Clinical trial protocol execution monitoring method and system based on clause semantic parsing

Abstract

The present application relates to the technical field of clinical trial monitoring, and particularly relates to a clinical trial plan execution monitoring method and system based on clause semantic analysis, comprising: splitting a plan text, identifying operational clauses and reference units, extracting element information containing operation objects and judgment parameters, identifying reference, premise, modification, exception correlation between clauses; constructing a dependency graph with clauses and reference units as nodes and correlation as directed edges, backfilling along backfill edges and covering judgment parameters according to modification relationship to obtain parameter-completed element information; generating structured judgment rules according to matched judgment strategy templates, determining a unique effective rule based on priority edges when multiple rules are hit; and comparing execution records with rules to obtain deviation identification results. The technical scheme of the present application can reduce dependence on artificial experience, improve the automation degree of plan execution deviation identification and the overall monitoring efficiency.

Description

Technical Field

[0001] This invention relates to the field of clinical trial monitoring technology, and in particular to a method and system for monitoring the execution of clinical trial protocols based on clause semantic parsing. Background Technology

[0002] Currently, monitoring the compliance of clinical trial protocol execution mainly relies on clinical monitors manually comparing protocol clauses with execution records. This approach is inefficient, has limited coverage, and struggles to promptly detect deviations caused by cross-clause correlations. Existing automated protocols primarily use keyword or pattern matching to extract information such as visit windows and testing thresholds from the protocol text and generate static judgment rules. However, these methods treat each clause as an independent unit, failing to consider the structural characteristics of the clinical trial protocol itself.

[0003] In clinical trial protocols, the criteria for operational clauses are often not fully contained within the clause itself, but rather scattered throughout the visit schedule master table and its footnotes, laboratory normal value reference tables, protocol amendments, etc. Furthermore, there are often interrelationships between clauses, such as references, amendments, and exceptions. Existing methods that match each clause independently are prone to generating incomplete or inconsistent criteria when processing such clauses due to the lack of features within the clause text, leading to missed or incorrect judgments of deviation monitoring. Simultaneously, when the same execution record simultaneously matches a general rule and its modified or exception rules, existing methods lack a mechanism to determine the effective rule based on protocol relationships.

[0004] Patent publication number CN116864050A discloses a clinical trial quality control method for semi-quantitative assessment of protocol deviation. This method acquires data such as clinical trial source documents and reports, preprocesses them into structured data, extracts element information through semantic extraction, compares this element information with standard element information in preset rules, marks protocol deviation events, and determines corresponding intervention measures and impact assessments. However, the preset rules in this technical solution are directly derived from existing documents such as inclusion / exclusion criteria and standard operating procedures. When processing complex original text, incomplete or inconsistent judgment criteria can easily occur, making it difficult to improve the accuracy of deviation assessment results. Summary of the Invention

[0005] In view of this, the present invention provides a clinical trial protocol execution monitoring method and system based on clause semantic parsing, to solve the technical problem that existing monitoring methods are prone to incomplete or inconsistent judgment criteria when there are complex relationships such as references and modifications between clinical trial protocol clauses, which in turn affects the accuracy of execution deviation identification.

[0006] The technical solution of this invention is implemented as follows: On one hand, this invention provides a clinical trial protocol execution monitoring method based on clause semantic parsing, comprising the following steps: S1. The clinical trial protocol text is split to identify operational clauses and reference units. Elements are extracted from the operational clauses and reference units to obtain element information containing operational objects and judgment parameters. The relationships between operational clauses and between operational clauses and reference units are identified. The relationships include reference relationships, premise relationships, modification relationships and exception relationships. S2. Construct a dependency graph using operational clauses and parameter units as nodes and relationships as directed edges. The directed edges corresponding to reference relationships, modification relationships, and prerequisite relationships are used as backfill edges, and the directed edges corresponding to exception relationships are used as priority edges. Backfill the decision parameters in the feature information along the backfill edges corresponding to reference relationships and prerequisite relationships, and overwrite the backfilled decision parameters according to the order of the scheme amendment versions along the backfill edges corresponding to modification relationships to obtain the feature information with complete parameters. S3. Match the judgment strategy template with the element information completed by the parameters and generate structured judgment rules; when the same execution record hits multiple structured judgment rules at the same time, determine the effective priority of each structured judgment rule based on the priority edge, and determine the unique effective rule based on the effective priority. S4. Obtain clinical trial execution records, compare the execution records with structured judgment rules, and if a unique effective rule has been determined, the unique effective rule shall prevail. Obtain deviation identification results based on the comparison results.

[0007] Based on the above technical solutions, preferably, step S1 includes the following steps: S11. Segment the clinical trial protocol text according to clause number, table boundary and footnote number to obtain clause units, table units and footnote units, and record the position index of each unit in the original text; S12. Perform binary classification on the clause units, identify the operational clauses, and parse the table units and footnote units into reference units; S13. Extract elements from operational clauses and parameter units. The judgment parameters in the element information include constraint type and constraint value. The value range of the constraint type includes numerical constraint, time constraint, triggering condition and prior dependency. Obtain the operation object and identify the association relationship by combining the candidate guide word and the relationship classification model.

[0008] Based on the above technical solutions, preferably, the table unit is parsed to obtain the table reference unit corresponding to the visit plan master table and the numerical reference unit corresponding to the laboratory normal value reference table, the footnote unit is parsed to obtain the footnote reference unit, and the table reference unit and the footnote reference unit are labeled with reference relationships. The method of identifying association relationships by combining candidate keywords and relationship classification models specifically includes: According to the preset guide word list, the candidate related pairs are located between the operational clauses and the reference units and operational clauses. Among them, the reference guide words include clause number, table number and words indicating the action of looking up; the amendment guide words include amendment version number and words indicating parameter changes; the exception guide words include words indicating the limitation or exclusion of the scope of application; and the prerequisite guide words include words indicating the conditions for the completion of the preceding operation. The candidate association pairs are input into the relationship classification model. The relationship classification model outputs the probability distribution of each association type. The association type with the highest probability is taken as the output association, and the probability corresponding to the output association is used as the classification confidence. Candidate association pairs with a classification confidence lower than a preset threshold are not associated and are transferred to manual review.

[0009] Based on the above technical solutions, preferably, step S2 specifically includes: S21. Construct a dependency graph using operational clauses and supply units as nodes and relationships as directed edges; S22. Initialize the observation mask and values ​​for each node in the dependency graph. The observation mask component corresponding to the dimension for which the decision parameters have been extracted is set to 1 and the actual value is used as the initial value. The observation mask component corresponding to the dimension for which the decision parameters have not been extracted is set to 0 and the value is empty. S23. According to the iteration round, backfill the decision parameters along the backfilling edges corresponding to the reference relationship and the premise relationship, and cover the backfilled decision parameters according to the order of the scheme amendment version along the backfilling edges corresponding to the correction relationship, and update the observation mask, value and source table. S24. When the changes in the mask between two adjacent observation rounds meet the convergence condition or the number of iteration rounds reaches the preset maximum number of iterations, the iteration stops and the parameter-completed element information is obtained.

[0010] Based on the above technical solutions, preferably, in step S21, the node attributes in the dependency graph include location index and feature information. The directed edges corresponding to the reference relationship point from the referenced node to the referencing node and are marked as backfill edges. The directed edges corresponding to the correction relationship point from the operational clause corresponding to the revised version of the later scheme to the corrected operational clause and are marked as backfill edges. The directed edges corresponding to the exception relationship point from the operational clause corresponding to the general rule to the operational clause corresponding to the exception rule and are marked as priority edges.

[0011] Based on the above technical solution, preferably, in step S24, the convergence condition is: ; In the formula, For node indexing, For iteration rounds, For the next iteration round, For nodes In the The observation mask of the wheel, This is the preset convergence threshold.

[0012] Based on the above technical solutions, preferably, in step S3, generating structured decision rules includes the following steps: S31. For each operational clause, check whether all slot combinations of the judgment strategy template have been marked as observed in the observation mask after the operational clause converges, and determine the template that meets the condition as the instantiable template of the operational clause. S32. Based on the mapping relationship between the slots of the instantiable template and the corresponding judgment parameters in the parameter-completed element information, fill the judgment parameters into the slots to obtain candidate rules. S33. When multiple candidate rules corresponding to the same operational clause generate conflicting conclusions, calculate the matching evidence score of each candidate rule, retain the candidate rule with the highest matching evidence score for instantiation, and obtain the structured judgment rule.

[0013] Based on the above technical solution, preferably, in step S3, when the same execution record simultaneously hits multiple structured decision rules, the edges corresponding to the reference relationship and the edges corresponding to the already covered correction relationship are excluded. A priority graph is constructed using the priority edges corresponding to the exception relationship. Cycle detection is performed on the priority graph, and structured decision rules with cycles are marked as pending manual review. For priority graphs without cycles, the effective priority is determined based on the reachability relationship between the nodes corresponding to the hit multiple structured decision rules. Among the hit multiple structured decision rules, the rule whose corresponding node can reach all other hit rule corresponding nodes is determined as the unique effective rule. If there is a cycle in the priority graph, or the nodes corresponding to multiple hit rules are mutually unreachable, it is determined that no unique effective rule has been determined, and the process is transferred to manual review.

[0014] Based on the above technical solutions, the preferred method for calculating the matching evidence score is as follows: ; In the formula, For operational terms index, For template index, Operational terms With template The matching evidence score for the corresponding candidate rule. Operational terms The converged observation mask, template The weight vector corresponding to the slot. To and A vector of all 1s with the same dimension. This is a preset constant.

[0015] Based on the above technical solutions, preferably, step S4 includes: S41. Obtain clinical trial execution records and compare the execution records with the judgment parameters of the structured judgment rules one by one according to each judgment parameter dimension; S42. When determining that the execution record deviates from the structured judgment rule based on the comparison results, the deviation identification results are divided into three levels: general deviation, major deviation, and serious deviation according to the matching evidence score, deviation magnitude, and the importance of the operational clause that triggers the structured judgment rule. S43. Archive the operational clause location index, judgment parameter tracing, triggered structured judgment rules, and corresponding execution records associated with the deviation identification results.

[0016] On the other hand, the present invention also provides a clinical trial protocol execution monitoring system based on clause semantic parsing, comprising: The clause splitting and element extraction module is used to split the clinical trial protocol text, identify operational clauses and reference units, extract elements from the operational clauses and reference units to obtain element information containing operational objects and judgment parameters, and identify the relationships between operational clauses and between operational clauses and reference units. The relationships include reference relationships, amendment relationships and exception relationships. The dependency graph construction and parameter backfilling module is used to construct a dependency graph with operational clauses and parameter units as nodes and relationships as directed edges. Directed edges corresponding to reference relationships, modification relationships, and prerequisite relationships are used as backfilling edges, and directed edges corresponding to exception relationships are used as priority edges. Backfilling is performed on the judgment parameters in the feature information along the backfilling edges corresponding to reference relationships and prerequisite relationships. Overwriting is performed on the backfilled judgment parameters according to the order of the scheme amendment versions along the backfilling edges corresponding to modification relationships, so as to obtain the feature information with complete parameters. The rule generation and priority determination module is used to match the judgment strategy template based on the element information of the parameter completion and generate structured judgment rules. When the same execution record hits multiple structured judgment rules at the same time, the effective priority of each structured judgment rule is determined based on the priority edge, and a unique effective rule is determined based on the effective priority. The execution comparison module is used to obtain clinical trial execution records, compare the execution records with structured judgment rules, and if a unique effective rule has been determined, the unique effective rule shall prevail. The deviation identification result is obtained based on the comparison result.

[0017] The present invention has the following advantages over the prior art: (1) This invention automatically identifies operational clauses and their relationships by splitting and semantically parsing the clinical trial protocol text, and generates structured judgment rules for execution record comparison. Compared with monitoring methods that rely on manual comparison or keyword matching, it can reduce the reliance on human experience and improve the automation level of protocol execution deviation identification and overall monitoring efficiency.

[0018] (2) By constructing a dependency graph containing operational clauses and reference units, and backfilling and updating the judgment parameters along the backfilling edge corresponding to the reference relationship and according to the coverage order corresponding to the modification relationship, the judgment parameters that were originally scattered in different clauses, tables or amendments can be associated and integrated into the element information of the same operational clause, avoiding the situation where the judgment basis is missing or inconsistent due to the failure to identify the reference and modification relationship between clauses, thereby improving the completeness and consistency of the information on which the generated judgment rules are based.

[0019] (3) In the case of multiple rules hitting the same operation object at the same time, the present invention performs loop detection and reachability judgment based on the priority edge corresponding to the exception relationship, and determines the only effective rule that the corresponding node can be reached by the other hit rule nodes. In the case of the coexistence of general provisions and exception provisions in the scheme, the application order between rules can be clarified, and the deviation from the judgment result due to rule conflict or unclear application order can be avoided.

[0020] (4) The present invention generates structured judgment rules by matching judgment strategy templates according to whether the judgment parameters are complete, and classifies and archives them in conjunction with deviation indicators in the execution record comparison stage, so that the deviation identification results have distinguishable severity and traceable judgment basis, which makes it convenient for monitoring personnel to verify and process the identified deviations. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the clinical trial protocol execution monitoring method of the present invention; Figure 2 This is a schematic diagram of the dependency graph structure of the present invention; Figure 3 This is a schematic diagram of the iterative backfilling and overlay process of the determination parameters in this invention; Figure 4 This is a structural diagram of the clinical trial protocol execution monitoring system of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, this invention provides a clinical trial protocol execution monitoring method based on clause semantic parsing, comprising the following steps: S1. The clinical trial protocol text is split to identify operational clauses and reference units. Elements are extracted from the operational clauses and reference units to obtain element information containing operational objects and judgment parameters. The relationships between operational clauses and between operational clauses and reference units are identified. The relationships include reference relationships, premise relationships, modification relationships and exception relationships. S2. Construct a dependency graph using operational clauses and parameter units as nodes and relationships as directed edges. The directed edges corresponding to reference relationships, modification relationships, and prerequisite relationships are used as backfill edges, and the directed edges corresponding to exception relationships are used as priority edges. Backfill the decision parameters in the feature information along the backfill edges corresponding to reference relationships and prerequisite relationships, and overwrite the backfilled decision parameters according to the order of the scheme amendment versions along the backfill edges corresponding to modification relationships to obtain the feature information with complete parameters. S3. Match the judgment strategy template with the element information completed by the parameters and generate structured judgment rules; when the same execution record hits multiple structured judgment rules at the same time, determine the effective priority of each structured judgment rule based on the priority edge, and determine the unique effective rule based on the effective priority. S4. Obtain the clinical trial execution records and compare them with the structured judgment rules. If a unique effective rule has been determined, the unique effective rule shall prevail, and the deviation identification result shall be obtained based on the comparison result. If a unique effective rule has not been determined, the case shall be transferred to manual review.

[0025] In one embodiment of the present invention, step S1 includes the following steps: S11. Segment the clinical trial protocol text according to clause number, table boundary and footnote number to obtain clause units, table units and footnote units, and record the position index of each unit in the original text.

[0026] Specifically, the clinical trial protocol text to be monitored is obtained. This text includes the main text, a summary visit plan, a reference table of normal laboratory values, appendices, and amendments—all unstructured or semi-structured documents. The protocol text is split into clause units, table units, and footnote units according to structural markers such as chapter titles, clause numbers, table boundaries, and footnote numbers. The position index of each unit in the original text is recorded. After splitting, two text views are retained: one retains the cited phrases and clause numbers, serving as a view for extracting relationships; the other only removes non-semantic characters such as footnote superscripts, page breaks, headers, and footers, without deleting citation or amendment clue words such as "see," "according to," "as per Appendix X," and "corrected to," serving as a view for extracting elements.

[0027] S12. Perform binary classification on the clause units, identify the operational clauses, and parse the table units and footnote units into reference units.

[0028] For each clause unit, a binary classification process is performed based on a pre-trained language model to identify operational clauses that clearly specify the execution requirements for procedures, examinations, medication, and follow-ups, distinguishing them from descriptive clauses such as background information and ethical statements. Clause units not identified as operational clauses are retained as descriptive clauses in the original text and do not participate in subsequent element extraction and rule generation. Table units and footnote units are directly parsed into reference units and do not participate in binary classification. Specifically, table units are parsed to obtain table reference units corresponding to the visit plan master table and numerical reference units corresponding to the laboratory normal value reference table, while footnote units are parsed to obtain footnote reference units. For the master visit plan table, a two-dimensional table structure is identified, with row names representing inspection or operation items and column names representing visit names and specified time points in the plan. For cells containing execution markers, a table reference unit is generated. The operation object is a combination of the row item and the column visit, the constraint type is time constraint, and the constraint value is the time point in that column combined with the time window in the corresponding footnote. The trigger condition is empty. The footnote text corresponding to the footnote number on the cell is parsed separately as a footnote reference unit, and a reference relationship is marked between it and the table reference unit, enabling the two-hop reference relationship between the main text clause referencing the table and the table unit referencing the footnote to be propagated on the dependency graph. For the laboratory normal value reference table, the rows are taken as inspection items, the columns as applicable groupings, and the cells as reference upper and lower limits, parsed as numerical reference units. The reference units only provide parameters and do not generate separate structured decision rules.

[0029] S13. Extract elements from operational clauses and parameter units. The judgment parameters in the element information include constraint type and constraint value. The value range of the constraint type includes numerical constraint, time constraint, triggering condition and prior dependency. Obtain the operation object and identify the association relationship by combining the candidate guide word and the relationship classification model.

[0030] For operational clauses, a named entity recognition and relation extraction model fine-tuned for the clinical trial field is used to extract element information. Entity types include at least the operation object, numerical value, comparison operator, protocol-specified time point, time window tolerance, preceding events, and clause or table number. The annotation system adopts BIO sequence annotation, where the protocol-specified time point refers to the planned execution time point given in the protocol, and the time window tolerance refers to the allowed offset before and after that time point. When there are multiple parallel or nested elements in the same operational clause, semantic role annotation is used to distinguish primary and secondary constraints, forming multiple sets of element information. Each set includes the operation object, constraint type, and constraint value. The value range of the constraint type is numerical constraint, time constraint, triggering condition, and preceding dependency. For parameter units, no clause-level named entity recognition is performed. Instead, the corresponding element information is directly generated from table structure parsing, so that parameters in the table such as visit windows and reference values ​​have the same representation form as the element information of the operational clauses.

[0031] Furthermore, the method of identifying relationships by combining candidate guidance words with a relationship classification model specifically includes: locating candidate relationship pairs between operational clauses and reference units, and operational clauses, according to a pre-set guidance word list. The pre-set guidance word list includes citation guidance words, amendment guidance words, exception guidance words, and prerequisite guidance words. Citation guidance words include clause numbers, table numbers, and words indicating the action of consulting, such as "see," "refer to," "according to," and "according to the appendix." Amendment guidance words include amendment version numbers and words indicating parameter changes, such as "amended to," "revised to," and "changed from," etc., which imply amendment. Exception guidance words include words indicating limitations or exclusions in the scope of application, such as "except," "if," and "not applicable." Prerequisite guidance words include words indicating the conditions for completing a preceding operation, such as "must be completed after," "should be completed first," and "performed after." The candidate relationship pairs are then input into the relationship classification model. The relationship classification model uses a pre-trained language model fine-tuned for the clinical trial domain. Pre-trained language models include BERT, BioBERT (pre-trained for biomedical literature), ClinicalBERT (pre-trained for clinical text), and BERT-wwm (pre-trained for Chinese text), etc., with fine-tuning of the classification layer on labeled clinical trial protocol text. The relationship classification model and the named entity recognition used for feature extraction share the same pre-trained language model, connected to the relationship classification head and named entity recognition head respectively. Using the two text units of a candidate association pair and their context as input, the classification layer outputs the probability distributions for citation relationships, modification relationships, exception relationships, premise relationships, and no-relationship categories. The category with the highest probability is taken as the classification result, and this highest probability is used as the classification confidence score. For classification confidence scores not lower than a preset threshold, corresponding associations are established according to the classification results; candidate association pairs with classification confidence scores lower than the preset threshold are not established and are transferred for manual review. For association pairs identified as modification relationships, the amendment version number or effective date is recorded; for association pairs identified as exception relationships, the pointing relationship between the operational clauses corresponding to the general rules and the operational clauses corresponding to the exception rules is clarified. Prerequisite relations are only used for filling in the parameters of the preceding dependency dimension and are not included in the priority determination.

[0032] In one embodiment of the present invention, such as Figure 2 and Figure 3 As shown, step S2 specifically includes: S21. Construct a dependency graph using operational terms and supply units as nodes and relationships as directed edges.

[0033] A pre-built, extensible decision strategy template library is constructed. This embodiment uses four common template types as examples: single-value threshold, interval range, conditional branch, and dependent on preceding events. The template library is not limited to these four types; templates can be added later based on new constraint patterns emerging in clinical trial protocols, and the feature dimensions and the required prototypes of the templates can be expanded accordingly. Definitions are aligned with the decision elements. dimensional feature space, dimensional index In this embodiment, the corresponding elements are, in order: operation object, numerical constraint, time constraint, triggering condition, and prior dependency. When a new template requires slots beyond the above five dimensions, such as set constraints for combined medication or multi-level dosage adjustment levels, the corresponding dimension is appended to the end of the feature space, and a mandatory dimension weight vector is configured for the new template. There are templates Take on the newly added dimension .template Using demand prototype vectors express, This indicates that the template was instantiated at the first... The dimension is a required matching dimension, given a priori by the template slot. Required slots include: single-value threshold type (required operation object and numerical constraint); interval range type (required operation object and time constraint, or upper and lower bounds of the operation object and numerical constraint); in either combination, the required slots are covered by the observation mask, allowing instantiation; conditional branch type (required operation object, triggering condition, and one of the numerical or time constraints); and dependent precedent event type (required operation object and its preceding dependency).

[0034] Using all operational clauses and parameter units as the node set, with node attributes including location index and feature information, a corresponding directed edge is established for each labeled relationship, forming a dependency graph. Specifically, directed edges corresponding to reference relationships point from the referenced node to the referencing node and are labeled as backfill edges; directed edges corresponding to modification relationships point from the operational clause corresponding to the later modified version of the scheme to the modified operational clause and are labeled as backfill edges; operational clauses corresponding to general rules refer to operational clauses that provide universally applicable requirements for a certain operational object; operational clauses corresponding to exception rules refer to operational clauses that provide different requirements for the same operational object under limited conditions. Directed edges corresponding to exception relationships point from the former to the latter and are labeled as priority edges. Directed edges corresponding to prerequisite relationships point from the node corresponding to the preceding operation to the operational clause dependent on that preceding operation and are labeled as backfill edges. Parameter units only serve as the starting point of edges and do not generate decision rules. Backfill edges include edges corresponding to reference relationships, modification relationships, and prerequisite relationships, used for parameter passing and coverage; priority edges only include edges corresponding to exception relationships, used to determine the order of effectiveness when both general and exception rules are simultaneously triggered. Equip the backfill edge with a transferable 3D mask The reference edge is selected based on the numerical constraints, time constraints, and preorder dependencies. Only decision parameter dimensions that have been observed by the cited node but are missing in this node are allowed to be added to this node, without overwriting existing values ​​of this node; the edges corresponding to the correction relationship are selected in the numerical constraint and time constraint dimensions. Amended values ​​can override the original values ​​of a node; exception edges do not participate in this override operation, and the triggering conditions and limited parameters extracted from the exception clause itself are used to generate independent structured decision rules; prerequisite edges are taken in the preceding dependency dimension. Only fill in the blanks, do not overwrite, and do not enter the priority graph. The above mask is given in advance as a configurable rule table for the domain.

[0035] S22. Initialize the observation mask and values ​​for each node in the dependency graph. The observation mask component corresponding to the dimension for which the decision parameters have been extracted is set to 1 and the actual value is used as the initial value. The observation mask component corresponding to the dimension for which the decision parameters have not been extracted is set to 0 and the value is empty.

[0036] For each node, establish a value table and a source table that correspond one-to-one with the decision parameter dimensions. The value table records the actual values ​​of each decision parameter dimension, and the source table records the source node, association type, and amendment version of each decision parameter dimension's value. Node at the... The wheel maintains three quantities simultaneously: observation mask. , quantity Indicates the first Whether the dimension has been observed; true value table , Store the The actual parameter value of the dimension, not Traceability Table , Record number The observation mask is used only for matching calculations. When filling in parameters, the actual value table must be read and the source traced back through the source table. The element information extracted from the node itself provides the initial observation mask, initial value, and initial source record: If a certain decision parameter dimension has been extracted, the observation mask component is 1, and the extracted actual value is written into the corresponding component of the value table as the initial value. The source table record shows the source node as this node, the relationship type as empty, and the amendment version as empty. For dimensions that have not been extracted, the observation mask component is 0, the corresponding component in the value table is empty, and the source table is empty. An empty value only indicates that the dimension has not been observed and does not mean that the dimension has no constraints. Operational clauses do not calculate template attribution during the initialization phase; parameter units do not calculate attribution or generate decision rules.

[0037] S23. According to the iteration round, backfill the decision parameters along the backfill edges corresponding to the reference relationship and the premise relationship, and perform coverage on the corresponding dimension of this node according to the order of the scheme amendment version along the backfill edges corresponding to the correction relationship. Regardless of whether the dimension has been backfilled, update the observation mask, value and source table.

[0038] Let the iteration round be... , The maximum number of iterations is preset, and the default value does not exceed the node size. This is acceptable in engineering practice. In each round, all operational clauses are filled in sequentially along the reference edges and prerequisite edges pointing to this node. Only filling in the blanks is performed; no overwriting is done. The filling dimension provided by neighboring nodes is determined by the following formula: ; In the formula, For the current node index, Indexing neighboring nodes, For feature dimension index, For iteration rounds, For nodes No. Vi in the The observation mask components of the wheel, For nodes Pointing to node The edge type in the first Transitive dimension mask on dimension 1 This is an indicator function. That is, only dimensions that are not yet observed by this node, have been observed by neighboring nodes, and whose edge type allows for propagation can be filled in. If multiple references or prerequisite sources exist for the same dimension and have the same value, that value is adopted; if the sources are the planned execution time point and the time window tolerance respectively, they are concatenated into a complete time window value and are not considered a conflict; if the values ​​conflict, this dimension is not automatically backfilled and is marked as awaiting manual confirmation. If the source is unique or the values ​​are consistent, then let... , , ,in, This represents the record in the traceability table for node i at dimension r. The default backfill step size is... After a successful backfill, the observation mask component corresponding to that dimension is set to 1, indicating that the dimension has been observed; if set to Then the mask is based on Updated, value retrieval and tracing are still the first times in this dimension... When the value is changed to positive, it is written, and subsequent filling will not rewrite the existing value.

[0039] After all nodes have completed this round of filling, a coverage operation is performed along the backfilling edges corresponding to the correction relationships (hereinafter referred to as correction edges). Coverage is only performed on correction edges and is conducted after all nodes have completed this round of filling to avoid overlapping with other filling operations. For the first... Dimension, collect all that is satisfied and The amendment neighbors are sorted in descending order of the version number or effective date of the corresponding amendment, with the most recently effective one being selected. Perform coverage on this dimension: ; In the formula, For the neighbors of the latest effective amendment, This refers to the version number or effective date of the amendment. When multiple amendments exist for the same dimension, the later-effective amendment covers the earlier-effective one, and the values ​​of multiple amendments are not summed or coexisted; dimensions without amendment neighbors retain the fill-in results. After the coverage is completed, the modified operational clause retains only one element information with updated parameters; the original parameters before the coverage are no longer retained as independent element information. During the fill-in and coverage processes, the source of the value of each decision parameter dimension is synchronously written to the corresponding component of the source table: when filling in along the backfill edge corresponding to the reference relationship, the corresponding component of the source table is updated to have the source node, reference relationship, and amendment version empty; when covering along the backfill edge corresponding to the correction relationship, the corresponding component of the source table is updated to have the source node, correction relationship, and the version number or effective date of the amendment. The source table is used to trace back the source node, association type, and amendment version of each decision parameter value after parameter completion.

[0040] The following description uses a specific example, and assumes operational clauses. The document should specify "The second visit will be shown on the visit schedule," and the corresponding form should be provided to the relevant unit. The value is Day 14, and its footnote provides the parameter unit. Values Heavens, the operational provisions corresponding to a certain amendment to the plan. The second visit window has been revised to Day 14. (Day). During initialization. Lacking a time constraint dimension; the first iteration follows the reference edge. , Enter Day 14 ( (day); then along the correction edge with The value range covers Day 14 ( (days); after coverage is complete. The time constraint dimension is Day 14 ( (sky), tracing the source respectively points to , , .

[0041] The following description uses a specific example, and assumes operational clauses. The document should specify "The second visit will be shown on the visit schedule," and the corresponding form should be provided to the relevant unit. The footnote parameter supply unit has been incorporated during the analysis. The value is Day 14 (±3 days), representing the operational clauses corresponding to a certain amendment to the plan. Revise the second visit window to Day 14 (±5 days). During initialization... Lacking a time constraint dimension; the first iteration follows the reference edge. Enter Day 14 (±3 days), this dimension points to Then along the correction edge The value range is covered for Day 14 (±5 days). After coverage is complete, The time constraint dimension is set to Day14 (±5 days), and this dimension only points to... The record shows the relationship type as a revision relationship and the corresponding amendment version.

[0042] S24. When the changes in the mask between two adjacent observation rounds meet the convergence condition or the number of iteration rounds reaches the preset maximum number of iterations, the iteration stops and the parameter-completed element information is obtained.

[0043] Preset convergence threshold When the change in the mask between two adjacent observation rounds is less than this threshold, or the number of iterations reaches the preset maximum number of iterations. When the iteration stops, the observation mask, values, and source table of each node are output as the element information for parameter completion. The convergence condition is determined by the following formula: ; In the formula, For node indexing, For iteration rounds, For nodes In the The observation mask of the wheel, Let the preset convergence threshold be denoted as . The converged value is then denoted as . The source tracing results are The observation mask result is , The values ​​for the decision parameters of the node are determined. This is a source table to record the origin nodes of the decision parameters. The backfill operation ensures that the observation mask components corresponding to each decision parameter dimension are monotonically non-decreasing and have an upper bound. The overwrite operation is a finite number of deterministic replacements in the order of amendment versions. The above iterations can satisfy the convergence condition within a finite number of steps.

[0044] Figure 3The diagram illustrates the iterative backfilling and covering process of decision parameters. The figure uses four types of decision parameters—object, value, time, and preceding order—as columns, and nodes A, B, C, and D as rows, showing the changes in the observation status of each node's parameters in three consecutive states during parameter completion. At time t0, the initial observation mask state is shown, where each node has only some parameters that have been observed (green), while the remaining parameters are in an unobserved null value state (gray hollow circles). At time t1, the state is represented by the state after parameter filling is completed along the backfilling edge in the dependency graph; some previously unobserved parameters have been backfilled and now have values ​​(yellow). At time t2, the final convergence state is achieved after further parameter covering along the correction edge. Parameters covered by the correction edge are marked with red diamonds. After covering, the observation mask no longer changes, and iteration stops. Unobserved dimensions remain null, and it is not required that all nodes be filled.

[0045] In one embodiment of the present invention, generating structured decision rules includes the following steps: S31. For each operational clause, check whether all slot combinations of the decision strategy template have been marked as observed in the observation mask after the operational clause has converged, and determine the template that meets the conditions as the instantiable template of the operational clause.

[0046] For each operational clause, examine its converged observation mask. Whether to cover all mandatory dimensions of a given decision strategy template. For each operational clause, check if any slot combination in the template library is covered by the observation mask. Templates meeting this condition are identified as instantiable templates for that operational clause. Whether to generate structured decision rules, not based on whether the matching evidence score exceeds a preset threshold, but based on whether all mandatory slots of the decision strategy template are covered by the converged observation mask of the operational clause. If no template meets this condition, the operational clause does not generate structured decision rules separately; it is only processed as a feature provision node. If the converged observation mask of a pure amendment clause still cannot cover the mandatory slots of any template, then no structured decision rules are generated separately; they are only used to cover the parameters of the amended clause.

[0047] S32. Based on the mapping relationship between the slots of the instantiable template and the corresponding judgment parameters in the parameter-completed element information, fill the judgment parameters into the slots to obtain candidate rules.

[0048] Based on the mapping relationship between the slots of the instantiable template and the corresponding decision parameters of the operational clause, the decision parameters are filled into the slots to obtain candidate rules. The parameter filling is not only based on the observation mask taken from the original feature of this clause; if a dimension is backfilled or overwritten by neighbors, it must be based on the source table. The referenced source node reads the value from that source. The rules also record the source clause number, the template type used, and whether each parameter is an original clause or obtained by backfilling or overwriting from another clause.

[0049] S33. When multiple candidate rules corresponding to the same operational clause generate conflicting conclusions, compare the matching evidence scores calculated for each candidate rule, retain the candidate rule with the highest matching evidence score for instantiation, and obtain a structured decision rule.

[0050] When multiple instantiable templates exist for the same operational clause, and the corresponding candidate rules constrain different operational objects or complementary slots, all candidate rules are retained. When multiple candidate rules constrain the same operational object and their conclusions are mutually exclusive, the matching evidence score of each candidate rule is calculated using the following formula, and the one with the highest score is retained for instantiation: ; In the formula, For operational terms index, For template index, Operational terms With template The matching evidence score for the corresponding candidate rule. Operational terms The converged observation mask, template The weight vector corresponding to the slot. To and A vector of all 1s with the same dimension. This is a preset constant used to prevent the denominator from being zero. For templates with alternative slot combinations, such as two observation mask coverage methods of interval range type, calculations are performed separately for each combination. The highest score is taken as the matching evidence score for that template. If multiple candidate rules have the same matching evidence score, the operational clause is marked as pending manual confirmation.

[0051] S34. In cases where the same execution record simultaneously hits multiple structured decision rules, the effective priority of each structured decision rule is determined based on priority edges. Specifically, this includes: excluding edges corresponding to reference relationships and edges corresponding to already covered correction relationships, and constructing a priority graph only using priority edges corresponding to exception relationships; performing cycle detection on the priority graph, i.e., performing strong connected component detection on the subgraph derived from exception edges, excluding reference edges and correction edges; if a directed cycle exists in the priority graph, the structured decision rule containing the cycle is marked as awaiting manual review and will not be included in subsequent automatic comparisons. The directed edges corresponding to exception relationships point from the operational clauses corresponding to general rules to the operational clauses corresponding to exception rules. Therefore, in a priority graph without cycles, the node corresponding to the exception rule is the sink that the node corresponding to the general rule can reach along the priority edge. For any execution record simultaneously hitting multiple structured decision rules, the rule that is reachable from all other hit rule nodes among the hit rule nodes is determined as the unique effective rule; if the hit rule nodes are mutually unreachable, the unique effective rule is not output, and the process proceeds to manual review.

[0052] In one embodiment of the present invention, step S4 includes: S41. Obtain clinical trial execution records and compare the execution records with the judgment parameters of the structured judgment rules one by one according to each judgment parameter dimension.

[0053] Acquire clinical trial execution records, which include information such as the operation target, numerical results, occurrence time, preceding event identifiers, and relevant clinical context. For each newly generated execution record, filter the set of structured judgment rules that match the operation target; for rules with trigger condition slots, further determine whether the trigger condition is met; if not, do not proceed to comparison; and compare each judgment parameter dimension according to rule type: single-value threshold rules compare whether the numerical result exceeds the limit; interval range rules compare whether the occurrence time or value falls within the window range; conditional branch rules determine whether the actual path is consistent with the trigger condition of the rule based on the execution record context; and rely on preceding event rules to backtrack and verify whether the operation corresponding to the preceding event identifier has been completed as required. If the number of screened structured judgment rules is greater than 1 and a unique effective rule has been determined, the comparison result is obtained based on this unique effective rule; if a unique effective rule cannot be determined, no automatic deviation conclusion is output, and the process is transferred to manual review.

[0054] S42. When determining that the execution record deviates from the structured judgment rule based on the comparison results, the deviation identification results are divided into three levels: general deviation, major deviation, and serious deviation according to the matching evidence score, deviation magnitude, and the importance of the operational clause that triggers the structured judgment rule.

[0055] When the comparison results show that the execution record is inconsistent with the structured judgment rule, it is judged as a deviation event. The deviation level is divided into three indicators: matching evidence score, deviation magnitude, and clause importance. The matching evidence score is the SKP written when the structured judgment rule is instantiated. The lower the score, the weaker the template fit and the higher the deviation tendency. In terms of deviation magnitude, the out-of-bounds ratio is taken for numerical classes, the ratio of the number of days exceeding the window length to the window length is taken for time window classes, and the missing or inverted binary value is taken for precedence dependency classes. In terms of clause importance, if the protocol or monitoring plan has marked key clauses, the marking result is directly used. Otherwise, operational clauses involving inclusion / exclusion, informed consent, randomization, dose safety, and serious adverse event reporting are marked as high, operational clauses related to safety checks, endpoint indicators, dosing and visit windows are marked as medium, and the rest are marked as low. After being weighted by preset weights, the deviation identification results are divided into three levels according to two preset thresholds: general deviation, major deviation, and serious deviation. General deviation corresponds to a slight exceedance of the window limit or a lack of record-keeping that does not affect the safety of the subjects or the integrity of the data, and is only recorded as a trace. Major deviation corresponds to situations such as missed detection, wrong dose, or large exceedance of the window that may affect the reliability of the data or the rights of the subjects, and an early warning is generated and pushed to the monitoring and project management personnel. Serious deviation corresponds to situations that may affect the safety of the subjects or the integrity of key data, and an immediate early warning is issued.

[0056] S43. Archive the original clause text location index, parameter tracing, structured judgment rules used, and execution records that trigger the rules together with the deviation identification results to form a deviation evidence chain that can be traced back to the original text of the scheme layer by layer.

[0057] like Figure 4 As shown, the present invention also provides a clinical trial protocol execution monitoring system based on clause semantic parsing, including a clause splitting and element extraction module, a dependency graph construction and parameter backfilling module, a rule generation and priority determination module, and an execution comparison module.

[0058] The clause splitting and element extraction module is used to split the clinical trial protocol text, identify operational clauses and reference units, extract elements from operational clauses and reference units to obtain element information containing operational objects and judgment parameters, and identify the relationships between operational clauses and between operational clauses and reference units. The relationships include reference relationships, premise relationships, amendment relationships and exception relationships.

[0059] The dependency graph construction and parameter backfilling module is connected to the clause splitting and element extraction module. It receives element information and relationships, and constructs a dependency graph with operational clauses and parameter units as nodes and relationships as directed edges. The directed edges corresponding to reference relationships, modification relationships, and premise relationships are used as backfilling edges, and the directed edges corresponding to exception relationships are used as priority edges. Backfilling is performed on the judgment parameters in the element information along the backfilling edges corresponding to reference relationships and premise relationships. Overwriting is performed on the backfilled judgment parameters according to the order of the scheme amendment versions along the backfilling edges corresponding to modification relationships, so as to obtain the element information with complete parameters.

[0060] The rule generation and priority determination module is connected to the dependency graph construction and parameter backfilling module. It receives the element information of parameter completion, matches the judgment strategy template according to the element information of parameter completion, and generates structured judgment rules. When the same execution record hits multiple structured judgment rules at the same time, it determines the effective priority of each structured judgment rule based on the priority edge, and determines the unique effective rule according to the effective priority.

[0061] The execution comparison module is connected to the rule generation and priority determination module. It receives structured judgment rules, obtains clinical trial execution records, compares the execution records with the structured judgment rules, and if a unique effective rule has been determined, the unique effective rule shall prevail. The deviation identification result is obtained based on the comparison result.

[0062] The four modules are connected in sequence and receive the technical features output by the previous module at each level, together realizing the automated monitoring of the execution of the clinical trial protocol.

[0063] Furthermore, the system can also be configured with a feedback closed-loop module connected to the execution comparison module to receive manual review conclusions on deviation identification results. If the review conclusion is a false alarm originating from errors in element extraction or association identification, the original text of the corresponding operational clause or parameter unit, the original model output, and the manually corrected element information and association type are added to the incremental sample to perform incremental fine-tuning of the model used for element extraction and association identification. If the false alarm originates from errors in parameter backfilling or overwriting on the dependency graph, the transitive dimension mask of the corresponding edge type is adjusted according to the association type and judgment parameter dimension index recorded in the current traceability table. If the false alarm originates from the mandatory dimension setting of the judgment strategy template being too strict or too broad, the weight vector corresponding to the slot in the corresponding template is adjusted after a preset number of manual review judgments. When the scheme is amended and updated, the system only re-executes text splitting, element extraction, parameter backfilling, rule generation, and execution comparison for the newly added or changed operational clauses and the nodes affected by them, without needing to reprocess the entire scheme text.

[0064] This invention parses the visit plan master table, laboratory normal value reference table, and footnotes into parameter units and incorporates them into a dependency graph. Combined with iterative backfilling based on edge type and version-based coverage mechanisms, it allows decision parameters, originally scattered across different carriers such as tables, footnotes, and amendments, to be completed before matching the decision strategy template. Furthermore, by excluding reference edges and already covered correction edges, and arbitrating only on the priority graph formed by exception relationships based on reachability, it ensures a unique effective decision when both general and exception rules are simultaneously met. This technical solution, without altering the semantic understanding of the operational clause text itself, organizes the reference, correction, and exception relationships between clauses through a graph structure. It provides a calculable and traceable implementation method for completing decision parameters and determining the rule effectiveness order, helping to reduce deviations, omissions, or misjudgments caused by scattered clause text parameters or concurrent rules.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the execution of clinical trial protocols based on clause semantic parsing, characterized in that, Includes the following steps: S1. The clinical trial protocol text is split to identify operational clauses and reference units. Elements are extracted from the operational clauses and reference units to obtain element information containing operational objects and judgment parameters. The relationships between operational clauses and between operational clauses and reference units are identified. The relationships include reference relationships, premise relationships, modification relationships and exception relationships. S2. Construct a dependency graph using operational clauses and parameter units as nodes and relationships as directed edges. The directed edges corresponding to reference relationships, modification relationships, and prerequisite relationships are used as backfill edges, and the directed edges corresponding to exception relationships are used as priority edges. Backfill the decision parameters in the feature information along the backfill edges corresponding to reference relationships and prerequisite relationships, and overwrite the backfilled decision parameters according to the order of the scheme amendment versions along the backfill edges corresponding to modification relationships to obtain feature information with complete parameters. S3. Match the judgment strategy template with the element information completed by the parameters and generate structured judgment rules; when the same execution record hits multiple structured judgment rules at the same time, determine the effective priority of each structured judgment rule based on the priority edge, and determine the unique effective rule based on the effective priority. S4. Obtain clinical trial execution records and compare them with structured judgment rules. If a unique effective rule has been determined, the unique effective rule shall prevail, and the deviation identification result shall be obtained based on the comparison result. If a unique effective rule has not been determined, the case shall be transferred to manual review.

2. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 1, characterized in that, Step S1 includes the following steps: S11. Segment the clinical trial protocol text according to clause number, table boundary and footnote number to obtain clause units, table units and footnote units, and record the position index of each unit in the original text; S12. Perform binary classification on the clause units, identify the operational clauses, and parse the table units and footnote units into reference units; S13. Extract elements from operational clauses and parameter units to obtain operational objects, and simultaneously identify association relationships by combining candidate guide words with a relationship classification model. The judgment parameters in the element information include constraint type and constraint value. The constraint type includes numerical constraints, time constraints, triggering conditions, and prior dependencies.

3. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 2, characterized in that, The table unit is parsed to obtain the table reference unit corresponding to the visit plan master table and the numerical reference unit corresponding to the laboratory normal value reference table. The footnote unit is parsed to obtain the footnote reference unit. The annotation reference relationship between the table reference unit and the footnote reference unit is shown. The method of identifying association relationships by combining candidate keywords and relationship classification models specifically includes: According to the preset guide word list, candidate related pairs are located between operational clauses and reference units, and operational clauses. Among them, reference guide words include clause number, table number and words indicating the action of looking up; amendment guide words include amendment version number and words indicating parameter changes; exception guide words include words indicating the limitation or exclusion of the scope of application; and prerequisite guide words include words indicating the conditions for the completion of the preceding operation. The candidate association pairs are input into the relationship classification model. The relationship classification model outputs the probability distribution of each association type. The association type with the highest probability is taken as the output association, and the probability corresponding to the output association is used as the classification confidence. Candidate association pairs with a classification confidence lower than a preset threshold are not associated and are transferred to manual review.

4. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 1, characterized in that, Step S2 specifically includes: S21. Construct a dependency graph using operational clauses and supply units as nodes and relationships as directed edges; S22. Initialize the observation mask, value, and source table for each node in the dependency graph. The observation mask component corresponding to the dimension for which the decision parameters have been extracted is set to 1 and the actual value is used as the initial value. The source table component records the source node as this node, the association type as empty, and the amendment version as empty. The observation mask component corresponding to the dimension for which the decision parameters have not been extracted is set to 0 and the value is empty. S23. According to the iteration round, backfill the decision parameters along the backfilling edges corresponding to the reference relationship and the premise relationship, and cover the backfilled decision parameters according to the order of the scheme amendment version along the backfilling edges corresponding to the correction relationship, and update the observation mask, value and source table. S24. When the changes in the mask between two adjacent observation rounds meet the convergence condition or the number of iteration rounds reaches the preset maximum number of iterations, the iteration stops and the parameter-completed element information is obtained.

5. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 4, characterized in that, In step S21, the node attributes in the dependency graph include location index and feature information. The directed edges corresponding to the reference relationship point from the referenced node to the referencing node and are marked as backfill edges. The directed edges corresponding to the correction relationship point from the operational clause corresponding to the later scheme amendment version to the corrected operational clause and are marked as backfill edges. The directed edges corresponding to the exception relationship point from the operational clause corresponding to the general rule to the operational clause corresponding to the exception rule and are marked as priority edges.

6. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 4, characterized in that, In step S24, the convergence condition is: ； In the formula, For node indexing, For iteration rounds, For the next iteration round, For nodes In the The observation mask of the wheel, This is the preset convergence threshold.

7. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 4, characterized in that, In step S3, generating structured decision rules includes the following steps: S31. For each operational clause, check whether all slot combinations of the judgment strategy template have been marked as observed in the observation mask after the operational clause converges, and determine the template that meets the condition as the instantiable template of the operational clause. S32. Based on the mapping relationship between the slots of the instantiable template and the corresponding judgment parameters in the parameter-completed element information, fill the judgment parameters into the slots to obtain candidate rules. S33. When multiple candidate rules corresponding to the same operational clause generate conflicting conclusions, calculate the matching evidence score of each candidate rule, retain the candidate rule with the highest matching evidence score for instantiation, and obtain the structured judgment rule.

8. The clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 7, characterized in that, In step S3, when the same execution record hits multiple structured decision rules at the same time, the edges corresponding to the reference relationship and the edges corresponding to the correction relationship that has been covered are excluded. The priority graph is constructed with the priority edges corresponding to the exception relationship. Cycle detection is performed on the priority graph, and the structured decision rules with cycles are marked as pending manual review. For priority graphs without cycles, the effective priority is determined based on the reachability relationships between the nodes corresponding to the multiple hit structured decision rules. Among the multiple hit structured decision rules, the rule whose corresponding node can reach all the other hit rule corresponding nodes is determined as the unique effective rule. If there are cycles in the priority graph, or the nodes corresponding to multiple hit rules are mutually unreachable, it is determined that no unique effective rule has been determined, and the case is transferred to manual review.

9. A clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 7, characterized in that, The matching evidence score is calculated using the following formula: ； In the formula, For operational terms index, For template index, Operational terms With template The matching evidence score for the corresponding candidate rule. Operational terms The converged observation mask, template The weight vector corresponding to the slot. To and A vector of all 1s with the same dimension. This is a preset constant.

10. A clinical trial protocol execution monitoring method based on clause semantic parsing as described in claim 7, characterized in that, Step S4 includes: S41. Obtain clinical trial execution records and compare the clinical trial execution records with the judgment parameters of the structured judgment rules one by one according to each judgment parameter dimension; S42. When determining that the execution record deviates from the structured judgment rule based on the comparison results, the deviation identification results are divided into three levels: general deviation, major deviation, and serious deviation according to the matching evidence score, deviation magnitude, and the importance of the operational clause that triggers the structured judgment rule. S43. Archive the operational clause location index, judgment parameter tracing, triggered structured judgment rules, and corresponding execution records associated with the deviation identification results.

11. A clinical trial protocol execution monitoring system based on clause semantic parsing, characterized in that, The system is used to execute the clinical trial protocol execution monitoring method as described in any one of claims 1 to 10, including: The clause splitting and element extraction module is used to split the clinical trial protocol text, identify operational clauses and reference units, extract elements from the operational clauses and reference units to obtain element information containing operational objects and judgment parameters, and identify the relationships between operational clauses and between operational clauses and reference units. The relationships include reference relationships, premise relationships, amendment relationships and exception relationships. The dependency graph construction and parameter backfilling module is used to construct a dependency graph with operational clauses and parameter units as nodes and relationships as directed edges. Directed edges corresponding to reference relationships, modification relationships, and prerequisite relationships are used as backfilling edges, and directed edges corresponding to exception relationships are used as priority edges. Backfilling is performed on the judgment parameters in the feature information along the backfilling edges corresponding to reference relationships and prerequisite relationships. Overwriting is performed on the backfilled judgment parameters according to the order of the scheme amendment versions along the backfilling edges corresponding to modification relationships, so as to obtain the feature information with complete parameters. The rule generation and priority determination module is used to match the judgment strategy template based on the element information of the parameter completion and generate structured judgment rules. When the same execution record hits multiple structured judgment rules at the same time, the effective priority of each structured judgment rule is determined based on the priority edge, and a unique effective rule is determined based on the effective priority. The execution comparison module is used to obtain clinical trial execution records, compare the execution records with structured judgment rules, and if a unique effective rule has been determined, the unique effective rule shall prevail. The deviation identification result is obtained based on the comparison result.

Images