Clinical pathway deviation detection method and system

Abstract

The application discloses a kind of clinical path deviation detection method and system, it is related to medical information field, the present application takes single-disease standard clinical path cognitive knowledge graph as standard path knowledge base, and introduces diagnosis and treatment behavior node, time sequence relationship, cause-effect relationship, mutual exclusion relationship, substitution relationship and multilevel weight into deviation detection process;Meanwhile, the actual diagnosis and treatment behavior sequence is extracted from electronic medical record using large language model, and the correspondence between the actual diagnosis and treatment behavior and the standard path node is identified through semantic matching and path alignment mechanism, and then the automatic identification, quantitative evaluation and interpretable output of multiple deviations such as step missing, sequence disorder, replacement, addition, repetition are realized.

Description

Technical Field

[0001] This invention relates to the field of medical informatics, specifically to a method and system for detecting clinical pathway deviations. Background Technology

[0002] With the continuous improvement of medical informatization and intelligentization, clinical pathways have become an important tool for medical institutions to standardize diagnosis and treatment behaviors, control differences in treatment, improve medical quality, and support single-disease management. Single-disease clinical pathways typically revolve around the entire process of diagnosis and treatment of a specific disease, standardizing the organization of patient care from admission assessment, examination and testing, diagnosis confirmation, treatment, medication management, nursing assessment to discharge follow-up. In actual implementation, clinical pathways not only require that diagnosis and treatment behaviors conform to predetermined steps and stages, but also need to meet various constraints such as examination results, diagnostic conclusions, risk assessments, complication status, contraindications, and medication conditions. If the actual diagnosis and treatment process deviates from the standard pathway, it may manifest as missing key nodes, abnormal treatment sequence, execution without meeting conditions, duplicate examinations, abnormal medication use, pathway delays, premature pathway implementation, or unrecorded pathway variations. These situations are generally referred to as clinical pathway deviation.

[0003] In the actual management of medical institutions, clinical pathway deviation detection is a crucial link in clinical pathway management, medical quality control, medical insurance review, and supervision of treatment behavior. Traditional clinical pathway deviation identification relies heavily on manual sampling, rule configuration, form comparison, or fixed indicator statistics. While it can identify some clear and simple pathway execution problems, it struggles to comprehensively identify deviations implicit in complex treatment processes. With the continuous accumulation of data such as electronic medical records, medical orders, examination and test results, nursing documents, cost details, and pathway execution records, clinical pathway deviation detection has gradually acquired the foundation for datafication, automation, and intelligence. However, single-disease clinical pathways have obvious staged, conditional, and procedural characteristics, and actual treatment data is characterized by diverse sources, complex structures, inconsistent expressions, and overlapping time series, making deviation detection unsuitable for simple field comparisons or static rule judgments. Knowledge graph technology can organize medical knowledge such as diseases, diagnoses, examinations, treatments, medications, nursing, and assessments into nodes and relationships, providing a technological foundation for the structured expression of pathway knowledge.

[0004] Furthermore, cognitive knowledge graphs constructed for single-disease clinical pathways can not only express general medical entity relationships, but also temporal, causal, conditional, mutually exclusive, substitution, stage-based, and pathway variation relationships between diagnostic and treatment behaviors. Based on this type of cognitive knowledge graph, actual patient diagnostic and treatment behaviors can be matched, aligned, and reasoned with standard pathway knowledge, thus providing a new technical approach for detecting clinical pathway deviations.

[0005] To achieve medical informatization, existing technology 1 (Chinese Patent "Dynamically Optimized Intelligent Clinical Pathway, Disease Group Pathway Automatic Modification Method, Device and Storage Medium", Publication No. CN121075699A) mainly addresses the problem that traditional clinical pathway systems cannot respond to changes in individualized patient data in real time, proposing a dynamically optimized intelligent clinical pathway or disease group pathway automatic modification method. Its basic idea is: by constructing a medical knowledge graph, integrating knowledge such as treatment guidelines, medical insurance rules, disease group rules, and hospital rules; collecting dynamic data such as patient vital signs and examination results in real time; when the monitored data deviates from a preset threshold, a reinforcement learning-driven decision engine automatically generates a path modification plan, which is reviewed by a doctor before execution; the modified path plan can be saved as evidence in the path archive, thereby shortening the path adjustment response time and continuously optimizing the path library during ongoing use.

[0006] The existing technology has the following problems: (1) The key to this scheme is the "automatic modification" and "dynamic optimization" of the path, rather than being specifically designed for clinical path deviation detection.

[0007] The core objective of this approach is to adjust existing clinical pathways or disease group pathways as patient dynamic data changes, enabling the pathways to adapt to individualized patient changes. Its focus is on "how the pathway is modified" and "how the pathway is continuously optimized," rather than systematically identifying whether the patient's actual treatment behavior deviates from the standard pathway. Therefore, this approach lacks a dedicated process for detecting clinical pathway deviations, such as identifying missing key nodes, abnormal execution order, execution due to unmet conditions, co-occurrence of mutually exclusive behaviors, and unrecorded pathway variations.

[0008] (2) Although the scheme involves a knowledge base or medical knowledge graph, it does not disclose cognitive relationship modeling for deviation detection.

[0009] The knowledge base in this scheme primarily covers diseases, drugs, surgeries, medical insurance rules, disease group rules, and hospital rules, used to determine whether data changes affect the pathway and to assist in generating pathway modification plans. However, it does not explicitly disclose the modeling of diagnostic and treatment behaviors in single-disease clinical pathways as a cognitive knowledge graph structure with temporal relationships, causal relationships, conditional triggering relationships, mutually exclusive relationships, substitution relationships, stage affiliation relationships, and pathway variation relationships. In particular, the scheme does not systematically disclose relationship types that deviate from the required detection criteria, such as "whether a certain diagnostic and treatment behavior violates preconditions," "whether two diagnostic and treatment behaviors are mutually exclusive," and "whether a certain examination result causally triggers subsequent treatment."

[0010] (3) The scheme lacks a precise alignment mechanism between the patient’s actual diagnosis and treatment behavior and the standard path nodes.

[0011] Clinical pathway deviation detection requires matching the patient's actual medical actions, including medication administration, examinations and tests, treatments, nursing records, and medication use, against the nodes, stages, conditions, and constraints of the standard pathway. This particular solution primarily determines whether pathway modification is needed based on patient data updates, but it does not disclose in detail how to extract actual medical actions from multi-source medical data, nor how to semantically, temporally, and stage-wise align the actual medical trajectory with the standard pathway nodes. Therefore, it is insufficient to directly support sophisticated detection of pathway execution deviations.

[0012] (4) The scheme lacks structured output of path deviation type, risk level and deviation location.

[0013] In clinical pathway deviation detection scenarios, it is necessary not only to determine whether an anomaly exists, but also to output the deviation type, deviation node, relevant pathway stage, involved medical behaviors, violated pathway constraints, and risk level. While this solution can generate pathway modification plans when patient data deviates from a preset threshold, it does not explicitly disclose the categorization and structured representation of deviation results, making it difficult to meet the needs of quality control personnel, medical insurance reviewers, and clinicians to review the causes, locations, and risks of deviations.

[0014] (5) The scheme is not good at tracing and interpreting evidence of deviations from the test results.

[0015] The scheme mentions that path modification plans can be saved as evidence, but this evidence mainly records the results of path modification and does not disclose the corresponding graph reasoning links, standard path basis, actual patient treatment evidence, and rule triggering basis for each deviation judgment. For clinical path deviation detection, the detection results need to explain "why a deviation occurred," "which standard node was used," "what the patient's actual implementation was," and "which type of relational constraint was violated." The scheme does not provide a complete explanation and traceability mechanism for deviation detection results.

[0016] (6) The dynamic optimization bias path library update of the scheme does not highlight the closed-loop management of single disease deviation detection.

[0017] This scheme summarizes path optimization strategies by statistically analyzing the number of times the path is applied, modified, modified nodes, and executed, primarily serving the continuous optimization of the path library. This application, however, focuses on identifying deviations, detecting anomalies, tracing evidence, and providing feedback on results during the actual patient treatment process based on a single-disease clinical path cognitive knowledge graph. While both involve dynamic changes in the path, this scheme lacks a closed-loop detection mechanism encompassing "standard path knowledge graph—actual patient treatment trajectory—deviation detection—evidence interpretation—quality control feedback."

[0018] Prior art two (Chinese patent "System and Method for Identifying Deviations in Clinical Pathway Execution", publication number CN102651051A) mainly addresses the problem of deviations from standard clinical pathways during execution, proposing a system and method for identifying such deviations. Its basic idea is: pre-store the semantic relationships between clinical pathways and activities; receive user-inputted actual execution activities; determine whether the actual activity matches the activities defined in the clinical pathway based on the semantic relationships; further determine whether it meets prerequisites and execution time requirements; when the actual execution activity is inconsistent with the clinical pathway definition, the system generates a clinical pathway execution deviation report.

[0019] The existing technology has the following problems: (1) This scheme relies on user input to perform actual activities and lacks a mechanism to automatically extract treatment trajectories from multi-source medical data.

[0020] This solution primarily receives the activity name, execution time, and clinical pathway name input by healthcare professionals through a receiving component, and then identifies deviations. In real-world medical scenarios, patient treatment behaviors are typically scattered across systems such as electronic medical records, doctor's orders, test results, nursing documentation, expense details, and pathway execution records. If the solution still relies heavily on user-inputted activities, it is susceptible to issues related to the completeness, timeliness, and accuracy of the input. This solution does not disclose in detail how to automatically extract actual patient treatment behaviors from multi-source heterogeneous medical data, nor does it establish an automatic mechanism for constructing patient treatment trajectories.

[0021] (2) The scheme is based on pre-stored clinical pathways and semantic matching, but does not disclose the knowledge graph structure of single-disease clinical pathways.

[0022] While this solution includes clinical pathway storage and semantic relationship storage components, its focus is on storing the semantic relationships between standard clinical pathways, required activities, and medical activities. It does not disclose the construction of a single-disease clinical pathway into a cognitive knowledge graph that includes diagnostic and treatment behavior units, pathway stages, knowledge points, evidence sources, relationship weights, and version information. For the single-disease clinical pathway scenario addressed in this invention, this solution lacks a dedicated graph-based organizational structure for the entire diagnostic and treatment process of a specific disease.

[0023] (3) The semantic relationships in this scheme are mainly used for activity matching, and do not cover the complex cognitive relationships required for deviation detection.

[0024] This approach emphasizes using equivalence and inclusion relationships within the medical terminology system to determine whether actual activities match the expected activities. Its semantic relationships primarily serve the question of whether activities with different names but similar meanings match. However, clinical pathway deviation detection requires not only determining whether activities match, but also identifying relationships such as causal dependence, conditional triggering, mutual exclusion, alternative paths, stage affiliation, and path variation between diagnostic and treatment behaviors. This approach does not systematically disclose cognitive relationship modeling methods for deviation detection, including causal relationships, mutual exclusion relationships, substitution relationships, and conditional triggering relationships.

[0025] (4) This scheme focuses on matching individual activities with standard activities and lacks path reasoning for the complete diagnosis and treatment process.

[0026] This scheme primarily focuses on determining whether user input activities match the required activities in the clinical pathway, whether prerequisites are met, and whether time requirements are satisfied. Its deviation identification leans more towards verifying individual or localized activities, and it does not disclose how it performs pathway stage identification, key node link analysis, pre- and post-procedure dependency judgment, and multi-step path reasoning based on the patient's complete treatment trajectory. Therefore, for deviations that cross stages, nodes, and data sources in complex cases, this scheme cannot directly provide a complete graph reasoning link.

[0027] (5) The scheme lacks evidence binding for deviations from the results and an interpretable output mechanism.

[0028] Clinical pathway deviation detection results need to explain "which pathway node the deviation occurred at," "which type of pathway constraint was violated," "what standard pathway knowledge was used as the basis," and "where the patient's actual execution evidence came from." This solution primarily outputs a clinical pathway execution deviation report, but it does not explicitly disclose the mechanism for linking deviation results to the original electronic medical record, medical order execution records, examination and test results, nursing records, standard pathway clauses, and the graph reasoning chain. Therefore, the interpretability of its deviation report in clinical review, quality control audit, and medical insurance review scenarios remains insufficient.

[0029] In summary, the following problems still exist in the process of deviation detection based on single-disease clinical pathways: (1) Clinical pathway deviations are complex and difficult to fully identify using simple rules: Deviation from a single disease's clinical pathway encompasses not only whether a particular examination, treatment, or medication step was performed, but also the rationality of the execution sequence, whether the execution conditions were met, the existence of mutually exclusive behaviors, the availability of alternative pathways, the occurrence of pathway variations, and the availability of corresponding evidence. Relying solely on fixed rules or manual checks makes it difficult to comprehensively identify issues such as temporal deviations, conditional deviations, causal chain deviations, mutually exclusive conflicts, and unrecorded variations in the diagnostic and treatment process.

[0030] (2) The actual diagnosis and treatment behavior of patients is difficult to accurately align with the knowledge of standard pathways: Patient's actual medical data is scattered across different systems, including electronic medical records, doctor's orders, examination and test results, nursing records, expense details, and pathway forms. These different data sources differ in time granularity, field structure, terminology, and recording methods. The same medical behavior may appear under different names, codes, or text descriptions, making it difficult to accurately match the actual medical behavior with the medical nodes, pathway stages, and triggering conditions in the standard pathway, thus affecting the accuracy of deviation detection results.

[0031] (3) Deviations in test results lack interpretability and supporting evidence: Clinical pathway deviation assessment involves physician treatment behavior, individual patient differences, causes of pathway variation, and medical quality evaluation. The test results must not only indicate "whether a deviation occurred," but also specify "at which node the deviation occurred, which type of pathway relationship was violated, which standard pathway knowledge was used as a basis, and what the corresponding actual clinical evidence is." Test results lacking evidence tracing and graph-based reasoning support are insufficient for clinical review, quality control analysis, and management decision-making.

[0032] (4) Deviation detection is difficult to adapt to the dynamic changes in pathway knowledge and real-world diagnostic and treatment data: Clinical pathways for single diseases are constantly evolving with adjustments to treatment guidelines, optimization of hospital processes, updates to medical evidence, and the accumulation of real-world data. If deviation detection rules and standard pathway knowledge are not updated accordingly, problems such as outdated detection criteria, increased false positives and false negatives, and difficulty in tracing historical results can easily arise. Therefore, it is necessary to establish a deviation detection mechanism that can combine dynamic updates of knowledge graphs, version management, and evidence backtracking. Summary of the Invention

[0033] To address the aforementioned shortcomings in existing technologies, the clinical pathway deviation detection method and system provided by this invention solve the problems of high reliance on manual review, crude rule judgment methods, difficulty in automatically extracting actual medical behaviors, difficulty in accurately aligning standard pathways with actual medical behaviors, insufficient detail in identifying deviation types, difficulty in quantifying the degree of deviation, and lack of interpretive basis for detection results in existing clinical pathway deviation detection methods.

[0034] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for detecting clinical pathway deviation is provided, comprising the following steps: Obtain a single-disease standard clinical pathway cognitive knowledge graph; the single-disease standard clinical pathway cognitive knowledge graph is used to store the standard diagnosis and treatment process and its cognitive relationships for the target disease, the nodes of the clinical pathway cognitive knowledge graph are the diagnosis and treatment behaviors in the standard clinical pathway, and the edges of the clinical pathway cognitive knowledge graph are the relationships between the nodes; The actual medical treatment behaviors of patients are extracted from the patient data and sorted according to the execution time to form a sequence of actual medical treatment behaviors; Matching and mapping actual diagnosis and treatment behavior sequences with standard clinical pathway cognitive knowledge graphs to generate correspondences between actual diagnosis and treatment behaviors and standard clinical pathway nodes; Based on the correspondence between actual clinical behavior and standard clinical pathway nodes, and combined with the node importance weight, relationship type, relationship strength weight, time window, triggering conditions and / or exceptional evidence in the standard clinical pathway cognitive knowledge graph, it is determined whether the actual clinical behavior deviates from the standard clinical pathway.

[0035] A system for implementing a clinical pathway deviation detection method is provided, comprising: The data layer is used to obtain a cognitive knowledge graph of standard clinical pathways for single diseases; it extracts actual diagnosis and treatment behaviors from patient data and sorts them by execution time to form a sequence of actual diagnosis and treatment behaviors; The alignment layer is used to match and correspond actual diagnosis and treatment behavior sequences with standard clinical pathway cognitive knowledge graphs, generating correspondences between actual diagnosis and treatment behaviors and standard clinical pathway nodes; The detection layer is used to determine whether the actual diagnosis and treatment behavior deviates from the standard clinical pathway based on the correspondence between the actual diagnosis and treatment behavior and the nodes of the standard clinical pathway, combined with the node importance weight, relationship type, relationship strength weight, time window, triggering conditions and / or exceptional evidence in the cognitive knowledge graph of the standard clinical pathway. The application layer is used to provide output deviation reports, risk warnings, quality analysis, and path optimization suggestions.

[0036] The beneficial effects of this invention are as follows: 1. This invention uses a single-disease standard clinical pathway cognitive knowledge graph as a knowledge base, and combines a large language model to extract, sort and semantically match the actual diagnosis and treatment behaviors in the patient's electronic medical record. Through processes such as path alignment, deviation identification, deviation classification, deviation degree calculation, report generation and feedback optimization, it realizes the automation, refinement, quantification and interpretability of clinical pathway deviation detection.

[0037] 2. This invention transforms the traditional clinical pathway deviation detection process, which relies on manual sampling and comparison, into a data parsing, path alignment, and deviation judgment process that is automatically completed by the system, significantly improving the automation level of clinical pathway deviation detection.

[0038] Automatic parsing of medical record data: The electronic medical record parsing module automatically reads patient medical records, doctor's orders, examination and test results, nursing records and discharge records, etc., reducing the workload of manually flipping through medical records item by item.

[0039] Automatic extraction of diagnostic and treatment behaviors: Using a large language model, actual diagnostic and treatment behaviors such as examination, diagnosis, treatment, medication, evaluation, and discharge are extracted from unstructured text and structured records.

[0040] Automatic path comparison: The system automatically matches the patient's actual treatment behavior sequence with the standard clinical pathway knowledge graph, replacing manual node-by-node verification.

[0041] Automatic output of test results: The system automatically generates deviation type, deviation location, deviation degree score and improvement suggestions to improve the efficiency of quality control work.

[0042] 3. This invention no longer uses standard clinical pathways merely as text, tables, or process descriptions, but instead constructs them into a cognitive knowledge graph that is computable, matchable, and reasonable, thereby improving the structured expression ability of standard clinical pathway knowledge.

[0043] Node-based path representation steps: Transform standard path content such as diagnosis, examination, treatment, medication, assessment, and events into graph nodes.

[0044] Relational expression path logic: Transform the temporal, causal, mutually exclusive, substitution, and conditional triggering relationships between diagnostic and treatment behaviors into graph edges.

[0045] Weighted representation of node importance: Assign importance weights to nodes on different paths to distinguish between critical nodes, general nodes, and optional nodes.

[0046] Intensified expression of relational constraints: Set relational strength weights for graph relations to determine the impact of temporal constraints, causal dependencies, mutual exclusion substitutions, and other relations on the degree of deviation.

[0047] 4. This invention utilizes a large language model to perform semantic parsing on electronic medical record text, which can handle problems such as free text, non-standard expressions, abbreviations, aliases, and compound sentences that exist in actual medical records, thereby improving the accuracy of identifying actual diagnosis and treatment behaviors in electronic medical records.

[0048] Supports unstructured text parsing: It can identify actual medical treatment behaviors from free text such as admission records, medical records, and discharge summaries.

[0049] Supports execution time extraction: It can extract the time when the diagnosis and treatment behavior occurs and form a sequence of actual diagnosis and treatment behaviors in chronological order.

[0050] Supports behavior type recognition: It can distinguish between different types of behaviors such as examination, diagnosis, treatment, medication, assessment, care, and events.

[0051] Supports binding of source evidence: It can link the extracted actual diagnosis and treatment behavior with original medical record fragments, medical order records or examination and test results, which facilitates subsequent traceability.

[0052] 5. This invention solves the problem of inconsistency between the actual description of diagnosis and treatment behavior and the standard path node name by means of path alignment and semantic matching mechanism, thereby improving the accuracy of pre-matching for deviation detection.

[0053] Exact match: Directly match behaviors and nodes that are exactly the same in name, encoding, or standard terminology.

[0054] Semantic matching: For behaviors with different names but similar meanings, the matching relationship is calculated using a large language model or semantic similarity model.

[0055] Replacement matching: When the actual behavior replaces the standard node, the judgment is made in combination with the mutual exclusion relationship, substitution relationship or relationship strength weight in the graph.

[0056] Duplicate matching: Identifying situations where the same standard node is matched by multiple actual behaviors, providing a basis for duplicate deviation judgment.

[0057] Unmatched identification: Marking cases where standard nodes are not executed or where the actual behavior cannot correspond to standard nodes, providing a basis for identifying missing deviations and new deviations.

[0058] 6. This invention no longer outputs only a binary conclusion of "whether it deviates or not", but subdivides the deviation into multiple types, making it easier for clinical quality control personnel to understand the reasons for deviation and the direction of improvement, and realizing the refined identification of clinical pathway deviation types.

[0059] Missing Step Deviation: Identifying situations where essential steps specified in standard clinical pathways are not present in the actual clinical procedure sequence.

[0060] Disordered sequence: Identify situations where the actual sequence of medical procedures is inconsistent with the temporal relationship in the standard pathway.

[0061] Replacement deviation: Identify situations where the actual execution behavior replaces standard path nodes, and further determine whether it is a reasonable replacement.

[0062] New deviations: Identify situations where diagnostic and treatment behaviors not specified in the standard pathway are actually performed, and determine whether they may be beneficial additions, redundant additions, or risky additions.

[0063] Repeated deviation: Identify situations where the same standard path node is executed repeatedly or matched multiple times, and further determine whether it is a necessary repetition.

[0064] 7. This invention can distinguish between reasonable deviations and unreasonable deviations, avoiding the simplistic identification of all behaviors that are not entirely consistent with the standard path as abnormal.

[0065] Reasonable substitution identification: When actual medication, examination or treatment substitutes for standard path nodes, the system can judge its rationality based on substitution relationships, mutual exclusion relationships and relationship strength weights in the knowledge graph.

[0066] Beneficial new identification: Identify behaviors that are not specified in the standard pathway but may provide additional diagnostic and treatment value, avoiding simple judgment as serious deviations.

[0067] Necessary duplication identification: Distinguishing between repetitive behaviors that may be clinically necessary, such as follow-up examinations, efficacy assessments, and disease monitoring.

[0068] Abnormal behavior screening: Provide risk warnings for replacement, addition, and repetitive behaviors that lack medical basis, path support, or conflict with mutually exclusive relationships.

[0069] 8. This invention uses deviation type weight, node importance weight, and relationship strength weight to comprehensively calculate different deviation situations, thereby achieving a quantitative evaluation of the degree of deviation in clinical pathways.

[0070] Deviation types can be quantified: basic weight coefficients can be set for different deviation types such as missing steps, disordered order, replacement, addition, and duplication.

[0071] Node importance can be quantified: Based on the node importance weights in the standard clinical pathway knowledge graph, the different impacts of key steps and general steps on deviation scores are reflected.

[0072] Relationship strength weights can be quantified: Relationship strength weights such as temporal sequence, causality, mutual exclusion, and substitution are incorporated into the deviation score calculation to reflect the importance of path relationship constraints.

[0073] Deviation levels can be categorized: based on the overall deviation score, deviations are classified into minor deviations, moderate deviations, significant deviations, and severe deviations, which facilitates quality control ranking and risk classification.

[0074] 9. The deviation detection results generated by this invention not only include deviation conclusions, but also explain the basis for deviation, deviation location, involved nodes, actual behavior and source of evidence, thereby improving the interpretability and traceability of deviation detection results.

[0075] Clear deviation location: The report can pinpoint the specific node or stage in the standard path where the deviation occurred.

[0076] The deviation is clearly based on the report: the report can explain whether the deviation violates the node requirements, the temporal relationship, the causal relationship, the mutually exclusive relationship, or the substitution relationship.

[0077] Actual behavior is traceable: Every deviation can be linked to the patient's actual medical records, doctor's orders, examination and test results, or nursing records.

[0078] The standards are traceable: each deviation can be linked to the corresponding standard path node, graph relationship, and weight information.

[0079] The interpretation results are verifiable: Doctors, quality control personnel, or medical insurance reviewers can verify the results based on the evidence and reasoning in the report.

[0080] 10. This invention uses a feedback optimization mechanism to feed back deviation detection results, manual review results, and path execution statistics to the knowledge graph update module, thereby achieving continuous optimization of standard path knowledge.

[0081] Dynamic weight adjustment: Adjust the node importance weight and relationship strength weight based on the frequency of deviations, clinical review results, and path execution effectiveness.

[0082] Pathway node optimization: For newly added behaviors that occur frequently and have clinical rationale, it is recommended to include them as optional nodes in the standard pathway.

[0083] Substitution relationship optimization: For alternatives that have been recognized as reasonable in the long term, their substitution relationship can be strengthened or their deviation weight can be reduced.

[0084] Optimize execution reminders: For critical nodes that are frequently missing, path execution reminders or management interventions can be triggered.

[0085] Version evolution: Supports continuous updates of standard pathway maps as diagnostic and treatment guidelines, medical evidence, and real-world data change. Attached Figure Description

[0086] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a structural block diagram of the system. Detailed Implementation

[0087] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0088] Example 1: like Figure 1 As shown, this clinical pathway deviation detection method includes the following steps: S1. Obtain the cognitive knowledge graph of the standard clinical pathway for a single disease; the cognitive knowledge graph of the standard clinical pathway for a single disease is used to store the standard diagnosis and treatment process of the target disease and its cognitive relationships. The nodes of the cognitive knowledge graph of the clinical pathway are the diagnosis and treatment behaviors in the standard clinical pathway, and the edges of the cognitive knowledge graph of the clinical pathway are the relationships between the nodes. In the specific implementation process, this step responds to clinical pathway deviation detection requests by obtaining the unique identifier of the patient to be tested, the target disease, the detection time range, and the detection scenario. The detection scenario may include clinical pathway quality control, medical insurance review, medical record verification, departmental quality analysis, and pathway execution evaluation. In this embodiment, based on the patient identifier, data related to the patient is retrieved from business systems such as the hospital information system, electronic medical record system, medical order system, examination and testing system, and nursing system. Simultaneously, based on the target disease, the corresponding standard clinical pathway cognitive knowledge graph is invoked.

[0089] Patient data includes electronic medical records, medical orders, examination and test results, nursing records, and discharge summaries; each actual medical procedure includes the name of the medical procedure, the type of medical procedure, the time of execution, the source record, and evidence fragments.

[0090] For example, for a patient whose target disease is "community-acquired pneumonia", this embodiment can read the patient's admission record, initial medical record, medical order record, laboratory report, imaging examination record and discharge summary according to the patient's hospital number, and at the same time call the "community-acquired pneumonia" standard clinical pathway cognitive knowledge graph as the deviation detection standard.

[0091] Specifically, the single-disease standard clinical pathway cognitive knowledge graph is used to store the standard diagnosis and treatment process and cognitive relationships of the target disease. The nodes of the clinical pathway cognitive knowledge graph are the diagnosis and treatment behaviors in the standard clinical pathway, and the edges of the clinical pathway cognitive knowledge graph are the relationships between the nodes.

[0092] The diagnostic and treatment behaviors in the standard clinical pathway include diagnosis, examination, treatment, medication, assessment, nursing, discharge, and events; each node in the clinical pathway cognitive knowledge graph is assigned a node importance weight; The relationships between diagnostic and treatment behaviors include temporal relationships, causal relationships, mutually exclusive relationships, substitution relationships, conditional triggering relationships, stage attribution relationships, and path variation relationships; each edge of the clinical pathway cognitive knowledge graph is configured with a relationship strength weight. The weight is used to represent the importance of nodes, the relationship strength weight, and the basic severity of deviation types.

[0093] The structural design of this single-disease standard clinical pathway cognitive knowledge graph aims to transform standard clinical pathways, originally existing in the form of text, tables, or process descriptions, into a computable, matchable, and reasonable knowledge structure. When accessing the graph, all pathway nodes, node types, node importance weights, pathway stages, relationship types, relationship strength weights, applicable conditions, and version information under the target disease can be retrieved. In this step, the source and version information of the weights can be retrieved simultaneously. Node importance weights are calculated based on expert scores, guideline evidence levels, clinical key event relevance, and historical case impact; relationship strength weights are calculated based on the degree of strong constraint of the relationship type, evidence level, time window constraints, and historical consistency; deviation type weights are configured and calibrated based on quality control rules, medical insurance review focus points, clinical risk levels, and manual review conclusions. Each weight records the calculation source, calculation time, applicable disease, applicable version, and manual confirmation record to ensure traceability of subsequent deviation reports.

[0094] For example, in the community-acquired pneumonia pathway, the "48-72 hour efficacy assessment" node is a key post-treatment assessment node with high expert scores and guideline recommendations, so its importance weight can be higher than that of ordinary follow-up examination nodes; the temporal relationship between "antibiotic treatment - efficacy assessment" has a clear time window, so its relationship strength weight is higher than that of ordinary weak temporal relationships; "simultaneous occurrence of mutually exclusive drugs" is a high-risk deviation type, and its deviation type weight is higher than that of general repeated examinations.

[0095] It should be noted that while this embodiment constructs a single-disease standard clinical pathway cognitive knowledge graph based on clinical guidelines, expert consensus, in-hospital clinical pathway texts, pathway forms, and expert knowledge, other alternative solutions can also utilize data such as historical high-quality cases from hospitals, single-disease quality control indicators, DRG / DIP management rules, medical insurance review rules, clinical diagnosis and treatment guidelines, drug instructions, or departmental treatment routines as knowledge sources. Data from different sources, after standardization, can all be used to construct the standard pathway knowledge structure for the target disease.

[0096] While this embodiment utilizes a large language model to semantically parse the path text and extract diagnostic behavior nodes and path relationships, alternative solutions include manual annotation, rule templates, medical event extraction models, knowledge engineering methods, process modeling tools, or a hybrid approach of "manual review + machine extraction" to construct the graph. For path forms with relatively fixed structures, template extraction can be used; for path specifications with complex text and numerous conditions, a large language model can be used to assist in construction.

[0097] In this embodiment, the standard clinical pathway cognitive knowledge graph includes at least temporal relationships, causal relationships, and mutually exclusive relationships. Alternatively, relationship types can be added or adjusted according to the characteristics of the disease, such as conditional trigger relationships, substitution relationships, stage affiliation relationships, evidence support relationships, complication trigger relationships, contraindication relationships, and priority relationships. Any approach that can express the logical constraints between diagnostic and treatment behaviors and be used for deviation detection can be considered an extension or alternative to the relationship modeling method of this invention.

[0098] This embodiment assigns importance weights to standard clinical pathway nodes and relationship strength weights to pathway relationships. Alternatively, weights can be determined by expert scoring, or based on historical case execution frequency, deviation frequency, impact on clinical outcomes, impact on medical insurance payments, importance of quality control indicators, or a multi-factor weighted algorithm. Weights can be expressed as continuous values ​​between 0 and 1, or as discrete levels such as high, medium, and low.

[0099] As can be seen from the above description, the standard clinical pathway cognitive knowledge graph mentioned in this embodiment is not an ordinary medical knowledge graph, but a specialized graph constructed for the execution logic of clinical pathways and deviation detection requirements. The core features of this standard clinical pathway cognitive knowledge graph include: The diagnostic and treatment behaviors are used as nodes in the graph, and the node types include diagnosis, examination, treatment, medication, assessment, nursing, and events. The logical relationships between diagnostic and treatment behaviors are used as graph edges, and the types of relationships include temporal relationships, causal relationships, mutually exclusive relationships, substitution relationships, conditional triggering relationships, and stage affiliation relationships. Configure importance weights for path nodes to indicate the criticality of the node in the standard path for a single disease; Configure relationship strength weights for path relationships to represent the constraint strength of time sequence, causality, mutual exclusion, substitution and other relationships on deviation judgment; The nodes and relationships in the graph can be further linked to disease types, pathway stages, applicable conditions, sources of evidence, and version information.

[0100] S2. Extract the actual medical treatment behaviors from the patient data and sort them according to the execution time to form a sequence of actual medical treatment behaviors; In this embodiment, for structured data, field information such as medical order name, execution time, test item name, and examination report time can be read directly; for unstructured text, a large language model can be used for semantic parsing to identify treatment behaviors, behavior types, and execution times from texts such as medical records, admission records, and discharge summaries.

[0101] The actual medical actions extracted in this step should include at least the action name, action type, execution time, source data, and fragments of original evidence. Action types can include examination, diagnosis, treatment, medication, assessment, nursing care, and events.

[0102] For example, a pneumonia patient's medical record states: "On March 1, 2025, a complete blood count, C-reactive protein (CRP), procalcitonin (PCP), and chest CT scan were performed; on March 2, 2025, the patient was diagnosed with community-acquired pneumonia and given moxifloxacin intravenously; on March 4, 2025, a repeat complete blood count and CRP test were performed; the patient was discharged on March 5, 2025." This can be extracted to form the following sequence of actual medical procedures: complete blood count, CRP test, PCP test, chest CT scan, diagnosis of community-acquired pneumonia, moxifloxacin intravenous infusion, repeat complete blood count, repeat CRP test, discharge.

[0103] In this embodiment, the actual diagnosis and treatment sequence is mainly extracted from electronic medical records, medical orders, examination and test results, nursing records, and discharge summaries. Alternatively, data sources such as surgical anesthesia systems, imaging reporting systems, pathology systems, cost detailing systems, clinical pathway execution systems, medical insurance settlement systems, mobile nursing systems, and patient follow-up systems can be further integrated to obtain a more complete actual diagnosis and treatment trajectory.

[0104] In this embodiment, a large language model is preferentially used to extract diagnostic and treatment behaviors from electronic medical record text. As alternatives, named entity recognition models, medical event extraction models, rule matching, regular expressions, medical dictionary matching, sequence labeling models, classification models, or structured field mapping methods can also be used for extraction. For highly structured data such as medical orders, tests, and examinations, extraction can be performed directly through field mapping; for free text such as medical records and discharge summaries, large language models or event extraction models can be used for extraction.

[0105] In this embodiment, actual medical procedures are prioritized and sequenced according to their execution time. Alternatively, execution time can be derived from the time of prescription issuance, prescription execution, report release, record writing, examination completion, nursing record, or system operation log. For data with conflicting times, the final time can be determined based on business priority rules; for example, execution time can be prioritized, followed by report time, and then record time.

[0106] In this embodiment, diagnostic and treatment behaviors are primarily categorized into types such as diagnosis, examination, treatment, medication, assessment, and events. Alternatively, they can be further refined according to the application scenario into types such as laboratory tests, imaging examinations, pathological examinations, surgical procedures, nursing interventions, rehabilitation measures, risk assessments, discharge follow-ups, and cost items. The behavior type can be determined based on the hospital's internal dictionary, national standard codes, departmental business rules, or model classification results.

[0107] S3. Match and correspond the actual diagnosis and treatment behavior sequence with the standard clinical pathway cognitive knowledge graph to generate the correspondence between the actual diagnosis and treatment behavior and the standard clinical pathway nodes; This step requires first obtaining the standard clinical pathway cognitive knowledge graph corresponding to the target disease, as well as the patient's actual diagnosis and treatment behavior sequence. The standard clinical pathway cognitive knowledge graph includes information such as diagnosis and treatment behavior nodes, node types, node importance weights, temporal relationships, causal relationships, mutually exclusive relationships, substitution relationships, and relationship strength weights; the actual diagnosis and treatment behavior sequence includes behavior names, behavior types, execution times, and source evidence extracted from data such as electronic medical records, medical orders, examination and test results, and nursing records.

[0108] The purpose of the above operations is to prepare two types of data simultaneously: "standard pathway data" and "actual implementation data," enabling subsequent item-by-item matching around the same target disease. The standard pathway map is used to provide a basis for judgment, while the actual behavior sequence is used to reflect the patient's actual diagnosis and treatment process.

[0109] For example, for cases of community-acquired pneumonia, the standard pathway map may include nodes such as "complete blood count", "C-reactive protein test", "procalcitonin test", "chest X-ray", "etiological diagnosis", "antibiotic treatment", "48-72 hour efficacy assessment", and "discharge"; the actual treatment sequence may include behaviors such as "complete blood count", "C-reactive protein test", "procalcitonin test", "chest CT scan", "moxifloxacin intravenous infusion", "complete blood count re-examination", and "discharge".

[0110] In this embodiment, the actual diagnosis and treatment behavior sequence is first standardized and arranged in chronological order to form an actual behavior sequence that can be compared with the standard clinical pathway cognitive knowledge graph (i.e., the standardized actual diagnosis and treatment behavior sequence). This operation can also be performed in step S2.

[0111] It should be noted that the standardization process includes cleaning up behavior names, unifying medical terminology, labeling behavior types, parsing execution time, binding source evidence, and arranging by time sequence. For situations where the same behavior has different descriptions, they can be mapped to a unified name; for multiple behaviors occurring on the same day, they can be arranged according to the order of recording time, doctor's order time, reporting time, or pathway stage.

[0112] This step reduces the impact of inconsistencies in the expression of medical record text, doctor's orders, and examination item names on matching results, avoiding missed matches due to name differences. The standardized behavioral sequence should retain original evidence so that subsequent test reports can be traced back to the original medical record or doctor's order.

[0113] For example, in medical records, "blood count," "complete blood count," and "blood cell analysis" can be grouped under "complete blood count examination"; "moxifloxacin intravenous drip" and "moxifloxacin intravenous infusion" can be grouped under "moxifloxacin medication"; and "chest CT scan" and "chest CT examination" can be grouped under "chest CT examination." After standardization, the actual behavior sequence is formed according to the date.

[0114] In practice, candidate standard nodes can be selected from the standard clinical pathway cognitive knowledge graph based on the actual behavior type, target disease, pathway stage, execution time, and semantic features. The purpose of generating the candidate set is to reduce interference from irrelevant nodes and improve matching efficiency and accuracy.

[0115] Preferably, candidate node screening can be based on the following conditions: similar or identical behavior types, for example, examination-related behaviors are preferentially matched with examination nodes; similar path stages, for example, admission examinations are preferentially matched with admission assessment stage nodes; similar semantic categories, for example, antibacterial drugs are preferentially matched with medication nodes; similar time windows, for example, follow-up examinations within 48-72 hours after treatment are preferentially associated with efficacy assessment nodes.

[0116] For example, candidate criteria for the actual behavior "chest CT examination" may include "chest X-ray examination" and "imaging examination"; candidate criteria for the actual behavior "moxifloxacin intravenous infusion" may include "amoxicillin clavulanate potassium", "respiratory quinolones", and "antibiotic therapy"; and candidate criteria for the actual behavior "blood routine re-examination" may include "treatment efficacy evaluation" related criteria in addition to "blood routine examination".

[0117] In the actual matching process, a precise matching judgment can be made between the actual diagnosis and treatment behavior and the candidate standard nodes. When the actual behavior name, standard code, terminology mapping result or normalized name is completely consistent with the standard path node, a precise matching relationship is generated, and the matching standard node, matching method, matching confidence and evidence source are recorded.

[0118] Exact matching is typically suitable for examinations, tests, medications, and event-based behaviors with well-defined names and a high degree of standardization. Actual medical procedures that achieve an exact match generally do not proceed to the semantic matching process, but their temporal sequence and stage information relative to the standard node can still be retained for subsequent temporal deviation assessments.

[0119] For example, if the actual behavior "complete blood count" is completely consistent with the standard node "complete blood count", an exact match is generated, and the match confidence can be recorded as 1.0; if the actual behavior "C-reactive protein test" is completely consistent with the standard node "C-reactive protein test", an exact match is also generated.

[0120] For actual medical behaviors that fail to pass an exact match, their semantic similarity with candidate standard nodes can be further calculated. Semantic similarity can be calculated using large language models, medical terminology vector models, thesaurus, knowledge graph adjacency relationships, or combined scoring methods. When the similarity reaches a preset threshold, a semantic matching relationship is generated; when the similarity is below the threshold, the actual behavior is not temporarily identified as the corresponding standard node, and enters the replacement, duplication, or non-match judgment process.

[0121] Semantic matching is primarily used to address discrepancies between actual medical record descriptions and standard pathway descriptions. It can simultaneously record similarity scores, matching reasons, and matching evidence to support subsequent manual review.

[0122] For example, the standard node is "chest X-ray examination," while the actual procedure is "chest CT scan." Both are chest imaging examinations, but the methods differ. If the system calculates a semantic similarity of 0.75 and sets the threshold to 0.7, a semantic matching relationship can be generated and labeled as "semantic matching for imaging examinations." Similarly, if the standard node is "pathogen diagnosis," and the actual treatment is "diagnosis of community-acquired pneumonia," a semantic matching relationship can be generated based on medical semantics to determine a diagnostic correlation between the two.

[0123] For cases where actual medical practices do not completely align with standard nodes but may constitute alternative actions, further judgment can be made based on substitution relationships, mutual exclusion relationships, causal relationships, or relationship strength weights within the standard clinical pathway cognitive knowledge graph. If an actual medical practice has a substitutable relationship with a certain standard node, or a reasonable substitution relationship with the category to which the standard node belongs, a substitution matching result is generated, and the reasonableness of the substitution is recorded.

[0124] This operation is mainly used to distinguish between "reasonable substitution" and "unreasonable substitution". For reasonable substitution, it can reduce the degree of deviation; for substitution behavior without a basis or that conflicts with the standard path, it can be judged as unreasonable substitution deviation in subsequent deviation detection.

[0125] For example, the standard pathway recommends the use of "amoxicillin-clavulanate potassium," while the actual medical record records the use of "moxifloxacin intravenous infusion." If there is a substitution or mutual exclusion relationship between "amoxicillin-clavulanate potassium" and "respiratory quinolones" in the standard clinical pathway cognitive knowledge graph, and the relationship strength weight is 0.85, it can be determined that "moxifloxacin" constitutes a reasonable substitution match for the standard medication; if the actual drug used has no substitution relationship with the standard drug and there is a contraindication conflict, it can be determined as an unreasonable substitution candidate.

[0126] After completing exact matching, semantic matching, and replacement matching, the remaining diagnostic and treatment behaviors and nodes are identified as duplicate or unmatched. If multiple actual diagnostic and treatment behaviors correspond to the same standard node, a duplicate matching marker is generated; if an actual diagnostic and treatment behavior cannot correspond to any standard path node, it is marked as an unmatched actual diagnostic and treatment behavior; if a standard path node is not matched by any actual diagnostic and treatment behavior, it is marked as an unmatched standard clinical path node.

[0127] Among these, duplicate matches may correspond to duplicate deviations, or they may be part of necessary follow-up examinations or efficacy assessments; non-matching actual diagnostic and treatment behaviors may correspond to new deviations; non-matching standard clinical pathway nodes may correspond to deviations involving missing steps. In this operation, a final deviation conclusion is not directly drawn, but rather a structured alignment result is generated for further judgment by the subsequent deviation identification module.

[0128] For example, the actual behavior "complete blood count" can form an exact match with the standard node "complete blood count"; the actual behavior "chest CT scan" can form a semantic match or substitution match with the standard node "chest X-ray"; the actual behavior "moxifloxacin intravenous infusion" can form a substitution match with the standard node "amoxicillin clavulanate potassium"; the actual behavior "re-examination of complete blood count" can form a duplicate match with the already matched standard node "complete blood count"; if the standard node "severity assessment of condition" has no corresponding behavior in the actual diagnosis and treatment behavior sequence, then the standard node is marked as an unmatched standard clinical pathway node.

[0129] For example, a patient undergoes a "complete blood count" upon admission and a "re-examination of the complete blood count" after treatment. If both matches the standard node "complete blood count", it is marked as a duplicate match; if the standard path node "severity assessment of the condition" has no corresponding actual behavior, it is marked as a non-matching standard node; if the actual behavior "chest CT scan" does not correspond to any standard node, it can be marked as a non-matching actual behavior and subsequently identified as a new deviation candidate.

[0130] Summarize the above matching results to generate a correspondence table between actual medical behaviors and standard path nodes. This correspondence table should include at least the name of the actual medical behavior, the actual execution time, the behavior type, the matching standard node, the matching method, the similarity score, the basis for the relationship, the node importance weight, the relationship strength weight, the matching status, and the source of evidence.

[0131] Matching methods can include exact matching, semantic matching, substitution matching, duplicate matching, non-matching actual clinical behavior, and non-matching standard nodes of clinical pathway sections. The matching status will serve as direct input for subsequent deviation identification and classification. For example, non-matching standard nodes can lead to step-missing deviation identification, duplicate matching can lead to duplicate deviation identification, substitution matching can lead to substitution rationality judgment, and non-matching actual behavior can lead to new deviation identification.

[0132] For example, the following alignment results can be generated: "Routine Blood Test - Standard Node N001 - Exact Match - Similarity 1.0"; "Chest CT Scan - Standard Node N004 Chest X-ray - Semantic Match - Similarity 0.75"; "Moxifloxacin Intravenous Infusion - Standard Node N007 Amoxicillin Clavulanate Potassium - Replacement Match - Relationship Strength Weight 0.85"; "Disease Severity Assessment - No Matching Standard Node". These results will be used subsequently to determine if there are any deviations such as missing steps, replacements, additions, or duplications.

[0133] In this embodiment, precise matching is prioritized when the actual medical behavior is completely consistent with the standard node name, code, or standard terminology. Alternatively, matching can be performed using in-hospital item codes, medical insurance item codes, drug codes, examination and testing item codes, clinical pathway node numbers, or medical order dictionary codes. For cases where multiple coding systems exist for the same behavior, a mapping table can be used to unify them.

[0134] In this embodiment, the similarity between the actual behavior and the standard node is preferentially calculated using a large language model or a semantic similarity model. Alternatively, semantic matching can be performed using vector embedding models, medical thesaurus, edit distance, keyword matching, knowledge graph proximity relationships, medical ontology mapping, or multi-model fusion scoring. The similarity threshold can also be dynamically adjusted based on the disease type, node type, or manual review results.

[0135] In this embodiment, replacement matching can be determined based on mutual exclusion relationships, substitution relationships, or relationship strength weights in the knowledge graph. As alternatives, they can also be determined based on available options in clinical guidelines, drug indications and contraindications, patient allergy history, examination substitution rules, medical insurance payment rules, expert rule bases, or high-frequency reasonable substitution behaviors in historical cases. For example, when a standard medication is replaced by a similar alternative drug due to a patient's allergy, its reasonableness can be determined by jointly using contraindication rules and alternative drug rules.

[0136] In this embodiment, a step-by-step matching method is preferred to generate the correspondence between actual behaviors and standard nodes. Alternatively, dynamic time warping, sequence alignment algorithms, edit distance algorithms, longest common subsequence algorithms, graph traversal algorithms, shortest path algorithms, or sequence-to-graph matching algorithms can be used for overall alignment. For diseases with long diagnostic and treatment processes and numerous path branches, a combination of sequence alignment and graph reasoning can be used to improve matching accuracy.

[0137] S4. Based on the correspondence between actual clinical behavior and standard clinical pathway nodes, and combined with the node importance weight, relationship type, relationship strength weight, time window, triggering conditions and / or exceptional evidence in the standard clinical pathway cognitive knowledge graph, determine whether the actual clinical behavior deviates from the standard clinical pathway.

[0138] In this embodiment, the deviation judgment rule set is invoked mainly based on the alignment matching results to identify deviations of the actual diagnosis and treatment behavior from the standard clinical pathway, and the deviation types are classified. The deviation types include at least step missing deviation, sequence disorder deviation, condition not met deviation, causal chain interruption deviation, mutual exclusion conflict deviation, substitution deviation, addition deviation, and duplication deviation.

[0139] Missing step deviation refers to the absence of a necessary node specified in the standard pathway in the actual treatment sequence without reasonable exception evidence. Disordered sequence deviation refers to the inconsistency between the execution order of actual treatment actions and the strong temporal relationships, pathway stages, or time windows in the standard pathway diagram. Unmet condition deviation refers to the execution of an action before the triggering condition is met, or the failure to execute within the specified time window after the triggering condition is met, including deviations from execution before condition is met and deviations from non-execution after condition is triggered. Disrupted causal chain deviation refers to the occurrence of a preliminary diagnosis, examination result, or risk event, but the subsequent execution node is not executed or is executed too late. Mutually exclusive conflict deviation refers to the simultaneous occurrence of two mutually exclusive treatment actions without a basis for conversion. Substitution deviation refers to the actual executed treatment action replacing a node in the standard pathway. Addition deviation refers to the actual execution of a treatment action not specified in the standard pathway. Repetition deviation refers to the repeated execution of a node in the standard pathway that typically only needs to be executed once during the actual treatment process.

[0140] In the deviation classification process, this invention further identifies whether the deviation is reasonable and incorporates the reasonableness judgment result into the deviation degree calculation. For example, substitution deviations can be further divided into reasonable substitutions and unreasonable substitutions; new addition deviations can be further divided into beneficial additions, redundant additions, or risky additions; duplicate deviations can be further divided into necessary duplicates and unnecessary duplicates; condition-triggered and mutually exclusive conflict deviations can be further divided into recorded causes of variation and unrecorded causes of variation.

[0141] For example, the standard clinical pathway cognitive knowledge graph is denoted as G=(V,E); the sequence of actual patient treatment behaviors is denoted as A={a1,a2,…,an}; and the set of correspondences between actual treatment behaviors and standard clinical pathway nodes is denoted as M; where V is the set of standard clinical pathway nodes, E is the set of relationships between nodes, and an represents the actual treatment behaviors of the bottom n patients; the deviation judgment rule set includes at least the following: R1 Step Missing Rule: If a necessary node v in the standard path satisfies Required(v)=1, and there is no corresponding actual action for v in the matching relation M, and no exceptional evidence such as path variation, patient refusal, contraindications, or physician explanations is found, then it is judged as a step missing deviation; if v is a critical node and Wv is higher than a preset threshold, then the risk level of this deviation is increased. Required(v) is a node necessity flag function, taking the value 1 when node v is a mandatory node under the current disease, path stage, and applicable conditions, and 0 otherwise.

[0142] R2 Sequence Disorder Rule: If a temporal edge e=(vbefore,vafter) exists in the standard clinical pathway cognitive knowledge graph, and both have been matched with actual actions abefore and aafter, but the actual execution time T is different... (a前) Later than T (a后) If the time edge exceeds the stage sequence and allowed time window δ specified in the graph, it is judged as a disordered deviation; if the time edge belongs to a strong time relationship or is related to key behaviors such as discharge, surgery, or medication initiation, the relationship strength weight is increased.

[0143] R3 Condition Not Met Rule: If there is a condition trigger edge e=(c,v) in the standard clinical pathway cognitive knowledge graph, it means that node v should only be executed when condition c is met. However, if the actual behavior a has been executed but the corresponding examination results, diagnosis conclusions, risk assessments, or complication records cannot prove that c is met, it is judged as a condition not met execution deviation. If v is not executed within the limited time window ΔT after condition c is met, it is judged as a condition triggered not executed deviation.

[0144] R4 Causal Chain Disruption Rule: If there is a causal edge e=(vcause, vef) in the standard clinical pathway cognitive knowledge graph, and vcause has already occurred in the actual sequence and triggered subsequent diagnosis and treatment requirements, but vef has not been executed, executed too late, or lacks evidence, it is judged as a deviation from causal chain interruption; if vef is executed in advance before vcause occurs or is confirmed, it is judged as a deviation from causal basis inadequacy.

[0145] R5 Mutual Exclusion Conflict Rule: If there is a mutually exclusive edge e=(v1,v2) in the standard clinical pathway cognitive knowledge graph, and the corresponding behaviors of v1 and v2 appear simultaneously in the actual sequence, and no explanatory evidence such as medical order adjustment, medication switch, allergy contraindication, change of condition or expert consultation is found, it is judged as a mutual exclusion conflict deviation; if the mutually exclusive edge is a strong constraint relationship, a high-risk warning is directly triggered.

[0146] R6 Replacement Reasonableness Rule: If the actual behavior a does not form an exact match with the standard node v, but has a substitution relationship with v or the semantic similarity reaches the substitution threshold θ, then it enters the replacement judgment; when the substitution strength, indication, contraindication or patient individualized evidence meets the preset conditions, it is judged as a reasonable substitution and the deviation weight is reduced; otherwise, it is judged as an unreasonable substitution deviation.

[0147] For example, the preferred formula for calculating the substitution intensity St is: St = 0.30 × E 证据 +0.25×C 类别 +0.20×I 适应证 +0.15×H 历史 +0.10×P 个体 Among them, E 证据 C indicates the level of evidence for the source of the vicarious relationship. 类别 Indicates the degree of consistency between actual behavior and standard nodes in terms of diagnostic and treatment purposes and categories, I 适应证 H indicates the degree of indication matching. 历史 This indicates the proportion of historical cases where the alternative treatment has been manually reviewed and confirmed to be reasonable, P 个体 The value represents the degree of support from individualized evidence such as the patient's allergy history, liver and kidney function, and complications. Each component is normalized to the [0,1] interval. The substitution strength threshold λ is determined based on a sample set of reasonable and unreasonable substitutions labeled by clinical experts, preferably taking the threshold corresponding to the largest F1 value or the largest Youden index. When there is a clear contraindication conflict, it is not affected by the St value and is directly judged as an unreasonable substitution.

[0148] R7 adds a new behavior rule: If the actual behavior 'a' cannot match any standard path node, its nature is determined based on the semantic relevance of 'a' to the current path stage, diagnostic conclusion, risk event, and neighboring nodes in the graph; when the relevance is high and there is supplementary diagnostic, risk control, or efficacy assessment evidence, it is judged as a beneficial addition; when the relevance is low or there is a lack of medical evidence, it is judged as a redundant addition; when 'a' is related to a strong mutually exclusive relationship, contraindication, or high-risk medication, it is judged as a risky addition deviation.

[0149] R8 Repeat Execution Rule: If the same standard node v is repeatedly matched by multiple actual behaviors a, then check whether there is an allowed relationship for re-examination, follow-up, efficacy evaluation or stage repetition in the atlas; if there is an allowed relationship and the repetition time meets the time window requirements, it is determined to be a necessary repetition; if there is no allowed relationship or the repetition behavior lacks clinical evidence, it is determined to be an unnecessary repetition deviation.

[0150] In specific judgments, this invention generates a structured record for each deviation, consisting of "rule number - triggering condition - hit evidence - exception evidence - preliminary conclusion - review status". For example, if the standard node Required(v)=1 and does not match the actual behavior, the R1 step missing rule is triggered first, and then the existence of contraindications, refusal of treatment, path variation records, or physician explanations is searched. If valid exception evidence exists, it is marked as an interpretable missing step and the score is reduced. If no exception evidence exists, it is marked as an uninterpreted step missing step and enters the high-weight calculation.

[0151] For example, if the actual "discharge" occurs earlier than the "48-72 hour efficacy assessment," the R2 order disorder rule is triggered based on the time sequence edge and time window; if "moxifloxacin" is actually used to replace the standard recommended drug, the R6 substitution rationality rule is triggered based on the substitution edge, patient allergy record, etiological results, and drug indications; if the same examination item is performed repeatedly within a short period of time, the R8 repeated execution rule is triggered based on the re-examination permission relationship, efficacy assessment stage, and repeated time window.

[0152] If the standard pathway requires a "severity assessment of the condition," but no corresponding record is found in the patient's medical history, it is identified as a step-missing deviation. If the standard pathway requires "examination—diagnosis—treatment—efficacy assessment—discharge," but the actual pathway is "examination—diagnosis—treatment—discharge," it is identified as a lack of efficacy assessment. If the standard pathway recommends "amoxicillin-clavulanate potassium," but "moxifloxacin" is actually used, and there is a substitution or mutual exclusion relationship between the two in the knowledge graph, it is identified as a substitution deviation, and further assessment is needed to determine whether it is a reasonable substitution. If a "chest CT scan" is actually performed, while the standard pathway only specifies a "chest X-ray," it can be identified as a new deviation or a substitution deviation.

[0153] After obtaining the deviation information, this embodiment calculates the score for each deviation based on the deviation type weight Wt, node importance weight Wv, relationship strength weight We, evidence credibility Ce, and reasonableness adjustment coefficient Ar. Here, Wt represents the basic severity of different deviation types, Wv represents the importance of the standard path node involved in the deviation within the target disease pathway, We represents the constraint strength of the temporal, causal, mutually exclusive, substitution, or conditional triggering relationships violated by the deviation, Ce represents the reliability of the original evidence bound to the deviation judgment, and Ar represents the reduction coefficient for reasonable substitution, beneficial addition, necessary duplication, or the existence of exceptional evidence.

[0154] The deviation type weight Wt can be configured according to disease type and management objectives and calibrated with historical data. Preferably, strong mutual exclusion conflict, execution without condition fulfillment, missing key steps, interrupted causal chain, serious sequence disorder, unreasonable substitution, new risk, and unnecessary duplication are assigned higher to lower base weights respectively; reasonable substitution, beneficial new addition, and necessary duplication are considered as explainable deviations and can be assigned lower weights or not included in high-risk deviations.

[0155] The preferred method for calculating the node importance weight Wv is: Wv = 0.35 × S 专家 +0.25×S 指南 +0.20×S 事件 +0.20×S 历史 The result is then normalized to the interval [0,1]. Where S... 专家 This represents the score given by clinical experts to this node, a value in the range [0,1]. The node's importance can be independently rated by at least three clinical experts on a scale of 1 to 5, and the average score is then calculated. 专家 =(average score) 1) / 4 normalization; only when tiered scoring data is lacking, a binary substitution is used, with key nodes set to 1 and non-key nodes set to 0; S 指南 Indicates the strength of the recommendation for this node from the guidelines or hospital pathways; S 事件 This indicates the degree of correlation between this node and key events such as diagnosis confirmation, risk assessment, treatment initiation, and discharge safety. 历史 This indicates the degree of impact on quality control conclusions, length of hospital stay, costs, or adverse outcomes when this node is missing or abnormal in historical cases.

[0156] The preferred method for calculating the relationship strength weight We is: We = 0.40 × R 类型 +0.25×R 证据 +0.20×R 时间 +0.15×R 一致性 The results are then normalized to the [0,1] interval. Where R... 类型 Used to distinguish different constraint strengths such as mutual exclusion, conditional triggering, causal relationships, strong temporal relationships, weak temporal relationships, and substitution relationships; R 证据 Used to indicate the credibility of the relationship derived from national guidelines, hospital pathways, expert consensus, or real-world data; R 时间 Used to indicate whether the relationship has a defined execution time window; R 一致性 Used to indicate the degree of stability of the relationship as clinically confirmed in historical cases.

[0157] The input parameters R-type, R-evidence, R-time, and R-consistency for the relation strength weight We are all scores within the range of [0,1]. Preferably, R-type is scored by clinical experts according to the degree of strong constraint of the relation, using a five-level scale and mapped to 0, 0.25, 0.50, 0.75, and 1.00; R-evidence is mapped to the same scale based on the evidence levels of hospital pathways, expert consensus, industry guidelines, and national guidelines; R-time is scored based on the existence of a clear time window and the consequences of exceeding the window; R-consistency is the proportion of historical cases in which the relation has been manually verified and confirmed to be valid. In the formula, 0.40, 0.25, 0.20, and 0.15 are preferred initial coefficients, and cross-validation or grid search can be used based on expert-annotated case sets to perform disease-specific calibration with the goal of maximizing the deviation detection F1 value.

[0158] For example, the calculation expression for the score Di of the deviation of the i-th item is: Di=100×Wt_i×Wv_i×We_i×Ce_i×Ar_i; The deviation type weight Wt_i of the i-th deviation is configured based on the disease and management objectives and calibrated with historical data; the node importance weight Wv_i involved in the i-th deviation is jointly determined by the expert path importance score, guideline / standard evidence level, clinical key event correlation, and historical case execution impact; the relationship strength weight We_i involved in the i-th deviation is jointly determined by the degree of relationship type constraint, medical evidence level, time window constraint, and historical consistency; the evidence credibility Ce_i involved in the i-th deviation is used to reflect the reliability of the deviation judgment evidence, such as structured medical orders, laboratory test reports, original medical records, nursing records, and manual review results, which can be set with different credibility levels; the reasonableness adjustment coefficient Ar_i involved in the i-th deviation is used to reflect individualized diagnosis and treatment and path variation factors, such as reasonable substitution, beneficial addition, necessary duplication, or the existence of path variation records, taking a reduction value less than 1, and taking an amplification value of 1 or greater than 1 when there is unreasonable substitution, risk addition, or lack of explanatory evidence.

[0159] The overall degree of deviation for each patient can be obtained based on the score for each deviation, expressed as D = 100 × [1]. ∏(1 [Di / 100]; When the management scenario requires balanced evaluation by path stage, the weighted average method of D=Σ(Di×W stage_i) / ΣW stage_i can also be used. The former method is used to highlight the comprehensive risk after the accumulation of multiple deviations, while the latter method is used to conduct balanced evaluation of different path stages. Among them, "path stage" is the process unit divided by the standard clinical path of the target disease according to the diagnosis and treatment sequence, including admission assessment, diagnosis confirmation, treatment implementation, efficacy monitoring, discharge assessment and follow-up; W stage_i represents the weight of the i-th deviation to which the stage belongs. The stage weight is determined and normalized to the [0,1] interval based on the stage criticality, risk level and historical deviation impact.

[0160] This embodiment can classify the degree of deviation into slight deviation, moderate deviation, significant deviation, and severe deviation based on D. Preferably, D < 20 indicates slight deviation, 20 ≤ D < 50 indicates moderate deviation, 50 ≤ D < 75 indicates significant deviation, and D ≥ 75 indicates severe deviation. If a single deviation involves strong constraints such as strong mutual exclusion conflicts, unprocessed critical values, or missing key treatments, a severe deviation or high-risk warning can be directly triggered even if the overall score does not reach the threshold.

[0161] The above operations provide a mechanism for quantifying the degree of deviation, enabling deviation detection results not only to be classified, but also to be quantified and compared.

[0162] Specifically, the value of the deviation type weight Wt can be jointly determined by hospital quality control rules, medical insurance review priorities, and clinical risk levels, and can be continuously calibrated based on manual review results. For example, mutually exclusive conflicts, execution without met conditions, and missing key steps can be assigned high weights; unreasonable substitutions and new risks can be assigned medium to high weights; unnecessary duplications can be assigned medium to low weights; and reasonable substitutions, beneficial additions, and necessary duplications can be assigned low weights or reduced.

[0163] The credibility of evidence (Ce) reflects the reliability of the evidence used to determine deviations. When the deviation evidence comes from structured medical order execution records, laboratory reports, or nursing execution records and is consistent with the medical record text, Ce takes a higher value; when it comes only from unstructured text extraction and lacks corroborating structured records, Ce takes a lower value and prompts manual review. The rationality adjustment coefficient (Ar) reflects individualized treatment and pathway variation factors. When there are contraindications such as allergies, patient refusal, physician explanations, consultation opinions, or pathway variation records, Ar can be reduced according to the rules; when explanatory evidence is lacking and high-risk relationships are involved, Ar can be 1 or greater than 1.

[0164] For example, if a case exhibits deviations such as "missing assessment of disease severity," "missing assessment of efficacy within 48-72 hours," "appropriate substitution of standard medication," and "repeated blood routine tests," the importance weight and correlation strength weight of each node can be retrieved to calculate the overall deviation score. For instance, if the calculated result is 61.3%, the case can be rated as "significant deviation," and the contribution of each deviation item to the total score can be listed in the report.

[0165] It should be noted that this implementation can be modified to include deviations such as time delay, early execution, unmet conditions, taboo conflicts, missing evidence, unrecorded path variations, incorrect stage attribution, and abnormal costs, depending on business needs. Different deviation types can correspond to application scenarios such as quality control, medical insurance, and clinical safety.

[0166] In this embodiment, when a standard node does not match a corresponding action in the actual diagnosis and treatment sequence, it is preferentially identified as a missing step. As an alternative, the judgment can also be made by combining the importance of the node, the path stage, the patient's applicable conditions, and the exclusion conditions. For optional nodes or nodes that are not applicable to the patient, even if they are not executed, they may not be judged as missing deviations; for critical nodes or mandatory nodes, if they are not executed, they can be directly judged as high-risk missing deviations.

[0167] In this embodiment, the order of actual treatment actions is primarily determined based on the temporal relationships in the atlas to determine whether they conform to the standard path. Alternatively, the determination can be made based on time windows, path stages, business rules, minimum time intervals, maximum allowable delay times, or the order in which clinical events are triggered. For example, efficacy assessment 48-72 hours after treatment can be determined based on the time difference between the treatment start time and the assessment record time.

[0168] In this embodiment, behaviors not matched with actual medical treatment are prioritized as new deviation candidates, and behaviors matched multiple times with the same standard node are prioritized as repeated deviation candidates. As an alternative, new behaviors can be further determined as beneficial, redundant, or risky new behaviors based on clinical rationality, cost impact, complexity of the patient's condition, and expert rules; repeated medical treatment behaviors can be determined as necessary or unnecessary repetitions based on the purpose of follow-up examination, changes in the patient's condition, the need for efficacy evaluation, and time intervals.

[0169] In this embodiment, the comprehensive deviation score is calculated primarily based on the deviation type weight, node importance weight, and relationship strength weight. Alternatively, rule-based scoring, expert scoring, analytic hierarchy process (AHP), weighted summation models, risk matrix models, Bayesian models, machine learning classification models, or regression models can be used to calculate the deviation severity. Different medical institutions can set different scoring rules according to their quality control requirements.

[0170] In this embodiment, the importance weights and relationship strength weights of deviation steps are derived from the standard clinical pathway cognitive knowledge graph. Alternatively, the weights can be dynamically determined based on clinical expert scores, historical case outcomes, quality control deduction standards, medical insurance review rules, medical safety risk levels, patient risk stratification, or actual implementation data. For high-risk diseases, the weights of key nodes and strong constraints can be increased.

[0171] In this embodiment, the degree of deviation is preferentially classified into slight deviation, moderate deviation, significant deviation, and severe deviation. Alternatively, a classification method such as low risk, medium risk, high risk, and critical risk can be used; or different thresholds can be set according to the management requirements of medical institutions. For medical insurance review scenarios, the classification can be based on the level of cost risk; for medical safety scenarios, the classification can be based on the degree of impact on patient safety.

[0172] In this embodiment, the overall path deviation of a single patient is calculated first. Alternatively, deviation scores for individual nodes, stages, categories, departments, diseases, or time periods can be calculated separately to meet the analytical needs of different management levels.

[0173] As can be seen from the above description, this method does not output "compliant / non-compliant" after detecting clinical pathway deviations. Instead, it performs fine-grained classification of deviations using a rule set, and binds each deviation to a triggering rule, a standard node, actual evidence, and exceptional evidence, thus transforming deviation detection from a coarse judgment to a refined classification. Its core features include: When a necessary node in the standard path does not appear in the actual diagnosis and treatment sequence, it is identified as a step missing deviation. When the actual sequence of medical procedures is inconsistent with the temporal relationship in the atlas, it is identified as a disordered or deviated sequence. When actual medical practice replaces a step in the standard pathway, it is identified as a substitution deviation. When a diagnostic or treatment procedure not specified in the standard pathway is actually performed, it is identified as a new deviation. When the same standard path node is executed repeatedly or matched multiple times, it is identified as a duplicate deviation; Further distinguish between reasonable and unreasonable substitutions based on substitution deviations; Further differentiate between beneficial additions, redundant additions, or risky additions in the newly added deviations; Further distinguish between necessary and unnecessary repetitions in the repetition deviation; Further distinguish between "execution if condition is not met" and "execution if condition is met" in conditional triggering relationships; Further distinguish between "broken causal chain" and "insufficient causal evidence" in causal relationships; Further distinguish between "mutually exclusive co-occurrence with switching basis" and "mutually exclusive conflict without explanatory evidence" in the mutual exclusion relationship; Output the rule number, triggering condition, evidence of hit, evidence of exception, and preliminary risk level for each deviation.

[0174] This embodiment utilizes path cognition relationships such as causal relationships, mutual exclusion relationships, substitution relationships, and conditional triggering relationships in a standard clinical pathway cognitive knowledge graph to judge the reasonableness of deviations. Its core features include: Based on causal relationships, determine whether a certain diagnostic or treatment behavior is based on prior diagnosis, examination results, or medical condition; Determine whether the actual execution order violates the standard path logic based on the timing relationship; Determine whether two diagnostic and treatment actions should not occur simultaneously or whether they constitute substitution based on mutual exclusion; Based on the substitution relationship, determine whether the actual diagnosis and treatment behavior can reasonably replace the standard path nodes; Determine whether a certain diagnostic or treatment action should be performed only after certain conditions are met, based on conditional triggering relationships; The impact of deviation behavior on path execution quality is determined based on relation strength weighting. Reduce the severity of deviations for replacements, additions, and repetitions with reasonable medical evidence; Increase the risk warning level for behaviors that violate strong binding relationships.

[0175] Example 2: The purpose of this embodiment is to propose a system for implementing the clinical pathway deviation detection method provided in Embodiment 1, such as... Figure 2 As shown, this system includes: The data layer is used to obtain a cognitive knowledge graph of standard clinical pathways for single diseases; it extracts actual diagnosis and treatment behaviors from patient data and sorts them by execution time to form a sequence of actual diagnosis and treatment behaviors; The alignment layer is used to match and correspond actual diagnosis and treatment behavior sequences with standard clinical pathway cognitive knowledge graphs, generating correspondences between actual diagnosis and treatment behaviors and standard clinical pathway nodes; The detection layer is used to determine whether the actual diagnosis and treatment behavior deviates from the standard clinical pathway based on the correspondence between the actual diagnosis and treatment behavior and the nodes of the standard clinical pathway, combined with the node importance weight, relationship type, relationship strength weight, time window, triggering conditions and / or exceptional evidence in the cognitive knowledge graph of the standard clinical pathway. The application layer is used to provide output deviation reports, risk warnings, quality analysis, and path optimization suggestions.

[0176] Specifically, the data layer is the foundational layer of this system. This layer mainly includes a standard clinical pathway knowledge graph, actual diagnosis and treatment behavior sequences, a knowledge graph construction module, an electronic medical record parsing module, clinical guidelines / expert knowledge / pathway text data sources, and electronic medical records / medical orders / examination and test records data sources.

[0177] The actual clinical behavior sequence represents the behaviors that a target patient has engaged in during actual clinical treatment. This sequence is extracted by the electronic medical record parsing module from data such as the patient's electronic medical record, medical orders, examination and test results, nursing records, and discharge summaries. Each actual clinical behavior includes at least the behavior name, behavior type, execution time, source record, and evidence fragment, and is sorted according to execution time to form temporally sequenced behavioral data that can be compared with the standard clinical pathway knowledge graph.

[0178] The knowledge graph construction module is used to construct a cognitive knowledge graph of standard clinical pathways for single diseases based on data such as clinical guidelines, expert knowledge, hospital pathway texts, standard clinical pathway forms, historical case databases, and manual review records. This module utilizes a large language model to perform semantic parsing of the pathway text, identifying treatment behaviors, pathway stages, sequence, causal dependencies, conditional triggers, mutual exclusion constraints, alternative solutions, causes of pathway variation, and key nodes, and converting these into graph nodes, relationships, rule conditions, and weight parameters. The module also includes a weight initialization and calibration submodule, used to initialize, normalize, and version-manage the node importance weights, relationship strength weights, and deviation type weights based on expert scores, guideline evidence levels, historical deviation frequencies, and review results.

[0179] The electronic medical record parsing module is used to extract actual medical behaviors from patients' real medical records. This module can call a large language model to perform semantic understanding on unstructured electronic medical record text, identify behaviors such as examinations, diagnoses, treatments, medications, assessments, and discharges in the text, and extract their execution time and source evidence. For structured medical orders, test results, and other data, the system can directly read fields and perform standardized transformations; for free text, the system forms a sequence of medical behaviors through semantic parsing and structured extraction.

[0180] The alignment layer is a crucial processing step before deviation detection. This layer mainly includes a sequence parsing module, a semantic matching module, and a correspondence generation module.

[0181] The sequence parsing module is used to standardize the sequence of actual medical actions, identifying the time sequence, type, stage of treatment, and source evidence for each action. For example, it can parse "blood routine and C-reactive protein tests on March 1, 2025, and moxifloxacin intravenous infusion on March 2" into multiple time-ordered action nodes. This module can also sort and group multiple actions within the same day, providing a clear actual execution trajectory for subsequent matching.

[0182] The semantic matching module is used to calculate the matching relationship between actual diagnostic and treatment behaviors and standard path nodes. For behaviors with exactly the same name, the system performs exact matching; for behaviors with different names but similar meanings, the system uses a large language model or semantic similarity model for semantic matching; for cases where the actual behavior substitutes for a standard path node, the system combines mutually exclusive relationships, substitution relationships, or relationship strength weights in the graph for substitution matching. For example, if the standard path node is "chest X-ray examination" and the actual behavior is "chest CT examination," the system can determine whether the two are semantically related or substitutional.

[0183] The correspondence generation module generates correspondences between actual behaviors and standard path nodes based on the matching results. Correspondences can include exact matches, semantic matches, substitution matches, duplicate matches, unmatched actual behaviors, and unmatched standard nodes. Unmatched actual behaviors may introduce new deviations, unmatched standard nodes may indicate missing steps, duplicate matches may result in repeated deviations, and substitution matches may be further judged as reasonable or unreasonable substitutions. The output of this module serves as the direct input for the detection layer to identify deviation types.

[0184] The detection layer is the core processing layer of this system. This layer does not merely match and judge individual activities, nor does it simply output deviation reports; instead, it forms a closed-loop detection process: "rule triggering—evidence binding—weight calculation—level determination—interpretable output." This layer mainly includes a deviation identification module, a deviation classification module, a degree calculation module, and a report generation module.

[0185] The deviation identification module determines whether the actual medical behavior deviates from the standard clinical pathway based on the correspondence between the actual treatment behavior and the standard pathway nodes, combined with the node importance weights, relationship types, relationship strength weights, time windows, triggering conditions, and exceptional evidence in the standard clinical pathway cognitive knowledge graph. This module denotes the standard pathway graph as G=(V,E), where V is the set of standard pathway nodes and E is the set of relationships between nodes; the sequence of the patient's actual treatment behavior is denoteed as A={a1,a2,…,an}; and the matching relationships output by the alignment layer are denoteed as M. The system executes deviation judgment rules one by one based on M, V, E, and A, and saves the triggering node, triggering relationship, actual evidence, and judgment reason for each rule.

[0186] The deviation classification module is used to categorize identified deviations. Deviation types include at least four categories: missing step deviation, disordered sequence deviation, substitution deviation, addition deviation, and duplication deviation. A missing step deviation indicates that a necessary step specified in the standard pathway was not executed; a disordered sequence deviation indicates that the actual sequence of actions does not conform to the temporal relationship specified in the standard pathway; a substitution deviation indicates that the actual action replaced a step in the standard pathway; an addition deviation indicates that a diagnostic or treatment action not specified in the standard pathway was performed; and a duplication deviation indicates that a step in the standard pathway that typically only needs to be performed once was repeated. For deviations such as substitution, addition, and duplication, the system can further determine their reasonableness, such as reasonable substitution, beneficial addition, or necessary duplication.

[0187] The degree calculation module is used to score each deviation identified by the detection layer individually and further generate a comprehensive deviation degree score at the patient level, stage level, node level, or disease level. The calculation input of this module includes at least the deviation type weight Wt, node importance weight Wv, relationship strength weight We, evidence credibility Ce, and reasonableness adjustment coefficient Ar.

[0188] The report generation module converts deviation detection results into structured reports. The report content includes at least the target disease, patient identifier, detection time, deviation type, deviation location, deviation description, involved standard nodes, involved actual behaviors, trigger rule number, deviation type weight, node importance weight, relationship strength weight, evidence credibility, reasonableness adjustment coefficient, individual deviation score, overall deviation score, deviation level, and improvement suggestions. For each deviation, the system can display its matching status, graph nodes, relationship basis, original evidence fragments, exceptional evidence, and the reason for the judgment, making the deviation detection results interpretable and verifiable.

[0189] The application layer is used to provide deviation detection results output by the detection layer to clinical, quality control, medical insurance, management, and research scenarios. This layer mainly includes deviation report display, risk warning push, quality analysis dashboard, and path optimization suggestions.

[0190] Deviation reports are used to present structured deviation detection reports to doctors, quality control personnel, medical insurance reviewers, or administrators. Users can view a patient's deviation details, overall deviation score, deviation level, and improvement suggestions, and can also trace back to the corresponding standard path nodes and actual treatment behaviors.

[0191] The deviation details should include at least the deviation type, deviation location, involved standard nodes, involved actual behavior, matching status, deviation description, node importance weight, relationship strength weight, deviation score contribution, and recommended handling measures. Reports can be exported through the system page, quality control platform, doctor workstation, medical insurance review system, or report export interface.

[0192] For example, for cases of community-acquired pneumonia, the system-generated report may include: "Deviation type: Step missing deviation; Deviation location: Severity assessment node; Deviation description: The standard pathway requires CURB-65 or PSI scoring, but no corresponding assessment record was found in the medical record; Recommendation: Supplement the severity assessment or improve the medical history record." For substitution deviations, the report may be described as: "The standard pathway recommends amoxicillin-clavulanate potassium, but moxifloxacin was actually used. A substitution relationship exists in the knowledge graph, which is determined to be a reasonable substitution, but it is recommended to record the reason for the substitution."

[0193] Risk warning push notifications are used to alert users to serious or high-risk deviations. When the system detects missing critical steps, serious sequence errors, unreasonable substitutions, or high-risk new behaviors, it can push warning information to clinicians, quality control personnel, or managers, prompting them to review and handle the issues promptly.

[0194] The quality analysis dashboard is used to statistically analyze deviations from multiple patients, departments, or time periods. The system can statistically analyze the frequency of different deviation types, the distribution of deviations for different diseases, the missing rate of key nodes, the use of alternative solutions, and the implementation status of departmental pathways, providing data support for hospital clinical pathway management and medical quality improvement.

[0195] Path optimization suggestions are used to identify areas for improvement in standard paths based on deviation detection results. For example, when a standard node is missing for a long period, the system can prompt for enhanced node execution reminders; when a certain alternative drug appears frequently and is often judged as a reasonable alternative, the system can suggest including it in the optional path; when a new test appears repeatedly in a large number of cases and has clinical rationale, the system can suggest revising the standard path.

[0196] In its implementation, the system also includes a self-updating layer. This layer acquires deviation detection results, manual review results, deviation frequency statistics, and path execution data from other layers / modules, thereby updating the cognitive knowledge graph of the standard clinical pathway for a single disease and / or the criteria used to determine whether actual treatment behavior deviates from the standard clinical pathway. The content used for updating may include high-frequency missing nodes, high-frequency substitution behaviors, high-frequency newly added behaviors, high-risk deviation types, manual confirmation results, and quality control processing results.

[0197] For recurring deviations that are confirmed to be clinically reasonable, the system can suggest adjustments to the standard pathway graph, such as adding optional nodes, adjusting substitution relationships, reducing the deviation weight of reasonable substitutions, or supplementing pathway descriptions. For recurring deviations that are deemed unreasonable, the system can increase the warning level, enhance node execution reminders, or increase the importance weight of the corresponding nodes. Through this process, the system can form a closed-loop mechanism of "deviation detection—manual review—feedback optimization—graph update—re-detection," enabling the standard clinical pathway knowledge graph to be continuously optimized as clinical practice and medical evidence change.

[0198] In this embodiment, deviation detection results are preferentially fed back to the self-updating layer. As an alternative, doctor review results, quality control rectification results, medical insurance review results, patient outcome data, cost data, departmental feedback, and expert review opinions can also be included as feedback data to optimize the importance weight of graph nodes, relationship strength weight, and path rules.

[0199] In this embodiment, the standard clinical pathway knowledge graph is adjusted primarily based on feedback results. Alternatively, updates can be made through manual review followed by updates, periodic batch updates, automatic incremental updates, version updates, canary releases, or multi-expert review. For pathway adjustments with high clinical risk, manual approval can be implemented; for low-risk weight adjustments, automatic or semi-automatic updates can be used.

[0200] In this embodiment, revision suggestions are prioritized based on high-frequency deviation trigger paths. Alternatively, optimization suggestions can be generated based on indicators such as deviation trends, patient outcomes, cost changes, departmental implementation rates, key node missing rates, and the frequency of alternative use. For example, if a newly added examination is consistently deemed reasonable, it can be suggested to include it in the optional pathway; if a key assessment is consistently missing, it can be suggested to add system reminders or mandatory completion items.

[0201] In summary, in the management and quality control of clinical pathways in medical institutions, the assessment of pathway implementation typically relies on manual methods such as random checks of medical records, verification of medical orders, examination and test results, and pathway forms. This method is labor-intensive, inefficient, and highly subjective, making it difficult to continuously, comprehensively, and in real-time detect pathway deviations from a large number of cases. Furthermore, different reviewers may have different criteria for determining whether the same treatment behavior constitutes a deviation, affecting the consistency of deviation detection results. This invention constructs an automated clinical pathway deviation detection process. By automatically acquiring electronic medical record data of target patients, using a large language model to parse the actual treatment behavior sequence, and automatically matching and comparing it with a standard clinical pathway cognitive knowledge graph, it reduces the workload of manual item-by-item verification and improves the coverage, processing efficiency, and consistency of clinical pathway quality control.

[0202] Clinical pathways for single diseases not only include the necessary steps such as examination, diagnosis, treatment, medication, and assessment, but also the sequence of these steps, causal dependencies, mutual exclusion constraints, alternatives, the importance weight of key nodes, and the requirements of each stage of the pathway. Traditional clinical pathways, expressed in text, tables, or fixed rules, struggle to fully represent these complex relationships, limiting systems to simple "do or not" judgments and hindering deviation detection based on pathway logic. This invention constructs a standard clinical pathway cognitive knowledge graph for the target disease, using diagnostic and treatment behaviors as nodes and temporal, causal, and mutually exclusive pathway relationships as edges. Multi-level weights are assigned to nodes and relationships, transforming the standard clinical pathway from static text into a computable, matchable, and reasonable knowledge structure, providing a foundation for subsequent pathway alignment, deviation identification, and deviation quantification.

[0203] Patients' actual medical treatment behaviors are typically scattered across various electronic medical record contents, such as admission records, progress notes, doctor's orders, examination and test results, and discharge summaries. These records include both structured fields and a large amount of unstructured text. Traditional rule-based or field-matching methods struggle to accurately identify the medical behaviors, behavior types, and execution times within the text, and also find it difficult to organize these scattered records into a complete sequence of actual patient medical behaviors. This invention utilizes a large language model to perform semantic parsing on electronic medical record text, extracting actual medical behaviors such as diagnoses, examinations, treatments, medications, assessments, and events, along with their execution times, and forming a sequence of actual medical behaviors in chronological order. This method transforms the originally scattered and unstructured medical records into temporally sequenced behavioral data that can be matched with standard paths.

[0204] In actual clinical practice, the same clinical action may be described in multiple ways. For example, the standard pathway specifies "chest X-ray examination," but the actual medical record may record it as "chest X-ray," "chest imaging examination," or "chest CT examination." Similarly, a medication step in the standard pathway may be replaced by a similar drug or alternative in actual practice. Relying solely on identical names for matching can easily lead to missed matches, mismatches, or failure to identify substitution relationships. This invention establishes a path alignment and matching mechanism between the sequence of actual clinical actions and the cognitive knowledge graph of the standard clinical pathway. For actions with identical names, precise matching is performed; for actions with different names but similar meanings, semantic matching is performed using a large language model or semantic similarity calculation; for cases where the actual action replaces a standard node, the strength weights of mutual exclusion, substitution, or correlation relationships in the knowledge graph are used for judgment, thereby generating a correspondence between the actual action and the standard path node.

[0205] Traditional clinical pathway deviation detection methods often only provide a "compliant" or "non-compliant" result, or simply determine whether a certain node was executed, making it difficult to differentiate the types of deviations. Deviations in real-world clinical scenarios can manifest in various ways, such as the failure to execute key steps, inconsistencies between the treatment sequence and the standard pathway, actual actions replacing standard steps, execution of actions not specified in the standard pathway, and repeated execution of a certain step. Without distinguishing the types of deviations, it is difficult to accurately analyze the causes, severity, and subsequent improvement measures. This invention, based on pathway alignment results, classifies and identifies deviations of actual treatment actions from the standard clinical pathway, including at least: step-missing deviations, sequence disordered deviations, substitution deviations, addition deviations, and duplicate deviations. Specifically, step-missing deviations identify situations where necessary steps in the standard pathway are not executed; sequence disordered deviations identify situations where the actual execution order conflicts with the temporal relationship of the pathway; substitution deviations identify situations where actual treatment actions replace standard steps; addition deviations identify actions not specified in the standard pathway but actually executed; and duplicate deviations identify situations where the same standard node is executed multiple times or matched repeatedly.

[0206] Deviations from clinical pathways are not necessarily irrational. For example, a standard pathway may recommend a certain medication, but if a patient has allergies, contraindications, or special conditions, an alternative medication may be used in practice; a standard pathway may specify a chest X-ray, but a chest CT scan may provide more detailed imaging information; some follow-up examinations, although seemingly repetitive, may be part of the efficacy evaluation. Traditional detection methods struggle to distinguish between reasonable substitutions, beneficial additions, necessary repetitions, and irrational deviations. This invention uses mutual exclusion relationships, relationship strength weights, node importance weights, and deviation type weights in a cognitive knowledge graph to determine the rationality of substitutions, additions, and repetitions. For example, when an actual medication substitutes for a standard medication, the strength weight of the mutual exclusion or substitution relationship between related drugs in the knowledge graph can be used to determine whether it is a reasonable substitution; when an actual action does not belong to a standard pathway node, it can be determined whether it is a beneficial addition or a redundant addition based on medical semantics and pathway context; when an examination is repeated, it can be determined whether it is a necessary repetition by combining the pathway stage and the efficacy evaluation relationship.

[0207] This invention does not simply rely on fixed rules to determine whether a certain examination or treatment has been performed, nor does it simply compare the patient's actual diagnosis and treatment behavior with the standard path text manually. Instead, it constructs a complete technical link around the characteristics of the diagnosis and treatment process of a single disease clinical path, namely, "standard clinical path cognitive knowledge graph - actual diagnosis and treatment behavior sequence - path alignment and matching - deviation identification and classification - deviation degree calculation - deviation report generation - feedback optimization".

[0208] This invention uses a single-disease standard clinical pathway cognitive knowledge graph as the standard pathway knowledge base, and introduces diagnostic and treatment behavior nodes, temporal relationships, causal relationships, mutually exclusive relationships, substitution relationships, and multi-level weights into the deviation detection process. At the same time, it uses a large language model to extract the actual diagnostic and treatment behavior sequence from electronic medical records, and identifies the correspondence between the actual diagnostic and treatment behavior and the standard pathway nodes through semantic matching and path alignment mechanisms. This enables automatic identification, quantitative evaluation, and interpretable output of various deviations such as missing steps, disordered order, substitution, addition, and repetition.

[0209] Unlike clinical pathway quality control methods that rely solely on manual review or fixed rule-based judgment, this invention not only determines whether a patient's treatment process deviates from the standard pathway, but also further identifies the type, location, involved nodes, relationship basis, and severity of the deviation. Unlike ordinary clinical pathway execution deviation identification methods, this invention introduces a single-disease clinical pathway cognitive knowledge graph, using temporal, causal, and mutually exclusive pathway relationships as the basis for deviation detection. It also combines a large language model to automatically extract and semantically align treatment behaviors from electronic medical record text, thereby improving the automation level of deviation detection, semantic understanding ability, deviation classification precision, and result interpretability.

Claims

1. A method for detecting deviation from clinical pathways, characterized in that, Includes the following steps: Obtain a single-disease standard clinical pathway cognitive knowledge graph; the single-disease standard clinical pathway cognitive knowledge graph is used to store the standard diagnosis and treatment process of the target disease and its cognitive relationships. The nodes of the clinical pathway cognitive knowledge graph are the diagnosis and treatment behaviors in the standard clinical pathway, and the edges of the clinical pathway cognitive knowledge graph are the relationships between the nodes. The actual medical treatment behaviors of patients are extracted from the patient data and sorted according to the execution time to form a sequence of actual medical treatment behaviors; Matching and mapping actual diagnosis and treatment behavior sequences with standard clinical pathway cognitive knowledge graphs to generate correspondences between actual diagnosis and treatment behaviors and standard clinical pathway nodes; Based on the correspondence between actual clinical behavior and standard clinical pathway nodes, and combined with the node importance weight, relationship type, relationship strength weight, time window, triggering conditions and / or exceptional evidence in the standard clinical pathway cognitive knowledge graph, it is determined whether the actual clinical behavior deviates from the standard clinical pathway.

2. The clinical pathway deviation detection method according to claim 1, characterized in that, The diagnostic and treatment behaviors in the standard clinical pathway include diagnosis, examination, treatment, medication, assessment, nursing, discharge, and events; each node in the clinical pathway cognitive knowledge graph is assigned a node importance weight; The relationships between diagnostic and treatment behaviors include temporal relationships, causal relationships, mutually exclusive relationships, substitution relationships, conditional triggering relationships, stage attribution relationships, and path variation relationships; the edges of the clinical pathway cognitive knowledge graph are all configured with relationship strength weights.

3. The clinical pathway deviation detection method according to claim 1, characterized in that, Patient data includes electronic medical records, medical orders, examination and test results, nursing records, and discharge summaries; each actual medical procedure includes the name of the medical procedure, the type of medical procedure, the time of execution, the source record, and evidence fragments.

4. The clinical pathway deviation detection method according to claim 1, characterized in that, The specific method for matching and mapping actual clinical behavior sequences with standard clinical pathway cognitive knowledge graphs to generate the correspondence between actual clinical behavior and standard clinical pathway nodes includes the following steps: The actual diagnosis and treatment sequence is standardized to identify the time sequence, type of diagnosis and treatment, execution time, source record and evidence fragments of each actual diagnosis and treatment action; The matching relationship between standardized actual medical behaviors and standard clinical pathway nodes is calculated as follows: For medical behaviors with identical names, exact matching is performed; for medical behaviors with different names but similar meanings, semantic matching is performed using a large language model or semantic similarity model; for cases where the actual medical behavior substitutes for a standard clinical pathway node, substitution matching is performed by combining mutually exclusive relationships, substitution relationships, or relationship strength weights in the standard clinical pathway cognitive knowledge graph; when the same standard clinical pathway node is matched with a standardized actual medical behavior more than once, it is recorded as a duplicate match; when a standard clinical pathway node is not successfully matched by a standardized actual medical behavior, it is recorded as a non-matched actual medical behavior; when a standardized actual medical behavior is not matched with a standard clinical pathway node, it is recorded as a non-matched standard clinical pathway node; that is, the corresponding relationships include exact matching, semantic matching, substitution matching, duplicate matching, non-matched actual medical behavior, and non-matched standard clinical pathway node.

5. The clinical pathway deviation detection method according to claim 4, characterized in that, Specific methods for determining whether actual clinical practice deviates from the standard clinical pathway include: Let the knowledge graph of standard clinical pathways be denoted as G=(V,E); let the sequence of actual patient treatment behaviors be denoted as A={a1,a2,…,an}; let the set of correspondences between actual treatment behaviors and standard clinical pathway nodes be denoted as M; where V is the set of standard clinical pathway nodes, E is the set of relationships between nodes, and an represents the actual treatment behaviors of the bottom n patients. If there is no actual diagnostic or treatment behavior in M ​​that corresponds to the necessary node in the standard clinical pathway, and no exceptional evidence is found, it is judged as a step-missing deviation. If there exists a temporal edge e = (v before, v after) in G, and both are matched with actual treatment actions a before and a after, but the actual execution time T is different... (a前) Later than T (a后) If the execution time exceeds the time specified in G and the allowed time window δ, it is judged as a sequence disorder deviation; where v is a node; If node v should only be executed when condition c is met, but the actual medical treatment a has been executed but the corresponding examination results, diagnosis conclusions, risk assessments or complication records cannot prove that c is met, then it is determined that the condition is not met and execution is deviated. If node v should only be executed when condition c is met, and the actual diagnosis and treatment behavior a has been executed and condition c is met, but v is not executed within the limited time window ΔT after condition c is met, then it is determined as a deviation of not being executed after the condition is triggered. If there exists a causal edge e=(vcause, vef) in G, and the vcause has already occurred in the actual medical treatment sequence and triggered subsequent medical treatment requirements, but the vef has not been executed, has been executed too late, or lacks evidence, then it is determined to be a deviation from the causal chain. If there exists a causal edge e=(vcause,veffect) in G, and vcause has already occurred in the actual medical treatment sequence and triggered subsequent medical treatment requirements, and vef is executed in advance before vcause has occurred or been confirmed, then it is judged as a deviation due to insufficient causal basis. If there exists a mutually exclusive edge e=(v1,v2) in G, and the actual medical behavior sequence shows medical behaviors corresponding to v1 and v2 at the same time, and no explanatory evidence is found, then it is judged as a mutually exclusive conflict deviation. If the actual medical treatment behavior a does not form an exact match with node v in G, but has a substitution relationship with v or the semantic similarity reaches the substitution threshold θ, and the substitution strength, indications, contraindications and patient individualization evidence do not meet the preset conditions, it is judged as a substitution deviation. If the actual medical behavior 'a' cannot match any node in G, its nature is determined based on the semantic relevance of 'a' to the current path stage, diagnostic conclusion, risk event, and neighboring nodes in the graph. When the relevance reaches a preset value and there is supplementary diagnostic, risk control, or efficacy evaluation evidence, it is judged as a beneficial addition. When the relevance is lower than the preset value or there is a lack of medical evidence, it is judged as a redundant addition. When 'a' is related to a strong mutually exclusive relationship, contraindication, or high-risk medication, it is judged as a deviation from the addition. If the same node v in G is repeatedly matched by multiple actual medical behaviors a, check whether there is an allowed relationship of re-examination, follow-up, efficacy evaluation or stage repetition in G; if there is an allowed relationship and the repetition time meets the time window requirements, it is determined to be necessary repetition; if there is no allowed relationship or the repetition behavior lacks clinical evidence, it is determined to be repetition deviation. The score for each deviation is calculated based on the deviation type weight Wt, node importance weight Wv, relationship strength weight We, evidence credibility Ce, and reasonableness adjustment coefficient Ar. The overall degree of deviation for each patient is obtained based on the score for each deviation.

6. The clinical pathway deviation detection method according to claim 5, characterized in that, The expression for calculating the score Di of the i-th deviation is: Di=100×Wt_i×Wv_i×We_i×Ce_i×Ar_i; The deviation type weight Wt_i of the i-th deviation is configured according to the disease and management objectives and calibrated with historical data; The node importance weight Wv_i involved in the deviation of the i-th term is jointly determined by the expert path importance score, the level of evidence of the guideline / norm, the relevance of key clinical events, and the impact of historical case implementation; The relation strength weight We_i involved in the i-th deviation is jointly determined by the relation type strong constraint degree, medical evidence level, time window constraint and historical consistency. The credibility of the evidence involved in the deviation of the i-th item, Ce_i, is used to reflect the reliability of the evidence used to judge the deviation; The reasonableness adjustment coefficient Ar_i involved in the deviation of the i-th term is used to reflect individualized diagnosis and treatment and pathway variation factors.

7. The clinical pathway deviation detection method according to claim 6, characterized in that, The specific methods for obtaining the overall degree of deviation for a corresponding patient based on the score of each deviation include: The overall deviation score D is obtained based on the score of each deviation of the patient, and its expression is: D=100×[1 ∏(1 Di / 100)]; Based on the overall deviation score D, the patient's overall deviation degree is divided into slight deviation, moderate deviation, significant deviation, and severe deviation.

8. A system for implementing the clinical pathway deviation detection method according to any one of claims 1 to 7, characterized in that, include: The data layer is used to obtain a cognitive knowledge graph of standard clinical pathways for single diseases; The actual medical treatment behaviors of patients are extracted from the patient data and sorted according to the execution time to form a sequence of actual medical treatment behaviors; The alignment layer is used to match and correspond actual diagnosis and treatment behavior sequences with standard clinical pathway cognitive knowledge graphs, generating correspondences between actual diagnosis and treatment behaviors and standard clinical pathway nodes; The detection layer is used to determine whether the actual diagnosis and treatment behavior deviates from the standard clinical pathway based on the correspondence between the actual diagnosis and treatment behavior and the nodes of the standard clinical pathway, combined with the node importance weight, relationship type, relationship strength weight, time window, triggering conditions and / or exceptional evidence in the cognitive knowledge graph of the standard clinical pathway. The application layer is used to provide output deviation reports, risk warnings, quality analysis, and path optimization suggestions.

9. The system according to claim 8, characterized in that, Deviation reports include basic patient information, target disease, testing time, standard pathway version, deviation overview, deviation details, deviation severity score, deviation grade, source of evidence, and / or improvement suggestions; The deviation details include deviation type, deviation location, involved standard nodes, involved actual diagnosis and treatment behavior, matching status, deviation description, node importance weight, relationship strength weight, deviation score contribution, and suggested handling measures.

10. The system according to claim 8, characterized in that, Also includes: The self-updating layer updates the cognitive knowledge graph of the standard clinical pathway for a single disease based on deviation detection results, manual review results, deviation frequency statistics, and pathway execution data, and / or serves as the criterion for determining whether actual diagnosis and treatment behavior deviates from the standard clinical pathway.

Images