Clinical test data report management system and method

Abstract

The invention discloses a clinical test data report management system and method, and relates to the field of data report management, and the method comprises the steps: carrying out the deep semantic analysis of original clinical test table data, and carrying out the multi-dimensional semantic mining in combination with a preset medical analysis rule set, therefore, key discovery points with remarkable clinical significance can be automatically identified and screened out. Furthermore, the discovery points are subjected to priority ranking and pruning to form a structured core discovery point sequence, and the structured core discovery point sequence is injected into a prompt template to construct accurate guidance for a large language model. And finally, converting the guide instruction into an abstract text with clear logic and highlighted key points by utilizing a large language model. In this way, automatic conversion from an original data table to a high-quality narrative abstract can be achieved, and therefore the writing efficiency and the intelligent level of a clinical test report are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of data reporting management, and more specifically, to a clinical trial data reporting management system and method. Background Technology

[0002] In the field of modern medicine and drug development, clinical trials are the cornerstone for evaluating the safety and efficacy of new drugs and therapies. Their rigor and standardization directly impact public health and well-being, as well as the innovative development of the pharmaceutical industry. Clinical trials generate massive amounts of multi-dimensional data, covering aspects such as baseline characteristics of participants, adverse events, laboratory indicators, and efficacy endpoints. After statistical analysis, this data must be presented in a clear, accurate, and regulatory-compliant manner in the clinical trial report, serving as a crucial basis for drug approval, clinical practice guidance, and subsequent research. Therefore, establishing an efficient and precise clinical trial data reporting management system is irreplaceable for accelerating new drug development, enhancing data value, and safeguarding public health.

[0003] However, the generation of current clinical trial data reports, especially in the summary description of complex data tables, still faces significant challenges. Existing methods generally rely on manual statistical analysis and medical writing, or use automated tools based on preset rules and templates for data extraction and preliminary statistics. These methods often struggle to achieve efficient and deep semantic understanding when dealing with massive amounts of heterogeneous clinical trial data. Manual writing is time-consuming and labor-intensive, prone to errors and subjective biases, and makes it difficult to ensure the consistency of report content. While rule-based automated tools can improve efficiency, they lack a deep understanding of the inherent logic and clinical significance of the data, only able to perform simple listing or preset information extraction, failing to capture the deeper clinical meaning and potential correlations behind the data. Especially when dealing with complex tabular data such as safety event summary tables containing dozens of rows and columns, existing technologies are inadequate in extracting insightful and summary descriptions from raw data. For example, existing methods cannot automatically generate conclusions from safety event tables that require comprehensive analysis and inductive reasoning, such as "the most common adverse events are... and most are mild to moderate" or "the incidence of SAEs is not significantly different among groups," and still rely heavily on expert judgment.

[0004] Therefore, there is an urgent need for an optimized clinical trial data reporting management system and method. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, a method for managing clinical trial data reporting is provided, comprising: Based on table metadata, semantic parsing of the original table data is performed to obtain structured table objects; Based on the set of analysis rules, multi-dimensional semantic mining is performed on structured table objects to obtain a list of semantic discovery objects; The list of semantic discovery objects is sorted by discovery point priority and pruned to obtain a sorted list of discovery point objects. The sorted list of discovered point objects is injected into the prompt template to obtain the final generated prompt; The final generated prompts are input into a large language model to obtain a draft of the summary text.

[0007] According to another aspect of this application, a clinical trial data reporting management system is provided, comprising: The table semantic parsing module is used to perform semantic parsing on the original table data based on the table metadata to obtain a structured table object; The multi-dimensional semantic mining module is used to perform multi-dimensional semantic mining on structured table objects based on the analysis rule set to obtain a list of semantic discovery objects; The discovery point optimization module is used to sort and prune the list of semantic discovery objects according to the priority of discovery points to obtain a sorted list of discovery point objects. The prompt generation module is used to inject the sorted list of discovered point objects into the prompt template to obtain the final generated prompt; The initial text generation module is used to input the final generated prompts into the large language model to obtain an initial draft of the summary text.

[0008] Compared with existing technologies, the clinical trial data report management system and method provided in this application automatically identify and screen key findings with significant clinical significance by performing deep semantic analysis on the original clinical trial table data and combining it with a preset set of medical analysis rules for multi-dimensional semantic mining. Further, these findings are prioritized and pruned to form a structured sequence of core findings, which is then injected into a prompt template to provide precise guidance to a large language model. Finally, the large language model is used to transform this guidance into a logically clear and focused summary text. This enables automated transformation from raw data tables to high-quality narrative summaries, significantly improving the efficiency and intelligence of clinical trial report writing. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a clinical trial data reporting management method according to an embodiment of this application.

[0011] Figure 2 This is a data flow diagram of the clinical trial data reporting management method according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of sub-step S1 of the clinical trial data reporting management method according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of sub-step S11 of the clinical trial data reporting management method according to an embodiment of this application.

[0014] Figure 5 This is a flowchart of sub-step S2 of the clinical trial data reporting management method according to an embodiment of this application.

[0015] Figure 6 This is a flowchart of sub-step S21 of the clinical trial data reporting management method according to an embodiment of this application.

[0016] Figure 7 This is a flowchart of sub-step S4 of the clinical trial data reporting management method according to an embodiment of this application.

[0017] Figure 8 This is a flowchart of sub-step S41 of the clinical trial data reporting management method according to an embodiment of this application.

[0018] Figure 9 This is a block diagram of a clinical trial data reporting management system according to an embodiment of this application. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] To address the problems mentioned above in the background art, this application proposes a method for managing clinical trial data reports. Figure 1 This is a flowchart of a clinical trial data reporting management method according to an embodiment of this application. Figure 2 This is a data flow diagram of a clinical trial data reporting management method according to an embodiment of this application. For example... Figure 1 and Figure 2As shown, the clinical trial data report management method includes the following steps: S1, based on table metadata, performing semantic parsing on the original table data to obtain structured table objects; S2, based on the analysis rule set, performing multi-dimensional semantic mining on the structured table objects to obtain a list of semantic discovery objects; S3, prioritizing and pruning the list of semantic discovery objects to obtain a sorted list of discovery objects; S4, injecting the sorted list of discovery objects into a prompt template to obtain a final generated prompt; S5, inputting the final generated prompt into a large language model to obtain a draft of the abstract text.

[0021] In the aforementioned clinical trial data reporting management method, step S1 involves performing semantic parsing on the original table data based on table metadata to obtain structured table objects. It should be understood that, to address the problem of low efficiency and ambiguity in subsequent analysis caused by heterogeneous formats, ambiguous field semantics, and chaotic data relationships in original clinical trial table data, this application combines table metadata with semantic parsing to transform unstructured raw data into structured table objects with clear semantics and explicit relationships. This provides an accurate and consistent data foundation for subsequent multi-dimensional semantic mining based on analysis rule sets, thereby improving the reliability and efficiency of the entire data reporting management process.

[0022] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S1 of the clinical trial data reporting management method according to an embodiment of this application. Figure 3 As shown, step S1 includes: S11, performing metadata-driven column semantic mapping and normalization on the original table data based on table metadata to obtain a normalized data frame; S12, constructing a hierarchical tree structure based on row content on the normalized data frame to obtain a hierarchical data tree; S13, performing object encapsulation and integrity verification on the hierarchical data tree to obtain the structured table object.

[0023] Specifically, in step S11, based on table metadata, the original table data undergoes metadata-driven column semantic mapping and normalization to obtain a normalized data frame. It should be understood that due to the heterogeneous semantic representation and inconsistent data format of the column fields in the original clinical trial table, directly involving them in the analysis would lead to a chaotic baseline for subsequent semantic mining. Therefore, this application further performs column semantic mapping and normalization based on table metadata to clarify the clinical semantic meaning of each column and unify the data format standard. This transforms discrete and semantically ambiguous column data into a normalized data frame with clear medical semantics and a unified format standard, thus laying a consistent data baseline for the subsequent hierarchical structure analysis of row content and effectively avoiding analytical biases caused by column semantic ambiguity.

[0024] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S11 of the clinical trial data reporting management method according to an embodiment of this application. Figure 4 As shown, step S11 includes: S111, performing metadata deconstruction and mapping rule classification on the table metadata to obtain static mapping rules and research group metadata objects; S112, generating dynamic mapping rules based on the research group metadata objects to obtain dynamic mapping rules; S113, merging the static mapping rules and dynamic mapping rules, and applying them to the original table data to obtain the normalized data frame.

[0025] More specifically, step S111 involves deconstructing the table metadata and classifying and generating mapping rules to obtain static mapping rules and study group metadata objects. It should be understood that, since clinical trial table metadata simultaneously carries both general table structure definitions and study group-specific data configurations, it is impossible to directly distinguish between general semantic rules and study-specific adaptation logic in a mixed-encoding state. Therefore, this application further deconstructs and classifies the table metadata to separate the general metadata layer and the study group-specific metadata layer, thereby extracting universal static mapping rules and encapsulating them into study group metadata objects containing trial-specific logic. This provides a clear data benchmark for the subsequent generation of dynamic rules and ensures the logical boundaries and accuracy of general and specific rules during merging, significantly improving the scalability of semantic mapping.

[0026] Specifically, in one possible embodiment, step S111 is implemented as follows: First, the hierarchical structure of the table metadata is parsed, extracting a general metadata layer containing column data types, basic format constraints, and industry standard field identifiers, and a research group metadata layer containing experimental group identifiers, custom field semantic descriptions, and intra-group data association rules. Then, semantic rules are extracted from the general metadata layer to generate static mapping rules, such as uniformly mapping "AE" to the standard term "adverse event" and forcing date formats to match the ISO8601 standard. Next, the research group metadata layer is objectified and encapsulated to construct a research group metadata object, containing experimental group coding rules and the hierarchical relationship of specific observation indicators. Finally, the static mapping rules and the research group metadata object are output.

[0027] More specifically, step S112 involves generating dynamic mapping rules based on the study group metadata object. It should be understood that static mapping rules only cover general medical data semantics and cannot adapt to study group-specific data definitions, such as trial-customized observation indicators and grouping logic, leading to deviations in the semantic mapping of study-specific data. Therefore, this application generates dynamic mapping rules based on the study group metadata object. By parsing the study-specific data dimensions, relationships, and encoding logic, it transforms the custom rules in the trial protocol into executable semantic mapping logic, thereby filling the adaptation blind spots of static rules, ensuring consistency between the semantic parsing of data within the study group and the trial design logic, and improving the mapping accuracy of complex clinical trial data.

[0028] Specifically, in one possible embodiment, step S112 is implemented as follows: First, the custom field descriptions in the study group metadata object are extracted, such as SAE grading standards, trial subgroup coding rules, and intragroup data relationships, such as the relationship logic between subject enrollment time and observation period. Then, based on the data definitions in the trial protocol, the semantic transformation rules of the custom fields are applied, such as the correspondence between SAE grading levels 1-3 and mild-severe conditions, as well as the mapping logic of the relationships. Finally, the dynamic mapping rules are associated and bound to the study group metadata object through a rule engine, ultimately generating a set of dynamic mapping rules for the study group.

[0029] More specifically, in step S113, the static mapping rules and dynamic mapping rules are merged and then applied to the original tabular data to obtain the normalized data frame. It should be understood that because the static and dynamic mapping rules exist independently, they cannot collaboratively complete the semantic normalization of the original tabular data, resulting in a disconnect between the mapping logic for general and specific data. Therefore, this application further merges the static and dynamic mapping rules and applies the merged rule set to the original tabular data. This integrates general semantic standards with research-specific logic, enabling the original tabular data to simultaneously meet industry standards and experimental customization requirements in a single processing flow. This eliminates semantic conflicts and generates a standardized data frame with a unified format and clear semantics, providing a precise data foundation for the subsequent construction of a hierarchical tree structure.

[0030] Specifically, in one possible embodiment, step S113 is implemented as follows: First, a rule merging engine is constructed, and conflict resolution strategies are defined, such as dynamic mapping rules prioritizing the coverage of conflicting semantic items in static rules. Next, the static and dynamic mapping rules are input into the rule merging engine, and a unified set of merged mapping rules is generated through syntax parsing, conflict detection, and priority fusion. Then, the semantic adaptation process for the original table data is initiated, and each column of data is semantically annotated according to the merged mapping rule set. For example, a custom field X is mapped to an experimental subgroup identifier according to dynamic rules, format compliance is verified according to static rules, and format conversion is performed, such as converting the experimental-specific date representation to a standard format. Finally, a standardized data frame with clear column semantics and a standardized format is output.

[0031] Specifically, step S12 involves constructing a hierarchical tree structure based on row content for the normalized data frame to obtain a hierarchical data tree. It should be understood that because the row data in the normalized data frame is flatly distributed, it cannot reflect the inherent hierarchical relationships within the clinical trial data, such as the subordinate relationship between subject groups and adverse events, or the event evolution relationship under a time series. Therefore, this application further constructs a hierarchical tree structure for the row content of the normalized data frame. By mining the logical relationships between row data, the flat row data is organized into a hierarchical data tree with a hierarchical recursive relationship, thus intuitively presenting the subordinate and related relationships between data. This provides structured relational support for subsequent semantic mining and facilitates rapid location of data association paths during multi-dimensional analysis.

[0032] Specifically, in one possible embodiment, step S12 is implemented as follows: First, key dimensions of the row data in the normalized data frame are extracted, such as subject ID, timestamp, and event type. Then, based on the hierarchical logic of the clinical data, such as trial group, subject, observation time, and adverse events, hierarchical association rules are constructed. Next, the row data is clustered and hierarchically divided according to the association rules, with the top-level dimension as the root node of the tree structure. Association dimensions are decomposed layer by layer downwards to form a hierarchical data tree containing root nodes, branch nodes, and leaf nodes. The leaf nodes correspond to specific row data records, the branch nodes correspond to data classification dimensions, and the root node corresponds to the subject domain of the entire data frame.

[0033] Specifically, step S13 involves object encapsulation and integrity verification of the hierarchical data tree to obtain the structured table object. It should be understood that since the hierarchical data tree only reflects the hierarchical structure of the data and does not form an entity object that can be directly used for semantic mining, and the tree structure may contain missing data or logical conflicts, such as empty key fields or broken relationships. Therefore, this application further encapsulates and verifies the integrity of the hierarchical data tree to endow it with structured object attributes and methods, while verifying the integrity and logical consistency of the data, thereby transforming the hierarchical data tree into a structured table object with clearly defined attributes and method interfaces. Integrity verification eliminates data defects, ensuring the reliability and usability of the data in the subsequent semantic mining process, enabling the object to directly support multi-dimensional semantic mining.

[0034] Specifically, in one possible embodiment, step S13 is implemented as follows: First, a class model for the structured table object is defined, including node access methods for the hierarchical data tree and data association query interfaces. Then, based on the class model, each node of the hierarchical data tree is encapsulated as an object, mapping node data to object attributes and node relationships to object methods. Subsequently, an integrity verification process is initiated, traversing the object's attributes and relationships, verifying the non-emptiness of key clinical fields, such as adverse event names and incidence data, verifying the logical closure of hierarchical relationships (e.g., child nodes must belong to the parent node dimension), marking or repairing defective data, and finally generating a structured table object that conforms to clinical data specifications.

[0035] In the aforementioned clinical trial data reporting management method, step S2 involves performing multi-dimensional semantic mining on structured table objects based on an analysis rule set to obtain a list of semantic discovery objects. It should be understood that since structured table objects only complete the structured organization of data, they have not yet extracted the hidden clinical semantic relationships and statistical patterns. Medical insights such as the distribution of adverse event types and differences in incidence rates between groups remain hidden within the data. Therefore, this application further conducts multi-dimensional semantic mining on structured table objects based on an analysis rule set, analyzing medical logical relationships from multiple perspectives such as data, time, and inter-group relationships, extracting clinically significant discovery points, thereby transforming structured data into a list of semantic discovery objects that can directly support medical decision-making. This lays the foundation for subsequent priority ranking and summary generation, and overcomes the efficiency and depth bottlenecks of manual analysis in multi-dimensional relationship mining.

[0036] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S2 of the clinical trial data reporting management method according to an embodiment of this application. Figure 5As shown, step S2 includes: S21, calculating candidate indicators and enriching data for each leaf node in the structured table object based on the analysis rule set to obtain an enriched structured table object; S22, performing rule-based multidimensional scanning and discovery instantiation on each leaf node in the enriched structured table object based on the analysis rule set to obtain an intermediate discovery set; S23, performing aggregation analysis on the intermediate discovery set to obtain the semantic discovery object list.

[0037] Specifically, in step S21, based on the analysis rule set, candidate indicator calculations and data enrichment are performed on each leaf node of the structured table object to obtain an enriched structured table object. It should be understood that since the leaf nodes of the structured table object only contain raw observation data and lack derived indicators required for clinical trial analysis, such as event incidence and inter-group differences, directly using them for semantic mining would make it difficult to identify key clinical patterns. Therefore, this application further performs candidate indicator calculations and data enrichment on each leaf node based on the analysis rule set (e.g., including frequency rules, difference rules, severity rules, and special attention principles) to supplement the missing derived indicators in the original data, enriching the information dimensions of the leaf nodes. This ensures that they not only contain basic data but also cover key analytical indicators supporting clinical judgment, thus providing a sufficient data foundation for subsequent rule-based multidimensional scanning and ensuring that potential clinical findings are not overlooked.

[0038] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S21 of the clinical trial data reporting management method according to an embodiment of this application. Figure 6 As shown, step S21 includes: S211, extracting the first leaf node from the structured table object; S212, extracting the names of the control group and the experimental group from the analysis rule set; S213, based on the names of the control group and the experimental group, extracting the corresponding incidence percentages from the first leaf node to obtain the incidence percentages of the control group and the experimental group; S214, calculating the overall incidence rate and the difference value based on the incidence percentages of the control group and the experimental group; S215, adding the overall incidence rate and the difference value as new key-value pairs to the first leaf node.

[0039] More specifically, step S211 involves extracting the first leaf node from the structured table object. It should be understood that since the structured table object exists as a hierarchical tree structure, containing root nodes, intermediate nodes, and leaf nodes, the leaf nodes are the basic units carrying specific clinical trial data. Without extracting the leaf nodes, subsequent calculations and analyses of specific data would lack direct operational objects. Therefore, this application further extracts the first leaf node from the structured table object to obtain raw data records that can be directly used for indicator calculations, transforming the abstract data framework into a concrete data carrier. This ensures that the application of subsequent analysis rules has a clear data focus, laying the foundation for data enrichment.

[0040] More specifically, step S212 involves extracting the names of the control group and experimental group from the analysis rule set. It should be understood that since one of the core analytical logics of clinical trial data is inter-group comparison, such as the difference between the drug group and the placebo group, and leaf nodes may contain data from multiple groups, if the specific names of the control group and experimental group are not clearly defined, it will be impossible to accurately locate the target data to be compared, leading to a loss of direction in the inter-group analysis. Therefore, this application further extracts the names of the control group and experimental group from the analysis rule set to clarify the objects of inter-group comparison, providing an identification basis for the subsequent accurate extraction of data from the leaf nodes. This ensures the relevance of data extraction, avoids analytical errors caused by group confusion, and provides an accurate object benchmark for the calculation of inter-group indicators.

[0041] More specifically, in step S213, based on the names of the control group and the experimental group, the corresponding incidence percentages are extracted from the first leaf node to obtain the incidence percentages of the control group and the experimental group. It should be understood that since the first leaf node contains data from multiple groups, such as various indicators from different experimental and control groups, and the core basis for inter-group comparison is key indicators such as incidence, if the corresponding incidence is not accurately extracted based on the group names, direct data for inter-group analysis cannot be obtained, resulting in a lack of input for subsequent calculations. Therefore, this application further extracts the corresponding incidence percentages from the first leaf node based on the names of the control group and the experimental group to obtain the core raw data for inter-group comparison, providing a basis for calculating the overall incidence and difference values. This ensures that the extracted data strictly corresponds to the target group, avoiding data mismatch and providing accurate raw input for the calculation of subsequent derived indicators.

[0042] Specifically, in one possible embodiment, step S213 is implemented as follows: First, the group data field of the first leaf node is examined, and the group name is matched with the control group and experimental group names extracted from the analysis rule set. After finding a successfully matched field, the incidence percentage data corresponding to that field is located, such as the "incidence" attribute corresponding to the "placebo group" field. The specific value of this data is extracted and used as the incidence percentage of the control group and the incidence percentage of the experimental group, respectively, to ensure the accuracy and completeness of the values.

[0043] More specifically, step S214 calculates the overall incidence rate and the difference value based on the incidence rate percentages in the control group and the experimental group. In a specific example of this application, calculating the overall incidence rate and the difference value based on the incidence rate percentages in the control group and the experimental group includes: calculating the overall incidence rate and the difference value using the following formula: ;in, This represents the percentage of incidence in the control group. The percentage of incidence in the experimental group. To obtain the maximum value, The overall incidence rate, To take the absolute value, This refers to the difference value. In other words, based on the percentage incidence of the two groups, the overall incidence and the difference value are calculated to supplement derived indicators reflecting the overall characteristics of the event and the strength of the inter-group contrast. This enriches the analytical dimensions of the data, provides more comprehensive indicator support for subsequent semantic mining, and makes the identification of discovery points more aligned with clinical decision-making needs. More specifically, in step S215, the overall incidence rate and the difference value are added as new key-value pairs to the first leaf node. It should be understood that since the calculated overall incidence rate and difference value are important supplements to the original data, if they are not stored in association with the original data, subsequent semantic mining will require repeated calculations or separate calls, which will reduce analysis efficiency and easily lead to data association breaks. Therefore, this application further adds these two indicators as new key-value pairs to the first leaf node to achieve the association storage of the original data and derived indicators, forming enriched data units. This allows the leaf node to contain more comprehensive analytical information, and subsequent semantic mining can directly obtain the original data and derived indicators from the same node, improving analysis efficiency while ensuring the tightness and completeness of data association.

[0044] Specifically, in step S22, the leaf nodes of the enriched structured table object are subjected to rule-based multidimensional scanning and discovery instantiation based on the analysis rule set to obtain an intermediate discovery set. It should be understood that although the leaf nodes of the enriched structured table object contain abundant data, meaningful associations and patterns have not yet been filtered out based on clinical logic. Simply piling up data cannot directly form discoveries that can be used for reporting. Therefore, this application performs multidimensional scanning and discovery instantiation on the enriched leaf nodes based on the analysis rule set. This is to identify clinically significant phenomena from the data, such as high-frequency events and significant inter-group differences, based on medical logic and statistical standards, and transform them into structured discovery instances. In this way, implicit clinical patterns can be transformed into explicit discovery instances, providing specific analytical units for subsequent aggregation analysis and ensuring that the discovered discoveries are highly consistent with the core concerns of clinical trials.

[0045] Specifically, in one possible embodiment, step S22 is implemented as follows: First, the system loads a preset set of analysis rules. This set of rules typically exists in the form of a configuration file, and its preset process includes defining the types of findings to be scanned, such as high-frequency events, significantly different events, and serious adverse events, and setting specific, quantifiable triggering conditions for each rule type. For example, the frequency rule is configured to identify events with an "occurrence rate ≥ 10%" as high-frequency events, the difference rule is configured to identify events with an "absolute difference in incidence rate between two groups ≥ 5% and a P-value < 0.05" as significantly different events, and the severity rule is judged based on the value of the event severity field. This set of rules also includes a "list of adverse events of special concern" maintained by experts to achieve targeted monitoring of specific risks. Each rule is accompanied by metadata, such as rule ID and description, to ensure its traceability. After loading this set of rules, the system performs a multi-dimensional scan of the leaf nodes of the enriched structured table object based on these clear judgment criteria, checking whether the data meets the rule judgment conditions from dimensions such as event type, severity, and inter-group comparison. Next, the phenomena that meet the criteria are discovered and instantiated, and then encapsulated into instance objects containing the discovery content, data basis, and rule source, ultimately forming an intermediate discovery set.

[0046] Specifically, step S23 involves performing an aggregation analysis on the intermediate discovery set to obtain the semantic discovery object list. It should be understood that since instances in the intermediate discovery set may be duplicated, closely related, or belong to the same clinical theme, directly using them in subsequent processes would lead to information redundancy or logical disorganization, affecting the conciseness and organization of the report. Therefore, this application further performs an aggregation analysis on the intermediate discovery set to integrate closely related discovery instances, eliminate redundant information, and extract holistic clinical findings. This transforms scattered discovery instances into a well-organized and logically connected list of semantic discovery objects, ensuring that the subsequently generated report summary is focused, hierarchical, and meets the information integration requirements of clinical trial reports.

[0047] Specifically, in one possible embodiment, step S23 is implemented as follows: First, all instance objects in the intermediate discovery set are analyzed for correlation to identify those belonging to the same event type and the same comparison dimension, such as adverse events of a certain system organ, or instances with causal relationships. Next, the associated instances are aggregated according to clinical themes, merging duplicate or highly similar findings to extract comprehensive conclusions, such as "the three most common types of events in adverse events of a certain system organ and their inter-group differences." Finally, the aggregated findings are encapsulated into semantic discovery objects containing core conclusions, supporting instances, and correlation dimensions, arranged in logical order to form a list of semantic discovery objects.

[0048] In the aforementioned clinical trial data reporting management method, step S3 involves prioritizing and pruning the semantic discovery object list to obtain a sorted list of discovery points. It should be understood that because the discovery points in the semantic discovery object list differ in clinical importance, statistical significance, and relevance to the core objectives of the trial, and may contain duplicate or secondary information, directly using them for subsequent prompt generation would lead to a blurred summary focus and redundant information interfering with the core conclusions. Therefore, this application further prioritizes and prunes the semantic discovery object list to differentiate the importance of each discovery point, retaining key information and eliminating secondary or duplicate content. This results in a sorted list of discovery points that focuses on core clinical findings and has a clear logical hierarchy, providing a precise and concise content foundation for subsequent prompt generation, ensuring that the final generated summary text highlights key points and meets the requirements of clinical trial reports for conveying core information.

[0049] Specifically, in one possible embodiment, step S3 is implemented as follows: First, priority ranking dimensions are defined, covering clinical importance (e.g., serious adverse events are prioritized over minor events), statistical significance (e.g., inter-group differences with a p-value less than 0.05 are prioritized over insignificant differences), and relevance to the trial endpoint (e.g., findings related to the primary efficacy endpoint are prioritized over secondary endpoints). Next, ranking rules are constructed based on these dimensions, assigning a comprehensive priority score to each finding. Then, findings are pruned according to a preset threshold, removing secondary findings with scores below the threshold and redundant findings with repetition exceeding the set standard. Finally, the remaining findings are arranged in descending order of priority, forming a ranked list of finding objects.

[0050] In a particular, in another possible preferred embodiment, prioritizing and pruning the list of semantic discovery objects to obtain a sorted list of discovery objects includes: using a unified scoring function to convert the relevant numerical values ​​of each semantic discovery object into a single score representing the importance of the clinical narrative; wherein the unified scoring function includes a base category weight and a dynamic amplitude modulator, and the score is the product of the base category weight and the dynamic amplitude modulator; the base category weight is a mapping from discovery type to weight preset according to expert knowledge; the dynamic amplitude modulator is set as a generalized logic function for mapping key numerical values ​​within the discovery point to a standardized interval; and sorting and pruning the list of semantic discovery objects based on the score.

[0051] Specifically, when prioritizing and pruning the list of semantic discovery objects, assigning base weights to each type of semantic discovery object and adjusting these weights based on their internal values ​​raises the issue of comparability between different types. Specifically, it involves ensuring that scores for high-frequency events can be meaningfully compared with scores for events with significant differences on the same scale. Furthermore, since clinical importance is typically not linear—for example, an increase in the incidence of an adverse event from 1% to 2% evokes far greater clinical concern than an increase from 41% to 42%—adjusting weights needs to capture non-linear diminishing marginal returns or threshold effects. Simultaneously, if the dynamic adjustment rules are too complex and fragmented—for example, setting up a separate If-Else logic for each type—the entire system becomes difficult to maintain and verify.

[0052] Therefore, it is desirable to prioritize and prune discovery points based on a unified scoring function. That is, the final priority of discovery points is calculated using a standardized scoring function. This application employs a unified scoring function to transform the relevant numerical values ​​of each semantic discovery object into a single score representing the importance of the clinical narrative. The unified scoring function includes a base category weight and a dynamic amplitude modulator, and the score is the product of the base category weight and the dynamic amplitude modulator.

[0053] Specifically, the unified scoring function includes a base category weight W and a dynamic amplitude adjuster M, where W represents the inherent importance of the category to which the discovery belongs. For example, it is a mapping from discovery type to weight pre-defined based on expert knowledge. The SAE summary, AESI incidence, and significant difference value correspond to the values ​​of W as 10, 9, and 7, respectively.

[0054] The dynamic amplitude adjuster M is set as a normalized nonlinear function to map the values ​​within the detection point to a standardized interval, for example... or This is used to dynamically adjust the base weights. Furthermore, to address nonlinearity and comparability issues, the dynamic amplitude adjuster M described here is configured as a generalized logistic function, also known as a Richards curve, used to map key values ​​within the discovery point to a standardized interval. Here, the generalized logistic function (or Richards curve) is used in biostatistics and machine learning to simulate threshold effects and saturation phenomena, consistent with clinical significance, and includes: Initial flat zone: When the value is very small, such as when the difference in incidence is less than 1%, its importance increases slowly; Rapid growth zone: When the value reaches a certain "critical point", such as when the difference rate is between 5% and 15%, its importance increases sharply. Saturation zone: When the value becomes very large, such as when the difference rate exceeds 40%, its importance growth slows down again because the event is already significant enough, and the additional boost from increasing the value is diminishing.

[0055] For example, for a given discovery point Its key internal values ​​are For example, for differential scans, It is the absolute difference in incidence between groups; for frequency scans, If the highest occurrence rate of the event is given, then the calculation formula for the dynamic amplitude regulator is: ;in, express The theoretical minimum value of the output value. express The theoretical maximum value of the output value Indicates the initial value parameter. Indicates the steepness coefficient. This indicates the key internal values ​​of the discovery point. This indicates the inflection point or half of the effective concentration. Indicates the asymmetric factor. This represents the dynamic amplitude regulator function.

[0056] Based on clinical significance, it can be taken and Simplified to: ;in The inflection point, or half-maximal effective concentration, is the midpoint of the steepest part of the Richards curve, representing the midpoint of the clinically significant threshold. For example, for differential scanning, it can be set to... This means that when the difference in incidence reaches 10%, the region is considered to have grown in importance the fastest and is the most worthy of attention.

[0057] is the steepness coefficient, which controls the steepness of the curve. The larger the curve is, the more... The more drastic the change in the vicinity, the more sensitive or stringent the system becomes in judging whether a significance threshold has been reached. For example, a large k value means that the importance score will jump dramatically from 9% to 11%.

[0058] thus, The output range is That is, its output value is a multiplier, when much smaller When the output approaches 1, it means the score is essentially the base weight W. When Much larger When the output approaches 2, it means the score has been amplified by at most 100%. This bounded design prevents a single, extremely large discovery from disproportionately dominating the entire sort.

[0059] Finally, the semantic discovery object list is sorted and pruned based on the scores. Specifically, the semantic discovery object list is sorted from highest to lowest score, placing high-importance discoveries at the top. Then, pruning is performed by setting a score threshold, such as 6 based on clinical needs, removing discoveries with scores below this threshold, and merging duplicate discoveries with similar scores to form a sorted list of discovery objects. This unifies the scale for prioritizing and pruning the semantic discovery object list, solving the comparability problem between heterogeneous data by mapping all discoveries to a single clinical significance score. For example, a high-base-weighted but low-value SAE event can be quantitatively compared with a low-base-weighted but extremely high-value common event. Furthermore, the use of generalized logistic functions incorporates the inherent non-linear thinking process of human experts when judging importance, and the parameters... and This makes expert experience configurable and interpretable, resulting in more reasonable rankings that better reflect clinical practice. For example, an AESI event with an incidence of 1.5% (such as myocarditis), even if its incidence is much lower than a common event with an incidence of 25% (such as nausea), may receive a higher final score and thus be prioritized for inclusion in the abstract.

[0060] In the aforementioned clinical trial data reporting management method, step S4 involves injecting the sorted list of discovery points into a prompt template to obtain the final generation prompt. It should be understood that since the sorted list of discovery points is merely a structured set of discovery points, it lacks clear guidance on the format specifications, logical framework, and medical expression requirements for the large language model's summary generation. Directly inputting it into the model may result in chaotic output text format, non-standard terminology, or logical gaps. Therefore, this application further injects the sorted list of discovery points into a prompt template to transform discrete discovery points into a set of instructions that conforms to clinical trial report specifications, is logically coherent, and accurately guides model generation. This provides the large language model with a final generation prompt that combines core content with format constraints, ensuring that the output summary text meets the requirements of clinical trial reports in terms of medical professionalism, logical rigor, and structural standardization, significantly improving the quality and applicability of the generated content.

[0061] In particular, in one specific embodiment, Figure 7 This is a flowchart of sub-step S4 of the clinical trial data reporting management method according to an embodiment of this application. Figure 7 As shown, step S4 includes: S41, performing atomic rendering of each discovery point object in the sorted discovery point object list to obtain an instruction statement list; S42, performing structured assembly of each instruction statement in the instruction statement list to obtain a formatted instruction block; S43, injecting the formatted instruction block into the prompt template to obtain the final generated prompt.

[0062] Specifically, in step S41, the instruction statements of each discovery point object in the sorted discovery point object list are atomically rendered to obtain an instruction statement list. It should be understood that since the discovery point objects in the sorted discovery point object list are structured data entities containing multi-dimensional information, such as discovery content, data basis, and related dimensions, directly inputting them as instructions into the large language model would lead to misunderstandings due to the ambiguity of information hierarchy, failing to accurately capture the core expressive requirements of each discovery point. Therefore, this application further performs atomic rendering of instruction statements on each discovery point object to decompose the multi-dimensional information of each discovery point into single, clear, and directly executable instruction statements, ensuring that the core requirements of each discovery point are clearly conveyed. This results in a list composed of independent and precise instruction statements, providing standardized basic units for subsequent structured assembly, avoiding instruction ambiguity caused by mixed information, and ensuring the accuracy of the large language model's understanding of each discovery point.

[0063] In particular, in one specific embodiment, Figure 8 This is a flowchart of sub-step S41 of the clinical trial data reporting management method according to an embodiment of this application. Figure 8As shown, step S41 includes: S411, extracting a first semantic discovery object from the sorted list of discovery point objects; S412, performing template parsing and data path identification on the first semantic discovery object based on the instruction template mapping table to obtain a parsed template string and a data path list; S413, performing hierarchical traversal and value extraction based on the data path list to obtain a path-to-value mapping table; S414, performing template replacement and statement synthesis on the parsed template string based on the path-to-value mapping table to obtain a first instruction statement.

[0064] More specifically, step S411 involves extracting a first semantic discovery object from the sorted list of discovery point objects. Specifically, this application extracts the first semantic discovery object from the sorted list of discovery point objects to transform batch processing into atomic operations that process each object individually, ensuring that subsequent template parsing and data extraction can focus on the complete semantic information of a single discovery point. This allows for the generation of independent and precise instruction statements for each discovery point, avoiding rendering errors caused by information overlap, thereby guaranteeing the accuracy and atomicity of the instruction statements.

[0065] More specifically, step S412 involves performing template parsing and data path identification on the first semantic discovery object based on the instruction template mapping table to obtain the parsed template string and data path list. It should be understood that since the information of the first semantic discovery object is structured, including event types, inter-group data, and statistical results, directly generating instruction statements would result in a chaotic expression format due to the lack of a unified template, failing to meet the requirements of subsequent structured assembly. Therefore, this application further performs template parsing and data path identification on the first semantic discovery object based on the instruction template mapping table to match the preset instruction template corresponding to the discovery point and locate the specific path of its internal data in the hierarchical structure. This provides a standardized framework for subsequent data extraction and instruction generation, ensuring that different types of discovery points are converted into instruction statements according to unified logic, improving the standardization and consistency of instruction generation.

[0066] Specifically, in one possible embodiment, step S412 is implemented as follows: First, the corresponding preset template is matched in the instruction template mapping table using the type identifier of the first semantic discovery object; the string structure of the template is parsed to identify the fixed expression part and the placeholder to be filled, such as "[event name] has a [difference] in the incidence rate between the [experimental group] and the [control group]". Then, based on the hierarchical data structure of the first semantic discovery object, the specific data path corresponding to each placeholder is identified, such as "event name" corresponding to the "eventName" field in the object attribute, and "difference" corresponding to the "statistics.difference" field, finally obtaining the parsed template string and the list of data paths.

[0067] More specifically, step S413 involves performing hierarchical traversal and value extraction based on the data path list to obtain a path-to-value mapping table. Specifically, this application further performs hierarchical traversal and value extraction on the hierarchical data structure of the first semantic discovery object based on the data path list to obtain the actual data value corresponding to each path, providing specific content for template filling. This transforms abstract path information into concrete numerical or textual information, ensuring that the instruction statement has actual targeting rather than a vague framework.

[0068] Specifically, in one possible embodiment, step S413 is implemented as follows: First, the hierarchical data structure of the first semantic discovery object is loaded, such as containing levels like "basic information," "statistical data," and "association groups." Then, based on each path in the data path list, such as "statistical data.occurrence rate.experimental group," the system traverses layer by layer from the top level to the target node. Next, the actual value of each target node is extracted, and the path and corresponding value are associated and recorded one by one, ultimately forming a path-to-value mapping table.

[0069] More specifically, step S414 involves performing template replacement and statement synthesis on the parsed template string based on the path-to-value mapping table to obtain the first instruction statement. It should be understood that since the parsed template string contains placeholders, without replacement, it remains an incomplete framework and cannot be directly used as an instruction to guide the large language model. Therefore, this application further performs template replacement and statement synthesis on the parsed template string based on the path-to-value mapping table to replace placeholders with actual data values, forming a complete and specific instruction statement. This allows for the combination of standardized templates and specific data to generate atomic instruction statements that conform to the standard format and contain actual content, providing effective units for the subsequent construction of the instruction statement list.

[0070] Specifically, in one possible embodiment, step S414 is implemented as follows: First, each placeholder in the parsed template string is matched with a path in the path-to-value mapping table. Then, based on the matching results, the placeholder is replaced with its corresponding value, such as replacing "[experimental group incidence rate]" with "15.3%". Next, the replaced string undergoes syntax validation and sentence optimization to ensure fluent expression and adherence to the rigor of medical instructions, such as "The incidence rate of a certain adverse event in the experimental group is 15.3%, and its difference from the control group needs to be emphasized," ultimately yielding the first instruction statement.

[0071] Specifically, step S42 involves structuring and assembling each instruction statement in the instruction statement list to obtain formatted instruction blocks. It should be understood that since the statements in the instruction statement list are independent and lack the hierarchy and order consistent with the logic of clinical trial reports, directly injecting the prompt template would result in logically scattered prompt content that fails to conform to the narrative pattern of clinical reports, which prioritizes core information over secondary information and safety over efficacy. Therefore, this application further structures and assembles the instruction statement list to sort and integrate the instruction statements according to the inherent logic of clinical trial reports, forming an instruction set with internal logical coherence. This transforms scattered instruction statements into hierarchically structured and logically ordered formatted instruction blocks, enabling the large language model to follow the conventional narrative structure of clinical reports when generating summaries, ensuring that the output content is clear, progressively focused, and conforms to the reading habits of medical reports.

[0072] Specifically, in one possible embodiment, step S42 is implemented as follows: First, an assembly logic framework is constructed based on the core themes of the clinical trial report. Statements in the instruction statement list are divided according to theme categories, and within the same category, they are prioritized according to the priority of the discovery points, such as instructions related to serious adverse events taking precedence over those related to minor events. The sorted statements are then connected using logical connectors to eliminate any sense of separation between statements, ultimately integrating them into a coherent, formatted instruction block, ensuring that the statements within the block progress in an orderly manner from core to secondary, and from the overall to the detailed.

[0073] Specifically, in step S43, a formatting instruction block is injected into the prompt template to obtain the final generated prompt. It should be understood that since the formatting instruction block only contains core expression requirements and logical order, lacking format constraints, medical terminology standards, and reporting style requirements for the content generated by the large language model, directly inputting it into the model would result in the generated summary not conforming to the standard specifications of clinical trial reports. Therefore, this application further injects the formatting instruction block into the prompt template to combine the instruction block with preset format constraints, terminology standards, and style requirements, forming a complete prompt that combines core content with formal specifications. This provides the large language model with a final generated prompt that has both a clear expression task and strict standard boundaries, ensuring that the output summary text conforms to the clinical report style in structure, meets medical professional standards in terminology, and maintains objectivity and rigor in style, fully meeting the professional requirements of clinical trial data reporting.

[0074] Specifically, in one possible embodiment, step S43 is implemented as follows: First, a preset prompt template is loaded. This template includes format requirements for clinical trial reports (e.g., "The abstract should describe safety and efficacy in paragraphs, with each part not exceeding 300 words"), a medical terminology lookup table (e.g., uniformly using "Serious Adverse Events (SAE)" instead of other common terms), and style guidelines. A specific area for embedding instruction blocks is located within the template, such as "Core Statement Task: [Embed instruction block here]". The formatted instruction block is then fully embedded in this area, making the template's specifications and the instruction block's content requirements an organic whole, ultimately generating a final prompt containing constraints across all dimensions of content, format, terminology, and style.

[0075] In the aforementioned clinical trial data reporting management method, step S5 involves inputting the final generated prompts into a large language model to obtain a draft summary text. It should be understood that while the final generated prompts are a set of instructions integrating core findings, logical frameworks, and formatting specifications, they are not directly usable narrative text for reporting. Manually converting these instructions into a professional summary is time-consuming and struggles to guarantee coherence and standardization. Therefore, this application further inputs the final generated prompts into a large language model. This model, based on the Transformer architecture of deep learning, possesses powerful language understanding and generation capabilities through pre-training on massive amounts of text data. In this application, this model can automatically convert highly structured and semantically rich instruction sets into natural language clinical trial report summaries. This rapidly generates logically coherent, terminologically standardized, and key-point-focused draft summary texts, significantly shortening the manual writing cycle while ensuring the text meets reporting requirements in terms of medical professionalism and information completeness, providing a high-quality foundation for subsequent manual review and optimization.

[0076] Specifically, in one possible embodiment, step S5 is implemented as follows: First, a large language model adapted for medical text generation is selected. This model is trained on clinical report corpora and has the ability to understand medical terminology and clinical logic. Then, the final generated prompt is completely input into the large language model. The priority, format constraints, and expression requirements of the key findings contained in the prompt are parsed by the model. Model generation parameters are set, such as the output length range, which must conform to the standard length of clinical trial abstracts, and a temperature coefficient to ensure the stability and accuracy of the generated content. Based on its understanding of the prompt, the model organizes the language according to the instructions, transforming the key findings into coherent narrative text. For example, the instruction "focus on describing the difference in the incidence rate of a certain adverse event between groups" is transformed into professional expressions such as "the incidence rate of a certain adverse event in the experimental group is X%, and in the control group it is Y%, the difference between groups is statistically significant, and the event is mostly mild to moderate." Finally, a draft abstract text containing all core findings, with a clear logical hierarchy, and conforming to medical report standards is generated.

[0077] In summary, the clinical trial data report management method based on the embodiments of this application is explained. It automatically identifies and filters key findings with significant clinical significance by performing deep semantic analysis on the original clinical trial table data and combining it with a preset set of medical analysis rules for multi-dimensional semantic mining. Furthermore, these findings are prioritized and pruned to form a structured sequence of core findings, which is then injected into a prompt template to provide precise guidance to a large language model. Finally, the large language model is used to transform this guidance into a logically clear and focused summary text. This achieves automated transformation from raw data tables to high-quality narrative summaries, significantly improving the efficiency and intelligence of clinical trial report writing.

[0078] Figure 9 This is a block diagram of a clinical trial data reporting management system according to an embodiment of this application. Figure 9 As shown, the clinical trial data report management system 100 according to an embodiment of this application includes: a table semantic parsing module 110, used to perform table semantic parsing on the original table data based on table metadata to obtain a structured table object; a multi-dimensional semantic mining module 120, used to perform multi-dimensional semantic mining on the structured table object based on an analysis rule set to obtain a list of semantic discovery objects; a discovery point optimization module 130, used to prioritize and prune the semantic discovery object list to obtain a sorted list of discovery point objects; a prompt generation module 140, used to inject the sorted list of discovery point objects into a prompt template to obtain a final generated prompt; and a text draft generation module 150, used to input the final generated prompt into a large language model to obtain a draft of the abstract text.

[0079] As described above, the clinical trial data reporting management system 100 according to embodiments of this application can be implemented in various wireless terminals, such as servers with clinical trial data reporting management algorithms. In one possible implementation, the clinical trial data reporting management system 100 according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the clinical trial data reporting management system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the clinical trial data reporting management system 100 can also be one of many hardware modules of the wireless terminal.

[0080] Alternatively, in another example, the clinical trial data reporting management system 100 and the wireless terminal may be separate devices, and the clinical trial data reporting management system 100 may connect to the wireless terminal via wired and / or wireless networks and transmit interactive information in accordance with an agreed data format.

[0081] Those skilled in the art will understand that the specific operations of each step in the aforementioned clinical trial data reporting management system have been referenced above. Figures 1 to 8 The clinical trial data reporting management method is described in detail in the description, and therefore, its repeated description will be omitted.

Claims

1. A method for managing clinical trial data reports, characterized in that, include: Based on table metadata, semantic parsing of the original table data is performed to obtain structured table objects; Based on the set of analysis rules, multi-dimensional semantic mining is performed on structured table objects to obtain a list of semantic discovery objects; The list of semantic discovery objects is sorted by discovery point priority and pruned to obtain a sorted list of discovery point objects. The sorted list of discovered point objects is injected into the prompt template to obtain the final generated prompt; The final generated prompts are input into a large language model to obtain a draft of the summary text.

2. The clinical trial data reporting management method according to claim 1, characterized in that, Based on table metadata, semantic parsing is performed on the original table data to obtain structured table objects, including: Based on table metadata, the original table data is subjected to metadata-driven column semantic mapping and normalization to obtain normalized data frames; A hierarchical tree structure based on row content is constructed from the normalized data frame to obtain the hierarchical data tree; The hierarchical data tree is encapsulated and its integrity is verified to obtain the structured table object.

3. The clinical trial data reporting management method according to claim 2, characterized in that, Based on table metadata, the original table data undergoes metadata-driven column semantic mapping and normalization to obtain a normalized data frame, including: Metadata deconstruction and mapping rule classification of table metadata are performed to generate static mapping rules and research group metadata objects; Dynamic mapping rules are generated based on the research group's metadata objects. After merging the static mapping rules and dynamic mapping rules, they are applied to the original table data to obtain the normalized data frame.

4. The clinical trial data reporting management method according to claim 1, characterized in that, Based on the analysis rule set, multi-dimensional semantic mining is performed on structured table objects to obtain a list of semantically discovered objects, including: Based on the set of analysis rules, candidate indicators are calculated and data enrichment is performed on each leaf node in the structured table object to obtain an enriched structured table object. For the analysis rule set, rule-based multidimensional scanning and discovery instantiation are performed on each leaf node in the enriched structured table object to obtain an intermediate discovery set; Aggregate analysis is performed on the intermediate discovery set to obtain the list of semantic discovery objects.

5. The clinical trial data reporting management method according to claim 4, characterized in that, The set of analysis rules includes frequency rules, difference rules, severity rules, and special attention principles.

6. The clinical trial data reporting management method according to claim 4, characterized in that, Based on the analysis rule set, candidate indicators are calculated and data enrichment is performed on each leaf node of the structured table object to obtain an enriched structured table object, including: Extract the first leaf node from the structured table object; Extract the names of the control group and experimental group from the set of analysis rules; Based on the names of the control group and the experimental group, the corresponding incidence percentages are extracted from the first leaf node to obtain the incidence percentages of the control group and the experimental group. The overall incidence rate and the difference value were calculated based on the incidence rate percentages in the control group and the experimental group. Add the overall incidence rate and the difference value as new key-value pairs to the first leaf node.

7. The clinical trial data reporting management method according to claim 6, characterized in that, Based on the incidence percentages in the control group and the experimental group, the overall incidence and the difference value are calculated, including using the following formula: ;in, This represents the percentage of incidence in the control group. The percentage of incidence in the experimental group. To obtain the maximum value, The overall incidence rate, To take the absolute value, The difference value.

8. The clinical trial data reporting management method according to claim 1, characterized in that, The sorted list of discovered point objects is injected into the hint template to obtain the final generated hint, including: Atomic rendering of each discovery point object in the sorted discovery point object list is performed to obtain the instruction statement list. The instruction statements in the instruction statement list are structured and assembled to obtain a formatted instruction block; Inject the formatting instruction block into the prompt template to obtain the final generated prompt.

9. The clinical trial data reporting management method according to claim 8, characterized in that, Atomic rendering of command statements is performed on each discovery point object in the sorted discovery point object list to obtain a command statement list, including: Extract the first semantic discovery object from the sorted list of discovery point objects; Based on the instruction template mapping table, template parsing and data path identification are performed on the first semantic discovery object to obtain the parsed template string and data path list; A hierarchical traversal and value extraction are performed based on the data path list to obtain a path-to-value mapping table. Based on the path-to-value mapping table, the parsed template string is subjected to template replacement and statement synthesis to obtain the first instruction statement.

10. A clinical trial data reporting management system, characterized in that, include: The table semantic parsing module is used to perform semantic parsing on the original table data based on the table metadata to obtain a structured table object; The multi-dimensional semantic mining module is used to perform multi-dimensional semantic mining on structured table objects based on the analysis rule set to obtain a list of semantic discovery objects; The discovery point optimization module is used to sort and prune the list of semantic discovery objects according to the priority of discovery points to obtain a sorted list of discovery point objects. The prompt generation module is used to inject the sorted list of discovered point objects into the prompt template to obtain the final generated prompt; The initial text generation module is used to input the final generated prompts into the large language model to obtain an initial draft of the summary text.

Images