System for automatic conversion of clinical trial sdtm dataset to adam dataset

Abstract

The present application provides an automatic conversion system from a clinical trial SDTM dataset to an ADaM dataset, and relates to the technical field of medical care information processing. The present application first constructs a double-index knowledge base supporting keyword and vector retrieval, which includes a clinical trial data statistical analysis plan, SDTM metadata and general knowledge, and then generates a clinical analysis dataset using a large language model. The process includes a multi-stage, step-by-step confirmation automatic workflow, which drives the LLM to sequentially perform the following tasks, and the output of each stage needs to be submitted to the user for confirmation before the next stage can be performed: in the first stage, based on the ADaM dataset specified by the user and the knowledge base, a target variable list is generated. In the second stage, based on the confirmed target variable list, the logical derivation steps for each target variable are generated. In the third stage, based on the confirmed logical derivation steps, executable data conversion program code is generated.

Description

Technical Field

[0001] This invention relates to the field of healthcare information processing technology, and in particular to an automatic conversion system for clinical trial SDTM datasets to ADaM datasets. Background Technology

[0002] Standardization of clinical trial data is the cornerstone of modern new drug development and approval. The CDISC (Clinical Data Interchange Standards Consortium) has developed a series of globally recognized data standards, among which the Study Data Tabulation Model (SDTM) and the Analysis Dataset Model (ADaM) are two of the most core standards in the data submission process. SDTM is used to organize and submit clinical trial data collected from original data sources; its core lies in "tabulation," that is, presenting the trial data in a structured manner.

[0003] SDTM data itself is not directly applicable to complex statistical analyses. For statistical analysis, SDTM data needs to be further processed into ADaM datasets. The core of ADaM lies in "analysis," requiring that the variables in the dataset can be directly used for statistical modeling and ensuring the traceability and reproducibility of the analysis results. Creating an ADaM is a complex derivation process, involving not only extracting data from multiple SDTM domains but, more importantly, calculating, deriving, and constructing new analytical variables, such as baseline markers, analytical values, population markers, and analysis periods, based on the complex analytical logic in SAP. This process requires practitioners to be proficient in both SDTM and ADaM standards, possess a deep understanding of statistical analysis methods and research protocols, and be able to translate these into precise programming logic. Therefore, ADaM creation is one of the most skill-intensive, time-consuming, and error-prone steps in clinical data processing, constituting a significant bottleneck restricting the efficiency of new drug development. Summary of the Invention

[0004] This invention provides an automatic conversion system for clinical trial SDTM datasets to ADaM datasets, in order to at least partially solve the above-mentioned problems.

[0005] The first aspect of this invention provides an automatic conversion system for clinical trial SDTM datasets to ADaM datasets, the system comprising:

[0006] The knowledge base construction module is used to parse clinical trial data statistical analysis plans, SDTM metadata and general knowledge, and build a dual-index knowledge base that supports keyword and vector retrieval;

[0007] The target variable generation module is used to call the large language model, perform hybrid retrieval on the knowledge base based on the first prompt word and the user-specified ADaM dataset, and generate a structured list of target variables.

[0008] The variable derivation logic generation module is used to call the large language model after the target variable list has been verified, perform a hybrid retrieval of the knowledge base based on the second prompt word, and generate structured logical derivation steps based on the business rules contained in the context information related to the target variable in the knowledge base.

[0009] The code generation module is used to translate the structured logical deduction steps into executable code based on the third prompt word after the logical deduction steps have been verified.

[0010] Optionally, the knowledge base construction module is used for:

[0011] The document-formatted clinical trial data statistical analysis plan is parsed to obtain text blocks;

[0012] Parse the SDTM metadata to obtain key metadata, which includes at least dataset definition, variable definition, value-level metadata, and code list;

[0013] The general knowledge is parsed to obtain structured rule entries or text blocks with metadata. The general knowledge includes: ADaM model specifications, ADaM implementation guidelines, specific analysis guidelines, ADaM compliance check rules, and expert experience knowledge.

[0014] Optionally, the target variable generation module is used to:

[0015] The large language model is invoked, with the fields of the user-specified ADaM dataset as the target, the clinical trial data statistical analysis plan, SDTM metadata and general knowledge in the knowledge base as background knowledge, and the traceability of the generated target variables as a constraint, to generate a structured list of target variables.

[0016] Optionally, the structured list of target variables includes: variable name, variable label, source type and source details, initially establishing traceability metadata for ADaM variables.

[0017] Optionally, the system further includes:

[0018] The verification information acquisition module is used for:

[0019] Receive modification information from the user regarding the target variable list, and obtain a verified target variable list;

[0020] The system receives modification information from the user regarding the logical derivation steps, and obtains verified logical derivation steps.

[0021] Optionally, the knowledge base construction module is further configured to:

[0022] The modification information sent by the user regarding the list of target variables and the modification information regarding the logical derivation steps are parsed and added to the knowledge base.

[0023] Optionally, the code generation module is configured to: execute the verified executable code in a secure sandbox environment after the executable code has been verified; if the execution fails, instruct the large language model to analyze the error and correct the code through debugging instructions, and execute it again until the code is successfully executed, and send the ADaM dataset generated after the code is successfully executed.

[0024] A second aspect of this invention provides an automatic conversion method from clinical trial SDTM datasets to ADaM datasets, the method comprising:

[0025] We analyze clinical trial data statistical analysis plans, SDTM metadata, and general knowledge to build a dual-index knowledge base that supports keyword and vector retrieval.

[0026] The large language model is invoked, and based on the first prompt word and the user-specified ADaM dataset, a hybrid retrieval is performed on the knowledge base to generate a structured list of target variables.

[0027] After the target variable list is validated, the large language model is invoked, and a hybrid retrieval is performed on the knowledge base based on the second prompt word. Based on the business rules contained in the context information related to the target variable in the knowledge base, a structured logical deduction step is generated.

[0028] After the logical deduction steps are verified, the structured logical deduction steps are translated into executable code based on the third prompt word.

[0029] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements an automatic conversion system for converting clinical trial SDTM datasets to ADaM datasets as described in the first aspect of the present invention.

[0030] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an automatic conversion system from clinical trial SDTM dataset to ADaM dataset as described in the first aspect of the present invention.

[0031] The fifth aspect of the present invention provides a computer program product, including a computer program / instructions, which are implemented by a processor as an automatic conversion system from clinical trial SDTM dataset to ADaM dataset as described in the first aspect of the present invention.

[0032] The automatic conversion system from clinical trial SDTM datasets to ADaM datasets proposed in this invention first constructs a dual-index knowledge base supporting keyword and vector retrieval. This knowledge base contains clinical trial data statistical analysis plans, SDTM metadata, and general knowledge. Then, it utilizes a Large Language Model (LLM) to generate the clinical analysis dataset. This process includes a multi-stage, step-by-step automated workflow. This workflow drives the LLM to perform the following tasks sequentially, and the output of each stage must be submitted to the user for review and confirmation before proceeding to the next stage: First stage: Based on the user-specified ADaM dataset and knowledge base, generate a list of target variables. Second stage: Based on the confirmed list of target variables, generate logical derivation steps for each target variable. Third stage: Based on the confirmed logical derivation steps, generate executable data conversion program code.

[0033] The automatic conversion system from clinical trial SDTM datasets to ADaM datasets proposed in this invention integrates user interaction, file processing, AI calling engine, and code execution environment. It realizes the above-mentioned multi-stage, step-by-step confirmation automated workflow and mixed information processing method to achieve the automation of data conversion from SDTM to ADaM. Attached Figure Description

[0034] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying 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.

[0035] Figure 1 This is a block diagram of the automatic conversion system from clinical trial SDTM dataset to ADaM dataset provided by the present invention;

[0036] Figure 2 This is a flowchart of the automatic conversion system from clinical trial SDTM dataset to ADaM dataset provided by the present invention;

[0037] Figure 3 This is a hardware structure diagram of any device with data processing capabilities where the automatic conversion system from clinical trial SDTM dataset to ADaM dataset provided in this embodiment of the invention is located. Detailed Implementation

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] This invention aims to address a long-standing critical bottleneck in the field of clinical trial data analysis: the process of converting SDTM to ADaM. Traditionally, this process has relied entirely on highly skilled clinical statisticians.

[0040] Converting SDTM to ADaM is not a simple "data transformation," but rather an automation of a highly complex workflow reliant on expert experience. Specifically, traditional processes require programmers to not only be proficient in both SDTM and ADaM standards and possess extensive project experience, but also to execute a series of tedious and error-prone manual steps:

[0041] 1. Analyze and interpret requirements: Study core documents such as the Statistical Analysis Plan (SAP) in depth to understand the complex analysis objectives and data derivation logic.

[0042] 2. Analysis Variable Design: Based on the understanding of the requirements, determine all analysis variables and their specific attributes that need to be included in the target ADaM dataset.

[0043] 3. Derivation rule construction: Design detailed variable mapping and derivation rules (often manifested as complex Excel mapping tables in practice), and clearly define how each ADaM variable is calculated or transformed from one or more SDTM variables.

[0044] 4. Programming Implementation: Manually write program code in languages ​​such as SAS or Python to transform the above rules into executable logic.

[0045] 5. Execution and Verification: Run the code and perform rigorous quality control and verification on the generated ADaM dataset.

[0046] The entire process is not only time-consuming, but also highly dependent on manual labor, making it difficult to guarantee the consistency and accuracy of the results, resulting in extremely high technical barriers and labor costs. Therefore, this invention proposes to utilize artificial intelligence (AI) technology to efficiently and compliantly automate the entire workflow from SDTM to ADaM.

[0047] Current technologies for automating clinical trial data analysis have improved data processing efficiency to varying degrees, but they generally suffer from a core flaw: they fail to achieve truly end-to-end automated analysis programming, rely heavily on manual pre-configuration, and fail to solve the industry challenge of fundamentally understanding the intent of scientific analysis and automatically generating complex derived logic code.

[0048] Based on this, this invention proposes combining the natural language understanding and code generation capabilities of a Large Language Model (LLM) with a specially constructed hybrid retrieval knowledge base to achieve a high degree of automation in the conversion process from the SDTM dataset to the ADaM dataset. This intelligently integrates multi-source information such as statistical analysis plans, SDTM metadata, and the CDISC standard, decomposing the complex standard conversion task into sub-tasks, and combining the efficiency of AI with the experience of human experts, thereby improving the generation efficiency and quality of the ADaM dataset.

[0049] Specifically, such as Figure 1 The diagram illustrates the structural block diagram of an automatic conversion system from clinical trial SDTM datasets to ADaM datasets provided in an embodiment of the present invention. The system includes:

[0050] The knowledge base construction module 101 is used to parse the clinical trial data statistical analysis plan, SDTM metadata and general knowledge, and build a dual-index knowledge base that supports keyword and vector retrieval;

[0051] The target variable generation module 102 is used to call the large language model, perform hybrid retrieval on the knowledge base based on the first prompt word and according to the user-specified ADaM dataset, and generate a structured list of target variables.

[0052] The variable derivation logic generation module 103 is used to call the large language model after the target variable list has been verified, perform a hybrid retrieval of the knowledge base based on the second prompt word, and generate structured logical derivation steps based on the business rules contained in the context information related to the target variable in the knowledge base.

[0053] The code generation module 104 is used to translate the structured logical deduction steps into executable code based on the third prompt word after the logical deduction steps have been verified.

[0054] In this embodiment of the invention, the knowledge base construction module can intelligently parse general knowledge such as clinical trial data statistical analysis plans, SDTM metadata, and CDISC standards to construct a dual-index knowledge base supporting keyword and vector retrieval. Next, based on the knowledge base, an LLM (Limited Language Management) list of target variables is generated and submitted to human experts for review. Then, after expert confirmation, the system calls the LLM again, combining it with business rules in the knowledge base to generate structured logical derivation steps for each variable. Finally, these logical derivation steps are translated into executable code and run. The LLM's unique self-correcting mechanism handles execution errors until a high-quality, deliverable ADaM dataset is successfully generated.

[0055] In this embodiment of the invention, through multi-stage intelligent invocation and processing, and by setting manual review and confirmation nodes between each automated stage, the complete workflow of clinical statistical programmers from SDTM to ADaM, which is highly dependent on expert experience (i.e., requirement understanding → rule design → programming implementation), is successfully transformed into a controllable, traceable, and efficient automated process.

[0056] Specifically, in the process of building a dual-index knowledge base, the knowledge base construction module needs to process and integrate all the knowledge sources required for the current clinical trial data statistical analysis project. For different types of data sources, the knowledge base construction module uses a dedicated parser to process them and builds two complementary indexes, keyword index and vector index, for each information unit.

[0057] In this embodiment of the invention, all knowledge sources required for clinical trial data statistical analysis include: the clinical trial data statistical analysis plan and SDTM metadata corresponding to the current project, as well as general knowledge. For each new project, the clinical trial data statistical analysis plan and SDTM metadata in the knowledge base can be updated while keeping the general knowledge unchanged.

[0058] Specifically, the knowledge base construction module 101 is used for:

[0059] The document-formatted clinical trial data statistical analysis plan is parsed to obtain text blocks;

[0060] Parse the SDTM metadata to obtain key metadata, which includes at least dataset definition, variable definition, value-level metadata, and code list;

[0061] The general knowledge is parsed to obtain structured rule entries or text blocks with metadata. The general knowledge includes: ADaM model specifications, ADaM implementation guidelines, specific analysis guidelines, ADaM compliance check rules, and expert experience knowledge.

[0062] Specifically, such as Figure 2 The diagram illustrates a flowchart of the automatic conversion system from the clinical trial SDTM dataset to the ADaM dataset in an embodiment of the present invention. Specifically, in the knowledge base construction phase, the knowledge base construction module 101 is used to: process input data sources such as SAP and Define.xlm, construct a dual index of keywords and vectors, and obtain the knowledge base.

[0063] In this embodiment of the invention, the SDTM metadata is obtained by organizing the clinical trial data collected from the original data source through the Research Data Tabulation Model (SDTM), presenting the clinical trial data in a structured manner.

[0064] SDTM is one of the core standards developed by CDISC to standardize the format, structure, and content of raw observational data from clinical trials submitted to regulatory agencies.

[0065] The core purpose of SDTM is to organize raw clinical data into a series of standardized datasets (called "domains") according to themes (such as demographics, adverse events, vital signs, etc.) to facilitate review by regulatory agencies.

[0066] In SDTM, variables are categorized into five roles: Identifier, Topic, Timing, Qualifier, and Rule. These roles collectively define the complete context of each observation. Variables are also classified into three categories based on their importance: Required, Expected, and Permissible. This categorization specifies which variables must be included when constructing the SDTM dataset.

[0067] All SDTM datasets are structured in two-dimensional tables, where rows represent observation records and columns represent variables. Each dataset must be accompanied by a Define.xml file, which serves as metadata and describes in detail the attributes of each variable in the dataset, such as name, label, data type, source, and role.

[0068] In this embodiment of the invention, the Clinical Trial Data Statistical Analysis Plan (SAP) is the core document guiding the entire process of clinical trial data statistical analysis. It aims to ensure that the analysis process is scientific, standardized, and transparent, and to guarantee the reliability and reproducibility of the results. Its content is consistent with the clinical trial protocol and mainly includes the following core elements:

[0069] Trial Background and Objectives: Briefly describe the research and development background and clinical needs of the investigational drug / intervention. Clearly state the primary (e.g., efficacy and safety verification) and secondary objectives of the clinical trial, echoing the research hypotheses in the protocol.

[0070] The purpose and scope of SAP: To standardize statistical analysis methods, processes, and variable definitions, ensuring the traceability of analysis results. To clearly define the applicable data scope (e.g., trial phase, centers, and subject population).

[0071] Key information about the experimental protocol: Key design elements of the experimental protocol (such as trial type, phases, design type: parallel control, crossover design, etc.).

[0072] Detailed experimental design: grouping method, experimental procedure, sample size calculation, etc.

[0073] Dataset definition: Clearly define the key datasets used in the analysis and explain the purpose of each dataset.

[0074] Data quality control: data verification methods and standards for database locking.

[0075] Statistical analysis content and methods: detailed baseline and demographic analysis methods, effectiveness analysis methods, security analysis methods, etc.

[0076] Statistical methods: setting statistical parameters, missing value handling strategies, outlier handling strategies, etc.

[0077] In this embodiment of the invention, the document parser built into the knowledge base construction module 101 can process clinical trial data statistical analysis plans in PDF format, extract plain text content, and segment it into text blocks according to logical units such as chapters, paragraphs, or tables.

[0078] In this embodiment of the invention, the user can upload the SDTM Define.xml file generated by SDTM. The knowledge base construction module 101 uses a dedicated XML parser to accurately extract key metadata such as dataset definition, variable definition, value-level metadata, and code list.

[0079] In this embodiment of the invention, the general knowledge comes from relevant knowledge documents such as the ADaM model specification, ADaM implementation guide, specific analysis guide and ADaM compliance check rules officially released by CDISC. These documents are preprocessed into structured rule entries or text blocks with metadata.

[0080] The CDISC has developed a series of international standards to regulate clinical research data processes, aiming to improve data quality, interoperability, and traceability. Key standards include:

[0081] Protocol Description Model (PRM): Provides a structured template for the design elements of clinical trial protocols (such as study objectives, inclusion and exclusion criteria, visitation plans, etc.), ensuring that different trial protocols maintain consistency in logical framework and key elements.

[0082] Clinical Data Acquisition Standard (CDASH): This standard specifies the naming rules, data types, and collection specifications for basic fields in the Case Report Form (CRF). Most variables can be directly mapped to SDTM to reduce the workload of subsequent data conversion.

[0083] Research Data Tabulation Model (SDTM) and Analysis Dataset Model (ADaM).

[0084] In this embodiment of the invention, after processing all content, the knowledge base construction module 101 can construct two complementary indexes for each information unit:

[0085] Keyword Index: This index is constructed using the industry-standard BM25 algorithm, which evaluates the relevance of query terms to documents by calculating term frequency and inverse document frequency. The advantage of this index lies in its ability to perform precise word matching. In clinical data processing scenarios, this is crucial for retrieving specific variable names, dataset names, function names, or specific rule numbers.

[0086] Vector Indexing: Employing an advanced pre-trained dual encoder model, text blocks are independently encoded into a fixed-dimensional dense vector. This vector captures the deep semantics of the text, enabling conceptual similarity retrieval.

[0087] When subsequent modules of the system (such as the variable generation module and the variable derivation logic generation module) need to obtain contextual information, they can perform a hybrid search of the knowledge base. The system simultaneously sends the query to the keyword index and the vector index, and weights and sorts the query scores to obtain the final candidate documents as the query results.

[0088] In this embodiment of the invention, differentiated processing strategies can be adopted according to the nature of the input information, keyword retrieval or semantic retrieval can be performed based on retrieval enhancement generation technology, and these can be combined for use by LLM.

[0089] In this embodiment of the invention, the complex ADaM generation task is decomposed into multiple phased tasks, and a human expert verification process is included. The goal of the first phase is to generate a list of target variables for the target ADaM dataset.

[0090] ADaM is a standard for standardizing statistical analysis datasets, designed to support efficient, reproducible statistical analysis and results reporting. Its relationship to SDTM and its core principles are as follows:

[0091] Traceability: This is the most fundamental requirement of ADaM. Every record and variable in an ADaM dataset must be clearly traceable to its source, primarily the SDTM dataset. This traceability is achieved by including source domains, source variables, and source sequence numbers in the ADaM dataset, ensuring a complete link from the analysis results to the original observation data.

[0092] Analysis-Ready: ADaM datasets are designed in an "analysis-ready" format, so that statistical analysts can generate final statistical charts and lists (TFLs) from ADaM data in one or very few steps.

[0093] Clear and Unambiguous: The metadata and supporting documentation of the ADaM dataset must clearly describe the derivation logic of all variables, especially those complex calculations and algorithms, without any ambiguity.

[0094] Its basic data structures fall into three categories:

[0095] The Subject-Level Analysis Dataset (ADSL) is the foundation of all ADaM datasets, with only one record for each subject. It contains key subject-level information such as demographic characteristics, treatment group, important baseline information, and key dates in the study.

[0096] Basic Data Structure (BDS): The BDS is used to store analytical data in sets of "one record per parameter, one analysis time point, and one record per subject," such as laboratory test results and vital signs. The BDS structure is flexible and can accommodate a large number of analytical parameters and time point data.

[0097] Occurrence Dataset (OCCDS): Used to describe the occurrence of an event or intervention, such as an adverse event (ADAE) or concomitant medication. Its structure is typically "one record per occurrence, one record per subject".

[0098] ADaM introduces a series of standardized analytical variables to support common statistical analysis needs, such as PARAM (analysis parameters), AVAL (analysis values), ABLFL (baseline record markers), CHG (change from baseline), and ANLzzFL (analysis population markers).

[0099] Specifically, the target variable generation module 102 is used for:

[0100] The large language model is invoked, with the fields of the user-specified ADaM dataset as the target, the clinical trial data statistical analysis plan, SDTM metadata and general knowledge in the knowledge base as background knowledge, and the traceability of the generated target variables as a constraint, to generate a structured list of target variables.

[0101] In this embodiment of the invention, the system, based on the user-specified ADaM dataset, utilizes the aforementioned hybrid retrieval method to obtain comprehensive contextual information from knowledge sources such as SAP, Define.xml, and the ADaM implementation guide. Subsequently, based on the pre-selected first prompt, the system guides the LLM to synthesize all information and generate a structured draft list of target variables.

[0102] In this embodiment of the invention, the first prompt word guides the LLM to take the fields of the ADaM dataset specified by the user as the target, the clinical trial data statistical analysis plan, SDTM metadata and general knowledge in the knowledge base as background knowledge, and the traceability of the generated target variables as a constraint, thereby generating a structured list of target variables.

[0103] Specifically, such as Figure 2 As shown, in the target variable generation stage, the target variable generation module 102 calls LLM to analyze the context in the knowledge base, generate a structured list of target variables, and submit it to the user for review.

[0104] Specifically, the structured list of target variables includes key columns such as variable name, variable label, source type (e.g., “Direct Copy”, “Derived”) and source details (e.g., “SDTM.DM.AGE” or “Derived per SAP Section 5.2.1”), thus initially establishing traceability metadata for ADaM variables.

[0105] The generated draft list of target variables is subsequently reviewed and verified by human experts (such as statistical programmers). Users can check, modify, add to, and delete items on the list through the system interface, and finally complete the verification.

[0106] In this embodiment of the invention, the user confirmation process serves as a critical quality control point, ensuring that the foundation for subsequent automated work is accurate and error-free.

[0107] In this embodiment of the invention, after obtaining the verified list of target variables, the system proceeds to the next stage, generating detailed, human-readable logical derivation steps for all target variables that need to be derived.

[0108] Specifically, such as Figure 2 As shown, in the derivation logic generation stage, the variable derivation logic generation module 103 calls the LLM to generate derivation steps for each derived variable in the target variable list based on the context in the knowledge base, and submits them for user review.

[0109] Specifically, for each variable in the list, the system can construct a specific query, call the hybrid retrieval knowledge base again, and obtain contextual information related to the derivation steps of that variable. Then, the LLM integrates the business rules contained in the contextual information based on the pre-defined second prompt words to generate structured logical derivation steps.

[0110] For example, the logic for generating the variable BMIBL (baseline body mass index) might be as follows:

[0111] Filter baseline weight: From the SDTM dataset VS, filter records where VSTESTCD == 'WEIGHT' and VSBLFL == 'Y', and take the VSORRES value.

[0112] Filter baseline height: From the SDTM dataset VS, filter records where VSTESTCD == 'HEIGHT' and VSBLFL == 'Y', and take the VSORRES value.

[0113] Unit conversion: If the height unit is 'cm', divide by 100 to convert to meters.

[0114] To calculate BMI: Calculate BMIBL = weight / (height in meters 2).

[0115] Formatting: Round the result to one decimal place.

[0116] In this embodiment of the invention, after the variable derivation logic generation module 103 obtains the set of logical derivation steps for all target variables, it can be submitted to the user for a second review. The user can modify any logical step. Only after all the logical derivation steps of the target variables have been finally confirmed by human experts will this set of logical derivation steps of the target variables serve as the basis for the next stage of executable code generation.

[0117] Specifically, the system also includes:

[0118] The verification information acquisition module is used for:

[0119] Receive modification information from the user regarding the target variable list, and obtain a verified target variable list;

[0120] The system receives modification information from the user regarding the logical derivation steps, and obtains verified logical derivation steps.

[0121] In this embodiment of the invention, the user can check, modify, add or delete the target variable list and logical derivation steps generated by LLM through the system interface. The verification information acquisition module receives the modification information and sends it to the corresponding target variable generation module 102 or variable derivation logic generation module 103 to adjust the target variable list or logical derivation steps, so as to obtain the verified target variable list and verified logical derivation steps.

[0122] Specifically, the knowledge base construction module 101 is also used for:

[0123] The modification information sent by the user regarding the list of target variables and the modification information regarding the logical derivation steps are parsed and added to the knowledge base.

[0124] In this embodiment of the invention, the knowledge base construction module 101 can parse the modification information of the target variable list and the modification information of the logical derivation steps sent by the user, and supplement them to the knowledge base as standardized rules, serving as general knowledge. In subsequent data analysis of new clinical trial data statistical analysis projects, this general knowledge is analyzed as background knowledge.

[0125] In this embodiment of the invention, the goal in the final stage of the workflow is to transform the user-verified logical derivation steps into executable and error-free code.

[0126] Specifically, the code generation module 104 is used to: execute the verified executable code in a secure sandbox environment after the executable code has been verified; if the execution fails, instruct the large language model to analyze the error and correct the code through debugging instructions, and execute it again until the code is successfully executed, and send the ADaM dataset generated after the code is successfully executed.

[0127] Specifically, such as Figure 2 As shown, during the code generation and execution phase, the code generation module 104 calls the LLM to translate the verified logical derivation steps into code and submits it for user review. After user approval, it is safely executed in a sandbox and undergoes self-correction. After successful code execution, the final ADaM dataset is submitted.

[0128] Specifically, the system instruction LLM acts as an expert programmer in a specified programming language (such as Python), accurately translating the complete set of logical rules approved in the previous stage into executable code that conforms to syntax specifications and industry best practices. Before execution, the generated code is presented to the user again for final review. After user confirmation, the code will be executed in a completely isolated and secure sandbox environment.

[0129] Furthermore, if the code fails to execute in the environment, the system automatically enters self-correction mode. A debugging prompt (containing user-approved logic rules, LLM-generated code, and error logs from execution failures) instructs the LLM to analyze the error and correct the code. This "execution-capture-analysis-correction" loop can iterate until the code executes successfully or the preset maximum number of attempts is reached.

[0130] Once the self-correcting loop successfully completes, the system delivers the generated ADaM dataset to the user after successful code execution. This closed-loop system significantly improves the robustness of the automated process and the reliability of the final output.

[0131] The automated conversion system from clinical trial SDTM datasets to ADaM datasets provided in this invention does not simply apply artificial intelligence technology to clinical data processing. Instead, it innovatively designs a complete closed-loop automated architecture with LLM as its core engine, which complies with regulatory standards. Its core advantage lies in the synergistic effect of "multi-stage LLM calls" and "hybrid information processing," and by setting mandatory manual review and confirmation nodes between each automated stage. This successfully transforms the highly experience-dependent workflow (i.e., requirements understanding → rule design → programming implementation) of clinical statistical programmers from SDTM to ADaM into a controllable, traceable, and efficient automated process.

[0132] The automatic conversion system from the SDTM dataset to the ADaM dataset provided in this invention maximizes the efficiency of large language models while ensuring the compliance and accuracy of results through human decision-making at key nodes, thus solving the fundamental problem of "unreliability and uncontrollability" of general AI solutions in serious medical scenarios.

[0133] The automated conversion system from SDTM to ADaM datasets for clinical trials provided in this invention heralds a paradigm shift in clinical trial data processing and has profound industry implications. Through AI-driven automation, this system compresses data preparation work, which traditionally takes weeks or even months, into hours, significantly shortening the analysis cycle and accelerating the overall drug development and market launch process, enabling innovative therapies to benefit patients sooner. The systematic process and embedded CDISC standard knowledge base ensure the high quality, consistency, and traceability of the analyzed datasets from the source, significantly improving the accuracy and compliance of data submitted to global regulatory agencies such as the China Food and Drug Administration and the U.S. Food and Drug Administration. In the long run, this invention is a key cornerstone for achieving end-to-end intelligent clinical research data flow. It efficiently combines complex business rules, expert knowledge, and cutting-edge AI capabilities, not only solving current data conversion challenges but also providing solid technical support for future fully automated statistical analysis, report generation, and even intelligent clinical development strategies, powerfully propelling the entire clinical research industry towards a more efficient and precise intelligent era.

[0134] An embodiment of the automatic conversion system from clinical trial SDTM datasets to ADaM datasets provided by this invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 3 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is used in the automatic conversion system from clinical trial SDTM dataset to ADaM dataset provided by this invention. (Except for...) Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0135] Based on the same inventive concept, this invention also provides an automatic conversion method from clinical trial SDTM datasets to ADaM datasets, the method comprising:

[0136] S201 analyzes clinical trial data statistical analysis plans, SDTM metadata, and general knowledge to build a dual-index knowledge base that supports keyword and vector retrieval;

[0137] S202, invoke the large language model, based on the first prompt word, and perform a hybrid retrieval of the knowledge base according to the user-specified ADaM dataset to generate a structured list of target variables;

[0138] S203, after the target variable list is validated, the large language model is invoked, and a hybrid retrieval is performed on the knowledge base based on the second prompt word. Based on the business rules contained in the context information related to the target variable in the knowledge base, a structured logical deduction step is generated.

[0139] S204, after the logical deduction step has been verified, the structured logical deduction step is translated into executable code based on the third prompt word.

[0140] Based on the same inventive concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements an automatic conversion system for converting clinical trial SDTM datasets to ADaM datasets as described in any of the above embodiments.

[0141] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the automatic conversion system from clinical trial SDTM dataset to ADaM dataset as described in any of the above embodiments.

[0142] Based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the automatic conversion system from clinical trial SDTM dataset to ADaM dataset as described in any of the above embodiments.

[0143] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (apparatus), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable terminal device, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable terminal device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions can also be loaded onto a computer or other programmable terminal device to cause a series of operational steps to be performed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0149] Finally, it should be noted that in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0150] The above provides a detailed description of an automatic conversion system for clinical trial SDTM datasets to ADaM datasets provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An automatic conversion system for clinical trial SDTM datasets to ADaM datasets, characterized in that, The system includes: The knowledge base construction module is used to parse clinical trial data statistical analysis plans, SDTM metadata, and general knowledge to build a dual-index knowledge base that supports keyword and vector retrieval. This includes: parsing document-formatted clinical trial data statistical analysis plans to obtain text blocks; parsing SDTM metadata to obtain key metadata, which at least includes dataset definitions, variable definitions, value-level metadata, and code lists; and parsing general knowledge to obtain structured rule entries or text blocks with metadata, including: ADaM model specifications, ADaM implementation guidelines, specific analysis guidelines, ADaM compliance check rules, and expert experience knowledge. The target variable generation module is used to call the large language model, perform hybrid retrieval on the knowledge base based on the first prompt word and the user-specified ADaM dataset, and generate a structured list of target variables. The variable derivation logic generation module is used to call the large language model after the target variable list has been verified, perform a hybrid retrieval of the knowledge base based on the second prompt word, and generate structured logical derivation steps based on the business rules contained in the context information related to the target variable in the knowledge base. The code generation module is used to translate the structured logical deduction steps into executable code based on the third prompt word after the logical deduction steps have been verified. This includes: after the executable code has been verified, executing the verified executable code in a secure sandbox environment; in the event of execution failure, instructing the large language model to analyze the error and correct the code through debugging instructions, and then executing it again until the code executes successfully, and sending the ADaM dataset generated after the code executes successfully.

2. The automatic conversion system from clinical trial SDTM dataset to ADaM dataset according to claim 1, characterized in that, The target variable generation module is used for: The large language model is invoked, with the fields of the user-specified ADaM dataset as the target, the clinical trial data statistical analysis plan, SDTM metadata and general knowledge in the knowledge base as background knowledge, and the traceability of the generated target variables as a constraint, to generate a structured list of target variables.

3. The automatic conversion system from clinical trial SDTM dataset to ADaM dataset according to claim 1, characterized in that, The structured list of target variables includes: variable name, variable label, source type, and source details, thus initially establishing traceable metadata for ADaM variables.

4. The automatic conversion system from clinical trial SDTM dataset to ADaM dataset according to claim 1, characterized in that, The system further includes: a verification information acquisition module, used for: Receive modification information from the user regarding the target variable list, and obtain a verified target variable list; The system receives modification information from the user regarding the logical derivation steps, and obtains verified logical derivation steps.

5. The automatic conversion system from clinical trial SDTM dataset to ADaM dataset according to claim 4, characterized in that, The knowledge base construction module is also used for: The modification information sent by the user regarding the list of target variables and the modification information regarding the logical derivation steps are parsed and added to the knowledge base.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the automatic conversion system from the clinical trial SDTM dataset to the ADaM dataset as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the automatic conversion system from the clinical trial SDTM dataset to the ADaM dataset as described in any one of claims 1-5.

8. A computer program product comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program / instruction implements the automatic conversion system from the clinical trial SDTM dataset to the ADaM dataset as described in any one of claims 1-5.

Images