Cost-aware knowledge enhanced multi-agent sequential diagnosis method and system

Abstract

The application provides a cost-aware knowledge enhanced multi-agent sequential diagnosis method and system, and belongs to the technical field of artificial intelligence and medical auxiliary diagnosis; the method comprises the following steps: constructing a medical diagnosis knowledge graph based on a target disease list; converting unstructured medical descriptions of a user into structured observation states through natural language processing, and anchoring symptoms and signs to corresponding nodes of the medical diagnosis knowledge graph; performing deterministic reasoning according to the current structured observation state and the medical diagnosis knowledge graph, to calculate confidence scores of each candidate disease, and determining the best action in the next step based on an utility function; and generating a natural language interactive response based on the result of the deterministic reasoning. Through construction of a high-quality knowledge graph, the application can reduce the diagnosis cost while improving the diagnosis accuracy.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and medical auxiliary diagnosis technology, and particularly relates to a cost-aware knowledge-enhanced multi-agent sequence diagnosis method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, large language models (LLMs) are widely used in medical examinations such as the USMLE; however, translating this static knowledge-based answering ability into the ability to perform sequential diagnosis in actual clinical practice still faces significant challenges.

[0004] Sequence diagnostics simulates a real clinical workflow: physicians collect information iteratively through multiple rounds of interaction (symptom inquiry, physical examination, laboratory tests, etc.), make decisions under uncertainty, and need to balance diagnostic accuracy with resource costs (including monetary expenses, patient discomfort, and time costs). This multi-round interactive process requires not only knowledge recall but also a systematic reasoning ability that coordinates information acquisition with hypothesis revision.

[0005] Existing LLM-based diagnostic methods suffer from a significant "knowledge-reasoning gap": despite possessing rich medical knowledge, LLM agents often struggle with systematic reasoning under cost constraints. Current research indicates that LLM agents tend to exhibit "defensive medicine" behavior in simulated diagnosis, ordering excessive and unnecessary tests to compensate for the uncertainty of reasoning, failing to effectively utilize discriminative features for efficient differential diagnosis. This behavior stems from the implicit and unstructured nature of the LLM reasoning process, lacking an explicit domain model to support systematic hypothesis testing. Specifically, existing technologies have significant shortcomings in the following two aspects:

[0006] On the one hand, there is a lack of knowledge graphs suitable for diagnosis and efficient methods for their construction. Existing methods for constructing medical knowledge graphs typically use quadruples or function models to express constraints, which rely on complex model fitting or manually defined rules, making it difficult to utilize the general knowledge of Large Language Models (LLMs) for large-scale, fully automated causal chain mining. Furthermore, medical concepts possess dynamic semantics (e.g., "heart failure" is both a disease and a symptom), and existing static ontology construction techniques struggle to handle this role transition.

[0007] On the other hand, there is a lack of multi-agent collaborative mechanisms to explicitly suppress "defensive medicine." Existing auxiliary diagnostic methods are mostly based on Bayesian networks or static rules to recommend examination items. Although this method takes probability into account, it often lacks a dynamic and fine-grained trade-off between "detection cost" and "information gain," leading to the recommendation of too many unnecessary examinations, i.e., the existence of defensive medicine. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a cost-aware knowledge-enhanced multi-agent sequence diagnosis method and system. By constructing a high-quality knowledge graph, it can improve the accuracy of diagnosis while reducing diagnostic costs.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0010] The first aspect of this invention provides a cost-aware knowledge-enhanced multi-agent sequence diagnosis method.

[0011] Cost-aware knowledge-enhanced multi-agent sequence diagnostic methods include:

[0012] Based on the target disease list, a medical diagnostic knowledge graph is constructed, which includes disease nodes, feature nodes, detection nodes, and edges connecting each node.

[0013] The system obtains the user's unstructured medical description, transforms it into a structured observation state through natural language processing, and anchors the symptoms and signs in the observation state to the corresponding nodes of the medical diagnostic knowledge graph.

[0014] Deterministic reasoning is performed based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease, and the optimal action for the next step is determined based on the utility function; the optimal action includes asking about symptoms, ordering tests, and giving a final diagnosis;

[0015] Based on the results of the deterministic reasoning, the system interacts with the user by generating corresponding natural language responses or specific diagnostic operation instructions.

[0016] Furthermore, the construction of the medical diagnostic knowledge graph is achieved through five stages: parallel knowledge extraction, quantitative typicality mapping, hybrid entity alignment, dynamic node upgrading, and attribute enrichment.

[0017] Furthermore, the implementation process of the parallel knowledge extraction and quantization typicality mapping includes: for each disease in the disease list, extracting its associated features using a large language model and verifying the detection items corresponding to the features; adopting a quantization typicality strategy to map the typicality of the features described in natural language into discrete weighted buckets as the weights of the edges connecting diseases and features; constructing an action-centered topology to explicitly establish verification edges between features and detection items.

[0018] Furthermore, the implementation process of the hybrid entity alignment includes: for newly extracted terms, recalling candidate nodes from the existing node set using a hybrid retrieval strategy; wherein, the hybrid retrieval strategy combines word matching based on overlap coefficient and semantic matching based on cosine similarity of embedding vectors; and using a large language model as a discriminator to determine whether the new term is merged into the candidate nodes or inserted as a new node.

[0019] Furthermore, the implementation process of the dynamic node upgrade includes: when a node that exists as a feature in the medical diagnostic knowledge graph is identified as an independent disease, it is promoted from the feature node set to the disease node set, while retaining its original incoming edge relationship and allowing it to establish new outgoing edge relationship, thereby automatically modeling the causal chain between diseases.

[0020] Furthermore, the confidence score for each candidate disease is calculated using the following formula:

[0021] ;

[0022] in, Indicates disease Confidence score, This represents the set of characteristic states that have been observed at the current moment; Representation of features status The state It includes both presence and absence; The feature scoring function is used to represent the disease. Corresponding features In state The rating at that time.

[0023] Furthermore, the optimal action for the next step is determined based on the utility function, including feature selection and detection selection; wherein, the feature selection is to prioritize features that are related to high-confidence diseases and can distinguish different candidate diseases by using the discriminative power scores of undetected features; the detection selection is to select by calculating the utility of candidate detection items and combining the value, cost and invasiveness of the verification features.

[0024] A second aspect of the present invention provides a cost-aware knowledge-enhanced multi-agent sequence diagnostic system.

[0025] A cost-aware, knowledge-enhanced multi-agent sequence diagnostic system includes:

[0026] The graph construction module is configured to: construct a medical diagnostic knowledge graph based on the target disease list, which includes disease nodes, feature nodes, detection nodes, and edges connecting each node;

[0027] The perceptual intelligent agent module is configured to: acquire the user's unstructured medical description, convert it into a structured observation state through natural language processing, and anchor the symptoms and signs in the observation state to the corresponding nodes of the medical diagnostic knowledge graph;

[0028] The reasoning agent module is configured to: perform deterministic reasoning based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease, and determine the optimal action for the next step based on the utility function; the optimal action includes asking about symptoms, ordering tests, and giving a final diagnosis;

[0029] The decision-making agent module is configured to interact with the user by generating corresponding natural language responses or specific diagnostic operation instructions based on the results of the deterministic reasoning.

[0030] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the cost-aware knowledge-enhanced multi-agent sequence diagnostic method as described in the first aspect of the present invention.

[0031] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the cost-aware knowledge-enhanced multi-agent sequence diagnostic method as described in the first aspect of the present invention.

[0032] The above one or more technical solutions have the following beneficial effects:

[0033] On the one hand, this invention constructs a medical diagnostic knowledge graph based on a target disease list through five stages: parallel knowledge extraction, quantitative typicality mapping, hybrid entity alignment, dynamic node upgrading, and attribute enrichment. Compared with existing technologies, this invention's fully automated pipeline-style medical diagnostic knowledge graph construction method based on LLM, especially the dynamic node upgrading mechanism, can adaptively handle the role transformation of medical concepts between "symptoms" and "diseases," constructing a high-quality graph without extensive manual intervention.

[0034] On the other hand, this invention constructs a Medical Diagnostic Knowledge Graph (MDKG) containing quantified typicality weights, combining the strong semantic understanding capabilities of a neural language model with the rigor of symbolic reasoning. This enables it to handle more complex unstructured clinical inputs, converge to the correct diagnosis more quickly, and significantly improve the diagnostic success rate. Furthermore, unlike existing cost-sensitive classification systems or simple symptom checkers, this invention introduces a specific medical utility function, explicitly using the invasiveness level as a penalty term in action planning, effectively suppressing the model's tendency to blindly prescribe expensive tests (defensive medicine) under uncertainty. Simultaneously, the reasoning process is based on explicit graph paths and deterministic scoring formulas, ensuring that every diagnostic decision can be traced back to specific evidence and weights, overcoming the problem of the invisibility of the end-to-end large-scale model reasoning process.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 This is a flowchart of the cost-aware knowledge-enhanced multi-agent sequence diagnosis method in Embodiment 1 of the present invention.

[0038] Figure 2 This is a flowchart of the automated construction of a medical diagnostic knowledge graph in Embodiment 1 of the present invention.

[0039] Figure 3 This is a timing diagram showing the collaborative interaction between the perceptual agent, the reasoning agent, and the decision-making agent in Embodiment 1 of the present invention.

[0040] Figure 4 This is a flowchart of evidence scoring and action planning based on a reasoning agent in Embodiment 1 of the present invention. Detailed Implementation

[0041] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0042] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0043] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0044] Example 1

[0045] This embodiment discloses a cost-aware knowledge-enhanced multi-agent sequence diagnosis method.

[0046] like Figure 1 As shown, the cost-aware knowledge-enhanced multi-agent sequence diagnosis method includes:

[0047] Step S1: Based on the target disease list, construct a medical diagnostic knowledge graph that includes disease nodes, feature nodes, detection nodes, and edges connecting each node.

[0048] Step S2: Obtain the user's unstructured medical description, convert it into a structured observation state through natural language processing, and anchor the symptoms and signs in the observation state to the corresponding nodes of the medical diagnostic knowledge graph;

[0049] Step S3: Perform deterministic reasoning based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease, and determine the optimal action for the next step based on the utility function; the optimal action includes asking about symptoms, ordering tests, and giving a final diagnosis;

[0050] Step S4: Based on the result of the deterministic reasoning, interact with the user by generating corresponding natural language responses or specific diagnostic operation instructions.

[0051] Based on the above process, this invention, through the construction of a high-quality knowledge graph, can improve diagnostic accuracy while reducing diagnostic costs. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0052] In step S1, based on the target disease list, a medical diagnostic knowledge graph (MDKG) is constructed, which includes disease nodes, feature nodes, detection nodes, and edges connecting each node, so that the graph contains quantified typicality weights and detection cost attributes.

[0053] like Figure 2 As shown, the construction of the medical diagnostic knowledge graph is achieved through five stages: parallel knowledge extraction, quantitative typicality mapping, hybrid entity alignment, dynamic node upgrading, and attribute enrichment.

[0054] The first stage is parallel knowledge extraction, which involves using a large language model to extract the associated features for each disease in the disease list and verifying the detection items required for those features.

[0055] In practice, the system receives a predefined list of diseases. As input. For each disease in the list. The system invokes large language models (such as GPT-4 or other finely tuned medical models) in parallel to perform knowledge queries. Unlike simple entity extraction, this embodiment requires the model to return structured triple information. ,in, Representative characteristics (symptoms, signs) This characteristic represents the disease Typical description, This is a list of detection methods required to verify the feature. This detection-centric topology (Action-CentricTopology) ensures that the constructed medical diagnostic knowledge graph not only contains declarative knowledge but also explicitly includes procedural knowledge, thereby supporting subsequent action planning.

[0056] The second stage involves quantifying typicality mapping. A quantified typicality strategy is employed to map the typicality of features described in natural language into discrete weighted buckets, which serve as the weights of the edges connecting diseases and features. An action-centric topology is constructed to explicitly establish validation edges between features and detection items.

[0057] In practical implementation, existing knowledge graphs often represent relationships solely through binary edges (i.e., presence or absence), neglecting the probabilistic differences in the occurrence of medical symptoms. Furthermore, directly using the continuous probability values ​​output by LLMs often introduces calibration errors. Therefore, this embodiment introduces a "quantitative typicality" strategy, mapping natural language descriptions to discrete weighted buckets. .

[0058] The specific mapping rule is as follows: the hallmark feature (Hallmark, ++) is mapped to a high positive weight. (e.g., +10), such features are almost sufficient conditions for the disease or are extremely high-frequency features; common features (+) are mapped to moderately positive weights. (e.g., +5) These features have a high pass rate but weak specificity; Rare features (-) are mapped to negative weights. (For example, -5), if this feature is observed, it actually reduces the probability of having the disease (exclusionary reasoning); mutually exclusive features (Impossible, --) are mapped to high negative weights. (For example, -10), such features are pathologically mutually exclusive with the disease. This discretization process acts as a low-pass filter, effectively stabilizing the subsequent inference process and avoiding decision oscillations caused by small probability fluctuations.

[0059] The third stage is hybrid entity alignment. For newly extracted terms, a hybrid retrieval strategy is used to recall candidate nodes from the existing node set. The hybrid retrieval strategy combines word matching based on overlap coefficient and semantic matching based on cosine similarity of embedding vectors. A large language model is used as a discriminator to determine whether the new term should be merged into the candidate nodes or inserted as a new node.

[0060] In practice, due to the large number of synonyms and variations in medical terminology (such as "fever" and "heat rash," "angina pectoris" and "chest pain"), direct merging can easily lead to redundancy or fragmentation of the retrieval map. Therefore, this embodiment employs a hybrid retrieval strategy to combine newly extracted terms... Align to existing node set The strategy first computes the candidate set. The formula is as follows:

[0061] ;

[0062] in, Candidate words with literal inclusion relationships (e.g., "diabetes" and "type 2 diabetes") are filtered based on the overlap coefficient (which is the size of the intersection divided by the size of the smaller set). Semantic synonyms are selected based on cosine similarity using dense vector embeddings. Subsequently, a large language model is used as the discriminator. Determine new terms based on the candidate set It should be merged into an existing node. It can be inserted into the graph as a completely new node.

[0063] The fourth stage is dynamic node upgrade. When a node that exists as a feature in the medical diagnostic knowledge graph is identified as an independent disease, it is promoted from the feature node set to the disease node set. At the same time, its original incoming edge relationships are preserved, and it is allowed to establish new outgoing edge relationships, thereby automatically modeling the causal chain between diseases.

[0064] In medical classification systems, some concepts may be symptoms of one disease in one context (e.g., "angina pectoris" is a characteristic of "coronary heart disease"), while in another context they are independent diagnostic entities. Existing static ontologies struggle to handle such dynamic transformations. To address this, this embodiment implements a dynamic node promotion mechanism. Specifically, if a node that originally existed as a feature... If a disease is identified as an independent disease during subsequent knowledge extraction (i.e., it appears in the input disease list), the system automatically performs the following operation: "promotes" it to a disease node. ; retain its original incoming edge (That is, to maintain its properties as a characteristic of other diseases), allowing it to establish new outbound boundaries. (That is, to endow it with the ability to possess its own characteristics as a disease). This mechanism enables complex causal chains (e.g., kidney disease). Heart failure (Difficulty breathing) can emerge automatically without the need for manual reconstruction of the atlas structure.

[0065] The fifth stage is attribute enrichment. The system utilizes the metadata enrichment module to enrich the attributes of each detection node in the graph. Infer and bind quantitative attributes .in, The currency cost level is (e.g., 1-5). The risk level is categorized as invasive (e.g., 0: non-invasive, 1: minimally invasive, 2: invasive). These attributes form the basis for subsequent calculations of testing utility.

[0066] like Figure 3 As shown, steps S2-S4 model the diagnostic process as a partially observable decision-making process, and realize the "Think-then-Act" closed loop through the collaborative cycle of the three agents of perception, reasoning, and decision-making.

[0067] In the perception phase of step S2, the user's unstructured medical description is received by the perception agent, transformed into a structured observation state through natural language processing, and the symptoms and signs in the observation state are anchored to the corresponding nodes of the medical diagnostic knowledge graph.

[0068] In the specific implementation process, due to the input of patient users For unstructured texts (such as spoken descriptions and medical report texts), the perceptual agent first utilizes extraction functions. Transform it into a structured observation state :

[0069] ;

[0070] in, Indicates the first Individual entities referencing text (e.g., "elevated white blood cell count") Indicates state; Indicates the first Patient user input, Indicates as of the date The history of dialogue between wheels This represents the number of observation items extracted in this round. The perceptual agent utilizes the aforementioned hybrid alignment mechanism to... Anchored to standard feature nodes in MDKG This updates the current symbolic state set. This standardization process effectively shields the interference of natural language diversity.

[0071] In the reasoning phase of step S3, the reasoning agent performs deterministic reasoning based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease and determine the optimal action for the next step based on the utility function; the optimal action includes asking about symptoms, ordering tests, and giving a final diagnosis.

[0072] like Figure 4 As shown, in the specific implementation process, this stage includes two core logic parts:

[0073] One aspect is evidence scoring. To overcome the instability of implicit reasoning in large models, this embodiment employs an explicit deterministic scoring mechanism. For any candidate disease... Its confidence score The calculation is as follows:

[0074] ;

[0075] ;

[0076] in, Indicates disease Confidence score, This represents the set of characteristic states that have been observed at the current moment; Representation of features status The state It includes both presence and absence. The feature scoring function is used to represent the disease. Corresponding features In state The score at that time, when the feature When it exists, its value is the edge weight. When features When missing, its value is ,in This is a punishment factor. The formula incorporates two-way reasoning logic: if the patient exhibits typical... Similar symptoms, scores increase; if the patient clearly lacks The key symptoms (i.e.) but If the score is negative, the score is severely penalized. This mechanism ensures the rigor of the diagnosis.

[0077] Secondly, there is cost-sensitive action planning. The system calculates the utility of the next action based on the current disease ranking. Since calculating the full Bayesian Value of Information (VoI) is too time-consuming, this embodiment uses a heuristic utility function to determine the optimal next action, specifically including feature selection and detection selection.

[0078] 1) Feature selection. Calculate undetected features. Discrimination score Prioritize features that are associated with high-confidence diseases and can distinguish different candidate diseases:

[0079] ;

[0080] in, For the front A set of candidate diseases This is the balance coefficient; Indicate candidate diseases Confidence score, Indicates disease With features Weights of related edges in a medical diagnostic knowledge graph Indicating in the candidate disease set Top weight The variance of the formula is calculated. The first term prioritizes features that are strongly correlated with high-confidence diseases (correlation), and the second term prioritizes features with large weight variances among candidate diseases (discrimination).

[0081] 2) Detection selection. Calculate candidate detection items. The utility Taking into account the value of its verification features, as well as its cost and invasiveness:

[0082] ;

[0083] in, Indicates the testing items The set of features that can be verified To test the effectiveness of the verification features, For the currency cost level to be tested, This represents the level of invasiveness risk detected. The design of the denominator in this formula explicitly discourages "defensive medical" practices that blindly perform high-cost examinations when information gain is low.

[0084] In the decision-making stage of step S4, the decision-making agent interacts with the user by generating corresponding natural language responses or specific diagnostic operation instructions based on the results of deterministic reasoning.

[0085] In practical implementation, the decision-making agent, acting as the "right brain," is responsible for generating structured summaries of the output from the reasoning agent. Transformed into natural language; among which, Represents the set of candidate diseases Results sorted by confidence score from highest to lowest. Indicates the score by discriminative power Sort by selecting the first A set of candidate features Indicated by detection utility Sort by selecting the first A set of candidate test items. It utilizes the generative capabilities of LLM to generate responses or instructions that conform to medical standards and are appropriately worded. Because the generation process is affected by The strong constraints prevented the hallucinations from occurring.

[0086] When the probability of the disease with the highest confidence level in the belief state exceeds the diagnostic threshold, or when all positive utility actions have been exhausted, the system terminates the interaction and outputs the final diagnosis.

[0087] Based on the methods described above, this embodiment provides a specific "chest pain" diagnostic scenario as an application example:

[0088] Initially, the patient complains of "chest pain, feeling like a stone is pressing on it." The perceptual agent identifies the core symptom "crushing chest pain" and activates relevant nodes in the MDKG. The inference agent quickly identifies candidate diseases such as "myocardial infarction," "angina pectoris," and "gastroesophageal reflux," at which point, due to limited information, the probabilities of each disease are similar.

[0089] The system then proceeded to the first round of interaction. Based on the graph structure, the decision-making agent discovered that "whether the pain radiates to the left shoulder" is a key feature distinguishing between cardiac and non-cardiac chest pain, and that obtaining this feature is extremely cost-effective (requiring only verbal inquiry). Therefore, the system generated the action "inquire about radiating pain".

[0090] The user replied, "I don't experience radiating pain, but I feel more uncomfortable when I lie down."

[0091] The perceptual agent updates its observations: "Radiation pain is absent," and "It worsens when lying flat." Based on this, the inference agent reduces the probability weight of "myocardial infarction" and significantly increases the weights of "angina pectoris" and "gastroesophageal reflux," because "It worsens when lying flat" is a typical feature of the latter.

[0092] The second round of interaction then begins, where the key to differentiating the remaining candidate diseases lies in the presence of "acid reflux and heartburn" or "abnormal electrocardiogram (ECG)." While an ECG is decisive evidence, it is relatively more costly than verbal inquiry (requiring a trip to the hospital). After calculating the system utility function, it is decided to first ask "whether there is acid reflux."

[0093] The user replied, "Yes, I often feel a sour taste in my mouth."

[0094] The perceptual agent confirmed the "acid reflux" characteristic. The inference agent determined that the confidence level of "gastroesophageal reflux" exceeded the diagnostic threshold of 85%.

[0095] Ultimately, the decision-making agent chose to terminate the consultation and output the following advice: "Based on your description, gastroesophageal reflux is the most likely cause. It is recommended that you consult a gastroenterologist and pay attention to regular eating habits."

[0096] In this process, the system completely avoids unnecessary electrocardiograms or CT scans, achieving high-precision triage solely through a low-cost consultation pathway. This embodiment fully demonstrates the beneficial effects of the invention in effectively controlling treatment costs and reducing patient burden while ensuring diagnostic accuracy.

[0097] Example 2

[0098] This embodiment discloses a cost-aware knowledge-enhanced multi-agent sequence diagnostic system.

[0099] A cost-aware, knowledge-enhanced multi-agent sequence diagnostic system includes:

[0100] The graph construction module is configured to: construct a medical diagnostic knowledge graph based on the target disease list, which includes disease nodes, feature nodes, detection nodes, and edges connecting each node;

[0101] The perceptual intelligent agent module is configured to: acquire the user's unstructured medical description, convert it into a structured observation state through natural language processing, and anchor the symptoms and signs in the observation state to the corresponding nodes of the medical diagnostic knowledge graph;

[0102] The reasoning agent module is configured to: perform deterministic reasoning based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease, and determine the optimal action for the next step based on the utility function; the optimal action includes asking about symptoms, ordering tests, and giving a final diagnosis;

[0103] The decision-making agent module is configured to interact with the user by generating corresponding natural language responses or specific diagnostic operation instructions based on the results of the deterministic reasoning.

[0104] Example 3

[0105] The purpose of this embodiment is to provide a computer-readable storage medium.

[0106] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the cost-aware knowledge-enhanced multi-agent sequence diagnostic method as described in Embodiment 1 of this disclosure.

[0107] Example 4

[0108] The purpose of this embodiment is to provide an electronic device.

[0109] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the cost-aware knowledge-enhanced multi-agent sequence diagnostic method as described in Embodiment 1 of this disclosure.

[0110] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0111] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0112] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A cost-aware knowledge-enhanced multi-agent sequence diagnosis method, characterized in that, include: Based on the target disease list, a medical diagnostic knowledge graph is constructed, which includes disease nodes, feature nodes, detection nodes, and edges connecting each node. The construction of the medical diagnostic knowledge graph is achieved through five stages: parallel knowledge extraction, quantitative typicality mapping, hybrid entity alignment, dynamic node upgrading, and attribute enrichment. The implementation process of parallel knowledge extraction and quantization typicality mapping includes: for each disease in the disease list, extracting its associated features using a large language model and verifying the detection items corresponding to the features; adopting a quantization typicality strategy to map the typicality of features described in natural language into discrete weight buckets, which serve as the weights of the edges connecting diseases and features; constructing an action-centered topology to explicitly establish verification edges between features and detection items. The system obtains the user's unstructured medical description, transforms it into a structured observation state through natural language processing, and anchors the symptoms and signs in the observation state to the corresponding nodes of the medical diagnostic knowledge graph. Deterministic reasoning is performed based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease, and the optimal action for the next step is determined based on the utility function, including feature selection and detection selection. Feature selection prioritizes features that are relevant to high-confidence diseases and can distinguish between different candidate diseases based on the discriminative power scores of undetected features. Detection selection involves calculating the utility of candidate detection items and selecting them based on the value, cost, and invasiveness of the verification features. The optimal action includes inquiring about symptoms, ordering tests, and providing a final diagnosis. Based on the results of the deterministic reasoning, the system interacts with the user by generating corresponding natural language responses or specific diagnostic operation instructions.

2. The cost-aware knowledge-enhanced multi-agent sequence diagnosis method as described in claim 1, characterized in that, The process of implementing the hybrid entity alignment includes: for newly extracted terms, recalling candidate nodes from the existing node set using a hybrid retrieval strategy; wherein the hybrid retrieval strategy combines word matching based on overlap coefficient and semantic matching based on cosine similarity of embedding vectors; and using a large language model as a discriminator to determine whether the new term is merged into the candidate nodes or inserted as a new node.

3. The cost-aware knowledge-enhanced multi-agent sequence diagnosis method as described in claim 1, characterized in that, The implementation process of the dynamic node upgrade includes: when a node that exists as a feature in the medical diagnostic knowledge graph is identified as an independent disease, it is promoted from the feature node set to the disease node set, while retaining its original incoming edge relationship and allowing it to establish new outgoing edge relationship, thereby automatically modeling the causal chain between diseases.

4. The cost-aware knowledge-enhanced multi-agent sequence diagnosis method as described in claim 1, characterized in that, The confidence score for each candidate disease is calculated using the following formula: ； in, Indicates disease Confidence score, This represents the set of characteristic states that have been observed at the current moment; Representation of features status The state It includes both presence and absence; The feature scoring function is used to represent the disease. Corresponding features In state The rating at that time.

5. A cost-aware knowledge-enhanced multi-agent sequence diagnostic system, characterized in that, include: The graph construction module is configured to: construct a medical diagnostic knowledge graph based on a list of target diseases, including disease nodes, feature nodes, detection nodes, and edges connecting each node; the construction of the medical diagnostic knowledge graph is achieved through five stages: parallel knowledge extraction, quantized typicality mapping, hybrid entity alignment, dynamic node upgrading, and attribute enrichment. The implementation process of parallel knowledge extraction and quantization typicality mapping includes: for each disease in the disease list, extracting its associated features using a large language model and verifying the detection items corresponding to the features; adopting a quantization typicality strategy to map the typicality of features described in natural language into discrete weight buckets, which serve as the weights of the edges connecting diseases and features; constructing an action-centered topology to explicitly establish verification edges between features and detection items. The perceptual intelligent agent module is configured to: acquire the user's unstructured medical description, convert it into a structured observation state through natural language processing, and anchor the symptoms and signs in the observation state to the corresponding nodes of the medical diagnostic knowledge graph; The reasoning agent module is configured to: perform deterministic reasoning based on the current structured observation state and the medical diagnostic knowledge graph to calculate the confidence score of each candidate disease, and determine the optimal action for the next step based on a utility function, including feature selection and detection selection; wherein, the feature selection prioritizes features that are related to high-confidence diseases and can distinguish different candidate diseases based on the discriminative power scores of undetected features; the detection selection selects features by calculating the utility of candidate detection items and considering the value, cost, and invasiveness of verification features; the optimal action includes inquiring about symptoms, ordering tests, and providing a final diagnosis; The decision-making agent module is configured to interact with the user by generating corresponding natural language responses or specific diagnostic operation instructions based on the results of the deterministic reasoning.

6. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the cost-aware knowledge-enhanced multi-agent sequence diagnostic method as described in any one of claims 1-4.

7. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the cost-aware knowledge-enhanced multi-agent sequence diagnostic method as described in any one of claims 1-4.

Images