Multi-source medical knowledge conflict processing method, decision assistance method, system and product

Abstract

The disclosure provides a multi-source medical knowledge conflict processing method, a decision assistance method, a system and a product, relates to the field of artificial intelligence and medical information technology, and in particular to intelligent medical treatment, conflict detection, and decision assistance technology. The method comprises the following steps: uniformly expressing medical knowledge fragments of multiple knowledge sources as structured knowledge units; constructing knowledge pairs based on different structured knowledge units, and detecting conflict relationships between the knowledge pairs; constructing a conflict graph by taking the structured knowledge units as nodes and the detected conflict relationships as edges; calculating the credibility weight of each node in the conflict graph; and based on the conflict graph and the credibility weight, eliminating the nodes with conflict relationships, and taking the remaining nodes after elimination as a knowledge subset. The disclosure can effectively identify and eliminate the logical conflicts between multi-source medical knowledge, and reduce the reasoning contradictions caused by the application of conflict knowledge to decision assistance.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and medical information technology, and in particular to smart healthcare, conflict detection, and decision support technology. Specifically, it relates to a method for handling multi-source medical knowledge conflicts, a decision support method, a system, an electronic device, a non-transitory computer-readable storage medium, and a computer program product. Background Technology

[0002] The knowledge sources relied upon for medical decision support are becoming increasingly diverse, typically including clinical guidelines, structured medical knowledge graphs, biomedical literature, historical case data, and / or the parameterized tacit knowledge of large models themselves. However, medical knowledge from different sources varies in terms of development time, disease spectrum coverage, level of evidence, and / or criteria used, often leading to knowledge conflicts in the same diagnostic scenario. This results in inconsistent, untraceable, or even unreliable decision support recommendations, posing safety risks. Summary of the Invention

[0003] In view of this, embodiments of the present disclosure provide a method for handling conflicts in multi-source medical knowledge, a decision support method, a system, an electronic device, a non-transitory computer-readable storage medium, and a computer program product.

[0004] According to one aspect of this disclosure, embodiments of this disclosure provide a method for handling conflicts in multi-source medical knowledge, comprising: uniformly expressing medical knowledge fragments from multiple knowledge sources as structured knowledge units; constructing knowledge pairs based on different structured knowledge units and detecting conflict relationships between knowledge pairs; constructing a conflict graph with structured knowledge units as nodes and detected conflict relationships as edges; calculating the credibility weight of each node in the conflict graph; and resolving nodes with conflict relationships based on the conflict graph and credibility weights, and retaining the nodes after resolution as a subset of knowledge.

[0005] According to another aspect of this disclosure, embodiments of this disclosure provide a decision support method, including: obtaining a knowledge subset based on a multi-source medical knowledge conflict resolution method; converting the knowledge subset into structured constraint hints; and controlling a medical decision support model to generate medical decision support information within the constraint boundaries of the structured constraint hints through the structured constraint hints.

[0006] According to another aspect of this disclosure, embodiments of this disclosure provide a multi-source medical knowledge conflict resolution system, comprising: a knowledge structuring module for uniformly expressing medical knowledge fragments from multiple knowledge sources as structured knowledge units; a conflict detection module for constructing knowledge pairs based on different structured knowledge units and detecting conflict relationships between knowledge pairs; a construction module for constructing a conflict graph using structured knowledge units as nodes and detected conflict relationships as edges; a calculation module for calculating the credibility weight of each node in the conflict graph; and a resolution module for resolving nodes with conflict relationships based on the conflict graph and credibility weights, and retaining the nodes after resolution as a subset of knowledge.

[0007] According to another aspect of this disclosure, embodiments of this disclosure provide a decision support system, including: an acquisition module for acquiring a subset of knowledge output, such as that from a multi-source medical knowledge conflict resolution system; a transformation module for transforming the knowledge subset into structured constraint hints; and a control module for controlling a medical decision support model to generate medical decision support information within the constraint boundaries of the structured constraint hints through the structured constraint hints.

[0008] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a multi-source medical knowledge conflict resolution method.

[0009] According to another aspect of this disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute a method for resolving conflicts in multi-source medical knowledge.

[0010] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements a method for resolving conflicts in multi-source medical knowledge.

[0011] This disclosure achieves explicit modeling of multi-source knowledge conflicts by uniformly representing multi-source medical knowledge fragments as structured knowledge units and constructing knowledge pairs. Based on the conflict relationships between the structured knowledge units and the detected knowledge pairs, a conflict graph is built, enabling the quantification and traceability of conflict relationships. Furthermore, by calculating node credibility weights and resolving nodes with conflict relationships based on the conflict graph and weights, a knowledge subset with optimal consistency is obtained. This effectively identifies and resolves logical conflicts between multi-source medical knowledge, reduces reasoning contradictions caused by directly applying conflicting knowledge to decision support, and helps improve the consistency, credibility, and traceability of decision support.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 A flowchart illustrating a method for handling conflicts in multi-source medical knowledge provided in an embodiment of this disclosure; Figure 2 A schematic flowchart of step S102 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure; Figure 3 A schematic flowchart of step S102 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the present disclosure; Figure 4 Another schematic diagram of the process of S102 in the multi-source medical knowledge conflict resolution method provided in the embodiments of this disclosure; Figure 5 This is another schematic flowchart of S102 in the multi-source medical knowledge conflict resolution method provided in the embodiments of this disclosure; Figure 6 A schematic flowchart of S104 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure; Figure 7 Another flowchart illustrating step S104 of the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure; Figure 8 This is another schematic flowchart of S104 in the multi-source medical knowledge conflict resolution method provided in the embodiments of this disclosure; Figure 9 This is another schematic flowchart of S104 in the multi-source medical knowledge conflict resolution method provided in the embodiments of this disclosure; Figure 10 A flowchart illustrating step S105 of the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure; Figure 11 A flowchart illustrating the decision support method provided in this embodiment of the disclosure; Figure 12 A structural block diagram of a multi-source medical knowledge conflict resolution system provided in this disclosure embodiment; Figure 13 A structural block diagram of a decision support system provided in an embodiment of this disclosure; Figure 14 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0014] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0015] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0016] Various modifications and variations can be made to this disclosure without departing from its spirit or scope, as will be apparent to those skilled in the art. Therefore, this disclosure is intended to cover modifications and variations falling within the scope of the corresponding claims (the claimed technical solutions) and their equivalents. It should be noted that the embodiments provided in this disclosure can be combined with each other without contradiction.

[0017] It should be noted that the collection, storage, use, processing, transmission, provision, and disclosure of various types of information (such as user personal information, device information, log data, etc.) involved in the technical solution disclosed herein all comply with the provisions and requirements of relevant laws and regulations, and the relevant processing methods and scope are open and transparent and do not violate public order and good morals.

[0018] Before describing the technical solutions provided in this disclosure, in order to facilitate understanding of this disclosure, the problems existing in the related technologies will be specifically explained below.

[0019] The knowledge sources relied upon for medical decision support are becoming increasingly diverse, typically including clinical guidelines, structured medical knowledge graphs, biomedical literature, historical case data, and / or the parameterized tacit knowledge of large models themselves. However, medical knowledge from different sources varies in terms of development time, disease spectrum coverage, level of evidence, and / or criteria used, often leading to knowledge conflicts in the same diagnostic scenario. This results in inconsistent, untraceable, or even unreliable decision support recommendations, posing safety risks.

[0020] Currently, the following methods are commonly used in related technologies to address the problem of multi-source knowledge conflicts: One approach is based on Retrieval Augmentation Generation (RAG), which retrieves knowledge fragments related to symptom descriptions from external knowledge bases, concatenates these fragments, and injects them directly into the context of a large model, allowing the model to handle conflicting information. However, this method only focuses on the relevance between knowledge and queries, lacking explicit identification and handling of logical conflicts in the retrieval results (such as two pieces of content giving contradictory conclusions on the same diagnostic criteria). Conflicting knowledge is directly input into the model, which can easily lead to contradictions in the model's reasoning.

[0021] Another approach is based on knowledge graph construction and fusion. This method extracts medical entities and relationships from different sources to build a unified knowledge graph, and then merges multi-source information through entity alignment and relationship mapping. However, this approach mainly addresses conflicts at the entity level (such as synonyms, aliases, etc.). It lacks explicit conflict relationship modeling and quantification for conflicts at the reasoning conclusion level between knowledge nodes (such as different sources giving different treatment recommendations under the same conditions), making it difficult to effectively resolve contradictions at the conclusion level.

[0022] All of the above methods suffer from insufficient mechanisms for detecting and resolving conflicts of multi-source knowledge, making it difficult to meet the requirements for the stability, traceability, and credibility of the auxiliary decision-making output.

[0023] To address at least one of the aforementioned technical problems, embodiments of this disclosure provide a method for handling multi-source medical knowledge conflicts, a decision support method, a system, an electronic device, a non-transitory computer-readable storage medium, and a computer program product.

[0024] The following section first introduces the method for handling conflicts in multi-source medical knowledge provided in the embodiments of this disclosure.

[0025] In the embodiments disclosed herein, the examples of drug names, disease names, etc., are fictitious or for illustrative purposes and do not represent any specific product or actual drug / disease name.

[0026] Figure 1 This is a schematic flowchart illustrating a method for handling conflicts in multi-source medical knowledge provided in an embodiment of this disclosure. Figure 1 As shown, the multi-source medical knowledge conflict handling method provided in this embodiment may include the following steps S101 to S105.

[0027] S101: Unify the expression of medical knowledge fragments from multiple knowledge sources into structured knowledge units.

[0028] In S101, multiple medical knowledge fragments can be obtained from multiple knowledge sources. Exemplarily, these multiple knowledge sources include, but are not limited to, clinical guideline databases, structured medical knowledge graphs, biomedical literature databases, historical case databases, and / or implicit knowledge from large models. Each medical knowledge fragment can then be uniformly expressed as a structured knowledge unit. The structured knowledge unit can adopt a preset data format and can at least contain the core elements of the knowledge, such as source, applicable conditions, conclusion, and / or level of evidence. Through a unified structured expression, heterogeneous knowledge from different sources and in different formats can be compared and processed within the same framework.

[0029] S102: Construct knowledge pairs based on different structured knowledge units and detect conflict relationships between knowledge pairs.

[0030] In S102, after uniformly representing knowledge fragments as structured knowledge units, knowledge pairs can be constructed based on different structured knowledge units. Each knowledge pair can include two structured knowledge units. By constructing knowledge pairs, originally independent structured knowledge units can be paired to form comparison units to be analyzed, facilitating subsequent conflict detection.

[0031] After constructing knowledge pairs, conflict detection is performed on each knowledge pair to determine the conflict relationship between the two structured knowledge units in each pair. Conflict detection is mainly used to determine whether there is a logical opposition or mutual exclusion relationship between the two structured knowledge units in a knowledge pair. Through the above conflict detection, previously scattered and implicit knowledge conflicts can be explicitly identified.

[0032] S103: Construct a conflict graph using structured knowledge units as nodes and detected conflict relationships as edges.

[0033] Specifically, each structured knowledge unit is treated as a node, and each detected conflict relationship is treated as an edge connecting two nodes, thus constructing a directed weighted conflict graph, or simply a conflict graph. In one example, the constructed conflict graph contains a set of nodes V that can include multiple structured knowledge units participating in conflict detection, and a set of edges E that can include conflict relationships between multiple structured knowledge units. Each edge can be labeled with a conflict type. By constructing a conflict graph, the originally scattered and implicit knowledge conflicts can be explicitly modeled as a computable graph structure, enabling conflict relationships to be quantified, computed, and traced.

[0034] S104: Calculate the credibility weight of each node in the conflict graph.

[0035] After constructing the conflict graph, a credibility weight can be calculated for each node (i.e., a structured knowledge unit). Credibility weights can be used to characterize the reliability of the evidence quality, source authority, and / or clinical suitability of a structured knowledge unit. Credibility weights reflect which node's knowledge is more worthy of adoption or use when there is conflict among multiple structured knowledge units.

[0036] By calculating the credibility weight of each node, accurate judgment criteria can be provided for conflict resolution in subsequent steps, enabling accurate choices to be made among multiple conflicting pieces of knowledge during conflict resolution.

[0037] S105: Based on the conflict graph and credibility weight, resolve nodes with conflicting relationships, and use the nodes retained after resolution as a subset of knowledge.

[0038] Based on the constructed conflict graph and the credibility weight of each node, nodes with conflicting relationships in the conflict graph can be resolved. For example, in some cases, for two nodes with conflicting conclusions, since they cannot both be logically true at the same time, the credibility weights of the two nodes can be compared, retaining the node with the higher credibility weight and excluding the node with the lower credibility weight, thereby resolving the conflict.

[0039] After conflict resolution, the nodes retained can be used as a subset of knowledge. This subset of knowledge can be used to assist in decision generation, for example, it can be converted into structured constraint hints to control the generation process of large models, or it can be used as a knowledge base for inference engines.

[0040] This disclosure achieves explicit modeling of multi-source knowledge conflicts by uniformly representing multi-source medical knowledge fragments as structured knowledge units and constructing knowledge pairs. Based on the conflict relationships between these structured knowledge units and detected knowledge pairs, a conflict graph is built, enabling the quantification and traceability of conflict relationships. Furthermore, by calculating node credibility weights and resolving conflicting nodes based on the conflict graph and weights, a knowledge subset with optimal consistency is obtained. This effectively identifies and resolves logical conflicts between multi-source medical knowledge, reducing inference contradictions caused by directly applying conflicting knowledge to decision support, and contributing to improved consistency, credibility, and traceability of decision support.

[0041] To facilitate understanding, the multi-source medical knowledge conflict resolution method provided in this disclosure will be described in detail below with reference to some specific embodiments.

[0042] Figure 2 This is a schematic flowchart of step S102 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 2 As shown, according to some embodiments of this disclosure, optionally, the structured knowledge unit includes at least an applicable condition field and a conclusion field. Exemplarily, the applicable condition field may include at least one of the following: symptoms, target object characteristics (such as age, gender, weight, physiological state, etc.), examination results, past medical history, or treatment history. The conclusion field may include at least one of the following: disease identification results, treatment recommendations, behavioral guidance, etc.

[0043] Accordingly, S102: Constructing knowledge pairs based on different structured knowledge units may include the following steps S201 and S202.

[0044] S201: Calculate the semantic similarity between the applicable condition field and the conclusion field of different structured knowledge units.

[0045] Since knowledge from different sources may differ in entity naming and concept boundaries, relying solely on precise matching of text characters to determine the relationship between two structured knowledge units often misses some semantically equivalent but differently expressed knowledge pairs, and may also introduce semantically irrelevant noise.

[0046] To address this issue, embodiments of this disclosure calculate the semantic similarity between the applicable condition fields and the semantic similarity between the conclusion fields of different structured knowledge units. Specifically, a pre-trained medical language model (such as the Med-BERT model) can be used to encode the applicable condition fields and conclusion fields of different structured knowledge units. Then, cosine similarity or other similarity measurement methods are used to calculate the semantic similarity between the applicable condition fields and the conclusion fields of different structured knowledge units. By calculating the semantic similarity in both the applicable condition and conclusion dimensions, the degree of association between different structured knowledge units in clinical scenarios and in identifying conclusions can be comprehensively assessed.

[0047] S202: Select structured knowledge units that meet the preset semantic similarity conditions to form knowledge pairs, and obtain a set of knowledge pairs containing at least one knowledge pair.

[0048] Based on the calculated semantic similarity results, structured knowledge units that meet preset semantic similarity conditions can be selected to form knowledge pairs, that is, semantically equivalent or similar structured knowledge units can be selected to form knowledge pairs. In this way, semantically related structured knowledge units can be paired, reducing the inclusion of irrelevant or weakly related structured knowledge units in the conflict detection scope, thereby improving the efficiency and accuracy of conflict detection.

[0049] Thus, by calculating the semantic similarity between the applicable condition field and the conclusion field of different structured knowledge units, and selecting structured knowledge units that meet the preset semantic similarity conditions to form knowledge pairs, it is possible to accurately filter out semantically related knowledge pairs from a large number of structured knowledge units, reducing the computational load and noise of conflict detection. Simultaneously, through dual screening along the two dimensions of applicable conditions and conclusions, the knowledge pairs included in conflict detection are substantially comparable in clinical scenarios and identification conclusions, improving the accuracy and efficiency of subsequent conflict detection.

[0050] According to some embodiments of this disclosure, optionally, S201: calculating the semantic similarity between the applicable condition field and the conclusion field of different structured knowledge units may include the following steps: For different structured knowledge units, the first semantic similarity between applicable condition fields is calculated based on a trained medical language model, and it is inferred whether the conclusion fields belong to the same clinical decision-making dimension. The same clinical decision-making dimension may include the same clinical problem type, the same medical behavior category, and / or the same diagnosis and treatment decision.

[0051] Specifically, when calculating the semantic similarity of the applicability condition fields, the applicability condition fields of any two structured knowledge units can be input into the trained medical language model. The medical language model converts the applicability condition fields of any two structured knowledge units into corresponding semantic embedding vectors, and then calculates the first semantic similarity between the two semantic embedding vectors. For example, the value range of the first semantic similarity can be between 0 and 1, and the larger the value, the closer the two applicability condition fields are semantically.

[0052] When determining whether conclusion fields belong to the same clinical decision-making dimension, the conclusion fields of any two structured knowledge units can be input into a trained medical language model. The medical language model can use rule-based judgment to determine whether the two conclusion fields belong to the same clinical problem type, the same medical behavior category, and / or the same treatment decision. If they belong to the same clinical problem type, the same medical behavior category, and / or the same treatment decision, then the two conclusion fields are determined to belong to the same clinical decision-making dimension. Otherwise, the two conclusion fields are determined not to belong to the same clinical decision-making dimension.

[0053] Accordingly, the preset semantic similarity conditions may include a first semantic similarity greater than a first threshold between two structured knowledge units, and conclusion fields belonging to the same clinical decision-making dimension. The size of the first threshold can be flexibly adjusted according to the actual situation, and this disclosure does not limit it. That is, structured knowledge units with a first semantic similarity greater than the first threshold and conclusion fields belonging to the same clinical decision-making dimension are selected to form knowledge pairs.

[0054] Through the above methods, the embodiments of this disclosure simultaneously consider the similarity of applicable conditions and the consistency of decision dimensions in the conclusions when constructing knowledge pairs, making the knowledge pairs truly comparable in a clinical sense. Specifically, similar applicable conditions ensure that the two structured knowledge units are applied to the same type of clinical scenario, and conclusions belonging to the same decision dimension ensure that the two structured knowledge units belong to the same type of clinical problem. The combination of these two factors gives the constructed knowledge pairs substantial comparative value, providing a high-quality candidate set for subsequent conflict detection and reducing invalid conflict detection caused by forcibly pairing knowledge units with different decision dimensions.

[0055] According to some embodiments of this disclosure, optionally, the multi-source medical knowledge conflict resolution method may further include the following steps before detecting conflict relationships between knowledge pairs: Use a set of logical rules in the medical field to perform rule constraint checks on knowledge pairs in a knowledge pair set; Based on the results of the rule constraints, filter out knowledge pairs in the knowledge pair set that do not conform to the logical rules of the medical field.

[0056] Specifically, in some embodiments, after constructing knowledge pairs, instead of directly performing conflict detection on multiple knowledge pairs, a set of logical rules from the medical domain is first introduced for pre-filtering. These medical domain logical rules can be a set of logical constraints compiled based on medical expertise, used to determine whether two structured knowledge units in a knowledge pair can be simultaneously valid or used in a medical logical sense. For example, for knowledge pairs involving the same indication but recommending two mutually exclusive drugs, or knowledge pairs that do not fit the patient's age range, the set of logical rules from the medical domain will determine that the knowledge pair contains a logical contradiction.

[0057] When performing rule constraint checks, the two structured knowledge units in each knowledge pair can be input into the rule engine. The rule engine then matches each rule entry in the medical domain logical rule set one by one to check whether the knowledge of the two structured knowledge units in the knowledge pair conforms to the medical domain logical rules.

[0058] Based on the results of the rule constraint check, knowledge pairs that do not conform to the logical rules of the medical field are filtered out, and only knowledge pairs that pass the rule check are retained for subsequent conflict detection. In this way, a batch of knowledge pairs that have obvious contradictions or are infeasible in medical logic can be eliminated in advance before conflict detection, reducing their interference in the conflict detection stage.

[0059] Thus, by introducing a set of logical rules from the medical domain to pre-filter knowledge pairs before conflict detection, we can identify and eliminate knowledge pairs that are clearly illogical in medical terms, thereby reducing the computational burden of subsequent conflict detection and improving overall processing efficiency. Furthermore, since the set of logical rules from the medical domain is built upon professional medical knowledge, its filtering results have high medical credibility and interpretability, effectively compensating for potential blind spots in medical logic that may exist in pure semantic similarity calculations, and improving the accuracy and reliability of conflict detection.

[0060] According to some embodiments of this disclosure, optionally, the set of logical rules in the medical field may include at least one of drug mutual exclusion rules, evidence level suppression rules, contraindication filtering rules, timeliness priority rules, and population adaptation rules.

[0061] Among these, the drug mutual exclusion rule can be used to determine whether there are known drug interactions or incompatibilities between drugs involved in two structured knowledge units. For example, two mutually exclusive drugs cannot be recommended simultaneously for the same indication. The evidence level suppression rule can be used to determine whether knowledge with a lower evidence level should not be used to negate knowledge with a higher evidence level when there are differences in evidence levels. The contraindication filtering rule can be used to determine whether the recommendations in a structured knowledge unit are applicable to the current subject. The timeliness priority rule is used to determine whether the more recently published knowledge should be prioritized when there are significant differences in the publication time of the knowledge sources. The population suitability rule is used to determine whether the recommendations in a structured knowledge unit are applicable to a specific population, such as children or the elderly. For example, adult dosages are not suitable for children.

[0062] The aforementioned medical logic rules can be used individually or in combination, and this disclosure does not impose any limitations on this. In practical applications, an appropriate set of rules can be flexibly selected based on the specific clinical scenario and the type of knowledge source.

[0063] The aforementioned medical logic rules can cover a variety of common medical logic constraints, enabling refined pre-filtering of knowledge pairs. Compared to conflict detection that relies solely on semantic similarity, combining these rules can more accurately identify knowledge pairs that contradict each other in medical logic, thereby improving the quality and efficiency of subsequent conflict detection.

[0064] According to some embodiments of this disclosure, the structured knowledge unit may optionally include an evidence level field, and the conflict relationship may include at least one of conclusion conflict, applicable condition conflict, and evidence level conflict.

[0065] Specifically, the evidence level field in the structured knowledge unit is used to characterize the strength of the scientific basis for the conclusions contained in the structured knowledge unit. The evidence level can be labeled based on GRADE ratings, evidence-based medicine grading, or other medical evidence grading standards, such as Grade A (high-quality evidence), Grade B (medium-quality evidence), and Grade C (low-quality evidence). By introducing the evidence level field, embodiments of this disclosure can further consider the dimension of evidence quality in conflict detection, in addition to considering the content and applicability of the conclusions.

[0066] Accordingly, during conflict detection, embodiments of this disclosure can detect at least one conflict relationship, such as conflict of conclusions, conflict of applicable conditions, and conflict of evidence levels.

[0067] A conclusion conflict occurs when two structured knowledge units have similar or identical semantics in their applicable condition fields, but opposite or mutually exclusive semantics in their conclusion fields. For example, if one structured knowledge unit recommends drug A as first-line treatment, while another recommends drug B, their different treatment recommendations under the same conditions constitute a conclusion conflict.

[0068] A conflict of applicable conditions is defined as a situation where two structured knowledge units have semantically similar or identical conclusion fields, but their applicable condition fields are logically mutually exclusive.

[0069] Evidence level conflict occurs when two structured knowledge units have semantically similar or identical conclusion fields, but the numerical difference in their evidence level fields exceeds a preset threshold. For example, one structured knowledge unit may be classified as Level A evidence, and another as Level C evidence. Both may lead to similar conclusions, but their evidence quality differs significantly, thus constituting an evidence level conflict. Evidence level conflict does not indicate that the two conclusions are logically mutually exclusive; rather, it suggests that the structured knowledge unit with the higher evidence level should be prioritized when resolving the conflict.

[0070] Through the above classification and detection, the embodiments of this disclosure can comprehensively identify conflict relationships between multi-source knowledge, providing an accurate classification basis for subsequent conflict resolution.

[0071] Thus, compared to existing technologies that only focus on conflicts at the entity level or a single type of conflict, the embodiments of this disclosure can more comprehensively capture various conflicts or contradictions that may exist between multi-source knowledge, providing richer and more accurate decision-making basis for subsequent conflict resolution, thereby improving the accuracy and clinical applicability of conflict resolution.

[0072] In some specific embodiments, the structured knowledge unit may optionally adopt a tuple data structure, which may specifically include a knowledge source field, an applicable condition field, a conclusion field, and an evidence level field.

[0073] Specifically, for example, a tuple data structure can be represented as a quadruple: K = (source, condition, conclusion, evidence_level). The knowledge source field (source) identifies the specific origin of the structured knowledge unit, such as guideline names, document identifiers, knowledge graph names, case database identifiers, or identifiers of implicit knowledge within a large model.

[0074] The "condition" field describes the clinical scenario or patient characteristics to which the structured knowledge unit applies, such as symptom description, examination indicators, patient age range, and comorbidities. The "conclusion" field describes the diagnostic recommendations, treatment suggestions, or prognostic judgments given by the structured knowledge unit. The "evidence_level" field characterizes the scientific strength of the conclusions contained in the structured knowledge unit, such as A / B / C grades or GRADE scores.

[0075] Thus, by designing structured knowledge units as tuple data structures containing four core fields—knowledge source, applicable conditions, conclusion, and level of evidence—standardized representation of multi-source medical knowledge fragments is achieved. This facilitates the comparison and processing of heterogeneous knowledge from different sources and in different formats within the same data structure framework, solving the problem of unified representation of heterogeneous knowledge. Furthermore, each field carries explicit semantic information, allowing subsequent conflict detection, weight calculation, and resolution processes to be directly based on the field content, improving processing efficiency and interpretability.

[0076] Figure 3 This is a schematic flowchart of step S102 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 3 As shown, according to some embodiments of this disclosure, optionally, the conflict relationship may include a conflict of conclusions. Accordingly, S102: detecting the conflict relationship between knowledge pairs may include the following steps S301 and S302.

[0077] S301: Calculate the first semantic similarity of the applicable condition field of the two structured knowledge units in the knowledge pair, and the second semantic similarity or semantic mutual exclusion of the conclusion field.

[0078] For two structured knowledge units in a knowledge pair, calculate the first semantic similarity between their applicable condition fields and the second semantic similarity or semantic mutual exclusion between their conclusion fields.

[0079] S302: If the first semantic similarity is greater than the second threshold, and the second semantic similarity is less than or equal to the third threshold, or the semantic mutual exclusion is greater than the fourth threshold, then the knowledge pair is determined to have a conclusion conflict.

[0080] When the first semantic similarity of the applicability condition fields of two structured knowledge units is greater than the second threshold, and the second semantic similarity of their conclusion fields is less than or equal to the third threshold, or the semantic mutual exclusion is greater than the fourth threshold, the knowledge pair is determined to have a conclusion conflict. That is, if the applicability condition fields of the two structured knowledge units in a knowledge pair are semantically similar, but their conclusions are dissimilar or mutually exclusive, the knowledge pair is determined to have a conclusion conflict. The values ​​of the second, third, and fourth thresholds can be flexibly adjusted according to actual circumstances, and this disclosure does not impose any limitations on them.

[0081] Thus, by calculating the first semantic similarity of the applicable conditions field and the second semantic similarity or semantic mutual exclusion of the conclusion field, and setting multiple threshold judgment conditions, knowledge pairs with similar applicable conditions but differing conclusions can be accurately identified, thereby achieving precise detection of conclusion conflicts. Compared to relying solely on a single dimension of similarity comparison, this embodiment considers both the similarity of applicable conditions and the differences in conclusions, effectively reducing the misjudgment of knowledge pairs with dissimilar applicable conditions but different conclusions as conclusion conflicts, thereby improving the accuracy and reliability of conclusion conflict detection.

[0082] Figure 4 This is another schematic flowchart of S102 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 4 As shown, according to some embodiments of this disclosure, optionally, the conflict relationship may include a conflict of applicable conditions. Accordingly, S102: detecting the conflict relationship between knowledge pairs may include the following steps S401 to S403.

[0083] S401: Calculate the second semantic similarity of the conclusion fields of two structured knowledge units in a knowledge pair.

[0084] For two structured knowledge units in a knowledge pair, a second semantic similarity is calculated between their conclusion fields. In some examples, such as encoding the conclusion fields using a pre-trained medical language model, the second semantic similarity can be calculated using methods such as cosine similarity. The second semantic similarity measures whether the conclusions of two knowledge units are semantically similar or identical.

[0085] S402: If the second semantic similarity is greater than the third threshold, then detect whether there is a logically mutually exclusive relationship between the applicable condition fields of the two structured knowledge units in the knowledge pair.

[0086] When the second semantic similarity of two structured knowledge units is greater than the third threshold, a logical mutual exclusion relationship detection is performed on the applicable condition fields. For example, the detection of logical mutual exclusion relationships can be done using rule-based judgment. For instance, it can be determined whether the two applicable condition fields respectively describe mutually exclusive patient populations, mutually exclusive clinical states, or mutually exclusive treatment stages. The size of the third threshold can be flexibly adjusted according to the actual situation.

[0087] S403: If a logically mutually exclusive relationship exists, then the knowledge pair is determined to have conflicting applicable conditions.

[0088] If there is a logical mutual exclusion relationship between the applicable condition fields, then the knowledge pair is determined to have an applicable condition conflict. An applicable condition conflict indicates that although two structured knowledge units give similar conclusions, their applicable clinical scenarios are mutually exclusive.

[0089] Thus, by calculating the second semantic similarity of the conclusion field as a precondition and then detecting the logical mutual exclusion relationship of the applicable condition field, knowledge pairs with similar conclusions but mutually exclusive applicable conditions can be accurately identified, thereby achieving accurate detection of applicable condition conflicts.

[0090] Figure 5 This is another schematic flowchart of S102 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 5 As shown, according to some embodiments of this disclosure, optionally, the conflict relationship may include a conflict of evidence levels. Accordingly, S102: detecting the conflict relationship between knowledge pairs may include the following steps S501 to S503.

[0091] S501: Calculate the second semantic similarity of the conclusion fields of two structured knowledge units in a knowledge pair.

[0092] Similar to the previous example, for two structured knowledge units in a knowledge pair, the second semantic similarity between their conclusion fields is calculated.

[0093] S502: Map the evidence levels of the two structured knowledge units in a knowledge pair to numerical values ​​and calculate the absolute value of the difference between the two numerical values.

[0094] In S502, the evidence level field of two structured knowledge units can be mapped to numerical values. Specifically, the numerical mapping relationship of evidence levels can be preset. For example, level A evidence can be mapped to 3, level B evidence to 2, and level C evidence to 1, or the numerical values ​​of the GRADE scores can be used directly. After the mapping is completed, the absolute value of the difference between the two numerical values ​​is calculated. This absolute value reflects the degree of difference in the evidentiary quality between the two structured knowledge units.

[0095] S503: If the second semantic similarity is greater than the third threshold and the absolute value is greater than the fifth threshold, then the knowledge pair is determined to have a conflict of evidence level.

[0096] When the second semantic similarity of the conclusion fields of two structured knowledge units is greater than the third threshold, and the absolute value of the difference in their evidence level values ​​is greater than the fifth threshold, the knowledge pair is determined to have an evidence level conflict. The size of the fifth threshold can be flexibly adjusted according to the actual situation, and this disclosure does not limit it. An evidence level conflict indicates that although the two knowledge units give similar conclusions, there is a significant difference in the quality of evidence. In subsequent resolution, the knowledge unit with the higher evidence level should be given priority.

[0097] Thus, by calculating the second semantic similarity of the conclusion field, mapping the evidence level to a numerical value and calculating the absolute value of the difference, it is possible to accurately identify knowledge pairs with similar conclusions but large differences in evidence quality, thereby achieving accurate detection of evidence level conflicts. Furthermore, when there are significant differences in evidence quality, knowledge units with higher evidence levels are considered for resolution, improving the reliability of the evidence quality of knowledge subsets and enhancing the scientific basis for decision support.

[0098] According to some embodiments of this disclosure, optionally, S103: constructing a conflict graph using structured knowledge units as nodes and detected conflict relationships as edges may include the following steps: Each structured knowledge unit is treated as a node. For every two nodes that have a conflict relationship, at least one directed edge is established between them, and each conflict edge is labeled with a conflict type label to obtain a conflict graph.

[0099] Specifically, after conflict detection is completed, all structured knowledge units involved in the conflict detection can be treated as nodes in a graph, with each node uniquely corresponding to a structured knowledge unit. The node set V is the set of all structured knowledge units.

[0100] Then, for every two nodes where a conflict is detected, a directed edge is established between them. The direction of the directed edge can represent the primary or secondary relationship of the conflict or the direction of its influence. For each directed edge, one or more conflict type labels are assigned, such as conclusion conflict, applicable condition conflict, and / or evidence level conflict.

[0101] In one example, if there are multiple conflict relationships between two nodes, an edge can be created and labeled with multiple conflict type labels, or multiple edges can be created, each labeled with a conflict type label; this disclosure does not limit this. When constructing a directed weighted conflict graph, each edge can also be assigned an initial conflict strength value to characterize the severity of the conflict.

[0102] Thus, by treating each structured knowledge unit as a node, the detected conflict relationships as directed edges, and labeling each edge with a conflict type, the originally scattered and implicit multi-source knowledge conflicts are transformed into a computable graph structure. This explicitly models and visualizes conflict relationships, allowing the specific location, type, and involved knowledge units of the conflict to be intuitively displayed, facilitating the resolution and tracing of knowledge conflicts. Secondly, the labeling of conflict type provides a classification basis for subsequent discrepancy resolution processing; for example, a forced resolution strategy is used for conclusion conflicts, and a priority ranking strategy is used for evidence level conflicts. Furthermore, the computable graph structure enables the application of graph theory algorithms for globally optimal knowledge subset selection, transforming the knowledge subset selection problem into a maximum weighted independent set problem. Standard graph theory algorithms are used to find a set of non-conflicting knowledge units with the highest overall credibility globally as the knowledge subset, ensuring that the final output knowledge subset has high overall consistency and credibility.

[0103] According to some embodiments of this disclosure, the credibility weight may optionally include at least one of the following: evidence strength weight, source reliability weight, and context adaptability weight.

[0104] Evidence strength weight can be used to characterize the evidentiary quality of structured knowledge units, reflecting the reliability of the medical evidence on which the structured knowledge unit is based.

[0105] Source reliability weights can be used to characterize the credibility of the knowledge source of a structured knowledge unit. For example, structured knowledge units from authoritative clinical guidelines have higher source reliability weights, while structured knowledge units from non-authoritative sources or those that have not been peer-reviewed have lower source reliability weights.

[0106] Context-adaptability weights can be used to characterize the degree of matching between structured knowledge units and the applicable conditions of target medical cases, reflecting the degree of matching between the knowledge unit and the current target medical case (such as the specific circumstances of the target object).

[0107] The three weights mentioned above can be used individually or in combination. When used in combination, the three weights can be weighted and summed to obtain a comprehensive credibility score. By introducing weights of multiple dimensions, the credibility of structured knowledge units can be comprehensively evaluated from different perspectives, providing a multi-dimensional basis for judgment in subsequent conflict resolution.

[0108] In this way, a multi-dimensional quantitative assessment of the credibility of structured knowledge units can be achieved from multiple dimensions, including evidence quality, source authority, and clinical suitability. Compared with assessment methods that rely on only a single dimension, the multi-dimensional weight design of this disclosure embodiment can more comprehensively and accurately reflect the overall credibility of knowledge units in clinical scenarios, thereby providing a more accurate decision-making basis for subsequent conflict resolution.

[0109] Figure 6 This is a schematic flowchart of step S104 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 6 As shown, according to some embodiments of this disclosure, optionally, the credibility weight may include an evidence strength weight. Accordingly, S104: Calculating the credibility weight of each node in the conflict graph may include the following steps S601 to S603.

[0110] S601: Obtain the level of evidence for structured knowledge units, the type of research on which they are based, and the size of the research sample.

[0111] For each structured knowledge unit, the level of evidence, the type of research on which it is based, and the size of the research sample are extracted from its corresponding fields. These three elements together constitute the basic data for assessing the strength of evidence.

[0112] S602: Map the level of evidence to the first score, the research type to the second score, and the research sample size to the third score.

[0113] Based on the preset mapping relationship between evidence levels and scores, the evidence level can be mapped to the first score. For example, in some examples, the following mapping relationship can be set: Level A evidence is mapped to 1.0, Level B evidence is mapped to 0.7, and Level C evidence is mapped to 0.4. Alternatively, the corresponding value can be directly used according to the GRADE score; this disclosure does not limit this approach.

[0114] Based on the preset mapping relationship between research types and scores, research types can be mapped to a second score. For example, in some examples, the following mapping relationship can be set: systematic review or meta-analysis is mapped to 1.0, randomized controlled trial is mapped to 0.9, cohort study is mapped to 0.7, case-control study is mapped to 0.5, case series is mapped to 0.3, and case report is mapped to 0.1. This disclosure does not limit this.

[0115] Based on a pre-defined normalization method, the size of the research sample can be normalized to a third score. For example, in some cases, a maximum and minimum normalization method can be used to map the ratio of the difference between the research sample size and the maximum and minimum sample sizes to a score between 0 and 1, i.e., the third score. The normalized third score reflects the contribution of sample size to the strength of evidence.

[0116] S603: The first score, the second score, and the third score are weighted and summed to obtain the evidence strength weight.

[0117] The three scores are weighted and summed to obtain the evidence strength weight. The weight coefficients corresponding to each score can be used to adjust the relative importance of different elements in the evidence strength assessment. The specific values ​​of the weight coefficients can be set according to the actual application scenario, and this disclosure does not limit them.

[0118] Thus, by acquiring three elements—evidence level, research type, and research sample size—of structured knowledge units, mapping them to calculable scores, and then weighting and summing them, the quantitative calculation of evidence strength weights is achieved. The calculation of evidence strength weights no longer relies on simple judgments of a single dimension, such as considering only the evidence level, but rather integrates multiple factors closely related to evidence quality, making the evaluation results more comprehensive and accurate. For example, incorporating research sample size effectively avoids small-sample studies receiving the same evidence weight as large-scale studies, improving the objectivity and scientific rigor of the evaluation.

[0119] Figure 7 This is another schematic flowchart of S104 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 7 As shown, according to some embodiments of this disclosure, optionally, the credibility weight may include a source reliability weight. Accordingly, S104: Calculating the credibility weight of each node in the conflict graph may include the following steps S701 to S705.

[0120] S701: Obtain the knowledge source, publication time, and target medical knowledge that matches the structured knowledge unit.

[0121] For each structured knowledge unit, the knowledge source, publication time, and target medical knowledge matching the structured knowledge unit are obtained. Target medical knowledge may include medical consensus, mainstream viewpoints, or standard knowledge matching the structured knowledge unit, but this disclosure does not limit this.

[0122] S702: Determine the source authority score corresponding to the knowledge source based on the preset authority scoring table.

[0123] Based on a pre-constructed authority rating scale, knowledge sources are mapped to corresponding source authority scores. In some examples, the authority rating scale can optionally be constructed based on professional evaluation standards in the medical field; for example, authoritative clinical guidelines are mapped to 1.0, systematic reviews or meta-analyses to 0.9, randomized controlled trials published in high-quality journals to 0.8, cohort studies to 0.6, case-control studies to 0.4, case reports to 0.2, and implicit knowledge within large models to 0.1. This disclosure does not impose limitations on this. The source authority score reflects the degree of authority of the knowledge source itself.

[0124] S703: Calculate the timeliness decay score based on the release time.

[0125] The timeliness decay score is calculated based on the release date of the structured knowledge unit. Generally, the newer the knowledge, the better its timeliness, and the higher the score should be; the older the knowledge, the worse its timeliness, and the lower the score should be. The timeliness decay score can be calculated using a time decay function, for example: (1) Where Δt represents the time difference between the current time and the publication time, and λ is the attenuation coefficient used to control the attenuation rate. Linear attenuation or other forms of attenuation functions can also be used; this disclosure does not limit this. Timeliness attenuation score s f The value range is between 0 and 1, and the more recent the release time, the closer the score is to 1.

[0126] S704: Calculate the consistency score between the knowledge of the structured knowledge unit and the target medical knowledge.

[0127] The target medical knowledge can be current mainstream medical consensus, standard textbooks, or standardized knowledge published by authoritative institutions. The consistency score can be calculated using semantic similarity measurement, the expression of which is as follows: (2) Among them, K u K represents the knowledge content of a structured knowledge unit. t Indicates target medical knowledge, This refers to a semantic similarity calculation function, such as the cosine similarity function. The consistency score is s. c The value range is between 0 and 1. The higher the score, the more consistent the structured knowledge unit is with the content of the target medical knowledge.

[0128] S705: The source reliability weight is obtained by weighting and summing the source authority score, timeliness decay score, and consistency score.

[0129] The source reliability weight is obtained by weighted summation of the three scores mentioned above. The specific calculation formula is as follows: (3) Where ωs represents the source reliability weight, s a The score indicates the authority of the source, s f Indicates the time-related decay score, s c Indicates the consistency score. These are configurable weighting coefficients used to adjust the relative importance of different elements in the source reliability assessment. The specific values ​​of the weighting coefficients can be set according to the actual application scenario, and this disclosure does not limit them.

[0130] Thus, the calculation of source reliability weights integrates multiple dimensions such as authority, timeliness, and consistency, making the evaluation results more comprehensive and objective. By introducing a timeliness decay score, the source reliability weights can dynamically adapt to the updating and iteration of medical knowledge, prioritizing the use of newer knowledge. By introducing a consistency score, the source reliability weights can assess the degree of consistency between knowledge units and mainstream medical consensus, minimizing the adoption of marginalized knowledge that deviates from mainstream viewpoints.

[0131] Figure 8 This is another schematic flowchart of S104 in the multi-source medical knowledge conflict resolution method provided in the embodiments of this disclosure. Figure 8 As shown, according to some embodiments of this disclosure, optionally, the credibility weight may include a context adaptability weight. Accordingly, S104: Calculating the credibility weight of each node in the conflict graph may include the following steps S801 to S803.

[0132] S801: Obtain contextual information of the target object, including at least one of age, gender, chief complaint, past medical history, and concomitant medications.

[0133] Obtain contextual information about the target subject (e.g., a patient). Contextual information includes, but is not limited to, age, gender, chief complaint, past medical history, concomitant medications, allergy history, family history, and other information relevant to diagnostic decisions. Contextual information can be obtained through user input, electronic medical record systems, or other data sources.

[0134] S802: Semantically match the context information of the target object with the applicable condition fields of the structured knowledge unit, and calculate the matching degree score.

[0135] The process involves semantically matching the contextual information of the target object with the applicability condition field of the structured knowledge unit, and calculating a matching score between the two. In some examples, the contextual information of the target object can be encoded as a first vector, and the applicability condition field of the structured knowledge unit can be encoded as a second vector. The cosine similarity between the two vectors is then calculated. This cosine similarity can be used as the matching score. In other examples, the contextual information of the target object and the applicability condition field can be input into a large language model, and the model can be prompted to output a matching score using prompt words. A higher matching score indicates a better match between the applicability condition of the structured knowledge unit and the contextual information of the target object.

[0136] S803: Use the matching degree score as the context adaptability weight.

[0137] The matching score calculated above is used as the context fit weight. The context fit weight reflects the degree of matching between the structured knowledge unit and the target object; the higher the weight, the better the matching between the structured knowledge unit and the target object.

[0138] Thus, by acquiring the contextual information of the target object and semantically matching it with the applicability field of the structured knowledge unit, the matching degree score is calculated as the contextual adaptability weight, thereby achieving a dynamic quantitative assessment of the clinical adaptability of the structured knowledge unit. In this embodiment, the contextual adaptability weight is no longer a static preset value, but is dynamically calculated based on the specific circumstances of the target object, enabling the credibility assessment of the structured knowledge unit to have individualized adaptability. Secondly, by incorporating individualized information such as the target object's age, gender, past medical history, and concomitant medications into the assessment, knowledge matching the target object's characteristics can be prioritized in complex clinical scenarios such as comorbid patients and special populations, reducing the clinical risks caused by overgeneralization recommendations.

[0139] Figure 9 This is another schematic flowchart of S104 in the multi-source medical knowledge conflict resolution method provided in the embodiments of this disclosure. Figure 9 As shown, according to some embodiments of this disclosure, optionally, S104: calculating the credibility weight of each node in the conflict graph may include the following steps S901 to S903.

[0140] S901: Weights for obtaining evidence strength, source reliability, and contextual adaptability.

[0141] The specific calculation methods for the evidence strength weight, source reliability weight, and context fit weight have been described above and will not be repeated here.

[0142] S902: The weighted sum of the evidence strength weight, source reliability weight, and context suitability weight is used to obtain the overall credibility score.

[0143] The overall credibility score can be obtained by weighting and summing the weights of evidence strength, source reliability, and context suitability. The weight coefficients of these three weights can be flexibly adjusted according to the actual situation, and this disclosure does not impose any limitations on them. In one example, the evidence strength weight might be 40%, the source reliability weight 30%, and the context suitability weight 30%. This disclosure does not impose any limitations on the specific values ​​of the weight coefficients and can be flexibly adjusted according to the actual application scenario.

[0144] S903: Use the overall credibility score as the credibility weight of the node.

[0145] The overall credibility score calculated above is used as the credibility weight of the node.

[0146] Thus, by acquiring three weights—evidence strength weight, source reliability weight, and context fit weight—and summing them to obtain a comprehensive credibility score, a multi-dimensional comprehensive evaluation of the credibility weight of structured knowledge units is achieved. Compared to methods relying on only a single dimension or simple superposition, this embodiment introduces configurable weight coefficients, allowing for flexible adjustment of the relative importance of different dimensions according to different clinical scenarios and application needs. For example, the proportion of evidence strength weight can be increased in clinical research scenarios, while the proportion of context fit weight can be increased in rapid decision-making scenarios. This configurable comprehensive evaluation mechanism has good adaptability and scalability, and can meet the conflict resolution needs in various clinical scenarios.

[0147] Figure 10 This is a schematic flowchart of step S105 in the multi-source medical knowledge conflict resolution method provided in this embodiment of the disclosure. Figure 10 As shown, according to some embodiments of this disclosure, optionally, S105: Based on the conflict graph and the confidence weight, the nodes with conflicting relationships are resolved, and the nodes retained after resolution are used as a knowledge subset, which may include the following steps S1001 and S1002.

[0148] S1001: Based on the conflict relationships in the conflict graph, construct constraints, including that there is no conflict between any two nodes in the knowledge subset.

[0149] Based on the conclusion conflict relationships marked in the conflict graph, constraints are constructed for selecting the knowledge subset. Specifically, for any two nodes with conflicting conclusion edges in the conflict graph, since they cannot logically be true simultaneously, the constraints require that the knowledge subset... A node cannot contain both of these nodes simultaneously. That is, for any pair of nodes with conflicting conclusions... Knowledge subset At most one node can be retained.

[0150] For example, constraints can be formally expressed as: (4) Here, Ec represents the set of all conflicting edges in the conflict graph, and S represents the subset of knowledge to be selected. This constraint ensures that there are no conflicting conclusions in the final output subset of knowledge, thus guaranteeing the consistency within the subset.

[0151] S1002: Solve for the knowledge subset under constraints with the goal of maximizing the sum of the credibility weights of each node in the knowledge subset.

[0152] Under the premise of satisfying the above constraints, to maximize the knowledge subset The objective is to find the optimal knowledge subset by summing the credibility weights of all nodes. This optimization problem can be formally expressed as: (5) Where V represents the set of all nodes in the conflict graph. Represents a node The overall credibility score is calculated. The goal of this optimization problem is to retain as many knowledge units with high credibility weights as possible while ensuring that there are no conflicting conclusions within the knowledge subset.

[0153] In specific solutions, algorithms such as greedy algorithms, dynamic programming, integer programming, or the maximum weight independent set algorithm based on graph theory can be used. This disclosure does not limit the specific solution algorithm. By solving the above optimization problem, a globally consistent optimal subset of knowledge can be obtained.

[0154] Thus, by constructing constraints based on the relationship of conflicting conclusions and optimizing the solution with the goal of maximizing the sum of the confidence weights of each node in the knowledge subset, the knowledge subset can be selected from a global perspective rather than a locally greedy one, effectively avoiding the local optimum problem that may be caused by simple confidence ranking. The introduction of constraints can effectively reduce the possibility of conflicting conclusions within the knowledge subset, that is, there will be no knowledge with the same applicable conditions but opposite conclusions in the same subset, thereby improving the logical consistency of the knowledge subset. In addition, with the goal of maximizing the sum of confidence weights, it tends to prioritize the retention of high-confidence knowledge units, which is in line with the clinical decision-making preference for high-confidence knowledge. While resolving the conflicting conclusions, it provides a relatively reliable knowledge basis for subsequent auxiliary decision-making.

[0155] According to some embodiments of this disclosure, optionally, each edge in the conflict graph may also be labeled with an initial conflict intensity value, which can be used to characterize the severity of the conflict between two nodes.

[0156] Accordingly, S1002: Solving for the knowledge subset under constraints with the objective of maximizing the sum of the credibility weights of each node in the knowledge subset may include the following steps: Using the initial conflict intensity value as a penalty term, and aiming to maximize the sum of the credibility weights of each node in the knowledge subset, the knowledge subset is solved under constraints.

[0157] Among them, the penalty term can be used to adjust the selection priority based on the initial conflict intensity value between nodes when multiple candidate nodes need to be selected, provided that the constraints are met. The higher the conflict intensity, the lower the priority of the selected node.

[0158] Specifically, the initial conflict intensity value can be obtained by weighting and summing the degree of overlap in applicable conditions, the degree of difference in conclusions, and the degree of difference in the levels of evidence. ρij The higher the value, the more severe the conflict between the two nodes.

[0159] During the optimization process, when selecting the target node from multiple candidate nodes that meet the constraints, a penalty term based on the initial conflict strength value is introduced. This term penalizes node pairs with high conflict strength, making it more likely that node pairs with strong conflict relationships will not be selected simultaneously. For example, a penalty term can be added to the optimization objective function: Add a penalty term to the objective function: (6) Here, μ is the penalty coefficient, used to control the strength of the penalty term. By introducing a penalty term, under the premise of satisfying the constraints, the system tends to select node pairs with lower conflict intensity, or to reduce the selection of nodes that have high conflict intensity with multiple nodes.

[0160] Thus, by introducing an initial conflict intensity value as a penalty, when selecting from multiple candidate nodes, not only the node's credibility weight but also the severity of the conflict between nodes is considered. This allows the optimization objective to better balance the relationship between retaining high-weight nodes and avoiding strong conflicts, improving the global rationality of the resolution result. Furthermore, the conflict intensity value can serve as an auxiliary criterion, helping to make more reasonable resolution decisions when credibility weights are similar.

[0161] According to some embodiments of this disclosure, optionally, S105: Based on the conflict graph and the confidence weight, the nodes with conflicting relationships are resolved, and the nodes retained after resolution are used as a knowledge subset, which may also include the following steps one and two.

[0162] Step 1: When there is a conflict of evidence level or applicable conditions between two nodes, add conflict labeling information to the nodes with conflict of evidence level or conditions.

[0163] Detected conflicts are categorized and processed. Conflicts in conclusions, due to their logical mutual exclusion, require mandatory resolution, meaning at most one node is retained within the knowledge subset. Conflicts in evidence level and applicable conditions, however, do not constitute logically mutually exclusive relationships.

[0164] A conflict in the level of evidence indicates that two knowledge units have similar conclusions but significantly different levels of evidence quality; however, they are not logically incompatible. A conflict in the applicable conditions indicates that two knowledge units have similar conclusions but mutually exclusive applicable conditions; however, they may not be logically contradictory. Therefore, for these two types of conflicts, this disclosure does not forcibly exclude either node, but rather adds conflict annotation information to the conflicting nodes to facilitate subsequent tracing.

[0165] The conflict labeling information may include at least one of the following: conflict type, identifiers of the conflicting parties, severity of the conflict, and resolution priority.

[0166] Step 2: Retain the nodes with added conflict annotations and their corresponding conflict annotation information in the knowledge subset.

[0167] Nodes with added conflict annotations are retained in the knowledge subset, along with their corresponding conflict annotation information. Subsequent use of the knowledge subset allows for differentiated processing based on the conflict annotation information. For example, when generating diagnostic suggestions, nodes with higher evidence levels are prioritized, or when presenting diagnostic evidence, the conflict annotation information of both nodes is displayed simultaneously for clinicians' reference and judgment. For conflicts in applicable conditions, the more suitable node can be dynamically selected based on the contextual information of the target object during the use of the knowledge subset.

[0168] Thus, by adopting a strategy of marked retention rather than mandatory exclusion for conflicts of evidence levels and applicable conditions, the loss of potentially clinically valuable knowledge due to mandatory exclusion is effectively reduced. For example, knowledge with low evidence levels may still be relevant in certain clinical scenarios. Secondly, by retaining conflict annotation information, users of the knowledge subset can understand the differences in quality or scope of application between nodes, facilitating flexible selection or reference based on actual scenarios in subsequent use, thereby improving the clinical adaptability and practicality of the knowledge subset.

[0169] According to some embodiments of this disclosure, the multi-source medical knowledge conflict resolution method may optionally include the following steps: When the difference in credibility weight between two conflicting nodes is lower than the preset sixth threshold, a human-machine collaborative review process is triggered. The conflict details and resolution plan are presented to the target personnel, and after receiving the confirmation instruction from the target personnel, the resolution or output is executed according to the confirmation instruction.

[0170] When the credibility weights of two nodes are close, i.e., the difference between them is less than or equal to a preset sixth threshold, the conflict details and resolution plan can be presented to the target personnel. For example, the conflict details may include the knowledge content, source, evidence level, overall credibility score, conflict type, and / or conflict severity of each of the two conflicting nodes. The resolution plan can be a suggestion calculated based on the current credibility weights, such as suggesting retaining node A and excluding node B, or retaining both nodes and marking them as conflicting. The size of the sixth threshold can be flexibly adjusted according to the actual situation, and this disclosure does not limit it.

[0171] Then, a confirmation instruction from the target personnel is received. This confirmation instruction can be to adopt the resolution plan, recalculate the credibility weights, or specify reserved nodes, etc., and this disclosure does not limit this. Then, resolution or output is performed according to the confirmation instruction. For example, if the confirmation instruction is to adopt the resolution plan, then resolution is performed according to the plan and a subset of knowledge is output.

[0172] Thus, this embodiment of the present disclosure adds a safety fallback mechanism of manual review on top of automatic conflict resolution. This proactively identifies and delegates decision-making authority to the target personnel when faced with highly uncertain conflicts, effectively reducing the clinical risks associated with forcibly outputting results when data is insufficient. Secondly, by providing the target personnel with complete conflict details and resolution plans, manual review is adequately supported by information, improving review efficiency and quality. Furthermore, the introduction of a human-machine collaborative review process helps meet the regulatory requirements for human supervision and safety assurance in smart healthcare.

[0173] Based on the same technical concept as the multi-source medical knowledge conflict resolution method provided in any of the above embodiments, this disclosure also provides a decision support method.

[0174] Figure 11 This is a schematic flowchart of a decision support method provided in an embodiment of this disclosure. Figure 11 As shown, according to some embodiments of this disclosure, optionally, the decision assistance method provided by the embodiments of this disclosure may include the following steps S1101 to S1103.

[0175] S1101: Obtain a subset of knowledge.

[0176] Specifically, a subset of knowledge can be obtained based on the multi-source medical knowledge conflict handling method provided in any of the above embodiments.

[0177] S1102: Transform a subset of knowledge into structured constraint hints.

[0178] The acquired knowledge subset is transformed into structured constraint hints. These structured constraint hints can use a pre-defined structured instruction format to convey constraint information to the medical decision support model. Each structured knowledge unit within the knowledge subset can be converted into one or more constraint statements, organized according to a unified format. By transforming the knowledge subset into structured constraint hints, the constraint information of the knowledge subset can be transmitted to the medical decision support model in a way that the model can understand.

[0179] S1103: Through structured constraint hints, control the medical decision support model to generate medical decision support information within the constraint boundaries of the structured constraint hints.

[0180] Structured constraint hints are passed as input to the medical decision support model. Under the constraints of these hints, the model is required to generate medical decision support information within the bounds of those constraints. In some specific examples, structured constraint hints can be injected into the model context as system-level instructions, strongly constraining the model's generative behavior.

[0181] Thus, when generating medical decision support suggestions, the medical decision support model must adhere to the constraint knowledge list in the structured constraint prompts; that is, it can only reason and generate suggestions within the scope specified by the constraint knowledge list. At the same time, the model cannot introduce content that contradicts the constraint knowledge list during the generation process.

[0182] Thus, by effectively identifying and resolving logical conflicts between multi-source medical knowledge, and reducing reasoning contradictions caused by directly applying conflicting knowledge to the medical decision-making support model, the consistency and credibility of the supported decisions are improved. Secondly, by transforming the knowledge subset after conflict resolution into structured constraint hints and using these to control the generation process of the medical decision-making support model, an effective connection is achieved between the knowledge resolution results and the model generation process. Under the constraints of the structured constraint hints, the model is effectively prevented from deviating from or ignoring the resolved knowledge subset during the generation process, resulting in high credibility of the medical decision-making support information output by the model. Furthermore, the structured constraint hints allow the content of the knowledge subset to be directly understood and executed by the model without additional adaptation or conversion, improving the efficiency of the overall process.

[0183] According to some embodiments of this disclosure, optionally, the structured constraint hints may include a constraint knowledge list and conflict resolution records. The constraint knowledge list may include at least one of the following: source identifier, evidence level, credibility weight, applicable conditions, and conclusion of the knowledge corresponding to the structured knowledge unit; the conflict resolution records may include the identifier of the resolved node and the reason for its resolution.

[0184] For example, the list of constraint knowledge is as follows: Item 1: Source: {source_id1}, Evidence level: {evidence_level1}, Credibility weight: {score1}, Applicable conditions: {condition1}, Conclusion: {conclusion1}; Item 2: Source: {source_id2}, Evidence level: {evidence_level2}, Credibility weight: {score2}, Applicable conditions: {condition2}, Conclusion: {conclusion2}; ...

[0185] For example, the conflict resolution record is as follows: The node to be resolved is {node_id1}. The reason for resolution is that it conflicts with the conclusion of node {node_id2} and has a low credibility weight. The node to be resolved is {node_id3}. The reason for resolution is that it conflicts with the conclusion of node {node_id4} and has a low credibility weight. ...

[0186] Each entry in the constrained knowledge list corresponds to a structured knowledge unit within the knowledge subset. Each entry may include the source identifier, evidence level, credibility weight, applicable conditions, and / or conclusions of the structured knowledge unit. By providing this information, the constrained knowledge list provides a traceable knowledge basis for medical decision support models.

[0187] The conflict resolution log records the knowledge nodes that were excluded or downgraded during the conflict resolution process, along with the reasons for their exclusion. For example, for a node excluded due to a conflicting conclusion, the conflict resolution log could state, "This node has a conflicting conclusion with another node, and its overall credibility score is low, therefore it was excluded." For nodes marked with a conflict label, the conflict resolution log could state, "This node has a conflicting level of evidence, but was not forcibly excluded." By providing the conflict resolution log, the model's output becomes fully traceable, allowing users or clinicians to clearly understand the processing of the knowledge upon which the model is based.

[0188] Thus, the constraint knowledge list provides a clear and complete knowledge boundary for the medical decision support model, giving the model's generative behavior a definite constraint basis. The conflict resolution record documents the complete conflict resolution process and decision-making basis, making the model's output diagnostic recommendations traceable and auditable, meeting the interpretability requirements of clinical scenarios. Simultaneously, each entry in the constraint knowledge list includes fields such as source identifier, evidence level, credibility weight, applicable conditions, and conclusion. The model can explicitly reference this information when generating output, enhancing the interpretability and credibility of the output. By incorporating the conflict resolution record into structured constraint prompts, the entire conflict resolution process can be shown to users or clinicians, helping to build user trust in AI diagnostic recommendations.

[0189] According to some embodiments of this disclosure, the decision support method may optionally include the following steps: When generating medical decision support information, the control medical decision support model explicitly adds source identifiers to the constrained knowledge it references. The source identifiers include at least one of the following: source name, publication time, and evidence level.

[0190] Specifically, when a medical decision support model generates diagnostic suggestions or treatment recommendations, it can explicitly instruct the model to cite source identifiers in the structured constraint prompts, or embed corresponding instructions in the model's prompts to require the model to explicitly label the knowledge sources it relies on in the output, and to explicitly add source identifiers to the cited constraint knowledge. These source identifiers include the source name, publication time, and / or level of evidence. In this way, each diagnostic suggestion or treatment recommendation is accompanied by its underlying knowledge source information, allowing users or clinicians to directly trace the source of each suggestion.

[0191] Thus, by controlling the medical decision-making support model to explicitly add source identifiers to the constrained knowledge referenced when generating medical decision-making support information, end-to-end traceability of the model's output is achieved. Users can directly trace the source of each medical decision-making support suggestion, understanding the AI's decision-making basis without additional queries. Information such as source name, publication time, and level of evidence allows users to quickly assess the reliability and timeliness of each suggestion, thereby making more accurate clinical decisions. Explicit source labeling helps build user trust in the AI ​​system, making AI-assisted decision-making more readily adopted in clinical scenarios.

[0192] According to some embodiments of this disclosure, the decision support method may optionally include the following steps: The processing process is recorded and archived, including one or more of the following: medical knowledge fragments from multiple knowledge sources, semantic similarity calculation results, conflict detection results, credibility weight calculation details, resolution decision records, structured constraint hints, and medical decision support information.

[0193] Specifically, each time a decision support process is executed, the entire process from input to output can be recorded and archived. These records can be stored in immutable storage media such as local databases, cloud storage, or blockchain to meet the data traceability and compliance requirements of medical AI systems.

[0194] In this way, by fully archiving the processing records, traceability of the entire decision support process is achieved. Complete processing records support post-event auditing and model behavior analysis, meeting regulatory requirements for traceability and compliance. The archived processing records can also serve as reference data for model training and tuning, improving the model's performance in subsequent tasks.

[0195] Based on the same technical concept as the multi-source medical knowledge conflict resolution method provided in any of the above embodiments, this disclosure also provides a multi-source medical knowledge conflict resolution system.

[0196] Figure 12 This is a structural block diagram of a multi-source medical knowledge conflict resolution system provided in an embodiment of this disclosure. Figure 12 As shown, the multi-source medical knowledge conflict resolution system 1200 provided in this embodiment may include the following modules: The knowledge structuring module 1201 is used to uniformly express medical knowledge fragments from multiple knowledge sources into structured knowledge units; The conflict detection module 1202 is used to construct knowledge pairs based on different structured knowledge units and detect conflict relationships between knowledge pairs; Module 1203 is used to construct a conflict graph with structured knowledge units as nodes and detected conflict relationships as edges. Calculation module 1204 is used to calculate the credibility weight of each node in the conflict graph; The resolution module 1205 is used to resolve nodes with conflicting relationships based on the conflict graph and credibility weights, and to retain the nodes after resolution as a subset of knowledge.

[0197] The specific working process of each of the above modules has been described above and will not be repeated here.

[0198] This disclosure achieves explicit modeling of multi-source knowledge conflicts by uniformly representing multi-source medical knowledge fragments as structured knowledge units and constructing knowledge pairs. Based on the conflict relationships between these structured knowledge units and detected knowledge pairs, a conflict graph is built, enabling the quantification and traceability of conflict relationships. Furthermore, by calculating node credibility weights and resolving conflicting nodes based on the conflict graph and weights, a knowledge subset with optimal consistency is obtained. This effectively identifies and resolves logical conflicts between multi-source medical knowledge, reducing inference contradictions caused by directly applying conflicting knowledge to decision support, and contributing to improved consistency, credibility, and traceability of decision support.

[0199] According to some embodiments of this disclosure, optionally, a structured knowledge unit includes at least an applicable condition field and a conclusion field. Specifically, the conflict detection module 1202 can be used to calculate the semantic similarity between the applicable condition field and the conclusion field of different structured knowledge units; select structured knowledge units that satisfy preset semantic similarity conditions to form knowledge pairs, thereby obtaining a knowledge pair set containing at least one knowledge pair.

[0200] According to some embodiments of this disclosure, optionally, the conflict detection module 1202 can be used to calculate the first semantic similarity between applicable condition fields based on a trained medical language model for different structured knowledge units, and to infer whether the conclusion fields are in the same clinical decision dimension; the preset semantic similarity conditions include that the first semantic similarity between two structured knowledge units is greater than a first threshold, and the conclusion fields are in the same clinical decision dimension, the same clinical decision dimension includes the same clinical problem type, the same medical behavior category and / or the same diagnosis and treatment decision.

[0201] According to some embodiments of this disclosure, optionally, the multi-source medical knowledge conflict handling system 1200 may include an inspection module for performing rule constraint checks on knowledge pairs in the knowledge pair set using a set of logical rules in the medical field; and filtering out knowledge pairs in the knowledge pair set that do not conform to the logical rules in the medical field based on the rule constraint results.

[0202] According to some embodiments of this disclosure, optionally, the set of logical rules in the medical field includes at least one of drug mutual exclusion rules, evidence level suppression rules, contraindication filtering rules, timeliness priority rules, and population adaptation rules.

[0203] According to some embodiments of this disclosure, the structured knowledge unit may optionally include an evidence level field, and the conflict relationship includes at least one of conclusion conflict, applicable condition conflict, and evidence level conflict.

[0204] According to some embodiments of this disclosure, optionally, the conflict relationship includes a conclusion conflict. Specifically, the conflict detection module 1202 can be used to calculate the first semantic similarity of the applicable condition field of the two structured knowledge units in the knowledge pair, and the second semantic similarity or semantic mutual exclusion of the conclusion field; if the first semantic similarity is greater than a second threshold, and the second semantic similarity is less than or equal to a third threshold, or the semantic mutual exclusion is greater than a fourth threshold, then it is determined that the knowledge pair has a conclusion conflict.

[0205] According to some embodiments of this disclosure, optionally, the conflict relationship includes a conflict of applicable conditions. Specifically, the conflict detection module 1202 can be used to calculate the second semantic similarity of the conclusion fields of the two structured knowledge units in the knowledge pair; if the second semantic similarity is greater than a third threshold, it detects whether there is a logically mutually exclusive relationship between the applicable condition fields of the two structured knowledge units in the knowledge pair; if a logically mutually exclusive relationship exists, it determines that there is a conflict of applicable conditions in the knowledge pair.

[0206] According to some embodiments of this disclosure, optionally, the conflict relationship includes evidence level conflict. Specifically, the conflict detection module 1202 can be used to calculate the second semantic similarity of the conclusion fields of two structured knowledge units in a knowledge pair; map the evidence levels of the two structured knowledge units in the knowledge pair to numerical values, and calculate the absolute value of the difference between the two numerical values; if the second semantic similarity is greater than a third threshold and the absolute value is greater than a fifth threshold, then it is determined that there is an evidence level conflict in the knowledge pair.

[0207] According to some embodiments of this disclosure, optionally, the construction module 1203 can be used to treat each structured knowledge unit as a node, establish at least one directed edge between each pair of nodes that have a conflict relationship, and label each conflict edge with a conflict type label to obtain a conflict graph.

[0208] According to some embodiments of this disclosure, optionally, the credibility weight includes at least one of the evidence strength weight, source reliability weight, and context adaptability weight; the evidence strength weight is used to characterize the evidence quality of the structured knowledge unit, the source reliability weight is used to characterize the credibility of the knowledge source of the structured knowledge unit, and the context adaptability weight is used to characterize the degree of matching between the structured knowledge unit and the applicable conditions of the target medical case.

[0209] According to some embodiments of this disclosure, optionally, the credibility weight includes an evidence strength weight. Specifically, the calculation module 1204 can be used to obtain the evidence level, the research type, and the research sample size of the structured knowledge unit; map the evidence level to a first score, the research type to a second score, and the research sample size to a third score; and weight and sum the first, second, and third scores to obtain the evidence strength weight.

[0210] According to some embodiments of this disclosure, optionally, the credibility weight includes a source reliability weight. The structured knowledge unit also includes a knowledge source field. Specifically, the calculation module 1204 can be used to obtain the knowledge source, publication time, and target medical knowledge matching the structured knowledge unit; determine the source authority score corresponding to the knowledge source according to a preset authority scoring table; calculate the timeliness decay score according to the publication time; calculate the consistency score between the knowledge of the structured knowledge unit and the target medical knowledge; and perform a weighted summation of the source authority score, timeliness decay score, and consistency score to obtain the source reliability weight.

[0211] According to some embodiments of this disclosure, optionally, the credibility weight includes a context adaptability weight. Specifically, the calculation module 1204 can be used to obtain context information of the target object, including at least one of age, gender, chief complaint, past medical history, and concomitant medications; semantically match the context information of the target object with the applicable condition fields of the structured knowledge unit, calculate a matching degree score; and use the matching degree score as the context adaptability weight.

[0212] According to some embodiments of this disclosure, optionally, the calculation module 1204 can be used to obtain the evidence strength weight, the source reliability weight, and the context adaptability weight; to perform a weighted summation of the evidence strength weight, the source reliability weight, and the context adaptability weight to obtain a comprehensive credibility score; and to use the comprehensive credibility score as the credibility weight of the node.

[0213] According to some embodiments of this disclosure, optionally, the resolution module 1205 can be used to construct constraints based on the conclusion conflict relationship in the conflict graph. The constraints include that there is no conclusion conflict between any two nodes in the knowledge subset; and to solve the knowledge subset under the constraints with the goal of maximizing the sum of the credibility weights of each node in the knowledge subset.

[0214] According to some embodiments of this disclosure, optionally, each edge in the conflict graph is also labeled with an initial conflict intensity value, which is used to characterize the severity of the conflict between two nodes. Specifically, the resolution module 1205 can be used to use the initial conflict intensity value as a penalty term to solve for the knowledge subset under constraints, with the goal of maximizing the sum of the credibility weights of each node in the knowledge subset; wherein, under the premise of satisfying the constraints, when multiple candidate nodes need to be selected, the penalty term is used to adjust the selection priority according to the initial conflict intensity value between nodes, with nodes having higher conflict intensity having lower priority for selection.

[0215] According to some embodiments of this disclosure, optionally, the resolution module 1205 can also be used to add conflict labeling information to the nodes with conflicting evidence levels or applicable conditions when there is a conflict between the two nodes; and retain the nodes with added conflict labels and their corresponding conflict labeling information in the knowledge subset.

[0216] According to some embodiments of this disclosure, optionally, the multi-source medical knowledge conflict resolution system 1200 may include an interaction module, which is used to trigger a human-machine collaborative review process when the difference in credibility weight between two nodes with conflicting relationships is lower than a preset sixth threshold, present the conflict details and resolution plan to the target personnel, and execute the resolution or output according to the confirmation instruction after receiving the confirmation instruction from the target personnel.

[0217] It should be noted that the multi-source medical knowledge conflict resolution system provided in this disclosure has the same or corresponding technical features as the multi-source medical knowledge conflict resolution method provided in any of the above embodiments, and produces the same technical effects. For the sake of brevity, further details are omitted here.

[0218] Based on the same technical concept as the decision support method provided in any of the above embodiments, this disclosure also provides a decision support system.

[0219] Figure 13 This is a structural block diagram of a decision support system provided in an embodiment of this disclosure. Figure 13 As shown, the decision support system 1300 provided in this embodiment may include the following modules: Module 1301 is used to acquire a subset of knowledge, such as that output by a multi-source medical knowledge conflict resolution system. Transformation module 1302 is used to transform a subset of knowledge into structured constraint hints; The control module 1303 is used to control the medical decision support model to generate medical decision support information within the constraint boundaries of the structured constraint prompts through structured constraint prompts.

[0220] The specific working process of each of the above modules has been described above and will not be repeated here.

[0221] Thus, by effectively identifying and resolving logical conflicts between multi-source medical knowledge, and reducing reasoning contradictions caused by directly applying conflicting knowledge to the medical decision-making support model, the consistency and credibility of the supported decisions are improved. Secondly, by transforming the knowledge subset after conflict resolution into structured constraint hints and using these to control the generation process of the medical decision-making support model, an effective connection is achieved between the knowledge resolution results and the model generation process. Under the constraints of the structured constraint hints, the model is effectively prevented from deviating from or ignoring the resolved knowledge subset during the generation process, resulting in high credibility of the medical decision-making support information output by the model. Furthermore, the structured constraint hints allow the content of the knowledge subset to be directly understood and executed by the model without additional adaptation or conversion, improving the efficiency of the overall process.

[0222] According to some embodiments of this disclosure, optionally, the structured constraint prompt includes a constraint knowledge list and a conflict resolution record. The constraint knowledge list includes at least one of the source identifier, evidence level, credibility weight, applicable conditions, and conclusion of the knowledge corresponding to the structured knowledge unit. The conflict resolution record includes the identifier of the node that was resolved and the reason for its resolution.

[0223] According to some embodiments of this disclosure, optionally, the control module 1303 can also be used to control the medical decision support model to explicitly add a source identifier to the referenced constraint knowledge when generating medical decision support information. The source identifier includes at least one of the source name, publication time, and evidence level.

[0224] Optionally, according to some embodiments of this disclosure, the decision support system 1300 may further include a recording module for recording and archiving the processing process. The processing process record includes one or more of the following: medical knowledge fragments from multiple knowledge sources, semantic similarity calculation results, conflict detection results, credibility weight calculation details, resolution decision records, structured constraint prompts, and medical decision support information.

[0225] According to embodiments of this disclosure, this disclosure also provides an electronic device, a non-transitory computer-readable storage medium storing computer instructions, and a computer program product.

[0226] Figure 14A schematic block diagram of an example electronic device 1400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0227] like Figure 14 As shown, the electronic device 1400 includes a computing unit 1401, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1402 or a computer program loaded from a storage unit 1408 into a random access memory (RAM) 1403. The RAM 1403 may also store various programs and data required for the operation of the device 1400. The computing unit 1401, ROM 1402, and RAM 1403 are interconnected via a bus 1404. An input / output (I / O) interface 1405 is also connected to the bus 1404.

[0228] Multiple components in device 1400 are connected to I / O interface 1405, including: input unit 1406, such as a keyboard, mouse, etc.; output unit 1407, such as various types of displays, speakers, etc.; storage unit 1408, such as a disk, optical disk, etc.; and communication unit 1409, such as a network card, modem, wireless transceiver, etc. Communication unit 1409 allows device 1400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0229] The computing unit 1401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1401 performs the various methods and processes described above, such as multi-source medical knowledge conflict resolution methods or decision assistance methods. For example, in some embodiments, the multi-source medical knowledge conflict resolution method or decision assistance method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1400 via ROM 1402 and / or communication unit 1409. When the computer program is loaded into RAM 1403 and executed by the computing unit 1401, one or more steps of the multi-source medical knowledge conflict resolution method or decision assistance method described above can be performed. Alternatively, in other embodiments, the computing unit 1401 may be configured, by any other suitable means (e.g., by means of firmware), to perform a multi-source medical knowledge conflict resolution method or a decision support method.

[0230] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0231] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0232] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0233] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0234] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0235] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0236] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0237] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for resolving conflicts in multi-source medical knowledge, comprising: Medical knowledge fragments from multiple knowledge sources are uniformly expressed as structured knowledge units; Knowledge pairs are constructed based on different structured knowledge units, and conflict relationships between knowledge pairs are detected. A conflict graph is constructed using structured knowledge units as nodes and detected conflict relationships as edges. Calculate the credibility weight of each node in the conflict graph; Based on the conflict graph and the credibility weight, nodes with conflicting relationships are resolved, and the nodes retained after resolution are used as a subset of knowledge.

2. The method according to claim 1, wherein, The structured knowledge unit includes at least an applicable condition field and a conclusion field. Knowledge pairs are constructed based on different structured knowledge units, including: Calculate the semantic similarity between the applicable condition field and the conclusion field of different structured knowledge units; Structured knowledge units that meet the preset semantic similarity conditions are selected to form knowledge pairs, resulting in a set of knowledge pairs containing at least one knowledge pair.

3. The method according to claim 2, wherein, Calculate the semantic similarity between the applicability condition field and the conclusion field of different structured knowledge units, including: For different structured knowledge units, the first semantic similarity between applicable condition fields is calculated based on the trained medical language model, and the conclusion fields are inferred to be in the same clinical decision dimension. The preset semantic similarity conditions include that the first semantic similarity between two structured knowledge units is greater than a first threshold, and the conclusion fields are in the same clinical decision dimension. The same clinical decision dimension includes the same clinical problem type, the same medical behavior category, and / or the same diagnosis and treatment decision.

4. The method according to claim 1, wherein, Before detecting conflict relationships between knowledge pairs, the method further includes: The knowledge pairs in the knowledge pair set are checked for rule constraints using a set of logical rules in the medical field. Based on the rule constraints, filter out knowledge pairs in the knowledge pair set that do not conform to the logical rules of the medical field.

5. The method according to claim 1, wherein, The structured knowledge unit also includes an evidence level field, and the conflict relationship includes at least one of conclusion conflict, applicable condition conflict, and evidence level conflict.

6. The method according to claim 5, wherein, The conflict relationships include conclusion conflicts, and the detection of conflict relationships between knowledge pairs includes: Calculate the first semantic similarity of the applicable condition field of the two structured knowledge units in the knowledge pair, and the second semantic similarity or semantic mutual exclusion of the conclusion field; If the first semantic similarity is greater than the second threshold, and the second semantic similarity is less than or equal to the third threshold, or the semantic mutual exclusion is greater than the fourth threshold, then the knowledge pair is determined to have a conclusion conflict.

7. The method according to claim 5, wherein, The conflict relationships include conflicts of applicable conditions, and the conflict relationships between detected knowledge pairs include: Calculate the second semantic similarity of the conclusion fields of the two structured knowledge units in the knowledge pair; If the second semantic similarity is greater than the third threshold, then it is detected whether there is a logical mutual exclusion relationship between the applicable condition fields of the two structured knowledge units in the knowledge pair; If a logically mutually exclusive relationship exists, then the knowledge pair is determined to have conflicting applicable conditions.

8. The method according to claim 5, wherein, The conflict relationships include conflicts of evidence levels and conflicts between pairs of detected knowledge, including: Calculate the second semantic similarity of the conclusion fields of the two structured knowledge units in the knowledge pair; Map the evidence levels of the two structured knowledge units in the knowledge pair to numerical values, and calculate the absolute value of the difference between the two numerical values; If the second semantic similarity is greater than the third threshold and the absolute value is greater than the fifth threshold, then the knowledge pair is determined to have a conflict of evidence level.

9. The method according to claim 1, wherein, A conflict graph is constructed using structured knowledge units as nodes and detected conflict relationships as edges, including: Each structured knowledge unit is treated as a node. For every two nodes that have a conflict relationship, at least one directed edge is established between them, and each conflict edge is labeled with a conflict type label to obtain a conflict graph.

10. The method according to claim 1, wherein, The credibility weight includes at least one of the following: evidence strength weight, source reliability weight, and context adaptability weight. The evidence strength weight is used to characterize the evidence quality of the structured knowledge unit, the source reliability weight is used to characterize the credibility of the knowledge source of the structured knowledge unit, and the context fit weight is used to characterize the degree of matching between the structured knowledge unit and the applicable conditions of the target medical case.

11. The method according to claim 10, wherein, The credibility weight includes an evidence strength weight. Calculating the credibility weight for each node in the conflict graph includes: To obtain the level of evidence for structured knowledge units, the type of research on which they are based, and the size of the research sample; The evidence level is mapped to a first score, the research type is mapped to a second score, and the research sample size is normalized to a third score. The first score, the second score, and the third score are weighted and summed to obtain the evidence strength weight.

12. The method according to claim 10, wherein, The credibility weight includes a source reliability weight, and the structured knowledge unit also includes a knowledge source field; Calculating the credibility weight of each node in the conflict graph includes: Obtain the knowledge source, publication time, and target medical knowledge that matches the structured knowledge unit; Based on a preset authority rating table, determine the source authority score corresponding to the knowledge source; Calculate the timeliness decay score based on the publication time; Calculate the consistency score between the knowledge of the structured knowledge unit and the target medical knowledge; The source reliability weight is obtained by weighting and summing the source authority score, the timeliness decay score, and the consistency score.

13. The method according to claim 10, wherein, The credibility weight includes a context adaptability weight. The credibility weight of each node in the conflict graph is calculated as follows: Obtain contextual information of the target object, wherein the contextual information includes at least one of age, gender, chief complaint, past medical history, and concomitant medications; The context information of the target object is semantically matched with the applicable condition fields of the structured knowledge unit, and a matching degree score is calculated. The matching score is used as the context adaptability weight.

14. The method of claim 10, wherein, Calculating the credibility weight of each node in the conflict graph includes: Obtain the evidence strength weight, the source reliability weight, and the context adaptability weight; The evidence strength weight, the source reliability weight, and the context suitability weight are weighted and summed to obtain the comprehensive credibility score; The overall credibility score is used as the credibility weight of the node.

15. The method according to claim 1, wherein, Based on the conflict graph and the credibility weights, nodes with conflicting relationships are resolved, and the nodes retained after resolution are taken as a subset of knowledge, including: Based on the conflict relationships in the conflict graph, constraints are constructed, including that there is no conflict between any two nodes in the knowledge subset. The knowledge subset is solved under the given constraints with the goal of maximizing the sum of the credibility weights of each node in the knowledge subset.

16. The method according to claim 1, wherein, Each edge in the conflict graph is also labeled with an initial conflict intensity value, which is used to characterize the severity of the conflict between two nodes. The goal is to maximize the sum of the credibility weights of all nodes in the knowledge subset. Solving for the knowledge subset under the given constraints includes: Using the initial conflict intensity value as a penalty term, and aiming to maximize the sum of the credibility weights of each node in the knowledge subset, the knowledge subset is solved under the constraints. The penalty term is used to adjust the selection priority based on the initial conflict intensity value between nodes when multiple candidate nodes need to be selected, provided that the constraint conditions are met. The higher the conflict intensity, the lower the priority of the selected node.

17. The method according to claim 15 or 16, wherein, Based on the conflict graph and the credibility weights, nodes with conflicting relationships are resolved, and the nodes retained after resolution are taken as a subset of knowledge. This also includes: When there is a conflict of evidence level or a conflict of applicable conditions between two nodes, add conflict labeling information to the nodes with conflict of evidence level or conditions. The nodes with added conflict annotations and their corresponding conflict annotation information are retained in the knowledge subset.

18. The method according to claim 1, wherein, The method further includes: When the difference in credibility weight between two conflicting nodes is lower than the preset sixth threshold, a human-machine collaborative review process is triggered. The conflict details and resolution plan are presented to the target personnel. After receiving the confirmation instruction from the target personnel, the resolution or output is executed according to the confirmation instruction.

19. A decision support method, comprising: A knowledge subset is obtained based on the method of any one of claims 1-18; Transform the aforementioned subset of knowledge into structured constraint hints; By using structured constraint prompts, the medical decision support model is controlled to generate medical decision support information within the constraint boundaries of the structured constraint prompts.

20. The method according to claim 19, wherein, The structured constraint hints include a constraint knowledge list and conflict resolution records. The constraint knowledge list includes at least one of the following: source identifier, evidence level, credibility weight, applicable conditions, and conclusion of the knowledge corresponding to the structured knowledge unit. The conflict resolution records include the identifier of the node that was resolved and the reason for its resolution.

21. The method according to claim 19, wherein, The method further includes: When generating medical decision support information, the medical decision support model explicitly adds a source identifier to the referenced constraint knowledge. The source identifier includes at least one of the following: source name, publication time, and evidence level.

22. The method according to claim 19, wherein, The method further includes: The processing process is recorded and archived, and the processing process record includes one or more of the following: medical knowledge fragments from multiple knowledge sources, semantic similarity calculation results, conflict detection results, credibility weight calculation details, resolution decision records, structured constraint prompts, and medical decision support information.

23. A multi-source medical knowledge conflict resolution system, comprising: The knowledge structuring module is used to uniformly express medical knowledge fragments from multiple knowledge sources into structured knowledge units; The conflict detection module is used to construct knowledge pairs based on different structured knowledge units and detect conflict relationships between knowledge pairs. The building module is used to construct a conflict graph using structured knowledge units as nodes and detected conflict relationships as edges. The calculation module is used to calculate the credibility weight of each node in the conflict graph; The resolution module is used to resolve nodes with conflicting relationships based on the conflict graph and the credibility weight, and to retain the nodes after resolution as a subset of knowledge.

24. A decision support system, comprising: The acquisition module is used to acquire a subset of knowledge output by the multi-source medical knowledge conflict resolution system as described in claim 23; The transformation module is used to transform the knowledge subset into structured constraint hints; The control module is used to control the medical decision support model to generate medical decision support information within the constraint boundaries of the structured constraint prompts, through structured constraint prompts.

25. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-22.

26. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-22.

27. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-22.

Images