Medical data processing method and system based on fusion of rule engine and generative artificial intelligence model

Abstract

The invention relates to a medical data processing method and system based on fusion of a rule engine and a generative artificial intelligence model. The method comprises the following steps: performing text analysis and medical term standardized linking on original medical text data to generate a structured medical information graph; based on the structured medical information graph, performing matching judgment through a rule engine and a medical business rule base, and generating a preliminary decision result with rule confidence; according to the rule confidence, screening items to be subjected to collaborative decision making, constructing prompt information, inputting the prompt information into a generative artificial intelligence model, and performing structured reasoning to obtain a model decision result with model confidence; and finally, performing consistency arbitration processing on the two based on double confidence coefficients to generate a target decision result and decision traceability information. By adopting the method, on the basis of ensuring high reliability and high interpretability of a processing result, the comprehensive processing capability of complex and changeable medical text data can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of smart healthcare and medical insurance payment technology, specifically relating to a medical data processing method and system for intelligent review and decision support under the Diagnosis-Intervention Packet (DIP) payment model. Background Technology

[0002] With the rapid development of medical informatics and artificial intelligence technologies, how to efficiently and accurately process massive, multi-source, and unstructured medical text data (such as electronic medical records, examination reports, and medical orders) to achieve intelligent decision support (such as disease-assisted diagnosis, medical insurance payment review, and clinical pathway recommendation) has become a core requirement of the industry. Currently, the mainstream automated processing technologies are mainly divided into two categories: one is deterministic processing based on predefined rule engines, and the other is generative intelligent processing based on generative artificial intelligence models.

[0003] In traditional technologies, rule-based approaches encode medical expertise, policies, regulations, and clinical guidelines into specific "if-then" logical rules, then match and judge structured or simply preprocessed data. This approach offers advantages such as reliable results and traceable processes when dealing with logically clear and well-defined scenarios. On the other hand, generative artificial intelligence models leverage their powerful natural language understanding and generation capabilities to directly process raw, unstructured text. Through semantic understanding and contextual reasoning, they output judgments or suggestions, demonstrating significant flexibility in handling complex contexts and ambiguous descriptions.

[0004] However, both of the aforementioned approaches have significant limitations. Rule engines, on the other hand, rely heavily on the completeness and timeliness of their rule base. When faced with the numerous variations in expression, complex combinations of complications, and novel situations not covered by existing rules in medical practice, they become rigid and lack adaptability, easily leading to missed judgments or inability to handle certain situations. Generative AI models, while highly flexible, suffer from uncontrollable risks of "illusions," factual errors, and logical inconsistencies in their generated results. In fields like medicine, where accuracy and safety are paramount, the reliability and credibility of their output are difficult to guarantee, and the decision-making process resembles a "black box," lacking the necessary interpretability to meet regulatory and auditing requirements. Therefore, existing technologies have failed to effectively address the core technical challenge of simultaneously achieving flexible and accurate processing of complex and variable medical text data while ensuring high reliability and interpretability of the processing results. Summary of the Invention

[0005] Therefore, it is necessary to provide a medical data processing method and system that integrates rule engines and generative artificial intelligence models to address the above-mentioned technical problems and simultaneously achieve flexible and accurate processing of complex and variable medical text data.

[0006] Firstly, this application provides a medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model, including:

[0007] S1. Perform text parsing on the received raw medical text data to obtain the preliminarily identified medical entities and relationships;

[0008] S2. Perform standardized linking of medical entities and relationships using medical terminology to generate a structured medical information graph; the structured medical information graph is used to represent medical entities linked to the standard terminology system and the relationships between entities;

[0009] S3. Based on the structured medical information graph, the system matches and judges the data with the rule engine and the pre-built medical business rule base to generate preliminary decision results with rule basis identifiers and rule confidence levels.

[0010] S4. Based on the rule confidence level in the preliminary decision results, filter to obtain the collaborative decision items;

[0011] S5. Based on the decision-making items to be coordinated and the corresponding structured medical information image segments, construct prompt information to obtain input prompt information for generative artificial intelligence models;

[0012] S6. Input the input prompt information into the generative artificial intelligence model for structured reasoning to obtain the model decision result with model confidence and reasoning basis;

[0013] S7. Based on the rule confidence in the preliminary decision results and the model confidence in the model decision results, perform consistency arbitration on the preliminary decision results and the model decision results to generate the target decision results and the corresponding decision tracing information.

[0014] In one embodiment, S3 includes:

[0015] S31. Based on the structured medical information graph, extract the linked entities and attribute relationships as matching facts;

[0016] S32. Match the matching facts with the rule conditions in the medical business rule base to obtain a set of rules that meet the trigger conditions;

[0017] S33. For each rule in the rule set, calculate the execution confidence of the rule based on the preset basic confidence of the rule and the identification confidence of the corresponding entity in the matching facts.

[0018] S34. Execute all actions defined in the rule set, obtain the execution results, aggregate the execution results, and generate preliminary decision results with rule basis identifiers and rule confidence levels.

[0019] In one embodiment, S5 includes:

[0020] S51. Based on the decision items to be coordinated, extract the relevant entities, relationships and time information through the structured medical information graph, perform formatted description processing, and generate information summary text;

[0021] S52. Retrieve rule clauses related to the decision-making item to be coordinated from the medical business rule base, and combine them with the judgment of the decision-making item to be coordinated in the preliminary decision results to generate rule context text;

[0022] S53. Combine the information summary text, rule context text, and preset instruction template to generate input prompts for generative artificial intelligence models.

[0023] In one embodiment, S7 includes:

[0024] S71. Obtain the rule judgment and rule confidence corresponding to the decision item to be coordinated from the preliminary decision results, and obtain the corresponding model judgment and model confidence from the model decision results;

[0025] S72. Based on the rule confidence level and the preset high confidence threshold, a first comparison result is obtained; when the first comparison result indicates that the rule confidence level is higher than the high confidence threshold and the rule judgment is clear, the rule judgment is determined as the target judgment, and the reasoning basis in the model decision result is determined as supplementary information.

[0026] S73. Based on the rule confidence level and the preset low confidence level threshold, a second comparison result is obtained; when the second comparison result indicates that the rule confidence level is lower than the low confidence level threshold or the rule judgment is unclear, the model confidence level judgment process is triggered.

[0027] S74. Based on the comparison between the model confidence level and the preset model confidence level threshold, a third comparison result is obtained; when the third comparison result indicates that the model confidence level is higher than the model confidence level threshold, the model judgment is determined as the target judgment, and the reasoning basis is subjected to compliance verification processing.

[0028] S75. When both the rule judgment and the model judgment are clear and conflicting, and both the rule confidence and the model confidence are higher than the preset conflict trigger threshold, a conflict adjudication is made based on the difference between the rule confidence and the model confidence, and an arbitration judgment is generated.

[0029] S76. Integrate and encapsulate the target judgments and corresponding judgment criteria of all decision items to generate target decision results and corresponding decision tracing information.

[0030] In one embodiment, conflict resolution is performed based on the difference between the rule confidence and the model confidence to generate an arbitration judgment, including:

[0031] S751. Calculate the absolute difference between the rule confidence and the model confidence;

[0032] S752. Compare the absolute difference with the preset conflict threshold. If the absolute difference is greater than the conflict threshold, generate a structured conflict report based on rule judgment, model judgment and their respective complete evidence chains, and trigger the external arbitration interface call.

[0033] S753. Receive the ruling result returned by the external arbitration interface call and use the ruling result as the arbitration judgment.

[0034] In one embodiment, the method further includes:

[0035] S8. Collect decision cases that have been confirmed as correct through consistency arbitration to form a positive training sample set;

[0036] S9. Collect model output cases that are judged to be incorrect after consistency arbitration to form a negative training sample set;

[0037] S10. Based on the positive training sample set and the negative training sample set, fine-tune the parameters of the generative artificial intelligence model to obtain the model parameter update amount;

[0038] S11. Update the parameters of the generative artificial intelligence model by updating the model parameters to obtain the updated generative artificial intelligence model.

[0039] Secondly, this application also provides a medical data processing system based on the fusion of a rule engine and a generative artificial intelligence model, including:

[0040] The medical text parsing module is used to parse the received raw medical text data to obtain the initially identified medical entities and relationships.

[0041] The medical terminology standardization linking module is used to standardize and link medical entities and relationships using medical terminology, generating a structured medical information graph. The structured medical information graph is used to represent the medical entities linked to the standard terminology system and the relationships between entities.

[0042] The rule engine matching module is used to perform matching and judgment based on the structured medical information graph, through the rule engine and the pre-built medical business rule library, to generate preliminary decision results with rule basis identifiers and rule confidence levels;

[0043] The decision item filtering module is used to filter the decision items to be collaboratively decided based on the rule confidence level in the preliminary decision results.

[0044] The prompt information construction module is used to construct prompt information based on the decision-making items to be coordinated and the corresponding structured medical information image segments, so as to obtain input prompt information for generative artificial intelligence models;

[0045] The generative artificial intelligence model reasoning module is used to input input prompts into the generative artificial intelligence model for structured reasoning, and obtain model decision results with model confidence and reasoning basis;

[0046] The consistency arbitration module is used to perform consistency arbitration on the preliminary decision results and the model decision results based on the rule confidence in the preliminary decision results and the model confidence in the model decision results, and generate the target decision results and the corresponding decision tracing information.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model as described in the first aspect.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model as described in the first aspect.

[0049] The aforementioned medical data processing method and system, which integrates a rule engine and a generative artificial intelligence model, performs deep analysis and standardizes medical terminology on the original medical text, constructing a structured medical information graph. Subsequently, the rule engine uses this graph to perform matching judgments, generating preliminary decisions with confidence assessments and filtering out those with insufficient confidence or those that cannot be processed for collaborative decision-making. For these collaborative decisions, intelligent prompts containing structured information and relevant rule context are constructed to guide the generative artificial intelligence model in structured reasoning and output model decisions with confidence and justification. A dual-confidence arbitration mechanism is used to integrate and select the preliminary rule results and model reasoning results, thereby generating the final decision and preserving a complete decision traceability chain. This method achieves effective synergy and complementarity between the determinism of the rule engine and the flexibility of the generative artificial intelligence model, significantly improving the processing capability and accuracy of complex and variable medical text data while ensuring high reliability and strong interpretability of the processing results. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model provided by this invention;

[0052] Figure 2 A flowchart illustrating an optional embodiment of the present invention for generating an arbitration determination method;

[0053] Figure 3 This is a schematic diagram of the structure of a medical data processing system based on the fusion of a rule engine and a generative artificial intelligence model, provided by the present invention. Detailed Implementation

[0054] 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.

[0055] In one embodiment, such as Figure 1As shown, a medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps S1 to S7:

[0056] S1. Perform text parsing on the received raw medical text data to obtain the initially identified medical entities and relationships.

[0057] Optionally, a hybrid parsing scheme combining dictionary matching and deep learning is adopted. The dictionary is derived from a set of medical terms in a localized experience knowledge base, including core terms such as disease names, surgical procedure names, drug names, and medical insurance item names. The deep learning model selected is the pre-trained language model BERT (Bidirectional Encoder Representations from Transformers), which is fine-tuned and optimized through medical domain corpora to be specifically adapted to medical text parsing scenarios. In practice, the original medical text is first preprocessed, including removing invalid characters, normalizing newline characters, and unifying capitalization. Then, it is segmented using the jieba word segmentation tool. After obtaining the segmentation results, a dictionary matching algorithm is used to quickly identify known medical entities. Simultaneously, the segmented text is converted into a token sequence that conforms to the input format of the BERT model. This sequence is then input into the fine-tuned BERT model. The model's entity recognition head outputs entity types (such as diagnosis, symptoms, and drugs) and entity boundaries, while the relation extraction head outputs the relationships between entities, such as "diagnosis-accompanying-symptoms" and "treatment-use-drugs". Finally, the medical entities and relationships identified by the two methods are fused, duplicates are removed, and entities with ambiguous boundaries are filled in, resulting in a complete set of medical entities and relationships. Entities include name, type, and text location information, while relationships include subject, predicate, and object entity information.

[0058] S2. Perform standardized linking of medical entities and relationships using medical terminology to generate a structured medical information graph; the structured medical information graph is used to represent medical entities linked to the standard terminology system and the relationships between entities.

[0059] Optionally, the standardization of medical terminology links is based on national and industry standard terminology systems, specifically including the International Classification of Diseases 10th Revision Clinical Modification (ICD-10-CM), the National Healthcare Security Administration DIP group terminology set, the generic drug name directory, and surgical procedure classification codes. The linking process combines string similarity matching and semantic similarity matching. String similarity is calculated using edit distance and cosine similarity to quickly match completely identical or highly similar terms. Semantic similarity matching involves inputting medical entities and standard terms into a pre-trained language model in the medical field, such as MedBERT, to generate fixed-dimensional semantic embedding vectors. The cosine similarity between these vectors is then calculated to handle variations in terminology expression, such as matching "lung infection" with "pneumonia." The structured medical information graph is constructed using a graph data model, with standardized medical entities as nodes. Node attributes include standard terminology code, terminology name, original entity name, and entity type. Edges are the relationships between entities, with edge attributes including relationship type, matching confidence, and the standard terminology system upon which the link is based. Specifically, the standardized entities and relationships after fusion are converted according to the graph data format, and the associations between entities are stored using an adjacency list structure. Graph databases (such as Neo4j) can be used for persistent storage.

[0060] S3. Based on the structured medical information graph, the system matches and judges the data with the rule engine and the pre-built medical business rule base to generate preliminary decision results with rule basis identifiers and rule confidence levels.

[0061] Optionally, the rule engine adopts an inference engine based on the Rete algorithm, which achieves efficient pattern matching by constructing a rule network and supports the rapid execution of complex rules. The medical business rule base is built on a localized experience knowledge base and includes DIP (Diagnosis-Intervention Packet) compliance verification rules, medical record homepage quality control rules, medical insurance payment review rules, etc. The rules are stored in the form of "if-then" production rules. Each rule includes attributes such as condition part, action part, rule identifier, and rule weight. The condition part is composed of multiple atomic conditions combined by logical operators (AND / OR / NOT). Atomic conditions involve entity attribute judgment (such as "whether the diagnosis code belongs to the DIP core disease group") and relation existence judgment (such as "whether there is a high-value consumable usage record that does not match the diagnosis"). In practice, the entities and relationships in the structured medical information graph are first transformed into factual data that the rule engine can recognize. The factual data is presented in the form of triples (subject-verb-object) and entity attribute key-value pairs. Then, the rule engine loads the medical business rule library, matches the factual data with the rule conditions using the Rete algorithm, traverses the pattern nodes in the rule network, activates the successfully matched rules, and generates preliminary decision results in the action part, such as "DIP enrollment is compliant" or "There are defects in the medical record front page". The calculation of rule confidence is based on the completeness of rule matching and rule weight. The confidence of a rule that completely matches all atomic conditions is the maximum value of the rule weight. For partial matching, the confidence is calculated by accumulating the weights of the matched atomic conditions. Finally, each preliminary decision result is associated with the corresponding rule identifier and rule confidence, ensuring that the decision-making process is traceable.

[0062] S4. Filter the results based on the rule confidence level in the preliminary decision results to obtain the collaborative decision items.

[0063] Optionally, the screening process adopts a dynamic threshold mechanism. The threshold is determined based on the statistical distribution of historical rule matching data. The basic threshold range is set by calculating the mean and standard deviation of the rule confidence. At the same time, it supports personalized adjustments according to specific business scenarios (such as DIP enrollment review and medical record quality control). For example, the threshold for DIP enrollment review is higher than the threshold for regular medical record data verification. Specifically, the process begins by collecting rule confidence data for all preliminary decision results from historical periods. Statistical analysis methods are then used to calculate the data distribution characteristics and determine a default threshold, such as setting a lower threshold based on the 3σ principle. Business personnel then adjust the threshold within the default range according to the rigor of the actual application scenario to obtain the final screening threshold. During screening, all preliminary decision results are traversed, and the rule confidence of each result is extracted and compared with the screening threshold. Preliminary decision results with rule confidence below the threshold are marked as collaborative decision items. Simultaneously, structured medical information image segments corresponding to the collaborative decision items are extracted. These segments contain entity nodes, relational edges, and original text association information related to the decision, ensuring that the subsequent generative artificial intelligence model can obtain complete contextual support during inference. Furthermore, the marked collaborative decision items are categorized and grouped according to business type (such as group entry review and quality control verification) to facilitate the batch construction of prompt information.

[0064] S5. Based on the decision-making items to be coordinated and the corresponding structured medical information image segments, construct prompt information to obtain input prompt information for generative artificial intelligence models.

[0065] Optionally, the prompt information is constructed using a combination of structured templates and dynamic content filling. The structured template has a predefined fixed format and includes a problem description module, a structured information module, a constraint module, and an output requirement module. The problem description module clarifies the core issue to be decided, such as "whether the DIP enrollment of this medical case is compliant". The structured information module is used to fill in the structured medical information image segments corresponding to the collaborative decision item. The constraint module clarifies the basis for reasoning, such as "reasoning is based only on the provided structured information and medical insurance DIP payment rules". The output requirement module specifies the format of the model output, such as "including decision conclusion, reasoning basis, and confidence level". First, select the corresponding structured template based on the business type of the decision-making item to be collaborated on. Then, extract the entity triples and entity attribute information related to the decision-making item from the structured medical information graph, and convert them into a form combining natural language description and structured triples, such as "Entity information: (patient, diagnosis, community-acquired pneumonia) (patient, use, cefuroxime); entity attribute: community-acquired pneumonia ICD-10 code J15.901"), and fill it into the structured information module of the template. Subsequently, supplement the constraints according to the business rules, clarify the boundaries and basis of the model reasoning, and finally integrate the content of each module of the template, remove redundant information, optimize the coherence and accuracy of the language expression, and form input prompt information for the generative artificial intelligence model. Preferably, the length of the prompt information is controlled within the input limit range of the generative artificial intelligence model to ensure that the model can completely parse it.

[0066] S6. Input the input prompt information into the generative artificial intelligence model for structured reasoning to obtain the model decision result with model confidence and reasoning basis.

[0067] Optionally, the generative artificial intelligence model selected is Qwen3.0 (Tongyi Qianwen 3.0), which has been fine-tuned with medical corpus and has the ability to understand medical knowledge and make logical reasoning. It can handle inputs that combine structured information with natural language. The pre-constructed input prompts are input into the model according to the required format (e.g., text string format). The model first uses the context understanding module to parse the structured medical information, decision-making problem, and constraints in the prompts, identifying core entities and key relationships. Then, it uses medical domain knowledge (e.g., DIP grouping rules, clinical treatment guidelines) and logical reasoning ability learned during pre-training to perform multi-step structured reasoning. During the reasoning process, the rationality of the relationships between entities is first verified, such as "the matching of community-acquired pneumonia with cefuroxime". Then, the compliance or rationality of the decision-making item is judged in combination with the constraints. At the same time, the key logical nodes in the reasoning process are recorded as the basis for reasoning. After the reasoning is completed, the model outputs a structured model decision result, including the decision conclusion (e.g., "DIP enrollment is compliant" "It is recommended to adjust the diagnostic code"), model confidence, and a text description of the reasoning basis. The model confidence is calculated by the softmax function of the model output layer, corresponding to the probability value of the decision conclusion. The reasoning basis is clearly associated with the structured medical information and medical knowledge in the input prompts, ensuring the interpretability of the decision-making process.

[0068] S7. Based on the rule confidence in the preliminary decision results and the model confidence in the model decision results, perform consistency arbitration on the preliminary decision results and the model decision results to generate the target decision results and the corresponding decision tracing information.

[0069] Optionally, the consistency arbitration process adopts a mechanism that combines weighted fusion and threshold judgment. The weight parameters are preset according to the business scenario. For example, in the medical insurance payment compliance decision, the weight of rule confidence is set to 0.6 and the weight of model confidence is set to 0.4. The weight setting is based on the importance of the objectivity of the rules and the flexibility of the model in different scenarios. The arbitration threshold is determined by verification through historical decision data and is used to judge the reliability of the comprehensive confidence. First, the rule confidence score of the preliminary decision result and the model confidence score of the model decision result are extracted. The overall confidence score is then calculated according to preset weights using the formula: Overall Confidence Score = Rule Confidence Score × Rule Weight + Model Confidence Score × Model Weight. If the preliminary decision result is consistent with the model decision result and the overall confidence score is higher than the arbitration threshold, this decision result is directly used as the target decision result. If the two results are consistent but the overall confidence score is lower than the threshold, it is marked as an item requiring manual review. If the two results are inconsistent, their respective overall confidence score contributions are calculated, and the decision result with the higher contribution score is selected as the target decision result. If the difference in contribution scores is less than the preset threshold, it is also marked as an item requiring manual review. The generation of decision traceability information integrates rule matching records (rule identifier, matching conditions, rule confidence score), model inference basis (inference nodes, model confidence score), and arbitration process (weight parameters, threshold, arbitration logic), storing them in a structured format (such as JSON) to ensure that every step of the target decision result is auditable and traceable. Furthermore, the traceability information is stored in association with the target decision result.

[0070] The aforementioned medical data processing method, which integrates rule engines and generative artificial intelligence models, fully combines the compliance advantages of rule engines with the complex reasoning capabilities of generative artificial intelligence models through a coherent technical approach from medical text parsing and terminology standardization to rule matching, model reasoning, and consistency arbitration. This approach ensures the standardization and reliability of medical decisions while enhancing adaptability to complex and variable scenarios. Furthermore, the generated decision traceability information meets the regulatory and auditing requirements of the medical field, ultimately achieving precision, efficiency, and traceability in medical data processing. This effectively solves the rigidity or unreliability issues inherent in traditional single technologies in medical data processing.

[0071] In one embodiment, S3 includes:

[0072] S31. Based on the structured medical information graph, extract the linked entities and attribute relationships as matching facts.

[0073] Optionally, the structured medical information graph uses a standard terminology system as its core association. Nodes store standardized linked medical entities, including entity name, standard code (such as ICD-10-CM diagnostic code, DIP group code), entity type (diagnosis, surgery, drug, etc.), and original text location. Edges store attribute relationships between entities, including relationship type (such as "diagnosis-accompanying-symptom" or "treatment-use-drug"), relationship matching confidence, and the name of the standard system on which the link is based. When extracting matching facts, a graph traversal algorithm is used to traverse the nodes and edges of the structured medical information graph. First, all entities that have completed standardized links are identified, and the core attributes of each entity are extracted and stored in the format of "entity identifier-attribute name-attribute value" triples. Then, the attribute relationships between entities are extracted and stored in the format of "subject entity identifier-relationship type-object entity identifier-relationship confidence". During the extraction process, entities with incomplete standardized links and relationships with confidence levels below a preset threshold are filtered out to ensure the reliability of the matching facts. At the same time, for DIP payment-related scenarios, entities and attribute relationships related to medical insurance settlement and medical treatment compliance are extracted first, such as the association between diagnosis and surgery, and the matching relationship between drugs and diagnoses, to form a structured set of matching facts.

[0074] S32. Match the facts with the rule conditions in the medical business rule base to obtain the set of rules that meet the trigger conditions.

[0075] Optionally, the medical business rule base is built on a localized experience knowledge base, including DIP compliance verification rules, medical record homepage quality control rules, and medical insurance payment review rules. Each rule consists of attributes such as a condition part, an action part, a rule identifier, a basic confidence level, and a rule weight. The condition part is composed of multiple atomic conditions combined through logical operators (AND / OR / NOT). The atomic conditions clearly point to the judgment criteria of entity attributes or relationships between entities, such as "the diagnosis code belongs to the DIP core disease group," "the surgical operation and the diagnosis have an adaptation relationship," and "the use of high-value consumables matches the severity of the condition." The matching process uses the Rete algorithm to achieve efficient pattern matching. This algorithm constructs a rule network, decomposes the atomic conditions in the rule conditions into network nodes, first preprocesses the extracted matching facts, converts the fact data in triple and quadruple format into fact objects that the rule engine can recognize, and then inputs the fact objects into the rule network. They are then matched and filtered by pattern nodes and connection nodes in sequence, triggering rules that fully satisfy all condition parts to form a rule set. For rule conditions containing "OR" logic, the condition is considered met as long as any atomic condition is satisfied; for rule conditions containing "NOT" logic, the corresponding fact must be confirmed to not exist before the rule can be triggered, to ensure the accuracy and completeness of the rule set.

[0076] S33. For each rule in the rule set, calculate the execution confidence of the rule based on the preset basic confidence of the rule and the identification confidence of the corresponding entity in the matching facts.

[0077] Optionally, the pre-set baseline confidence level for a rule is a fixed value set based on the rule's compliance priority, clinical practice acceptance, and policy alignment. This value is determined through joint review by clinicians, medical insurance experts, and technical personnel. For example, compliance rules directly related to the security of the medical insurance fund have a higher baseline confidence level than general quality control rules. The entity identification confidence level is generated during the medical entity extraction and standardized linking process, reflecting the reliability of entity identification and standardization. For instance, entities directly identified through dictionary matching have a higher identification confidence level than entities identified through model prediction, and entities with perfectly matched standardized links have a higher identification confidence level than entities with semantic matching. The execution confidence level is calculated using a weighted aggregation algorithm: for a single rule, the identification confidence levels of all entities involved in the rule's conditions are first extracted, and their average value is calculated as the entity's comprehensive confidence level. Then, the entity's comprehensive confidence level is multiplied by the pre-set baseline confidence level to obtain the rule's execution confidence level. If the rule conditions do not involve differences in entity identification confidence levels, such as rules based solely on fixed coding judgments, the rule's baseline confidence level is directly used as the execution confidence level to ensure that the execution confidence level objectively reflects the reliability of rule matching.

[0078] S34. Execute all actions defined in the rule set, obtain the execution results, aggregate the execution results, and generate preliminary decision results with rule basis identifiers and rule confidence levels.

[0079] Optionally, the action portion of each rule in the rule set specifies the concrete execution operation, including but not limited to "marking DIP enrollment as compliant," "notifying deficiencies in the medical record front page," "warning of medical insurance violation risks," and "suggesting adjustments to the treatment plan." When executing rule actions, they are executed sequentially in descending order of execution confidence to avoid conflicts between high-confidence and low-confidence rules. After execution, the results of all rules are collected and integrated using a result aggregation algorithm: for conflict-free execution results, the corresponding rule identifier and execution confidence are directly retained; for conflicting execution results, such as the same entity being marked as "compliant" and "non-compliant" by different rules, the result of the rule with higher execution confidence is retained, and the identifiers and confidence differences of the conflicting rules are recorded; for complementary execution results, such as different rules indicating quality control deficiencies in different dimensions, they are merged to form a complete set of execution results. Finally, the integrated execution results are packaged in the format of "decision conclusion - rule basis identifier - rule confidence (i.e. execution confidence) - matching fact source" to generate preliminary decision results. The rule basis identifier includes the unique number of the triggering rule and key matching conditions to ensure that the decision results are traceable, while meeting the actual needs of primary healthcare institutions for DIP payment review and treatment standard management.

[0080] In the above embodiments, reliable matching facts are extracted from the structured medical information graph, relevant rule sets are triggered through efficient rule matching, and execution confidence is calculated by combining the basic confidence of the rules and the confidence of entity recognition. Finally, the results are aggregated to form a preliminary decision result with traceability basis. This not only ensures the alignment of the decision with the needs of grassroots DIP payment reform, but also improves the accuracy and interpretability of the decision through quantitative confidence and rule traceability, effectively supporting the compliance management of medical insurance funds and the standardization of diagnosis and treatment behavior.

[0081] In one embodiment, S5 includes:

[0082] S51. Based on the decision items to be coordinated, extract the relevant entities, relationships and time information from the structured medical information graph, perform formatted description processing, and generate information summary text.

[0083] Optionally, the collaborative decision-making items clearly define core business scenarios requiring further reasoning, such as doubts about the compliance of DIP enrollment or deficiencies in the medical record cover sheet that need confirmation. When extracting relevant information, a graph-based association traversal algorithm is used. Starting with the core entity corresponding to the collaborative decision-making item, it traverses adjacent nodes and edges in the structured medical information graph, prioritizing the extraction of directly related entities (such as symptoms, surgeries, and medications corresponding to diagnoses), entity attributes (standard codes, types, and text descriptions), relationships between entities (such as "diagnosis-treatment-surgery" and "medication-matching-diagnosis"), and key time information (such as the time of prescription issuance, treatment implementation time, and medical record submission time). During the extraction process, association information with a relationship confidence level below a preset threshold is filtered to ensure the relevance and reliability of the information. The formatted description processing uses a fixed natural language template, organizing the content logically according to "core entity information → related entities and relationships → key time information". Core entities are clearly marked with standard terminology codes and types, relationships are clearly expressed with predicates, and time information is uniformly formatted as "YYYY-MM-DD". When there is no specific date, the time dimension is marked, such as "during diagnosis and treatment". Finally, a clear and complete information summary text is generated, ensuring that the generative artificial intelligence model can quickly capture core diagnosis and treatment related information.

[0084] S52. Retrieve rule clauses related to the decision-making item to be coordinated from the medical business rule base, and combine them with the judgment of the decision-making item to be coordinated in the preliminary decision results to generate rule context text.

[0085] Optionally, the medical business rule base includes structured rules such as DIP compliance verification, medical insurance payment review, and medical record homepage quality control. Each rule includes attributes such as a unique identifier, condition description, action instructions, and applicable scenarios. When retrieving relevant rule clauses, a combination of keyword matching and scenario classification filtering is used. The core business type of the decision-making item to be collaborated on (such as "DIP enrollment" and "coding compliance") and key entity type (such as diagnosis, surgery, and drugs) are used as keywords. The rule conditions and applicable scenario fields in the rule base are traversed to filter out highly relevant rule sets. Furthermore, the rules are sorted according to priority (such as rules related to medical insurance fund security having higher priority than general quality control rules), and the top 3-5 core rules are selected. When generating the rule context text, the core conditions and action instructions of the rules are presented one by one, while redundant technical details are hidden. Then, the judgment conclusion of the decision-making item to be collaborated on, the rule basis identifier, and the rule confidence level are added to the preliminary decision result. It is clearly stated that "the preliminary decision needs further collaborative verification due to insufficient rule confidence level", forming a logical chain of "rule clause → preliminary judgment → collaborative needs", providing clear rule reference and decision background for the generative artificial intelligence model.

[0086] S53. Combine the information summary text, rule context text, and preset instruction template to generate input prompts for generative artificial intelligence models.

[0087] Optionally, the preset instruction template is customized for medical scenarios, clearly defining the reasoning task, output requirements, and constraints of the generative AI model. Core content includes: "Based on the provided information summary and rule context, determine the compliance / reasonableness of the XX collaborative decision item; the output conclusion must include specific reasoning basis (related rule clauses and medical information) and model confidence; the reasoning must align with grassroots DIP payment reform practices, ensuring the conclusion is interpretable and complies with clinical and medical insurance policy requirements." During the assembly process, the preset instruction template is first embedded, then the information summary text and rule context text are sequentially assembled. Connecting words (such as "Specific treatment information is as follows:" and "Related business rules are as follows:") are used to achieve a smooth transition and remove duplicate information. If the rule identifier is already presented in the rule context, it will not be repeated. After assembly, the text length is validated. If it exceeds the input limit of the generative AI model, the principle of "prioritizing core information" is adopted. Basic rules already understood by medical staff are simplified in the rule context, while high-priority rules and controversial rules are retained first. This ensures that the input prompts fully cover key information and meet the model's processing requirements, ultimately generating standardized input text for the generative AI model.

[0088] In the above embodiments, by accurately extracting diagnosis and treatment-related information, filtering core business rules, and standardizing input prompts, a complete, focused, and grassroots medical scenario-compliant reasoning basis is provided for the generative artificial intelligence model. This avoids reasoning bias caused by information redundancy and ensures the policy compliance and clinical adaptability of the model's decisions. It effectively improves the accuracy and interpretability of the generative artificial intelligence model's output in complex medical decision-making scenarios, and provides efficient technical support for collaborative decision-making in DIP payment reform.

[0089] In one embodiment, S7 includes:

[0090] S71. Obtain the rule judgment and rule confidence level corresponding to the decision item to be coordinated from the preliminary decision results, and obtain the corresponding model judgment and model confidence level from the model decision results.

[0091] Optionally, the decision-making items to be collaboratively decided are associated with the preliminary decision results and model decision results through a unique identifier. This identifier includes the business scenario number corresponding to the decision item (such as DIP enrollment review, medical record quality control number) and the unique ID of the core entity (such as medical record number, patient treatment ID), ensuring the accuracy of data association. When extracting information from the preliminary decision results, the rule judgment and rule confidence level corresponding to the decision-making items to be collaboratively decided are accurately located. The rule judgment is a clear conclusion output by the rule engine after matching based on the medical business rule base, such as "DIP enrollment is compliant" or "There are defects in the medical record homepage". The rule confidence level is the previously calculated execution confidence level, reflecting the reliability of rule matching. When extracting information from the model decision results, the corresponding model judgment (such as "Agree to DIP enrollment" or "Suggest supplementing diagnostic coding") and model confidence level are obtained simultaneously. The model confidence level is calculated by the softmax function of the output layer of the generative artificial intelligence model, corresponding to the probability value of the model judgment. During the extraction process, data integrity is verified to ensure that the rule judgment, rule confidence, model judgment, and model confidence of each collaborative decision item are fully obtained. If any is missing, it is marked as data anomaly, and a manual review process is triggered to ensure the reliability of the data source for subsequent decision processing.

[0092] S72. Based on the rule confidence level and the preset high confidence threshold, a first comparison result is obtained; when the first comparison result indicates that the rule confidence level is higher than the high confidence threshold and the rule judgment is clear, the rule judgment is determined as the target judgment, and the reasoning basis in the model decision result is determined as supplementary information.

[0093] Optionally, the preset high-confidence threshold is determined based on the compliance priority and historical decision data of the DIP payment reform scenario. It is jointly reviewed and formulated by clinical experts, medical insurance managers, and technical personnel. Among them, the high-confidence threshold for rules directly related to the security of medical insurance funds and the core compliance of diagnosis and treatment is higher than that for general quality control rules. When comparing the rule confidence with the high-confidence threshold, a numerical comparison method is used. If the rule confidence is higher than the threshold and the rule judgment is a clear conclusion (not a vague statement), then the rule judgment is directly determined as the target judgment, because the rule judgment is based on a standardized medical business rule base and has strong policy compliance and traceability. At the same time, the reasoning basis in the model decision results is extracted as supplementary information. This supplementary information needs to be screened for clinical logic, localized diagnosis and treatment experience, etc., related to the rule judgment. For example, when the rule judgment is "DIP enrollment compliance", the "adaptability analysis of diagnosis and treatment items and core diseases in DIP grouping" in the model reasoning basis is supplemented. This ensures the authority of the target judgment and enhances the comprehensiveness and interpretability of the decision through supplementary information, which is in line with the dual requirements of "policy basis + clinical evidence" in the grassroots medical insurance review.

[0094] S73. Based on the comparison between the rule confidence and the preset low confidence threshold, a second comparison result is obtained; when the second comparison result indicates that the rule confidence is lower than the low confidence threshold or the rule judgment is unclear, the model confidence judgment processing is triggered.

[0095] Optionally, the preset low confidence threshold is the lowest critical value at which rule judgment has reference value. A value below this threshold indicates insufficient rule matching or the existence of special cases not covered by the rule. Its determination logic is consistent with the high confidence threshold, and the low confidence threshold is lower than the high confidence threshold. Ambiguous rule judgment refers to a vague statement in the rule output, such as "compliance cannot be determined" or "further verification is required." This is usually caused by partial rule matching, atomic condition conflicts, or special clinical scenarios not included in the rule base. After comparing the rule confidence with the low confidence threshold, if the rule confidence is lower than the threshold or the rule judgment is ambiguous, model confidence judgment processing is immediately triggered. This is because at this point, the rule judgment can no longer provide reliable support for decision-making. It is necessary to rely on the generative artificial intelligence model's understanding and reasoning ability of complex scenarios to supplement the limitations of the rules and ensure adaptability to complex and variable medical scenarios.

[0096] S74. Based on the comparison between the model confidence level and the preset model confidence level threshold, a third comparison result is obtained; when the third comparison result indicates that the model confidence level is higher than the model confidence level threshold, the model judgment is determined as the target judgment, and the reasoning basis is subjected to compliance verification processing.

[0097] Optionally, the preset model confidence threshold is determined based on the training effect of the generative AI model in the medical field and the reliability requirements of DIP compliance scenarios. It is calibrated by comparing historical model decision data with actual compliance results to ensure that the accuracy of the model's judgment meets the rigorous requirements of the medical industry. When comparing the model confidence with the model confidence threshold, if the model confidence is higher than the threshold, the model judgment is determined as the target judgment because the model inference has high reliability at this time. The compliance verification process needs to combine the localized medical business rule base, medical insurance payment policies, and clinical diagnosis and treatment guidelines to check the model's reasoning basis item by item. The focus is on verifying whether the reasoning basis complies with local DIP grouping rules, medical insurance fund payment scope, and clinical pathway standards. Content in the reasoning basis that conflicts with policies or contradicts clinical logic is excluded. If there is a conflict, the conflict point is marked and the original text of the model's reasoning basis is retained to ensure that the target judgment has both the ability to adapt to complex scenarios and complies with policy compliance requirements, avoiding decision risks caused by model "illusions".

[0098] S75. When both the rule judgment and the model judgment are clear and conflicting, and both the rule confidence and the model confidence are higher than the preset conflict trigger threshold, a conflict adjudication is made based on the difference between the rule confidence and the model confidence, and an arbitration judgment is generated.

[0099] Optionally, the preset conflict trigger threshold is a critical value indicating that both rule-based and model-based judgments possess high reliability. This threshold is higher than both the high-confidence threshold and the model confidence threshold, ensuring that conflict adjudication is triggered only when both conclusions are credible. Conflict between rule-based and model-based judgments refers to contradictory conclusions, such as a rule-based judgment of "compliant" versus a model-based judgment of "non-compliant." In this case, adjudication is based on the difference between the rule confidence and model confidence, with the difference threshold determined by historical conflict handling effectiveness and business priorities. If the difference between the rule confidence and model confidence exceeds the difference threshold, the judgment with higher confidence is selected as the arbitration decision. If the difference is less than or equal to the difference threshold, in line with the core objectives of DIP payment reform, rule-based judgment is prioritized as the arbitration decision, as it better aligns with the rigid requirements of medical insurance policies. In special scenarios (such as those involving complex complication treatments), priority weight adjustments can be made based on pre-set priorities by clinical experts, ensuring that arbitration decisions are based on quantitative data while also meeting the actual needs of grassroots medical insurance review, balancing policy compliance and clinical flexibility.

[0100] S76. Integrate and encapsulate the target judgments and corresponding judgment criteria of all decision items to generate target decision results and corresponding decision tracing information.

[0101] Optionally, the integration process first categorizes the target judgments according to the business type of the decision-making items to be coordinated (such as DIP enrollment, medical record quality control, and medical insurance settlement) to ensure the structured nature of the decision results. The judgment basis corresponding to the target judgment includes rule basis (rule identifier, matching conditions), model reasoning basis (content after compliance verification), and arbitration logic (if a conflict is triggered, the arbitration logic). It needs to be organized according to the logic of "core basis + supplementary basis". The core basis is the direct basis for determining the target judgment, and the supplementary basis is the auxiliary content to enhance the credibility of the decision. Standardized data format is used during encapsulation. The target decision result includes a unique identifier of the decision item, target judgment, business type, and decision status (such as "confirmed" or "requires manual review"). The decision traceability information includes rule-related information (rule judgment, rule confidence, rule basis identifier), model-related information (model judgment, model confidence, verified reasoning basis), conflict handling information (if any), and decision timestamp, ensuring that each target decision result can be traced back to the entire process of rule matching, model reasoning, and arbitration handling, which meets the auditability requirements of medical insurance supervision and adapts to the decision review needs of primary healthcare institutions.

[0102] In the above embodiments, by accurately extracting decision-related information, making hierarchical judgments based on multi-level thresholds, making quantitative adjudications for conflict scenarios, and integrating and tracing the source in a structured manner, the policy compliance of rule-based judgments and the complex scenario adaptability of model-based judgments are fully integrated. This ensures the accuracy and compliance of DIP payment-related decisions, while also improving the interpretability and auditability of decisions. It effectively adapts to the medical insurance management needs of grassroots county-level medical consortia and provides dual protection for the safety of medical insurance funds and the standardization of diagnosis and treatment.

[0103] In an optional embodiment, such as Figure 2 As shown, conflict resolution is performed based on the difference between the rule confidence and the model confidence, generating an arbitration judgment, including:

[0104] S751. Calculate the absolute difference between the rule confidence and the model confidence.

[0105] Optionally, the rule confidence is a quantified value calculated based on the rule's basic confidence and entity recognition confidence during rule execution, while the model confidence is the probability value of the corresponding decision conclusion output by the large model through structured inference. Both are in the numerical range of 0-1 and retain consistent decimal precision to ensure calculation accuracy. When calculating the absolute difference, standard absolute value operation logic is used, i.e., calculated using the mathematical formula "Absolute Difference = |Rule Confidence - Model Confidence|". Specifically, the original values ​​of the rule confidence and model confidence corresponding to the conflict resolution item are first extracted, ensuring that both data are unrounded original calculation results to avoid precision loss in intermediate steps. Then, the basic numerical calculation function is called to perform the absolute value operation, and the result retains the same number of decimal places as the original confidence, ensuring that the difference accurately reflects the degree of difference between the two confidence levels. This calculation process maintains consistency with the confidence calculation logic in previous steps, and the data sources all rely on specific business scenarios such as DIP compliance verification and medical record quality control to ensure the relevance and effectiveness of the difference calculation.

[0106] S752. Compare the absolute difference with the preset conflict threshold. If the absolute difference is greater than the conflict threshold, generate a structured conflict report based on rule judgment, model judgment and their respective complete evidence chains, and trigger the external arbitration interface call.

[0107] Optionally, the preset conflict threshold is a critical value determined based on the level of medical insurance fund security and the strictness of treatment guidelines in the DIP payment reform. It is jointly reviewed and formulated by clinical experts, medical insurance managers, and technical personnel to determine whether external professional intervention is required for adjudication of conflicts. During comparison, the calculated absolute difference is compared with the conflict threshold. If the absolute difference is greater than the conflict threshold, it indicates a significant difference between the rule-based judgment and the model-based judgment, with both having high confidence levels, requiring external arbitration. The generation of structured conflict reports must follow a fixed template, including a unique identifier for the decision-making item to be coordinated, a conflict scenario classification (e.g., DIP enrollment compliance conflict, medical record coding compliance conflict, etc.), rule-based judgment conclusions and a complete chain of evidence (rule identifier, matching medical business rule clauses, related entities and relational data in the structured medical information diagram), model-based judgment conclusions and a complete chain of evidence (reasoning basis, cited localized experience knowledge base content, clinical pathway standard clauses), absolute difference, conflict threshold, and other core information to ensure the report is comprehensive and logically clear. The external arbitration interface adopts a RESTful API architecture. The interface address is pre-configured and securely authenticated. When calling, authentication is performed through the API key. The structured conflict report is uploaded as a request parameter in JSON format. A three-retry mechanism is set up. If the call fails, a manual notification process is triggered to ensure that the arbitration request can be effectively delivered. The data transmission process complies with the relevant requirements for medical data privacy protection.

[0108] S753. Receive the ruling result returned by the external arbitration interface call and use the ruling result as the arbitration judgment.

[0109] Optionally, the ruling results returned by the external arbitration interface are standardized structured data. These data originate from a professional conclusion reached by an external arbitration panel composed of medical insurance experts, clinical staff, and technical personnel, based on a comprehensive review of received structured conflict reports. Upon receiving the ruling results, the returned data is first verified for completeness. This involves checking for required fields such as the ruling conclusion (e.g., supporting rule-based judgments, supporting model-based judgments, or corrective judgments), the ruling basis (expert review opinions, cited medical insurance policy clauses, and clinical treatment guidelines), arbitrator information, and the ruling timestamp. If any fields are missing or formatted incorrectly, an error message is immediately returned, triggering a manual review process. If the data is complete, the ruling conclusion is verified for legality, ensuring it complies with DIP payment reform policies and clinical treatment standards, excluding invalid or non-compliant rulings. After successful verification, the ruling conclusion is directly mapped to the arbitration judgment, and the ruling basis, arbitrator information, and ruling time are simultaneously recorded and incorporated into the decision-making traceability information system to ensure the traceability of the arbitration judgment. At the same time, the arbitration judgment will be transmitted to the subsequent results integration stage to ensure that it can be packaged in a unified manner with the target judgments of other decision items to form a complete target decision result that meets the auditable requirements of medical insurance supervision and the actual business needs of primary medical institutions.

[0110] In the above embodiments, by accurately calculating the absolute difference of confidence, standardizing the generation of conflict reports, standardizing the invocation of external arbitration interfaces, and verifying the receipt of ruling results, the fairness and compliance of conflict decision-making scenarios are ensured, and the serious discrepancies between rule judgment and model judgment are effectively resolved. This not only safeguards the security of medical insurance funds and the standardization of diagnosis and treatment in DIP payment reform, but also enhances the authority and traceability of decision results, providing a reliable conflict solution for complex decision-making scenarios in county-level medical consortia.

[0111] In one embodiment, the method further includes:

[0112] S8. Collect decision cases that have been confirmed as correct through consistent arbitration to form a positive training sample set.

[0113] Optionally, decision cases confirmed as correct through consistency arbitration are collected. During the collection process, the real medical data contained in the cases strictly adheres to the requirements of laws and regulations such as the Personal Information Protection Law, performing data anonymization processing, such as removing identifiable personal information like patient names and ID numbers. Following the principle of least privilege, only core medical information and decision-making process data relevant to the decision are extracted to form a positive training sample set. Decision cases confirmed as correct through consistency arbitration must contain complete business scenario context and decision-making process information, specifically including structured medical information image segments corresponding to the decision-making item to be collaboratively decided, input prompts for the generative artificial intelligence model, the model's original output results, rule judgment results, arbitration conclusions, and decision tracing information. During the collection process, the entire process data is first linked by the unique identifier of the decision result, and cases with an arbitration conclusion of "confirmed as correct" are screened, excluding cases requiring manual review or with incomplete data. Subsequently, the selected cases are structured, extracting core elements: standardized data of medical entities and relationships, relevant rules and clauses of the DIP policy, reasonable logical fragments in the model's reasoning basis, and the correct decision conclusion confirmed by arbitration, organized in the format of "input information - intermediate process - correct output". Simultaneously, the business type corresponding to each case (such as DIP enrollment compliance review, medical record homepage quality control) and the applicable localized rule categories are labeled to ensure that the samples are highly consistent with the grassroots DIP payment scenario. Finally, the structured cases are uniformly stored in JSON format to form a positive training sample set, providing high-quality data support that fits actual business for generative artificial intelligence models to learn correct decision-making logic.

[0114] S9. Collect model output cases that are judged to be incorrect after consistency arbitration to form a negative training sample set.

[0115] Optionally, for model output cases determined to be erroneous through consistency arbitration, the focus should be on the differences between the model output and the correct conclusion, as well as the reasons for the error. During collection, the entire process data is linked using a unique decision identifier. Cases with an arbitration conclusion of "model output error" are selected, with a focus on extracting the model output result, error type annotation (e.g., misunderstanding of DIP rules, misjudgment of medical entity relationships, coding logic errors, inappropriate policy adaptation), explanation of the error reasons pointed out by the arbitration, the correct decision conclusion, and its corresponding basis. In the structured processing stage, the input information and business scenario annotations consistent with the positive training samples must be retained. At the same time, the specific content of the model's erroneous output, the erroneous link (e.g., missing reasoning basis, semantic understanding deviation), and the direction of correction (e.g., strengthening localized DIP grouping rule learning, optimizing medical entity association judgment) must be clearly annotated. The case storage format remains consistent with the positive training samples. This negative training sample set will directly point to the weak links of the model in grassroots DIP scenarios, providing precise targets for targeted optimization.

[0116] S10. Based on the positive training sample set and the negative training sample set, fine-tune the parameters of the generative artificial intelligence model to obtain the model parameter update amount.

[0117] Optionally, parameter fine-tuning employs LoRA (Low-Rank Adaptation) technology. This technique freezes most of the original model's parameters by inserting a low-rank matrix into the Transformer layer of the generative AI model, training only the newly added low-rank matrix parameters. This reduces computational costs while ensuring the fine-tuning effect fits the specific scenario. Before fine-tuning, the positive and negative training sample sets are divided into training and validation sets according to a preset ratio. A cross-entropy loss function optimized for the medical field is used, with "input prompt information + scene label" as the model input and "correct decision conclusion" as the target output to construct the training task. During training, the learning rate is dynamically adjusted by combining DIP rules and clinical pathway specifications from the localized experience knowledge base. The error between the model's prediction result and the target output is calculated through backpropagation, and the low-rank matrix parameters are iteratively updated. After each training round, the validation set is used to evaluate the model's performance on key DIP decision indicators (such as rule fit, entity recognition accuracy, and decision conclusion accuracy). When the validation set indicators tend to stabilize, training is stopped, the parameter update data of the low-rank matrix is ​​extracted, and integrated to form the model parameter update quantity, ensuring that the update quantity can accurately correct the model's erroneous decision logic in basic DIP scenarios.

[0118] S11. Update the parameters of the generative artificial intelligence model by updating the model parameters to obtain the updated generative artificial intelligence model.

[0119] Optionally, the parameter update process needs to integrate the model parameter update (low-rank matrix parameters) into the original generative AI model. Specifically, through the model's parameter fusion interface, the low-rank matrix parameters are combined with the corresponding Transformer layer parameters of the original model according to preset weights, without reconstructing the entire model architecture, ensuring update efficiency. After the update, two rounds of verification are required: First, verify the model's compatibility in the domestic IT environment (Kylin OS + DM database + Kingdee middleware) to ensure compliance with the project's domestic adaptation requirements; second, use a reserved DIP scenario test set, which includes real medical record data from different county-level medical consortia and diverse DIP rule scenarios, to test the model's decision accuracy, rule fit, error rate, and other key indicators, confirming that the model performs better than before the update in new cases. If the verification passes, the updated generative AI model is saved; if there are insufficient adaptations in some scenarios, a small number of targeted samples can be added based on the test results, and the S10-S11 process can be repeated for secondary fine-tuning. The final updated generative AI model will be more in line with the DIP payment business needs of grassroots county-level medical consortia and will have more accurate decision-making capabilities.

[0120] In the above embodiments, by collecting positive and negative decision-making cases in grassroots DIP scenarios, the generative artificial intelligence model is optimized using targeted parameter fine-tuning technology. This enables the model to continuously learn correct decision-making logic and correct errors and deviations, significantly improving its decision-making accuracy and localization adaptability in scenarios such as compliance management of medical insurance funds and quality control of medical records, providing more reliable intelligent support for the DIP payment reform of county-level medical consortia.

[0121] The aforementioned medical data processing and system based on the fusion of a rule engine and a generative AI model centers on this integration. First, raw medical text is transformed into a structured medical infographic through text parsing and terminology standardization. Then, the rule engine matches the medical business rule base to generate preliminary decisions with supporting evidence and confidence levels. After screening items to be coordinated, precise prompts are input into the generative AI model to obtain inference results. Finally, a multi-threshold arbitration mechanism integrates the results of both systems to generate decision traceability information. This is further supported by refined rule execution, structured prompts, layered arbitration processes, and iterative model fine-tuning. This technical solution leverages the policy compliance and traceability advantages of rule engines, combined with the generative AI model's ability to understand and reason about complex scenarios. Furthermore, by utilizing confidence quantification and decision traceability mechanisms, it effectively solves the core technical problems of rigidity in traditional rule engines and unreliable and uninterpretable outputs of generative AI models. This achieves a unified approach to medical data processing, unifying high accuracy, strong interpretability, and adaptability to complex scenarios.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] Based on the same inventive concept, this application also provides a system for implementing the aforementioned medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the medical data processing system based on the fusion of a rule engine and a generative artificial intelligence model provided below can be found in the limitations of the medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model described above, and will not be repeated here.

[0124] In one exemplary embodiment, such as Figure 3 As shown, a medical data processing system 10 based on the fusion of a rule engine and a generative artificial intelligence model is provided, including:

[0125] The medical text parsing module 11 is used to parse the received raw medical text data to obtain the initially identified medical entities and relationships;

[0126] The medical terminology standardization link module 12 is used to perform medical terminology standardization linking on medical entities and relationships, and generate a structured medical information graph; the structured medical information graph is used to represent medical entities linked to the standard terminology system and the relationships between entities;

[0127] The rule engine matching module 13 is used to perform matching and judgment based on the structured medical information graph, through the rule engine and the pre-built medical business rule library, to generate preliminary decision results with rule basis identifiers and rule confidence.

[0128] The decision item screening module 14 is used to screen based on the rule confidence level in the preliminary decision results to obtain the decision items to be coordinated.

[0129] The prompt information construction module 15 is used to construct prompt information based on the decision-making items to be coordinated and the corresponding structured medical information image segments, so as to obtain input prompt information for generative artificial intelligence models;

[0130] The generative artificial intelligence model reasoning module 16 is used to input input prompts into the generative artificial intelligence model for structured reasoning, and obtain model decision results with model confidence and reasoning basis;

[0131] The consistency arbitration module 17 is used to perform consistency arbitration on the preliminary decision results and the model decision results based on the rule confidence in the preliminary decision results and the model confidence in the model decision results, and generate the target decision results and the corresponding decision tracing information.

[0132] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model as described above.

[0133] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a medical data processing method based on the fusion of a rule engine and a generative artificial intelligence model as described above.

[0134] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0135] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A medical data processing method based on a fusion of a rule engine and a generative artificial intelligence model, characterized by, The method comprises: S1, text analysis is performed on the received original medical text data to obtain preliminarily recognized medical entities and relationships; S2, medical terminology standardization linking processing is performed on the medical entities and relationships to generate a structured medical information graph; the structured medical information graph is used to represent medical entities linked to a standard terminology system and the association relationship between the medical entities; S3, based on the structured medical information graph, matching judgment is performed through a rule engine and a preconfigured medical business rule library to generate a preliminary decision result with rule basis identification and rule confidence; S4, the rule confidence in the preliminary decision result is used for screening to obtain a to-be-coordinated decision item; S5, input prompt information for a generative artificial intelligence model is constructed based on the to-be-coordinated decision item and the corresponding structured medical information graph to obtain input prompt information for the generative artificial intelligence model; S6, the input prompt information is input into the generative artificial intelligence model for structured reasoning to obtain a model decision result with model confidence and reasoning basis; S7, based on the rule confidence in the preliminary decision result and the model confidence in the model decision result, consistency arbitration processing is performed on the preliminary decision result and the model decision result to generate a target decision result and corresponding decision traceability information.

2. The method of claim 1, wherein, The S3 comprises: S31, based on the structured medical information graph, the linked entities and attribute relationships are extracted as matching facts; S32, the matching facts and rule conditions in the medical business rule library are matched to trigger a rule set in which the conditions are met; S33, for each rule in the rule set, the rule execution confidence is generated by calculating the rule preconfigured basic confidence and the recognition confidence of the corresponding entity in the matching facts; S34, the actions defined by all rules in the rule set are executed to obtain an execution result, and the execution result is aggregated to generate the preliminary decision result with rule basis identification and rule confidence.

3. The method of claim 1, wherein, The S5 comprises: S51, according to the to-be-coordinated decision item, the associated entities, relationships and time information in the structured medical information graph are extracted for formatted description processing to generate an information abstract text; S52, the rule clauses related to the to-be-coordinated decision item are searched in the medical business rule library, and the judgment on the to-be-coordinated decision item in the preliminary decision result is combined to generate a rule context text; S53, the information abstract text, the rule context text and a preconfigured instruction template are spliced to generate the input prompt information for the generative artificial intelligence model.

4. The method of claim 1, wherein, The S7 comprises: S71, the rule judgment and rule confidence corresponding to the to-be-coordinated decision item are obtained from the preliminary decision result, and the corresponding model judgment and model confidence are obtained from the model decision result; S72, compare the rule confidence and the preset high confidence threshold to obtain a first comparison result; when the first comparison result indicates that the rule confidence is higher than the high confidence threshold and the rule judgment is clear, the rule judgment is determined as the target judgment, and the reasoning basis in the model decision result is determined as supplementary information; S73, compare the rule confidence and the preset low confidence threshold to obtain a second comparison result; when the second comparison result indicates that the rule confidence is lower than the low confidence threshold or the rule judgment is not clear, trigger the model confidence judgment processing; S74, compare the model confidence and the preset model confidence threshold to obtain a third comparison result; when the third comparison result indicates that the model confidence is higher than the model confidence threshold, the model judgment is determined as the target judgment, and the reasoning basis is subjected to compliance verification processing; S75, when the rule judgment and the model judgment are both clear and conflict, and the rule confidence and the model confidence are both higher than the preset conflict triggering threshold, conflict arbitration is performed based on the difference between the rule confidence and the model confidence to generate an arbitration judgment; S76, integrate the target judgment of all decision items and the judgment basis corresponding to the target judgment to encapsulate the target decision result and the corresponding decision traceability information.

5. The method of claim 4, wherein, The conflict arbitration processing based on the difference between the rule confidence and the model confidence to generate an arbitration judgment comprises: S751, calculate the absolute difference between the rule confidence and the model confidence; S752, compare the absolute difference with a preset conflict threshold; when the absolute difference is greater than the conflict threshold, generate a structured conflict report based on the rule judgment and the model judgment and the respective complete evidence chain, and trigger an external arbitration interface call; S753, receive an arbitration result returned through the external arbitration interface call, and take the arbitration result as the arbitration judgment.

6. The method of claim 1, wherein, The method further comprises: S8, collect decision cases confirmed to be correct through the consistency arbitration processing to constitute a positive training sample set; S9, collect model output cases determined to be incorrect through the consistency arbitration processing to constitute a negative training sample set; S10, based on the positive training sample set and the negative training sample set, perform parameter fine-tuning on the generative artificial intelligence model to obtain a model parameter update amount; S11, perform parameter updating on the generative artificial intelligence model through the model parameter update amount to obtain an updated generative artificial intelligence model.

7. A medical data system based on the fusion of a rule engine and a generative artificial intelligence model, characterized in that, The system comprises: a medical text analysis module configured to perform text analysis on received original medical text data to obtain initially recognized medical entities and relationships; a medical terminology standardization linking module configured to perform medical terminology standardization linking processing on the medical entities and relationships to generate a structured medical information graph; the structured medical information graph is configured to represent medical entities linked to a standard terminology system and the association relationships between the medical entities; a medical text analysis module configured to perform text analysis on received original medical text data to obtain initially recognized medical entities and relationships; a medical terminology standardization linking module configured to perform medical terminology standardization linking processing on the medical entities and relationships to generate a structured medical information graph; the structured medical information graph is configured to represent medical entities linked to a standard terminology system and the association relationships between the medical entities; The rule engine matching module is configured to perform matching judgment based on the structured medical information graph by using a rule engine and a preset medical business rule library, and generate a preliminary decision result with a rule basis identifier and a rule confidence; The decision item screening module is configured to screen the rule confidence in the preliminary decision result to obtain a to-be-coordinated decision item; The prompt information construction module is configured to construct prompt information based on the to-be-coordinated decision item and a corresponding structured medical information picture segment, and obtain input prompt information for a generative artificial intelligence model; The generative artificial intelligence model reasoning module is configured to input the input prompt information into the generative artificial intelligence model to perform structured reasoning, and obtain a model decision result with a model confidence and a reasoning basis; The consistency arbitration module is configured to perform consistency arbitration processing on the preliminary decision result and the model decision result based on the rule confidence in the preliminary decision result and the model confidence in the model decision result, and generate a target decision result and corresponding decision traceability information.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the method in any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 6.

Images