Structured feature-based cross-guideline recommendation opinion consistency detection method and system

Abstract

This invention discloses a method and system for cross-guideline recommendation consistency detection based on structured features. First, it collects clinical practice guideline texts from multiple sources, parses and processes the recommendations within these texts to generate structured clinical semantic data. Then, based on this structured clinical semantic data, recommendations from different sources are paired to construct recommendation pairs, and a structured feature representation is generated for each pair, resulting in a set of candidate recommendation pairs with potential relationships. Next, using these recommendation pairs as consistency analysis objects, consistency inference analysis is performed under the constraints of structured semantic information. The results of the consistency relationship determination between recommendations and their corresponding analytical basis are output, along with a credibility assessment and rollback control of the output. This invention provides interpretable and controllable intelligent support for clinical guideline knowledge integration, guideline version comparison, evidence conflict identification, and intelligent decision support systems.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and intelligent decision support systems, and more specifically, to a method and system for cross-guideline recommendation consistency detection based on structured features, used for automated analysis and reliable determination of consistency and conflict relationships among recommendations in clinical practice guideline texts from multiple sources. Background Technology

[0002] Clinical practice guidelines are important achievements of evidence-based medicine, widely applied in disease diagnosis and treatment, drug use, risk assessment, and public health management. As medical research evidence accumulates, different countries, regions, and academic organizations often publish multiple clinical practice guidelines for the same disease or clinical problem. Due to differences in evidence sources, research design, expert consensus formation processes, and update cycles, different guidelines may exhibit inconsistencies or even contradictions in areas such as defining the applicable population, selecting interventions, setting thresholds, and expressing the strength of recommendations.

[0003] In practical applications, medical staff or guideline administrators often need to compare and analyze guideline recommendations from different sources to determine their consistency or discrepancies. Current technologies often rely on human experts to read and compare each recommendation line by line, or to match certain structured fields using pre-defined rules. While these methods are feasible for handling smaller, relatively standardized guideline texts, they often suffer from low efficiency, high subjectivity, and high maintenance costs when dealing with large amounts of natural language descriptions, complex semantic structures, and frequently updated guideline texts.

[0004] In recent years, with the development of natural language processing technology, some technical solutions have attempted to introduce automated text analysis methods to perform consistency analysis on recommendations in clinical guidelines. However, existing automated methods still have shortcomings in practical engineering applications. On the one hand, different guidelines express recommendations in significantly different ways, and key information such as applicable populations, intervention measures, and outcome indicators often appear in unstructured form, increasing the difficulty of semantic understanding and alignment. On the other hand, when dealing with fine semantics such as numerical thresholds, dosage ranges, or contraindications, existing methods struggle to guarantee the stability of the judgment results.

[0005] Furthermore, some model-based analysis methods often lack clear explanations of the criteria used to determine consistency, making it difficult to trace the reasoning process. When there is uncertainty or potential conflict risk in the model output, existing technologies also lack corresponding credibility assessment and risk control measures, which limits their practical application in clinical decision support and guideline management scenarios.

[0006] Therefore, how to improve the controllability, stability, and reliability of the consistency testing process for clinical recommendations while ensuring the efficiency of automated analysis remains a pressing technical problem to be solved in the current technology. Summary of the Invention

[0007] To address the common problems in the automatic processing of recommendations in existing clinical practice guidelines, such as scattered semantic information in recommendations, high difficulty in matching recommendations across guidelines, lack of unified constraints in the analysis process, and insufficient stability of consistency judgment results, this invention proposes a cross-guideline recommendation consistency detection method and system based on structured features and credibility backtracking.

[0008] This invention aims to improve the stability and interpretability of clinical guideline knowledge integration and intelligent decision support applications by unifying the recommendation processing flow at the system level, through semantic element extraction and standardized governance mechanisms, candidate pair screening and priority grading mechanisms based on structured features and risk gating, controlled consistency reasoning, and a closed-loop mechanism for credibility assessment and backoff control.

[0009] To achieve the above objectives, the first aspect of the present invention provides a method for cross-guideline recommendation consistency detection based on structured features, comprising: We collect clinical practice guideline texts from multiple sources, analyze and process the recommendations in the clinical practice guideline texts, and generate structured clinical semantic data that includes applicable populations, intervention measures, control protocols, outcome indicators, and recommendation strength or evidence levels. Based on the structured clinical semantic data, recommendations from different sources are combined in pairs to construct recommendation pairs, and a structured feature representation is generated for each recommendation pair. Based on the structured feature representation of the recommendation pairs, a set of candidate recommendation pairs with potential relationships is generated. Using the candidate recommendation pairs obtained from the screening as the consistency analysis objects, consistency reasoning analysis is performed under the constraints of structured semantic information. The results of the consistency relationship judgment between the recommendations and their corresponding analysis basis are output, and the credibility assessment and rollback control of the output are performed.

[0010] In one implementation, multiple sources of clinical practice guideline texts are collected, and the recommendations in the clinical practice guideline texts are parsed and processed, including: Obtain the original text of the recommendations in the clinical practice guidelines and the source identification information of the guidelines to which they belong; Automatically identify and locate recommendations from clinical practice guideline texts; Semantic analysis was performed on the recommendations to extract the applicable population, intervention measures, control group, outcome indicators, and recommendation strength or level of evidence. The extracted clinical semantic information is standardized, mapping similar semantic information in different forms of expression into a unified structured field representation.

[0011] In one implementation, generating a structured feature representation for each recommendation includes: Based on the applicable population field, semantic similarity of the recommendations is calculated on the dimension of population characteristics. Based on the intervention measures field, semantic similarity is calculated for the recommendations in the dimension of intervention strategies to measure the similarity between the two recommendations in terms of intervention methods; Based on the outcome index field, the semantic similarity of the recommendations in the clinical outcome dimension is calculated to measure whether the clinical outcomes or goals of the two recommendations are consistent. Based on the text content of the recommendations, the semantic correlation between the recommendations is analyzed to capture supplementary semantic information beyond the structured fields; The analysis results from the above dimensions are combined to form a structured feature representation of the recommendation pair.

[0012] In one implementation, based on the applicable population field, semantic similarity calculation is performed on the recommendation opinions along the population feature dimension, including:

[0013] in, For consistency judgment based on disease codes, a perfect match is recorded as 1, otherwise it is recorded as 0; The similarity of comorbidities is calculated using the multiple union-intersection ratio; Similarity is used for matching based on demographic information; , , These are the corresponding weights, used to control the contribution of different factors. This refers to semantic similarity based on the dimension of population characteristics.

[0014] In one implementation, a set of candidate recommendation pairs with potential relationships is generated based on the structured feature representation of the recommendation pairs, including: Based on preset scoring or screening rules, a weighted scoring model is constructed using logistic regression to comprehensively evaluate the multidimensional feature representations of recommendation pairs. The weighted scoring model is as follows: , ( ) is the recommendation consisting of recommendation i and recommendation j. The corresponding feature vector is w represents the feature weight vector, and σ(·) is the Sigmoid function. This is the overall scoring result; The comprehensive evaluation results are compared with the preset conditions, and the recommended opinion pairs that meet the conditions are selected to form a set of candidate recommended opinion pairs.

[0015] In one implementation, the pairs of recommendations in the selected candidate recommendation pair set are used as the consistency analysis objects. Consistency reasoning analysis is performed under structured semantic information constraints, outputting the consistency relationship determination results between recommendations and their corresponding analysis basis, including: The structured clinical semantic fields of candidate recommendation pairs, field-level evidence fragment indexes, and the original content of the recommendation opinions are combined to form a structured semantic input for consistency analysis; Under the structured semantic input constraints, the relationship type between candidate recommendation pairs is determined, and the relationship type includes at least consistent, inconsistent, and irrelevant; Generate the analytical basis or analytical path corresponding to the consistency relationship determination results.

[0016] In one implementation, the method further includes credibility assessment and rollback control of the consistency analysis results, specifically: The credibility of the consistency determination results is assessed based on the use of structured fields, sufficiency of evidence, consistency of analysis path, and stability of analysis conclusions during the analysis process. When the credibility assessment result is lower than the preset condition, a rollback process is triggered, and the consistency analysis of the candidate recommendation pair is re-executed, or the analysis is performed again after adjusting the prompt constraints, supplementing evidence fragments, rearranging the field expansion order, or switching to the conservative judgment mode.

[0017] Based on the same inventive concept, a second aspect of this invention provides a cross-guideline recommendation consistency detection system based on structured features, comprising: The clinical semantic parsing module is used to collect clinical practice guideline texts from multiple sources, parse and process the recommendations in the clinical practice guideline texts, and generate structured clinical semantic data that includes the applicable population, intervention measures, control schemes, outcome indicators, and recommendation strength or evidence level. The candidate recommendation pair generation module is used to combine recommendations from different sources in pairs based on the structured clinical semantic data to construct recommendation pairs, generate a structured feature representation for each recommendation pair, and generate a set of candidate recommendation pairs with potential relationships based on the structured feature representations of the recommendation pairs. The controlled consistency reasoning module is used to perform consistency analysis on pairs of recommendations in the selected candidate recommendation pair set, under the constraints of structured semantic information. It outputs the consistency relationship judgment results between recommendations and their corresponding analysis basis, and performs credibility assessment and rollback control on the output.

[0018] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is used to implement the above-mentioned method for detecting consistency of recommendations.

[0019] Based on the same inventive concept, a fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it is used to implement the above-mentioned method for detecting consistency of recommendations.

[0020] Compared to existing technologies that transform clinical guidelines into computable representations and handle conflicts based on rules or argumentation frameworks, emphasizing argumentative interpretation and computable guideline deployment, this invention focuses on consistency detection of cross-guideline recommendations. It highlights structured feature-driven candidate screening and risk grading, field-level evidence fragment location and traceable evidence chains, credibility assessment and backoff control loops, so as to trigger re-inference or manual review and retain audit logs when the model output is uncertain or conflicts with field evidence.

[0021] Compared with technologies that automatically synthesize evidence and generate recommendations based on large language models, which mainly address the generation process from evidence to recommendation, this invention addresses the detection process for determining the consistency and conflict relationships of recommendation pairs, and enhances the stability and traceability of cross-recommendation judgments through structured feature constraints and credibility back-off closed loops.

[0022] Furthermore, existing research on source tracing or source management in guidelines mostly focuses on the expression of proof or the management of rule sources, but it does not form a unified closed loop with structured feature candidate screening, controlled consistency reasoning, credibility assessment and backoff control.

[0023] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows: By extracting semantic elements and representing them in a structured manner from recommendations in clinical practice guidelines, the complexity of matching between different guidelines due to differences in expression is reduced. By introducing a structured feature and candidate recommendation pairing screening mechanism before consistency analysis, the scale of the analysis is effectively controlled, improving the overall processing efficiency of the system. By imposing structured semantic constraints on the consistency analysis process and outputting an evidence index, the analysis results have clear evidence and improved traceability. By introducing a credibility assessment and backtracking control mechanism for the consistency analysis results, the system has the ability to ensure result stability in complex semantic scenarios. This invention has a clear structure and strong systematicity, and is suitable for consistency detection and knowledge integration scenarios of recommendations from multiple sources and versions of clinical practice guidelines, demonstrating significant engineering application value. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the overall process of the consistency detection method for recommendations in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the semantic element parsing and standardization process of recommendations in an embodiment of the present invention; Figure 3 This is a schematic diagram of the candidate recommendation pair generation and risk gating process in an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the benefits of candidate recommendations on scale reduction and time consumption in an embodiment of the present invention; Figure 5 This is a schematic diagram of the consistency reasoning, credibility assessment, and rollback control process in an embodiment of the present invention; Figure 6 This is a comparison chart of the consistency relationship verification of Qwen3-8B recommendations in this embodiment of the invention; Figure 7 This is a comparison chart of the consistency relationship verification of GLM4-9B recommendations in this embodiment of the invention; Figure 8 This is a schematic diagram of the system hierarchy of the recommendation consistency detection system in an embodiment of the present invention. Detailed Implementation

[0026] Example 1 This embodiment provides a method for cross-guideline recommendation consistency detection based on structured features. Please refer to [link to relevant documentation]. Figure 1,include: S1: Collect clinical practice guideline texts from multiple sources, parse and process the recommendations in the clinical practice guideline texts, and generate structured clinical semantic data that includes the applicable population, intervention measures, control protocols, outcome indicators, and recommendation strength or evidence level.

[0027] S2: Based on the structured clinical semantic data, recommendations from different sources are combined in pairs to construct recommendation pairs, and a structured feature representation is generated for each recommendation pair. Based on the structured feature representation of the recommendation pairs, a set of candidate recommendation pairs with potential relationships is generated.

[0028] S3: Using the pairs of recommendations in the candidate recommendation pair set obtained from the screening as the consistency analysis objects, perform consistency reasoning analysis under the constraints of structured semantic information, output the consistency relationship judgment results between recommendations and their corresponding analysis basis, and perform credibility assessment and backoff control on the output.

[0029] Specifically, S1 is a semantic element governance and evidence localization method for recommendations from multiple sources of clinical practice guidelines; S2 is a candidate recommendation pair generation method based on structured features and risk gating strategies; and S3 is a closed-loop method for controlled consistency reasoning, credibility assessment, and backoff control.

[0030] Before executing the three functional modules of this invention, it is first necessary to acquire multi-source clinical practice guideline text data, including the full text of the guideline, the original text of the recommendations, and the source information of the guideline to which it belongs (such as guideline name, issuing institution, version number, and publication time). The guideline text data serves as the core input for subsequent analysis of recommendation elements, construction of structured features, and consistency reasoning analysis, and is used to trigger the candidate recommendation pair screening and consistency relationship determination process.

[0031] To enable cross-guideline traceability and auditing, this embodiment assigns a unique source identifier (source_id) to each guideline and a unique opinion identifier (rec_id) to each recommendation. It also establishes a relationship of "source - recommendation - structural fields - evidence fragments - reasoning output" to ensure that the consistency judgment result can be located at the original text basis.

[0032] Through the synergy of the above three parts, this invention achieves automated element parsing, structured field representation, structured feature construction, candidate pair compression and screening, and traceable consistency relationship determination for clinical practice guideline recommendations, forming an end-to-end, highly reliable, and highly adaptable clinical guideline knowledge integration and intelligent decision support method.

[0033] Specifically, S1 involves the collection of guideline data and the analysis of recommendation elements. S1 can be implemented in the following ways: S1.1: Obtain the original text of the recommendations in the clinical practice guideline and the source identification information of the guideline to which it belongs; S1.2: Automatically identify and locate recommendations from clinical practice guideline text; S1.3: Perform semantic analysis on the recommendations to extract the applicable population, intervention measures, control schemes, outcome indicators, and recommendation strength or level of evidence; S1.4: Standardize the extracted clinical semantic information and map similar semantic information in different forms of expression into a unified structured field representation.

[0034] In practice, multiple sources of clinical practice guideline texts can be automatically collected through preset data interfaces or storage paths to obtain the original text of recommendations and the source information of the guidelines to which they belong, including guideline name, issuing institution, version number, and publication date, and to establish the association between recommendations and guideline sources. Based on predefined text recognition rules, the original text of recommendations can be automatically located and extracted from guideline texts. Semantic parsing is performed on the extracted recommendation text to automatically extract clinical semantic elements such as applicable population, intervention measures, control protocols, outcome indicators, and recommendation strength or level of evidence. The extracted clinical semantic elements are standardized and normalized, automatically mapping similar elements in different forms of expression to a unified structured field representation and storing it as structured data. Finally, a field-level evidence fragment index is established for each recommendation.

[0035] like Figure 2 As shown, the overall implementation process of element parsing is divided into three core stages: guideline data access and preprocessing, recommendation location and extraction, and clinical element extraction and standardized storage. This invention supports integration with multiple sources such as guideline publishing platforms, medical databases, or internal institutional guideline repositories, automatically acquiring guideline text through interface calls or batch import methods. The integration process meets the following specifications: the interface protocol uses HTTPS to ensure data transmission security and supports standardized request formats; the core fields to be synchronously acquired include guideline name, publishing institution, version number, publication time, full text of the guideline, and identifiers of the recommendation paragraphs it contains; heterogeneous guideline data from different sources are uniformly converted into parsable structured or semi-structured formats for storage, and a unique source identifier is established for each guideline to support subsequent cross-guideline consistency analysis.

[0036] In one embodiment, after S1.1, this embodiment further performs format standardization and noise filtering on the collected guide text.

[0037] Specifically, to address common noise issues in guideline texts, such as table of contents interference, headers and footers, reference markers, footnote numbers, and duplicate paragraphs, the system performs text cleaning through a combination of rule matching and semantic filtering: removing irrelevant chapter content and formatting symbols; standardizing character encoding to UTF-8, removing redundant spaces and line breaks, and standardizing punctuation; and deduplicating duplicate paragraphs from different versions of the same guideline to ensure the stability of subsequent recommendation location and extraction. For guideline texts containing numerous abbreviations, synonyms, or mixed use of drug brand names and generic names, the system can also utilize a clinical terminology mapping dictionary and drug knowledge base for preliminary normalization to reduce the impact of expression differences on subsequent alignment analysis.

[0038] For the extraction of recommendation elements, this embodiment adopts a collaborative approach of "clinical field system constraints + semantic parsing" to extract core elements such as applicable population, intervention measures, control protocols, outcome indicators, and recommendation strength / level of evidence. The extraction results must be strictly aligned with the predefined field types in the system. (1) Extraction of applicable population elements: The system automatically identifies information such as disease type, stratification conditions, age range, comorbidities or special population restrictions, and outputs a unified structured field Population; (2) Extraction of intervention elements: The system automatically identifies information such as drug category, dosage and frequency, non-drug intervention, examination and monitoring or surgical operation, and outputs a unified structured field Intervention; (3) Extraction of comparison scheme elements: The system identifies the comparison objects or alternative scheme information involved in the recommendations and outputs a unified structured field Comparison; (4) Extraction of outcome indicator elements: The system automatically identifies information such as efficacy endpoint, risk reduction target, indicator threshold or event outcome, and outputs a unified structured field Outcome; (5) Extraction of recommendation strength or evidence level elements: The system identifies recommendation strength, evidence level or evidence-based classification information, outputs a unified structured field Level & Grade, and retains the original description of the original classification system for traceability.

[0039] To avoid ambiguity in extraction and ensure consistency in structured output, the system imposes uniform constraints on field definitions, as shown in Table 1: Table 1: Explanation of Structured Fields in Recommendations

[0040] In one implementation, the semantic parsing described above can be performed by a machine reading model, which can be a large language model or other equivalent semantic extraction model. For example, DeepSeekR1, Qwen, GLM, or other large language models or equivalent models can be invoked, combined with field extraction prompt templates to output structured field results. The choice of model is not limited; the system can select an equivalent model that meets the accuracy and cost constraints under different deployment conditions.

[0041] In one implementation, the method further includes verifying and correcting the extraction results. Specifically, to ensure the accuracy and consistency of clinical element extraction, the system employs a dual mechanism of "rule consistency verification + structural integrity verification" to correct extraction errors and ambiguities, thereby ensuring data quality. (1) Structural integrity verification rules: The system performs integrity checks on the structured fields of each recommendation. If key fields are missing (e.g., Population, Intervention, or Outcome is empty), it is marked as an incomplete sample and triggers a re-analysis process. If the recommendation strength or evidence level field cannot be identified, it is recorded with a default mark and the original evidence fragment is retained for supplementation in the subsequent reasoning stage.

[0042] (2) Element consistency correction mechanism: The system realizes the logical consistency check between elements through machine verification. After discovering contradictions, the correction process is triggered: normalize synonyms and abbreviations; verify the matching between population constraints and intervention measures; and standardize the threshold units in the description of outcome indicators to ensure the stability of subsequent similarity calculation and consistency reasoning.

[0043] In one implementation, the evidence location and indexing are achieved using a "field-driven evidence fragment segmentation—matching—scoring" process: (1) Evidence fragment segmentation: The system takes the paragraph where the original recommendation is located as the center, performs sentence-level segmentation in the full text of the guide according to the boundaries such as period / semicolon / line break, and constructs candidate evidence fragments based on the sliding window strategy. The window length can be selected from 80 to 240 characters. (2) Field anchor point generation: Generate a set of field anchor points for the structured fields Population, Intervention, Outcome, and Level & Grade respectively. The field anchor points include field keywords, standardized synonyms, coding items (such as standard codes for diseases / drugs / indicators) and numerical threshold units (numerical value + unit + direction). (3) Evidence fragment matching: Calculate the matching degree between each candidate evidence fragment and the set of field anchor points. Specifically, for the field... Anchor point set With candidate evidence fragments Calculate the overall matching degree:

[0044] in, , , , For preset weights or weight parameters optimized based on historical samples, the matching degree should include at least two or more of the following categories of indicators: a. Word literal matching score : Coverage or Jaccard coefficient is calculated based on the set of anchor keywords and the word segmentation results of candidate evidence fragments, for example:

[0045] Used to measure the proportion of direct occurrences of field keywords in evidence fragments.

[0046] b. Synonym normalization matching score : Terminology normalization (including synonym mapping, abbreviation expansion, and encoding standardization) is performed on both field anchors and candidate evidence fragments. Coverage is then calculated after normalization.

[0047] Used to identify semantically equivalent but literal expressions.

[0048] c. Numerical threshold consistency score : When a field anchor contains a numerical expression, the system extracts the numerical value, unit, and inequality direction, standardizes the units, and then calculates the interval overlap ratio or direction consistency score. For example:

[0049] If the numerical ranges are completely identical, the score is 1; if they partially overlap, the score is between 0 and 1; if the directions conflict, the score is 0.

[0050] d. Semantic embedding similarity score : Using a semantic representation model, vector representations of field anchor text and candidate evidence fragments are generated, and cosine similarity is calculated:

[0051] Used to measure overall semantic relevance.

[0052] The system can select candidate evidence fragments with a comprehensive matching degree higher than a preset threshold, or select the top K fragments with the highest scores as the field evidence set.

[0053] (4) Evidence scoring and selection: Select the top-K highest-scoring evidence fragments for each field as the field evidence set, and record the field-evidence fragment mapping relationship; (5) Index structure: A location index is established for each evidence fragment. The location index includes at least two items from the chapter number, paragraph number, period, page number or character offset, and records the evidence fragment source guide identifier source_id, recommendation identifier rec_id, field name, matching score and matching method mark.

[0054] During the controlled consistency inference phase, the structured fields of candidate recommendation pairs are compared and aligned field by field, and field-level conclusions are generated under the constraints of the structured output template. For each field, the inference output includes results such as consistent, partially consistent, inconsistent, or indeterminate. When the field-level conclusion is inconsistent or partially consistent, a field-level difference point object is generated, which includes: The data includes field names, difference types (e.g., differences in population scope, threshold range, intervention intensity, recommendation level, etc.), summaries of difference content, and an index set of evidence fragments supporting the difference judgment.

[0055] Among them, field-level differences are generated by the controlled consistency inference step and output as part of the consistency determination result.

[0056] When outputting the consistency judgment results, the overall consistency relationship and the conclusions of each field are output together; for fields with differences, the corresponding field-level difference points and their evidence fragment indexes are also output to achieve traceable positioning from the overall conclusion to the field differences and then to the original text basis.

[0057] S2 generates a structured feature representation for each recommendation, including: S2.1: Based on the applicable population field, perform semantic similarity calculation on the recommendation opinions in terms of population characteristics; S2.2: Based on the intervention measures field, perform semantic similarity calculation on the intervention strategy dimension of the recommendations to measure the similarity between the two recommendations in terms of intervention methods; S2.3: Based on the outcome index field, the semantic similarity of the recommendations in the clinical outcome dimension is calculated to measure whether the clinical outcomes or goals of the two recommendations are consistent. S2.4: Based on the text content of the recommendations, analyze the degree of semantic correlation between the recommendations to capture supplementary semantic information beyond the structured fields; S2.5: Combine the analysis results of the above dimensions to form a structured feature representation of the recommendation pair.

[0058] Specifically, the similarity of population characteristics, intervention measures, outcome indicators, and semantic similarity of the original text between recommendations is calculated based on structured fields. These similarity features are then combined to form a structured feature representation of the recommendation pair. Based on the structured feature representation, the recommendation pairs are ranked using a comprehensive scoring strategy, and candidate recommendation pairs with potential correlation are selected according to a preset threshold to reduce the computational scale of subsequent consistency inference.

[0059] In the specific implementation process, such as Figure 3 As shown, the structured feature construction and candidate selection process is divided into four stages: structured field alignment, multidimensional similarity calculation, feature set construction, and candidate pair scoring and selection. The specific implementation is as follows: Step 1: Align structured fields. Align the structured fields of recommendations from different sources according to a unified field system to ensure that the Population, Intervention, Outcome, and Level & Grade fields are comparable in semantics and format, and establish a mapping relationship between the unique identifier of the recommendation and the source information of the guide; Step 2: Multidimensional similarity calculation. The system calculates the following separately: (1) Population feature similarity: The similarity is generated based on the degree of matching of disease type, stratification conditions, age range and comorbidity restrictions in the Population field; (2) Intervention similarity: Similarity is generated based on the degree of matching between the intervention category, drug category, or management strategy in the Intervention field; Intervention similarity ( The similarity feature (SFR) measures the degree of similarity between two recommendations in terms of intervention methods, including drug name, treatment method, or intervention type. This feature is used to determine whether recommendations revolve around the same or similar interventions.

[0060] (3) Outcome index similarity: Similarity is generated based on the index name, threshold, and target direction in the Outcome field; Outcome index similarity ( This is used to measure whether two recommendations focus on the same clinical outcomes or goals, such as whether they target the same physiological indicators, clinical endpoints, or risk events.

[0061] (4) Original text semantic similarity: Supplementary similarity is generated based on the semantic representation of the original text in the recommendations to cover the limitations and exceptions not fully expressed by the structured fields. Original text semantic similarity ( Similarity is calculated based on the semantic representation of the original text of the recommendation opinions to capture supplementary semantic information beyond the structured fields and mitigate the impact of differences in the wording of different guidelines.

[0062] Among these, based on the applicable population field, semantic similarity calculation is performed on the recommendations in terms of population characteristics, including:

[0063] in, For consistency judgment based on disease codes, a perfect match is recorded as 1, otherwise it is recorded as 0; The similarity of comorbidities is calculated using the multiple union-intersection ratio; Similarity is used for matching based on demographic information; , , These are the corresponding weights, used to control the contribution of different factors. This refers to semantic similarity based on the dimension of population characteristics.

[0064] Based on intervention, outcome, and text fields, this implementation uses a medical semantic embedding model (SBERT-Base-Chinese-NLI) to map field content into dense semantic vectors, and employs cosine similarity to construct a general embedding similarity calculation function, as shown in the following formula:

[0065] in, These are two recommendations to be compared. The target field for which similarity is to be calculated. , These are the semantic embedding vectors of the fields corresponding to the two recommendations, and the fields are... The range of values ​​is , respectively.

[0066] Step 3: Feature Set Construction. The system combines the similarity features from the above dimensions to form a structured feature set for the recommendation pair, as shown in Table 2: Table 2: Explanation of the Structured Features in the Recommendations

[0067] Step 4: Candidate Pair Scoring and Screening. The system comprehensively evaluates the structured features using a comprehensive scoring strategy to generate a correlation score. The correlation score is compared with a preset threshold, and recommendation pairs that meet the threshold conditions are screened to generate a set of candidate recommendation pairs.

[0068] In one implementation, a set of candidate recommendation pairs with potential relationships is generated based on the structured feature representation of the recommendation pairs, including: Based on preset scoring or screening rules, a weighted scoring model is constructed using logistic regression to comprehensively evaluate the multidimensional feature representations of recommendation pairs. The weighted scoring model is as follows: , ( ) is the recommendation consisting of recommendation i and recommendation j. The corresponding feature vector is w represents the feature weight vector, and σ(·) is the Sigmoid function. This is the overall scoring result; The comprehensive evaluation results are compared with the preset conditions, and the recommended opinion pairs that meet the conditions are selected to form a set of candidate recommended opinion pairs.

[0069] Specifically, during the training phase, based on a small number of manually labeled recommendation pairs, each training sample consists of a recommendation pair and its corresponding features and labels, which is formally represented as ( , , , , , The supervision signal Y∈{0,1} indicates whether further consistency or inconsistency analysis is necessary for the recommendation pair. When Y=1, it indicates that the recommendation pair may have a potential relationship such as consistency or inconsistency, while when Y=0, it indicates that no obvious semantic or logical relationship has been observed. A learned weight vector w enables the weighted scoring model to distinguish between potentially related recommendation pairs and irrelevant recommendation pairs.

[0070] During the screening phase, the recommended opinions are binary-valued based on a preset threshold T:

[0071] in, The comprehensive score is the output of the weighted scoring model, where T is the pre-set screening threshold. When the comprehensive score... If the score is not lower than the threshold T, the recommendation pair is determined to have a potential semantic relationship and is included in the candidate recommendation pair set for subsequent consistency relationship identification; if the comprehensive score is lower than the threshold T, the recommendation pair is considered to lack obvious correlation and is filtered out in the candidate screening stage.

[0072] As a preferred approach, a risk gating strategy is introduced before candidate scoring to perform rapid filtering and risk classification on recommendations, ensuring that conflicting candidates are prioritized for subsequent consistent inference. The risk gating strategy consists of "hard gating rules + risk priority classification": (1) Hard gate filtering rules: When the intersection of the disease / scenario codes in the Population field is empty and the intervention categories (drugs / surgery / lifestyle / monitoring) in the Intervention field are completely different, it is judged as a low-association combination and filtered directly; when only one of them is satisfied, it enters the candidate but the priority is reduced; (2) Conflict Clue Identification Rules: The system identifies conflict triggering patterns in the Level & Grade fields and original evidence fragments. The patterns include at least polar words such as “contraindication / not recommended / avoid / use with caution / strongly recommended / must / should”, numerical threshold inequalities (such as ≥, ≤, <, >), and dose / frequency / treatment unit difference patterns. When any of the above patterns is identified, a conflict risk mark is assigned. (3) Risk level classification: The system outputs risk level labels Risk∈{L0,L1,L2} based on gating and conflict clues, where L0 represents regular candidates, L1 represents potential conflict candidates, and L2 represents high-risk conflict candidates; (4) Integration with candidate scores: The system adopts a “gating priority + score threshold” strategy, that is, first perform hard gating filtering, then calculate the comprehensive score of the retained samples and set thresholds or Top-N quotas according to risk level, so that L1 / L2 candidates can enter the consistency inference stage first under the same resource budget.

[0073] To illustrate the impact of the candidate selection mechanism on the scale of consistent inference and computational resource consumption, in one embodiment, taking an example with N=100 recommended opinions, a comparison is made between the full pairing method and the candidate selection optimization method. The comparison metrics include the number of candidate pairs, the number of inference calls, and the total processing time. The results are shown below. Figure 4 In the example configuration of this embodiment, the number of fully paired combinations is 4950. After candidate filtering optimization, the number of candidate combinations entering the consistent inference stage is reduced to 851, thereby reducing the number of inference calls and lowering the total processing time from 6631s to 1038s. It should be understood that the above values ​​are used to illustrate the scale reduction and time savings brought about by candidate filtering. The specific time and number of calls will vary depending on the data scale, model type, deployment method, concurrency strategy, and hardware environment.

[0074] By pre-filtering low-association combinations, the input range for subsequent consistent inference can be narrowed to a more potentially correlated candidate set, thereby reducing the computational cost of inference. For combinations not covered in the screening stage, supplementary processing can be carried out by adjusting thresholds and gating rules, expanding candidate quotas, or introducing manual review strategies.

[0075] Although there may be a trade-off between recall and compression during the candidate screening stage, the reasoning scale of irrelevant recommendation pairs can be significantly reduced while ensuring the availability of subsequent consistency inferences. This improves the overall efficiency of the system from an engineering perspective, and this trade-off is in line with the practical application needs of clinical guideline consistency analysis.

[0076] In one implementation, the pairs of recommendations in the selected candidate recommendation pair set are used as the consistency analysis objects. Consistency reasoning analysis is performed under structured semantic information constraints, outputting the consistency relationship determination results between recommendations and their corresponding analysis basis, including: S3.1: Combine the structured clinical semantic fields, field-level evidence fragment indexes, and original content of the candidate recommendation pairs to form a structured semantic input for consistency analysis; S3.2: Under the structured semantic input constraints, determine the relationship type between candidate recommendation pairs, wherein the relationship type includes at least consistent, inconsistent, and irrelevant; S3.3: Generate the analytical basis or analytical path corresponding to the consistency relationship determination results.

[0077] In practice, S3 involves consistent reasoning under structured semantic constraints, combined with credibility assessment and rollback control to form a closed-loop process. For example... Figure 5 As shown, the process first plans the consistency reasoning steps and analysis objectives, then integrates the structured fields and field-level evidence fragment indexes of candidate recommendation pairs with the original text to form structure-enhanced input; the structure-enhanced input is used as external constraint knowledge input to the reasoning model; the reasoning model is guided to output relationship judgment results and corresponding reasoning basis or reasoning path; structured output template constraints; output consistency verification and re-reasoning mechanism; credibility assessment and scoring are performed, and the control loop and auditable logs are rolled back.

[0078] The guiding reasoning model outputs relationship determination results and corresponding reasoning basis or reasoning path, including: Align and compare the Population field of candidate recommendation pairs to identify whether the applicable populations are consistent or conflicting. Align and compare the Intervention field of candidate recommendation pairs to identify whether the interventions are consistent, complementary, or mutually exclusive. Align and compare the Outcome field of candidate recommendation pairs to identify whether the outcome indicators are consistent or whether there are differences in the objectives; By combining the Level & Grade fields with the original evidence fragments, the difference between the recommendation strength and the level of evidence is determined, and a consistency conclusion is output. Examples of relation types are shown in Table 3. Table 3: Examples of Consistency Relationship Types in Recommendations

[0079] In one implementation, controlled consistent reasoning employs structured output templates to reduce the instability caused by the free generation of the reasoning model. The system provides structured semantic constraints to the reasoning model, requiring the output to include at least: (1) Field alignment conclusion table: Provide field-level conclusions for Population, Intervention, Outcome, and Level & Grade respectively. The field-level conclusions should include at least the following: consistent / inconsistent / irrelevant. (2) List of differences: For fields that are determined to be inconsistent or partially consistent, output a summary of differences (e.g., threshold differences, dosage differences, contraindication differences, differences in applicable population stratification, etc.). (3) Citation of evidence: Each point of difference must cite the corresponding field evidence fragment index (chapter / paragraph / page number / offset); (4) Overall Relationship Determination: Based on the field-level conclusions, output the overall relationship type (including at least consistent, inconsistent, and irrelevant).

[0080] In one implementation, the system performs consistency checks on the consistent inference output. When there is a preset logical inconsistency between the overall relationship determination result and the field alignment conclusion table, the inference output is determined to be abnormal output and a review inference process is triggered. The logical inconsistencies include at least: key inconsistencies exist in field-level conclusions while the overall relationship is determined to be consistent; all field-level conclusions are irrelevant while the overall relationship is determined to be consistent or inconsistent; or the field-level differences do not match the cited evidence fragments, etc.

[0081] In the review reasoning process, the system can reconstruct the reasoning inputs and constraints to regenerate consistency judgment results. Review reasoning strategies include at least one or more of the following: adjusting the order of field expansion and alignment comparisons; strengthening conflict triggering rule constraints and introducing more stringent structured output templates; supplementing or expanding the set of field-level evidence fragments; and enabling result self-consistency verification constraints to ensure consistency between field-level conclusions and the overall conclusion during the reasoning process. If logical inconsistencies cannot be eliminated after a preset number of review reasoning iterations, a review flag is output, and the process enters the manual review phase.

[0082] In one implementation, the method further includes credibility assessment and rollback control of the consistency analysis results, specifically: The credibility of the consistency determination results is assessed based on the use of structured fields, sufficiency of evidence, consistency of analysis path, and stability of analysis conclusions during the analysis process. When the credibility assessment result is lower than the preset condition, a rollback process is triggered, and the consistency analysis of the candidate recommendation pair is re-executed, or the analysis is performed again after adjusting the prompt constraints, supplementing evidence fragments, rearranging the field expansion order, or switching to the conservative judgment mode.

[0083] In practice, to ensure that the consistency relationship determination results have measurable credibility output and support subsequent rollback control, the system performs a credibility assessment on the consistency inference output, generating a credibility score (Conf). This credibility score can be normalized to the range of 0 to 1 and is obtained based on a computable indicator system, which includes at least the following: (1) Field reference integrity C1: Used to characterize the proportion of the number of fields in the inference output that are explicitly given a conclusion to the preset field set; (2) Sufficiency of evidence citation C2: This is used to characterize whether all points of difference are given corresponding evidence fragment indexes or location information. The higher the proportion of these criteria, the higher the score. (3) Conclusion self-consistency C3: used to characterize whether the overall relationship determination and the field-level conclusion meet the preset logical constraints. If there is a conflict, the score will be reduced. (4) Multi-configuration stability C4: This is used to characterize the degree of consistency of conclusions when performing multiple inferences on the same input under different inference configuration parameters or different inference constraint templates. The number of multiple inferences can be 2 to 5. The system can generate stability indicators using consistency ratio, fusion rules, or majority consistency principles. When the stability is lower than a preset threshold, backoff control or output verification flag is triggered.

[0084] In one implementation, the system can calculate a confidence score Conf based on a weighted fusion method. For example, Conf can be obtained by a weighted combination of C1, C2, C3, and C4, where each weight can be determined by preset empirical parameters or obtained through statistical learning from historical samples. When the confidence score Conf is lower than a preset threshold θ, the system triggers a fallback control.

[0085] In one implementation, rollback control includes a set of rollback strategies, which includes at least one or more of the following: supplementing or expanding the set of evidence fragments; rearranging the field expansion and alignment comparison order; enabling stricter structured output template constraints; adjusting the threshold or weight of conflict triggering rules to improve conflict detection sensitivity; and switching to a conservative judgment mode to avoid outputting high-risk conclusions. After adjusting the structure-enhanced input or inference constraints according to the rollback strategies, the system re-executes controlled consistency inference and records the rollback count, rollback strategy, input summary, output structure, evidence index, and timestamp to form an auditable log. When the rollback count reaches a preset upper limit but Conf is still below the threshold θ, the system outputs a review flag and retains the evidence chain and inference log to support subsequent manual review and audit traceability.

[0086] To demonstrate the technical effectiveness of the method in the recommendation consistency detection task, this embodiment constructs a test dataset based on multi-source clinical practice guidelines and sets different consistency analysis configurations. It compares the large language model analysis method driven solely by original text prompts with the analysis method of this embodiment, which combines "structured field constraints + candidate screening + consistency inference + confidence backoff". The experimental evaluation metrics include accuracy, macro-precision, macro-recall, and macro-F1 score. The results are as follows: Figure 6 , Figure 7 As shown, "the method of the present invention" indicates the implementation of the method described in this embodiment; the remaining comparison items are the baseline configurations under the same dataset and evaluation indicators.

[0087] The above results demonstrate that, under the same dataset and evaluation metrics, the method in this embodiment can improve the usability and stability of consistency relationship determination through the combination of structured field constraints and candidate recommendation pairing screening mechanism. At the same time, it enhances the interpretability of results through explicit field alignment and inference basis output, making it suitable for engineering application scenarios oriented towards the integration of multi-source clinical guideline knowledge and intelligent decision support.

[0088] Example 2 Based on the same inventive concept, this embodiment discloses a cross-guideline recommendation consistency detection system based on structured features, including: The clinical semantic parsing module is used to collect clinical practice guideline texts from multiple sources, parse and process the recommendations in the clinical practice guideline texts, and generate structured clinical semantic data that includes the applicable population, intervention measures, control protocols, outcome indicators, and recommendation strength or evidence level. The candidate recommendation pair generation module is used to combine recommendations from different sources in pairs based on the structured clinical semantic data to construct recommendation pairs, generate a structured feature representation for each recommendation pair, and generate a set of candidate recommendation pairs with potential relationships based on the structured feature representations of the recommendation pairs. The controlled consistency reasoning module is used to perform consistency analysis on pairs of recommendations in the selected candidate recommendation pair set, under the constraints of structured semantic information. It outputs the consistency relationship judgment results between recommendations and their corresponding analysis basis, and performs credibility assessment and rollback control on the output.

[0089] Specifically, the semantic element parsing module also retains the original text evidence fragments and source tracing information corresponding to each semantic element; the candidate recommendation pair generation module is used to perform field alignment and multi-dimensional similarity calculation based on field data, construct a structured feature representation of the recommendation pair, and generate a set of candidate recommendation pairs with potential semantic relevance through candidate screening and risk gating strategies to control the input scale of subsequent consistency inference; the controlled consistency inference module performs structural enhancement fusion of the fields of the candidate recommendation pair with the original text evidence fragments, and performs consistency inference analysis under the constraints of semantic elements, outputting the consistency relationship type, difference points, and evidence index.

[0090] The system also includes a credibility assessment and rollback control module and a result management and interaction module. The credibility assessment and rollback control module is used to assess the credibility of the consistent inference output and trigger rollback control when the credibility is insufficient. After reconstructing the structure to enhance the input or adjusting the field expansion order, consistent inference is executed again to improve output stability and engineering reliability. The result management and interaction module is used to structure, index, and visualize the consistency relationship determination results, and provides conflict prompts, evidence chain tracing, and manual review support interfaces.

[0091] In practical implementation, the cross-guideline recommendation consistency detection system based on structured features and credibility backtracking of this invention can be divided into three layers: a guideline data parsing layer, a feature calculation and inference service layer, and a user interface layer. The system layer architecture is as follows: Figure 8 As shown, the guide data parsing layer is used to implement the semantic element parsing module; the feature calculation and reasoning service layer is used to implement the candidate recommendation pair generation module, the controlled consistency reasoning module, and the credibility assessment and rollback control module; the user interface layer is used to implement the result management and interaction module, providing external support for querying, displaying, and reviewing.

[0092] Specifically, the guideline data parsing layer is used to automatically collect and preprocess clinical practice guideline texts from multiple sources, and to identify, locate, extract semantic elements, and standardize recommendations. The field-based data output by this layer includes at least: a unique identifier for the recommendation, a guideline source identifier, the original text of the recommendation, and fields for Population, Intervention, Comparison, Outcome, and Level & Grade, as well as an index of the corresponding original text evidence fragments. To ensure cross-source consistency analysis capabilities, the guideline data parsing layer also includes terminology standardization and field integrity verification mechanisms. These mechanisms are used to normalize synonyms, abbreviations, and differences in grading systems across different guidelines, and to mark or re-parse recommendations with missing key fields or obvious logical conflicts.

[0093] The feature calculation and inference service layer is used to realize the core capabilities of candidate recommendation pair generation and consistency analysis. The candidate recommendation pair generation module first aligns the recommendation fields from different sources, and then calculates multi-dimensional similarity features such as population feature similarity, intervention measure similarity, outcome indicator similarity, and original text semantic similarity, and combines them to form a structured feature representation of the recommendation pair. In the candidate screening stage, the module can use a comprehensive scoring strategy or a gating strategy to generate a set of candidate recommendation pairs. The gating strategy includes at least: directly filtering low-association combinations when the Population or Intervention fields do not overlap at all; and elevating combinations to high priority candidates when potential conflict clues appear in the Level & Grade or original text evidence fragments, so as to ensure that conflict-type combinations can enter subsequent inference analysis.

[0094] The controlled consistency inference module is used to analyze candidate recommendation pairs as consistency objects, performing consistency relationship determination under semantic element constraints. In specific implementation, the module merges the fields of the candidate recommendation pairs with original evidence fragments to form structure-enhanced input, and sets inference constraints, requiring the inference process to compare each field and explicitly indicate the source of consistency or difference. The inference output includes at least a consistency relationship type, which includes at least consistent, inconsistent, and irrelevant relationships; it also includes a summary of differences corresponding to the relationship type and an evidence index reference to support the interpretability and traceability of the results.

[0095] The credibility assessment and rollback control module is used to improve the stability of inference output. This module assesses the credibility of consistent inference output, based on at least the following factors: whether fields are fully referenced, whether the inference conclusion is consistent with the field content, whether the conclusion is stable under multiple inferences using the same input, and whether the evidence fragments are sufficiently referenced. When the credibility score is lower than a preset threshold, a rollback control mechanism is triggered, including but not limited to: reconstructing the structure to enhance the input, adjusting the field expansion order, supplementing evidence fragments, or changing the priority of candidate pairs before re-executing consistent inference. If a stable conclusion cannot be obtained after multiple rollbacks, the output needs to be manually reviewed and marked, and the candidate pairs, field differences, evidence fragment indexes, and inference logs are written to the results management module for manual review.

[0096] The user interface layer provides the ability to display and interact with the consistency relationship determination results. The results management and interaction module stores the consistency relationship determination results in a structured manner, supporting queries and aggregated displays by guideline source, disease topic, population stratification, intervention category, and outcome indicator. It provides conflict alerts and summaries of differences for recommendations determined to be inconsistent or high-risk conflicts. It also supports one-click navigation from the results page back to the original evidence fragment and its corresponding guideline section for auditing and review. In one implementation, this layer also provides a manual review and annotation interface for writing expert review results back to the system as a basis for subsequent adjustments to thresholds, gating strategies, or field mapping rules.

[0097] In summary, this invention combines semantic element parsing of cross-guideline recommendations with a structured feature-driven candidate recommendation pair screening mechanism, and performs controlled consistency inference under semantic element constraints. It also incorporates credibility assessment and a backtracking control loop to achieve stable determination and traceable output of consistency relationships among recommendations from multiple clinical practice guidelines. This provides interpretable and controllable intelligent support for clinical guideline knowledge integration, guideline version comparison, evidence-based conflict identification, and intelligent decision support systems. This invention is applicable to, but not limited to, clinical guideline knowledge base construction, guideline iteration difference analysis, cross-institutional guideline conflict alerts, and rule verification and updating for decision support systems.

[0098] Since the system in Embodiment 2 of this invention is the same system used in the cross-guideline recommendation consistency detection method based on structured features in Embodiment 1, those skilled in the art can understand the specific structure and variations of this system based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All systems used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0099] Example 3 Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a cross-guideline recommendation consistency detection method based on structured features, as described in Embodiment 1.

[0100] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the cross-guideline recommendation consistency detection method based on structured features in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0101] Example 4 Based on the same inventive concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0102] Since the computer device described in Embodiment 4 of this invention is the same computer device used to implement the cross-guideline recommendation consistency detection method based on structured features in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0103] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A method for detecting cross-guideline recommendation consistency based on structured features, characterized in that, include: We collect clinical practice guideline texts from multiple sources, analyze and process the recommendations in the clinical practice guideline texts, and generate structured clinical semantic data that includes applicable populations, intervention measures, control protocols, outcome indicators, and recommendation strength or evidence levels. Based on the structured clinical semantic data, recommendations from different sources are combined in pairs to construct recommendation pairs, and a structured feature representation is generated for each recommendation pair. Based on the structured feature representation of the recommendation pairs, a set of candidate recommendation pairs with potential relationships is generated. Using the candidate recommendation pairs obtained from the screening as the consistency analysis objects, consistency reasoning analysis is performed under the constraints of structured semantic information. The results of the consistency relationship judgment between the recommendations and their corresponding analysis basis are output, and the credibility assessment and rollback control of the output are performed.

2. The method for cross-guideline recommendation consistency detection based on structured features as described in claim 1, characterized in that, Collect clinical practice guideline texts from multiple sources, and analyze and process the recommendations in the clinical practice guideline texts, including: Obtain the original text of the recommendations in the clinical practice guidelines and the source identification information of the guidelines to which they belong; Automatically identify and locate recommendations from clinical practice guideline texts; Semantic analysis was performed on the recommendations to extract the applicable population, intervention measures, control group, outcome indicators, and recommendation strength or level of evidence. The extracted clinical semantic information is standardized, mapping similar semantic information in different forms of expression into a unified structured field representation.

3. The method for cross-guideline recommendation consistency detection based on structured features as described in claim 1, characterized in that, For each recommendation, a structured feature representation is generated, including: Based on the applicable population field, semantic similarity of the recommendations is calculated on the population feature dimension. Based on the intervention measures field, semantic similarity is calculated for the recommendations in the dimension of intervention strategies to measure the similarity between two recommendations in terms of intervention methods; Based on the outcome index field, the semantic similarity of the recommendations in the clinical outcome dimension is calculated to measure whether the clinical outcomes or goals of the two recommendations are consistent. Based on the text content of the recommendations, the semantic correlation between the recommendations is analyzed to capture supplementary semantic information beyond the structured fields; The analysis results from the above dimensions are combined to form a structured feature representation of the recommendation pair.

4. The method for cross-guideline recommendation consistency detection based on structured features as described in claim 3, characterized in that, Based on the target audience field, semantic similarity calculation is performed on the recommendations across the audience feature dimension, including: in, For consistency judgment based on disease codes, a perfect match is recorded as 1, otherwise it is recorded as 0; The similarity of comorbidities is calculated using the multiple union-intersection ratio; Similarity is used for matching based on demographic information; , , These are the corresponding weights, used to control the contribution of different factors. This refers to semantic similarity based on the dimension of population characteristics.

5. The method for cross-guideline recommendation consistency detection based on structured features as described in claim 1, characterized in that, Based on the structured feature representation of recommendation pairs, a set of candidate recommendation pairs with potential relationships is generated, including: Based on preset scoring or screening rules, a weighted scoring model is constructed using logistic regression to comprehensively evaluate the multidimensional feature representations of recommendation pairs. The weighted scoring model is as follows: , ( ) is the recommendation consisting of recommendation i and recommendation j. The corresponding feature vector is w represents the feature weight vector, and σ(·) is the Sigmoid function. This is the overall scoring result; The comprehensive evaluation results are compared with the preset conditions, and the recommended opinion pairs that meet the conditions are selected to form a set of candidate recommended opinion pairs.

6. The method for cross-guideline recommendation consistency detection based on structured features as described in claim 1, characterized in that, Using the selected candidate recommendation pairs as the consistency analysis objects, consistency reasoning analysis is performed under the constraints of structured semantic information. The results of the consistency relationship determination between the recommendations and their corresponding analysis basis are output, including: The structured clinical semantic fields of candidate recommendation pairs, field-level evidence fragment indexes, and the original content of the recommendation opinions are combined to form a structured semantic input for consistency analysis; Under the structured semantic input constraints, the relationship type between candidate recommendation pairs is determined, and the relationship type includes at least consistent, inconsistent, and irrelevant; Generate the analytical basis or analytical path corresponding to the consistency relationship determination results.

7. The method for cross-guideline recommendation consistency detection based on structured features as described in claim 1, characterized in that, The method also includes credibility assessment and rollback control of the consistency analysis results, specifically: The credibility of the consistency determination results is assessed based on the use of structured fields, sufficiency of evidence, consistency of analysis path, and stability of analysis conclusions during the analysis process. When the credibility assessment result is lower than the preset condition, a rollback process is triggered, and the consistency analysis of the candidate recommendation pair is re-executed, or the analysis is performed again after adjusting the prompt constraints, supplementing evidence fragments, rearranging the field expansion order, or switching to the conservative judgment mode.

8. A cross-guideline recommendation consistency detection system based on structured features, characterized in that, include: The clinical semantic parsing module is used to collect clinical practice guideline texts from multiple sources, parse and process the recommendations in the clinical practice guideline texts, and generate structured clinical semantic data that includes the applicable population, intervention measures, control schemes, outcome indicators, and recommendation strength or evidence level. The candidate recommendation pair generation module is used to combine recommendations from different sources in pairs based on the structured clinical semantic data to construct recommendation pairs, generate a structured feature representation for each recommendation pair, and generate a set of candidate recommendation pairs with potential relationships based on the structured feature representations of the recommendation pairs. The controlled consistency reasoning module is used to perform consistency analysis on pairs of recommendations in the selected candidate recommendation pair set, under the constraints of structured semantic information. It outputs the consistency relationship judgment results between recommendations and their corresponding analysis basis, and performs credibility assessment and rollback control on the output.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a cross-guideline recommendation consistency detection method based on structured features as described in any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a cross-guideline recommendation consistency detection method based on structured features as described in any one of claims 1 to 7.

Images