainty of the candidate diagnosis probability distribution; Step S5: Select the target follow-up question with the highest computational information value; Step S6: instruct a multi-granularity semantic parsing module to generate a natural language form of the target follow-up question; Step S7: sending the generated target follow-up question to the user and receiving the user's response; Step S8: If a preset termination condition is not met, the user response is used as new evidence and the process returns to step S3.

2. A medical service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that: The information value is calculated based on the expected reduction in Shannon entropy of the candidate diagnosis probability distribution.

3. A medical service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that: The multi-granularity semantic parsing module is configured to respond to instructions from the hybrid knowledge reasoning engine and perform dynamic decomposition on coarse-grained entities in previous user input to extract fine-grained attributes.

4. A medical service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that: The dynamic medical knowledge graph is updated in real time based on the user's responses during the current conversation session to include entities and relationships specific to the user.

5. The medical service method based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 1, characterized in that: The termination conditions include at least one of the following: The probability of a single diagnosis exceeds a predetermined threshold; or the maximum information value of all potential follow-up questions is lower than a predetermined threshold.

6. A medical service system based on multi-granularity semantic parsing and knowledge reasoning engine, characterized by: include: A data interface module, configured to receive user inquiries; a multi-granularity semantic parsing module, the multi-granularity semantic parsing module being operatively connected to the data interface module and configured to parse natural language queries at variable levels of granularity; A hybrid knowledge reasoning engine module is also operatively connected to the multi-granularity semantic parsing module, and the hybrid knowledge reasoning engine module further comprises: A dynamic medical knowledge graph to store medical entities and relationships; A probabilistic logic reasoning module is operatively connected to the dynamic medical knowledge graph, the probabilistic logic reasoning module is used to maintain a probability distribution of a set of candidate diagnoses, calculate the information value of potential follow-up questions based on the expected impact of the uncertainty of the probability distribution, select a target follow-up question based on maximizing the information value, and send instructions to the multi-granularity semantic parsing module based on the selected target follow-up question.

7. A medical service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that: The probabilistic logic reasoning module calculates the information value by calculating the expected reduction in Shannon entropy of the candidate diagnosis probability distribution.

8. The medical service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that: The multi-granularity semantic parsing module is configured to respond to instructions from the hybrid knowledge reasoning engine and perform dynamic decomposition on coarse-grained entities in previous user input to extract fine-grained attributes.

9. The medical service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that: The system also includes a large language model for generating a final natural language question, wherein the input of the large language model is constrained by the structured output from the hybrid knowledge reasoning engine to prevent information hallucination.

10. A medical service system based on multi-granularity semantic parsing and knowledge reasoning engine according to claim 6, characterized in that: The system is configured to terminate the diagnostic interaction when the probability of a single diagnosis exceeds a predetermined threshold, or when the maximum information value of all potential follow-up questions is below a predetermined threshold.

Images