Construction and evaluation method of applicant drug clinical test quality management system

Abstract

The invention discloses a construction and evaluation method for an applicant drug clinical test quality management system, and aims to solve the problems that traceable evidence chains are lacked among applicant clinical test quality system elements, evidence materials and evaluation conclusions, evaluation is difficult to quantify, and closed-loop improvement based on evaluation results is difficult. According to the method, quality system element data, evidence material data and evidence source data are obtained and standardized, and a time heterogeneous evidence graph comprising element nodes, evidence nodes, version nodes, responsibility nodes and evidence source nodes is constructed; executing graph neural network representation learning on the evidence graph to generate a relationship likelihood parameter and an evidence credibility likelihood parameter; constructing a Bayesian factor graph containing conflict factors, and executing belief propagation to obtain element compliance posterior distribution and evidence credibility posterior distribution; and further calculating a compliance score, an uncertainty score and a coverage rate score to form a three-dimensional score result, and extracting an evidence chain to generate traceable output, thereby achieving the technical effects of automatic establishment of the evidence chain, compliance quantitative evaluation, conflict evidence suppression and evaluation result driven closed-loop improvement.

Description

Technical Field

[0001] This invention relates to the field of quality management in pharmaceutical clinical trials, and more particularly to a method for constructing and evaluating a quality management system for sponsor-sponsored drug clinical trials. Background Technology

[0002] Drug clinical trials involve multiple stages, including protocol design, subject protection, data management, pharmacovigilance, and supplier management. Sponsors typically need to establish a comprehensive quality management system covering the entire process and create corresponding policy documents, process records, and evidence materials in accordance with regulations such as drug clinical trial guidelines to support regulatory inspections, audits, and internal quality improvement. With the increasing informatization and digitalization of clinical trials, sponsors generally build electronic document management systems, clinical trial management systems, electronic data acquisition systems, training management systems, standard operating procedure management systems, and corrective and preventive action management systems to achieve traceability of quality activities, document archiving, and process tracking. Simultaneously, the industry is gradually introducing risk-based quality management concepts, evaluating the operational status of the quality system and promoting continuous improvement by setting quality indicators, conducting monitoring and audits.

[0003] Existing technologies still have the following shortcomings in the construction and evaluation of sponsors' clinical trial quality management systems:

[0004] 1. The evidence chain mainly relies on manual compilation, and there is a lack of automatically generated and verifiable correspondence between quality system elements, evidence materials and evaluation conclusions; information such as the time attributes, version applicability, responsibility boundaries, authorization validity period, and audit trail of evidence is scattered in different systems or documents, resulting in high traceability costs and easy omissions.

[0005] 2. The assessment methods are mostly checklist-based or rule-based scoring, which makes it difficult to conduct a unified quantitative assessment of the credibility of evidence, the coverage of evidence, and the compliance of elements. It also makes it difficult to output the uncertainty of the assessment results, affecting the consistency and interpretability of the assessment.

[0006] 3. There is a lack of effective modeling and inference mechanisms for changes in credibility caused by differences in evidence sources and for situations where there are contradictions between evidence. As a result, the evaluation results are difficult to form a structured and traceable output and automatically drive subsequent improvement actions, resulting in low efficiency of closed-loop improvement.

[0007] Therefore, the construction and evaluation method of a drug clinical trial quality management system that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a method for constructing and evaluating a quality management system for sponsor drug clinical trials. In response to the problems in the prior art, the correspondence between quality system elements, evidence materials and evaluation conclusions relies on manual sorting, making it difficult to form a traceable evidence chain. The evaluation is difficult to quantify and lacks expression of uncertainty. It is also difficult to uniformly handle differences in the credibility of evidence sources and contradictions in evidence, and it is difficult to drive closed-loop improvement. Therefore, a technical solution based on time heterogeneous evidence graphs and probabilistic inference is proposed. This invention acquires and standardizes quality system element data, evidence material data, and evidence source data to construct a time-heterogeneous evidence graph containing element nodes, evidence nodes, version nodes, responsibility nodes, and evidence source nodes. It records time attributes for edges, validity periods for versions, authorization scope and validity periods for responsibility relationships, audit trail information for evidence, and source credibility for evidence sources. Graph neural network representation learning is performed on the evidence graph to generate relationship likelihood parameters and evidence credibility likelihood parameters. A Bayesian factor graph containing conflict factors is then constructed, and confidence propagation is performed to obtain the posterior distribution of element compliance and evidence credibility. Compliance scores, uncertainty scores, and coverage scores are calculated to form a three-dimensional scoring result. The evidence chain is extracted to generate traceable outputs and closed-loop improvement projects. This invention possesses the technical effects of automatic evidence chain construction, conflict evidence suppression, quantitative compliance assessment, and improvement-driven closed-loop output.

[0009] This invention provides a method for constructing and evaluating a sponsor's drug clinical trial quality management system, comprising:

[0010] S1. Obtain quality system element data, evidence material data, and evidence source data from the sponsor's drug clinical trial, and perform standardization processing. S2. Construct a time-heterogeneous evidence graph based on the standardized data, including element nodes, evidence nodes, version nodes, responsibility nodes, and evidence source nodes, as well as the relationship edges between element nodes and evidence nodes, the version applicability edges between evidence nodes and version nodes, the responsibility attribution edges between evidence nodes and responsibility nodes, and the source relationship edges between evidence nodes and evidence source nodes. Record time attributes for edges, record version validity attributes for version nodes, record authorization scope and authorization validity attributes for responsibility attribution edges, record audit trail information for evidence nodes, and set source credibility attributes for evidence source nodes. S3. Perform graph neural network representation learning on the time-heterogeneous evidence graph to generate likelihood parameters, including relationship likelihood parameters for the validity of relationships between element nodes and evidence nodes, and evidence credibility likelihood parameters for evidence nodes. S4. Construct a target Bayesian algorithm based on the time-heterogeneous evidence graph and likelihood parameters. The factor graph includes element factors, evidence factors, relationship factors, and conflict factors. Conflict factors are constructed based on the mutually exclusive or contradictory relationships between evidence nodes associated with the same element node. These mutually exclusive or contradictory relationships include conflicts determined by time attributes, version validity attributes, authorization scope attributes, authorization validity attributes, or audit trail information. The joint probability of conflict is reduced through factor potential functions. S5: Perform confidence propagation on the target Bayesian factor graph to obtain the posterior distribution of element compliance for each element node and the posterior distribution of evidence credibility for each evidence node. S6: Calculate compliance score and uncertainty score based on the posterior distribution of element compliance. Calculate coverage score based on predefined evidence requirements and the posterior distribution of evidence credibility. Combine the compliance score, uncertainty score, and coverage score to obtain a three-dimensional score result and generate an evaluation conclusion. S7: Determine the target element node based on the evaluation conclusion and extract the evidence chain associated with the target element node from the time-heterogeneous evidence graph, generating a traceable output associated with the evidence chain.

[0011] Optionally, S1 includes:

[0012] Obtain the quality system element data, which includes element identifier, element name, element scope of application, corresponding process or activity, corresponding responsible entity, and corresponding evidence requirements.

[0013] The evidence material data is obtained, which includes evidence identifier, evidence type, evidence content or evidence content index, evidence generation time, version information corresponding to the evidence, and responsibility information corresponding to the evidence, and includes audit trail information;

[0014] Obtain the evidence source data, which includes a source identifier, source system information, and verification information used to determine the credibility of the source;

[0015] Convert the above data fields into a preset data format, and establish the correspondence between element identifiers and evidence identifiers, and between source identifiers and evidence identifiers, according to the preset identifier alignment rules;

[0016] The time field is subjected to timestamp normalization processing, which includes unifying the time format and time zone, and missing time fields are filled or marked.

[0017] Optionally, S2 includes:

[0018] Element nodes are generated based on the standardized quality system element data;

[0019] Evidence nodes are generated based on the standardized evidence data, and the audit trail information is written into the evidence nodes;

[0020] Based on the standardized evidence data, a version node is generated and the version validity period attribute is written; a responsibility node is generated and the role or permission scope information is written.

[0021] The associated edge, version applicable edge, and responsibility attribution edge are generated based on the identification alignment relationship, and the time attribute is written to each edge.

[0022] Event nodes are generated based on the event information in the evidence material data, and event occurrence edges are generated between the evidence nodes and the event nodes, and time attributes are recorded for the event occurrence edges.

[0023] Furthermore, setting source credibility attributes for evidence source nodes includes:

[0024] Based on the computerized system verification status, audit trail integrity, and data integrity verification results, source credibility attributes are written to the evidence source nodes according to preset assignment rules, which include standardization and weighted summarization.

[0025] The source credibility attribute is then written into the evidence node connected to it as the evidence source credibility attribute of the evidence node.

[0026] When the same evidence node is connected to multiple evidence source nodes, the credibility attributes of the multiple sources are aggregated according to a preset aggregation rule to update the credibility attributes of the evidence source of the evidence node.

[0027] Optionally, S3 includes:

[0028] Node feature vectors are constructed for different types of nodes in the time heterogeneous evidence graph, and the time attributes of the edges and the version validity period attributes are encoded as time features.

[0029] Based on the node feature vector and the time feature, a graph neural network is used to perform multiple rounds of message passing and aggregation updates on the time heterogeneous evidence graph to obtain the element node representation vector and the evidence node representation vector.

[0030] The relationship likelihood parameter is calculated based on the element node representation vector and the evidence node representation vector;

[0031] The likelihood parameter of the evidence credibility is calculated based on the evidence node representation vector and the evidence source credibility attribute of the evidence node.

[0032] Optionally, S4 includes:

[0033] Set the compliance status variable of each element node in the time heterogeneous evidence graph, and set the credibility status variable of each evidence node in the time heterogeneous evidence graph.

[0034] Generate element factors based on element compliance status variables, and generate evidence factors based on evidence credibility status variables;

[0035] For the associated edges, generate relationship factors and connect the corresponding element compliance status variables and evidence credibility status variables;

[0036] Based on the relationship likelihood parameter and the evidence credibility likelihood parameter, factor potential functions are set for the relationship factor and the evidence factor, respectively.

[0037] For a set of evidence nodes associated with the same element node, mutual exclusion or contradiction relationships are determined based on the evidence node's time attribute, version validity period attribute, authorization scope attribute, authorization validity period attribute, and audit trail information. Conflict factors are then generated and connected to the corresponding evidence credibility status variables.

[0038] Optionally, S5 includes:

[0039] Set initial marginal distributions for compliance status variables of each element and credibility status variables of each piece of evidence;

[0040] Iterative message passing is performed in the target Bayesian factor graph, which includes message computation from factor to variable and message computation from variable to factor, and the messages are normalized after each iteration.

[0041] The iteration stops when the number of iterations reaches a preset threshold or when the change in messages between two adjacent iterations is less than a preset change threshold.

[0042] The compliance posterior distribution of the element and the credibility posterior distribution of the evidence are calculated based on the message at the time of termination of the iteration.

[0043] Optionally, S6 includes:

[0044] The posterior mean is calculated as the compliance score based on the posterior distribution of the aforementioned elements, and the posterior uncertainty is calculated as the uncertainty score. The posterior uncertainty is determined by the variance of the posterior distribution or by the width of the confidence interval under a preset confidence level.

[0045] Based on predefined sub-feature mapping relationships, feature nodes are mapped to at least two types of sub-features, and the denominator is determined for each type of sub-feature according to evidence requirements;

[0046] In the time-heterogeneous evidence graph, evidence nodes that are associated with element nodes and whose posterior distribution of evidence credibility meets a preset credibility threshold are retrieved as coverage evidence, and coverage scores are calculated.

[0047] The compliance score, uncertainty score, and coverage score are combined to generate a three-dimensional score result.

[0048] Optionally, the S7 includes:

[0049] Element nodes whose evaluation conclusions indicate that their compliance score is lower than the preset compliance score threshold, their uncertainty score is higher than the preset uncertainty score threshold, or their coverage score is lower than the preset coverage score threshold are identified as target element nodes.

[0050] In the time-heterogeneous evidence graph, the evidence node set is retrieved along the associated edge starting from the target element node, and the retrieval is further extended along the version application edge, responsibility attribution edge and source association edge to obtain the version node set, responsibility node set and evidence source node set to generate the evidence chain;

[0051] The identifiers of each node and edge in the evidence chain, along with their time attributes, version validity attributes, source credibility attributes, authorization scope attributes, authorization validity attributes, and audit trail information, are written into the traceable output.

[0052] The traceable output further includes closed-loop improvement projects generated based on gap types, which include missing evidence, insufficient evidence credibility, evidence not matching version validity, evidence not matching responsibility, or evidence conflict.

[0053] The beneficial effects of this invention are:

[0054] 1. By constructing a time-heterogeneous evidence graph that includes time attributes, version validity period, responsibility authorization validity period, audit trail information, and evidence source credibility attributes, and extracting evidence chains associated with target elements from the graph, automatic association and traceable output between quality system elements, evidence materials, and evaluation conclusions are achieved, reducing the cost of manual collection and review, and improving the efficiency and consistency of inspection preparation.

[0055] 2. By using a graph neural network on the evidence graph to generate relationship likelihood parameters and evidence credibility likelihood parameters, and constructing a Bayesian factor graph containing conflicting factors to perform confidence propagation, it is possible to make consistent inferences on element compliance and evidence credibility on a global scale. At the same time, it can suppress contradictory evidence caused by time, version applicability, authorization boundaries or audit trail anomalies, reduce omissions and misjudgments and improve the interpretability of the assessment.

[0056] 3. By calculating compliance scores, uncertainty scores, and coverage scores based on posterior distributions to form a three-dimensional scoring result, and using this result to locate weak elements and generate structured gap types and improvement projects, a quantitative assessment of the operational status of the quality management system and a closed-loop result-driven improvement are achieved, thereby improving the feasibility and verifiability of continuous improvement. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart illustrating the construction and evaluation method of a sponsor's drug clinical trial quality management system proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1 A method for constructing and evaluating a sponsor's drug clinical trial quality management system, comprising:

[0061] S1. Obtain quality system element data, evidence material data, and evidence source data from the sponsor's drug clinical trial, and perform standardization processing. S2. Construct a time-heterogeneous evidence graph based on the standardized data, including element nodes, evidence nodes, version nodes, responsibility nodes, and evidence source nodes, as well as the relationship edges between element nodes and evidence nodes, the version applicability edges between evidence nodes and version nodes, the responsibility attribution edges between evidence nodes and responsibility nodes, and the source relationship edges between evidence nodes and evidence source nodes. Record time attributes for edges, record version validity attributes for version nodes, record authorization scope and authorization validity attributes for responsibility attribution edges, record audit trail information for evidence nodes, and set source credibility attributes for evidence source nodes. S3. Perform graph neural network representation learning on the time-heterogeneous evidence graph to generate likelihood parameters, including relationship likelihood parameters for the validity of relationships between element nodes and evidence nodes, and evidence credibility likelihood parameters for evidence nodes. S4. Construct a target Bayesian algorithm based on the time-heterogeneous evidence graph and likelihood parameters. The factor graph includes element factors, evidence factors, relationship factors, and conflict factors. Conflict factors are constructed based on the mutually exclusive or contradictory relationships between evidence nodes associated with the same element node. These mutually exclusive or contradictory relationships include conflicts determined by time attributes, version validity attributes, authorization scope attributes, authorization validity attributes, or audit trail information. The joint probability of conflict is reduced through factor potential functions. S5: Perform confidence propagation on the target Bayesian factor graph to obtain the posterior distribution of element compliance for each element node and the posterior distribution of evidence credibility for each evidence node. S6: Calculate compliance score and uncertainty score based on the posterior distribution of element compliance. Calculate coverage score based on predefined evidence requirements and the posterior distribution of evidence credibility. Combine the compliance score, uncertainty score, and coverage score to obtain a three-dimensional score result and generate an evaluation conclusion. S7: Determine the target element node based on the evaluation conclusion and extract the evidence chain associated with the target element node from the time-heterogeneous evidence graph, generating a traceable output associated with the evidence chain.

[0062] In this specific embodiment, S1 includes:

[0063] The computer system retrieves data from the sponsor's quality management master data table, electronic document management system, clinical trial management system, electronic data acquisition system, and training management system through direct database connections. In the same batch processing task, three types of input datasets are generated. Each quality system element data entry includes an element identifier, element name, element scope, corresponding process or activity, corresponding responsible entity, and corresponding evidence requirements. Each evidence material data entry includes an evidence identifier, evidence type, evidence content or evidence content index, evidence generation time, corresponding version information, and corresponding responsibility information, and also includes audit trail information. Each evidence source data entry includes a source identifier, source system information, and verification information used to determine source credibility. The audit trail information is defined as a set of records of events such as the creation, modification, review, approval, and revocation of evidence materials, with each record including event type, event occurrence time, operator identifier, and a summary of the content before and after the change. The verification information is defined as a set of structured fields used for subsequent source credibility calculations and includes at least a computerized system verification status identifier, an audit trail function activation identifier, an audit trail integrity verification result field, and a data integrity verification result field.

[0064] The three types of input datasets are converted into preset data formats. The preset data formats are defined as a structured record format with unified character encoding as UTF-8, unified field names as fixed key sets, unified enumeration values ​​as preset code tables, unified text fields with invisible control characters removed, and unified index fields represented by URIs and carrying hash check values. The data source system identifier and extraction batch number are written to each record to support subsequent traceability.

[0065] A correspondence between element identifiers and evidence identifiers, and between source identifiers and evidence identifiers, is established based on preset identifier alignment rules. The preset identifier alignment rules are defined as follows: the element identifier reference field explicitly written in the evidence material data is used as the first alignment key to directly generate a mapping relationship table between element identifiers and evidence identifiers. When the reference field is missing, the evidence type listed in the evidence requirement field in the element data is matched with the evidence type in the evidence material data as the second alignment key. When the match is successful, a mapping relationship table is generated. The source system identifier field in the evidence material data is matched with the source identifier field in the evidence source data to generate a mapping relationship table between source identifiers and evidence identifiers. The two mapping relationship tables and the original records are stored in the database for direct use in subsequent steps.

[0066] Timestamp normalization is performed on all time fields. This normalization is defined as converting the evidence generation time and the event occurrence time in the audit trail information to a unified time format and time zone, and then writing the conversion result into the normalized time field. The unified time format is defined as a UNIX timestamp in milliseconds calculated from Coordinated Universal Time (UTC), while retaining the original time string field for verification. The unified time zone is defined as the UTC time zone and calculated according to the formula:

[0067] ;

[0068] in This represents the converted UTC timestamp in seconds. This represents the local timestamp in seconds corresponding to the original time. This represents the time difference in seconds between the local time zone and UTC, calculated from the time zone offset configured in the source system.

[0069] The missing time field is filled or marked. The filling or marking process is defined as follows: when the evidence generation time is missing, the earliest creation event time is extracted from the audit trail information corresponding to the same evidence identifier as the evidence generation time and written into the filling identifier field. When there is no event time that can be used for filling in the audit trail information, the evidence generation time is written into a null value, the missing identifier field is set to a preset missing code, and the missing reason is written into the missing reason field to ensure that the time information gap can be explicitly perceived and traceable conclusions can be generated in the subsequent inference and scoring stages.

[0070] In this specific embodiment, S2 includes:

[0071] The computer system uses a standardized data table and an identifier alignment table as its sole input to construct a temporally heterogeneous evidence graph in a graph database. The The system adopts an attribute graph storage structure and consists of a node table and an edge table. The node table contains six types of records: feature nodes, evidence nodes, version nodes, responsibility nodes, evidence source nodes, and event nodes. Each record is written with a globally unique node identifier, a node type field, and a creation batch number field.

[0072] The system generates element nodes one by one based on the quality system element data and writes the element identifier, element name, element scope of application, the process or activity corresponding to the element, the responsible entity corresponding to the element, and the evidence requirement field corresponding to the element.

[0073] The system generates evidence nodes one by one based on the evidence material data and writes the fields of evidence identifier, evidence type, evidence content or evidence content index, evidence generation time and audit trail information. The audit trail information is written into the evidence node in a structured array and the array elements include event type, event occurrence time, operator identifier and content summary before and after the change, so that the evidence node can be directly used for audit trail consistency verification and conflict identification in subsequent steps.

[0074] The system generates version nodes one by one based on the version information in the evidence material data and writes the version identifier, version number, version validity start time and version validity end time fields. The version validity start time and version validity end time both adopt the standardized UTC timestamp field of step S1.

[0075] The system generates responsibility nodes one by one based on the responsibility information in the evidence data and writes them into the responsibility identifier, role field and permission scope field. The permission scope field is represented by a structured permission list and the list elements are composed of permission object identifiers and permission operation sets.

[0076] The system generates evidence source nodes one by one based on the evidence source data and writes source identifier, source system information, and source credibility attribute. The source credibility attribute is obtained by standardizing and weighting three types of verification information—computerized system verification status, audit trail integrity verification result, and data integrity verification result—according to preset assignment rules, and the value range is [insert range here]. The verification status of the computerized system is mapped to standardized values. The audit trail integrity verification results are mapped to standardized values. Data integrity verification results are mapped to standardized values. And calculate the source credibility attribute:

[0077] ;

[0078] in This represents the source credibility attribute value of the evidence source node. The weight representing the verification status of the computerized system is 0.40. This represents the weight of the audit trail integrity verification result and its value is... The weight of the data integrity verification result is represented by a value of 1. This represents the standardized value obtained by mapping the verification status identifier, and is 1 when the verification status identifier is verified, 0.5 when it is being verified, and 0.5 when it is not verified. This represents the standardized value obtained by mapping the audit trail integrity check result field, where 1 is taken when the check passes and 0 is taken when the check fails. This represents the standardized value obtained by mapping the data integrity verification result field, and is 1 when the verification passes and 0 when the verification fails.

[0079] The system generates associated edges between feature nodes and evidence nodes based on the identifier alignment relationship table and writes an edge identifier, an edge type field, and a time attribute field for each associated edge. The time attribute field takes the evidence generation time of the evidence node to indicate the start time when the evidence generates a supporting relationship with the feature.

[0080] The system generates a version applicability edge between the evidence node and the version node and writes it into the time attribute field, wherein the time attribute field is taken as the evidence generation time. At the same time, the version validity start time and version validity end time are retained in the version node for subsequent version applicability determination.

[0081] The system generates a responsibility attribution edge between the evidence node and the responsibility node, and writes the authorization scope field, authorization validity period start time, and authorization validity period end time field into the responsibility attribution edge. The authorization scope field is taken from the authority scope field of the responsibility node, and the authorization validity period start time and authorization validity period end time are obtained by normalizing the authorization cycle recorded in the responsibility information.

[0082] The system generates source association edges between evidence nodes and evidence source nodes and writes them into a time attribute field, where the time attribute field is the evidence generation time. At the same time, the source credibility attribute of the evidence source node is written into the evidence node connected to it to form an evidence source credibility attribute field. When the same evidence node is connected to multiple evidence source nodes, the system uses a deterministic aggregation rule to update the evidence source credibility attribute field of the evidence node. The aggregation rule is to take the minimum value of the source credibility attribute of all connected evidence source nodes and write the list of source identifiers participating in the aggregation into the evidence node to ensure conservative credibility and traceability.

[0083] The system generates event nodes one by one based on event information in the evidence material data. Each event node contains an event identifier, event type, event occurrence time, and operator identifier fields. An event occurrence edge is generated between the evidence nodes and event nodes, and a time attribute field is recorded for each event occurrence edge. This time attribute field is the event occurrence time. Therefore, the audit trail information is simultaneously presented in two consistent ways: an embedded structure within the evidence nodes and an external event subgraph within the evidence nodes. This enables subsequent steps to perform unified retrieval and inference of time, version validity, authorization boundaries, and source credibility.

[0084] In this specific embodiment, S3 includes:

[0085] Computer systems with temporal heterogeneous evidence graphs This is used as input and to call a heterogeneous graphical neural network model with pre-trained and fixed parameters. Perform representation learning to output relation likelihood parameters and evidence credibility likelihood parameters;

[0086] First of all Different types of nodes in the system construct node feature vectors and unify them to the same dimension to facilitate cross-type message transmission. Among them, the feature of element nodes is obtained by concatenating the feature identifier, the scope of application of the element, the process or activity code, the responsible entity code, and the evidence requirement code; the feature of evidence nodes is obtained by concatenating the evidence type code, the evidence content index hash, the normalized value of the evidence generation time, the audit trail information statistics, and the credibility attribute of the evidence source; the feature of version nodes is obtained by concatenating the version number code and the normalized value of the version validity period start and end time; the feature of responsibility nodes is obtained by concatenating the role code and the multi-hot vector of the scope of authority; the feature of evidence source nodes is obtained by concatenating the source system type code, the computerized system verification status code, the audit trail integrity verification result code, and the data integrity verification result code; and the feature of event nodes is obtained by concatenating the event type code, the normalized value of the event occurrence time, and the operator identifier code.

[0087] The encoding method is defined as follows: for discrete fields, a trainable embedding table is used to perform lookup mapping and output a fixed-length vector; for continuous fields, min-max normalization is performed and a scalar is output; for the scope of permissions and evidence requirements, a multi-hot vector is used and expanded according to a preset field order, so that all node feature vectors have a definite field composition and a definite splicing order before entering the model.

[0088] The system encodes the time attributes of edges and the version validity period attributes of version nodes into time features and uses them as part of the edge input. The time features are defined as a relative time scalar calculated from the difference between the edge timestamp and the evaluation benchmark timestamp, a validity period matching scalar calculated from the inclusion relationship between the version validity period start and end time and the edge timestamp, and an authorization matching scalar calculated from the inclusion relationship between the authorization validity period start and end time and the edge timestamp. The time features and edge types are used together to distinguish message passing under different time semantics.

[0089] The heterogeneous graph neural network model Employing two-round message passing and aggregation updates, with the hidden representation dimension set to... Its parameters consist of the linear transformation matrix corresponding to each edge type, the linear mapping matrix of the time feature, and the input projection matrix corresponding to each node type. During the training phase, these parameters are learned using supervised samples formed from historical audit conclusions and evidence verification results. After training, the parameters are solidified and implemented in the following steps. Only forward computation is performed.

[0090] In the forward computation, the model generates neighbor messages for each round of message passing based on edge type and time characteristics. It then averages and aggregates the messages of the same type, concatenates them with its own representation, and updates the node representation through a linear transformation and ReLU activation function. This yields the feature node representation vector and the evidence node representation vector, denoted as follows: and ,in Representing feature nodes In the model output layer dimensional represents a vector. Indicates evidence node In the model output layer Dimension represents a vector;

[0091] The system then calculates the likelihood parameters based on the representation vector and writes them into a memory data structure for use in step S4 to directly construct the Bayesian factor graph. The likelihood parameters are calculated in one step using the following formula:

[0092] ;

[0093] in Representing feature nodes With evidence nodes The likelihood parameter for the relationship between the two is given, and its value range is given. Indicates evidence node The likelihood parameter of the credibility of evidence and its value range is This represents the logical Sigmoid function. This represents the transpose operation of a vector or matrix. The trainable weight matrix represents the relational likelihood and is stored in the fixed model parameters. A trainable weight vector representing the likelihood of evidence credibility is stored in the fixed model parameters. This represents the vector of evidence nodes. Credibility attribute of evidence source for evidence nodes Obtained by concatenating vectors in a fixed order dimensional vector, This indicates that step S2 writes the evidence source credibility attribute of the evidence node, and the value range is... And the stated With the These parameters are used as input parameters for the relational factor potential function and the evidence factor potential function in step S4, respectively, to achieve a learnable quantitative representation of the correlation strength and credibility of element-evidence.

[0094] In this specific embodiment, S4 includes:

[0095] Computer systems with temporal heterogeneous evidence graphs and relational likelihood parameters Likelihood parameter of evidence credibility Construct a target Bayesian factor graph for the input. The The system employs a factor graph data structure, consisting of a set of variable nodes and a set of factor nodes. For each feature node... Establish an element compliance status variable And regulations This indicates that the element is compliant. This indicates that the element is non-compliant, and the system will address each evidence node accordingly. Establish a credible state variable for evidence And regulations This indicates that the evidence is credible. This indicates that the evidence is unreliable;

[0096] The system for each Generate a feature factor and connect it as a unary factor to the corresponding feature compliance state variable. The system for each Generate an evidence factor and connect it as a unary factor to the corresponding evidence credibility state variable. The system Each element node With evidence nodes The associated edge between them generates a relation factor and is connected as a binary factor to it. and To express the "strength of support for the compliance of elements when the evidence is credible";

[0097] The system is based on the same element node The set of associated evidence nodes is used to construct conflict factors. The construction process involves considering any two evidence nodes within this set. and Calculate the conflict determination result and generate a conflict factor when a conflict is determined, and connect the conflict factor as a binary factor to the given result. and The conflict determination result consists of two parts: mutual exclusion determination and contradiction determination. A conflict is determined if either part is true. The mutual exclusion determination is completed through a built-in mutual exclusion rule table, which uses evidence type pairs as keys and mutual exclusion identifiers as values, and is fixed during deployment. The condition for a mutual exclusion relationship to be established is the evidence node. With evidence nodes An ordered pair of evidence type fields corresponds to a mutually exclusive flag that is true in the mutual exclusion rule table, and both are in the same feature node. The applicable intervals overlap, and the applicable intervals are jointly determined by the evidence generation time of the evidence node, the version validity period start and end time of the version node connected to it, and the authorization validity period start and end time of the responsibility attribution edge connected to it. The overlap determination adopts the deterministic rule of non-empty interval intersection.

[0098] The determination of contradictory relationships is completed through a deterministic consistency check of time attribute, version validity period attribute, authorization scope attribute, authorization validity period attribute and audit trail information. The failure of the check is regarded as the establishment of a contradictory relationship. The consistency check includes: the evidence generation time of the evidence node falling within the version validity period of the version node connected to it as the version consistency pass condition; the evidence generation time of the evidence node falling within the authorization validity period of the responsibility attribution edge connected to it as the authorization validity period consistency pass condition; the permission object identifier and permission operation set recorded in the responsibility information of the evidence node being included in the permission scope field of the responsibility node as the authorization scope consistency pass condition; and the existence of a unique creation event in the audit trail information of the evidence node and no modification event after the approval event as the audit trail consistency pass condition.

[0099] After creating variable nodes and factor nodes, the system writes each factor into a potential function table and stores the connection relationships between factors and variables in a sparse adjacency list, thus forming a computational graph that can be used for subsequent confidence propagation. The potential function table is arranged in the order of binary variable values. Expand into a fixed-length array and define it using the following formula:

[0100] ;

[0101] in The potential function representing the factor. The potential function representing the evidence factor. The potential function representing the relation factor. The potential function representing the conflict factor. Representing feature nodes The corresponding element compliance status variables, Indicates evidence node The corresponding evidence credibility state variable, Represents the prior probability of compliance of an element and takes the value of This represents the likelihood parameter indicating the credibility of evidence, and its value range is... Representing feature nodes With evidence nodes The likelihood parameter for the relationship between the two is given, and its value range is given. This represents the conflict suppression coefficient, with a value of 0.1, and is passed when the conflict factor is valid and both conflicting pieces of evidence are simultaneously inferred to be credible. Reduce the joint probability to suppress the interference of conflicting evidence on subsequent compliance inferences.

[0102] In this specific embodiment, S5 includes:

[0103] Computer systems using target Bayesian factor graphs Execute confidence propagation operation on the input to obtain the posterior distribution of compliance state variables and credible state variables of each element;

[0104] The system fixes the value range of each variable node as follows: and for Each "variable node-factor node" connection edge is assigned two directional message vectors. The message vectors are real number arrays of length 2, and the array indices correspond to state 0 and state 1, respectively.

[0105] The system provides compliance status variables for each element. With each piece of evidence, the state variables of credibility An initial marginal distribution is set and an initialization message is written. The initialization rule is defined as assigning a uniform distribution to the initial messages of all variables and their neighboring factors. And the initial messages of all unary factors to their connecting variables are normalized according to the two values ​​of the corresponding potential function, so that the factor... With evidence factors Before the iteration begins, and Apply prior and likelihood constraints;

[0106] The system employs a synchronous iterative approach for message passing. In each iteration, all "factor-to-variable" messages are updated first, followed by all "variable-to-factor" messages. After each message update, the message vector is normalized to ensure that the sum of the messages in both states equals 1. The message update is calculated using the following formula:

[0107] ;

[0108] in express The variable nodes in Taken from set and express Factor nodes in Taken from set Represents variable nodes The value of and Indicates from variable node Send to factor node And the corresponding values The message component, Indicates from factor node Send to variable node And the corresponding values The message component, Represents variable nodes exist The set of adjacency factor nodes in the data. This indicates removing factor nodes from the adjacency set. The set after, Represents factor nodes The set of connected variable nodes, This indicates removing variable nodes from the set of connected variables. The set after, express Any variable node in, Represents a set A joint assignment vector of the values ​​of all variables within it. Represents a set A joint assignment vector of the values ​​of all variables within it. Represents factor nodes In joint assignment The value and specific form of the potential function are determined by the potential function table written in step S4. This indicates a series of multiplication operations. Indicates to Since the enumeration and summation operation is performed across the entire space and all dependent variables are binary, the number of enumerations is determined by the order of the factors. This represents a normalization operator that processes input vectors by dividing each of the two components by the sum of the two components, such that the sum of the normalized two components is equal to... Represents variable nodes In taking values The posterior approximate marginal distribution components;

[0109] The system sets the iteration threshold to 1. Set the message change threshold to And calculate the total change in messages across the entire graph after each iteration. And will Defined as the maximum absolute difference between all connected edges, all directional messages, and the two state components, when the number of iterations reaches... or Stop iteration when the time is right;

[0110] After the system stops iterating, it checks the compliance status variables of each element. With each evidence credibility state variable According to the formula above Calculate the posterior approximate marginal distributions of its two states and and Write out the posterior distribution of element compliance and the posterior distribution of evidence credibility respectively for use in step S6 for scoring calculation and subsequent evidence chain extraction.

[0111] In this specific embodiment, S6 includes:

[0112] The computer system reads the posterior distribution of compliance of elements and the posterior distribution of evidence credibility, and analyzes the temporally heterogeneous evidence graph. Each feature node in Generate three-dimensional scoring results and evaluation conclusions;

[0113] The system will include element nodes The corresponding element compliance status variable is denoted as And denote the marginal probability of the "compliant" state in its posterior distribution as ,in Represents the index of the feature node and is related to The element identifiers in the text correspond one-to-one. Representing feature nodes The element compliance status variable and its value range is express The posterior probability at time and the range of values ​​is ;

[0114] The system is based on the above Calculate the compliance score and uncertainty score and write them into the element scoring table. The element scoring table uses the element identifier as the primary key and includes a compliance score field, an uncertainty score field, and a coverage score field.

[0115] The system simultaneously loads the pre-defined sub-element mapping relationship table. Evidence Requirements Table ,in Each element node Deterministic mapping to two types of sub-elements DOC represents the sub-element of institutional and procedural categories and is used to constrain the coverage calculation of institutional document evidence. REC represents the sub-element of operational records and is used to constrain the coverage calculation of process execution record evidence. by Requires storing evidence for keys and This indicates the evidence type identifier and stores the required quantity for each evidence type under this element and its sub-element. ,in This indicates the sub-feature category identifier, with a value of DOC or REC. This indicates the evidence type identifier and is consistent with the evidence type field of the evidence node. Indicate the type of evidence At the element node and sub-elements The required quantity must be a positive integer;

[0116] The system in China-Israel element nodes The set of associated evidence nodes is obtained by searching along the association edges between the feature nodes and the evidence nodes from the starting node. and to Each evidence node Read its evidence credibility state variables The marginal probability of a "credible" state in the posterior distribution ,in Represents the index of the evidence node and is related to The evidence markers in the document correspond one-to-one. Indicates evidence node The evidence is credible, and the range of values ​​is [missing information]. , express The posterior probability at time and the range of values ​​is ;

[0117] The system fixes the credibility threshold at 1. And accordingly, satisfy The evidence nodes are identified as candidates for covering evidence and are categorized and statistically analyzed according to the evidence type field of the evidence nodes. Simultaneously, for each... Calculate the number of hits The Defined as in a set The type of evidence in China is equal to And the posterior confidence level meets the threshold. The count of different evidence markers;

[0118] The system provides each element node with... Calculate the compliance score once using the following formula Uncertainty score With coverage score And generate a three-dimensional scoring result. To generate evaluation conclusions:

[0119] ;

[0120] in Representing feature nodes The compliance score and the range of values ​​are Representing feature nodes The uncertainty score is defined as a binary random variable. The variance under the posterior distribution and its range is: Representing feature nodes The coverage score and the range of values ​​are This represents a three-dimensional scoring result vector composed of compliance score, uncertainty score, and coverage score in a fixed order. Representing sub-feature categories Coverage weight and take and Indicating the requirements for evidence In the given All types of evidence appearing below Summation, This function represents the minimum of the two values ​​and is used to truncate the number of hits to the required number to avoid excessive coverage of a single type of evidence, which would lead to an inflated coverage rate.

[0121] The system completes After calculation, the results are written into the feature scoring table, and an evaluation conclusion field is generated simultaneously. The evaluation conclusion field is defined as follows: Using a fixed threshold and recording three judgment indicators, subsequent steps can be based on compliance scores below a certain threshold. or uncertainty score higher or coverage score lower than Perform consistent gap location and closed-loop processing on element nodes.

[0122] In this specific embodiment, S7 includes:

[0123] The computer system reads the feature scoring table and evaluation conclusion fields and determines the target feature node set. The According to the formula Build, in Represents the set of indexes for the target feature nodes. Represents the feature node index and is a time-heterogeneous evidence graph. The element identifiers in the text correspond one-to-one. Representing feature nodes Compliance score, Representing feature nodes Uncertainty score, Representing feature nodes Coverage score, Indicates the compliance score threshold and takes Indicates the uncertainty scoring threshold and takes This represents the coverage score threshold, set to 0.90. Represents a logical OR operation;

[0124] The system targets each element node. exist The process involves extracting a deterministic chain of evidence and generating a traceable output, with the extraction process based on element nodes. The set of evidence nodes is obtained by performing an adjacency search along the associated edges between the feature nodes and the evidence nodes, starting from the feature node. During retrieval, the edge identifier and time attribute field of each associated edge are written into the edge list to form a "feature-evidence" primary link.

[0125] The system then Starting from the version, we expand the search along the version application edge, the responsibility attribution edge, and the source association edge to obtain the version node set. Set of Responsible Nodes With the set of evidence source nodes For each version, an edge identifier and time attribute field are written, and the version identifier, version validity start time, and version validity end time fields are written to the version node connected to it. For each responsibility attribution edge, an edge identifier and time attribute field are written, along with the authorization scope field, authorization validity start time, and authorization validity end time field. The responsibility node connected to it is written with the responsibility identifier, role field, and permission scope field. For each source association edge, an edge identifier and time attribute field are written, and the evidence source node connected to it is written with the source identifier, source system information, and source credibility attribute fields. At the same time, the system reads the audit trail information field from each evidence node and writes the event type, event occurrence time, operator identifier, and summary of content before and after change in the audit trail information into the audit trail list in ascending order of event time. This ensures that the evidence chain output simultaneously includes node identifier, edge identifier, time attribute, version validity attribute, source credibility attribute, authorization scope attribute, authorization validity attribute, and audit trail information.

[0126] The system encapsulates the above node list and edge list into a traceable output object. The results are written into a traceability output table, which uses the element identifier as the primary key and includes the element's three-dimensional scoring results, a list of evidence chain nodes, a list of evidence chain edges, and an audit trail list. The threshold judgment flag is also written into the table to support repeatable judgment during review.

[0127] The system further generates closed-loop improvement projects based on the gap type and writes them into the improvement project table. The gap type is generated by the following deterministic rule: when the evidence requirement table in step S6... When the number of corresponding evidence types exceeds the number of hits, a missing evidence gap is generated and the missing evidence type identifier and gap number are written into the improvement project. The posterior probability of the evidence's credibility is lower than the credibility threshold. And its type of evidence belongs to When the required range is specified, a gap is generated indicating insufficient credibility of the evidence, and it is written into the corresponding evidence identifier list and its evidence source identifier list. When the evidence generation time of the evidence node does not fall within the start and end time range of the version validity period of the version node it is connected to, a gap is generated indicating a mismatch between the evidence and the version validity period, and it is written into the evidence identifier, the version identifier, and the time fields of both. When the evidence generation time of the evidence node does not fall within the start and end time range of the authorization validity period of the responsibility attribution edge it is connected to, or when the permission object identifier and permission operation set of the responsibility information of the evidence node are not included in the permission range field of the responsibility node, a gap is generated indicating a mismatch between the evidence and responsibility, and it is written into the evidence identifier, responsibility identifier, authorization range field, and authorization validity period field. When a conflict factor correspondence has been established for two evidence nodes associated with the same element node in step S4, or when the conditions of mutual exclusion or contradiction are met in the conflict determination in step S4, an evidence conflict gap is generated, and the conflict evidence identifier pair and the rule type that triggers the conflict are written into the gap.

[0128] Each improvement project record includes a project identifier, element identifier, gap type, list of related evidence identifiers, responsible party field, and due date field. The responsible party field is taken from the element node. The responsible entity field and the due date field are obtained by adding the evaluation batch timestamp to the result with a fixed duration of 30 days and storing it using UTC timestamp millisecond values. This enables the automatic generation of target element positioning, traceable evidence chain output, and gap-driven closed-loop improvement projects based on evaluation conclusions.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0130] This invention addresses the technical problems in the evaluation of sponsors' drug clinical trial quality management systems, namely, the difficulty in automatically associating system elements with evidence materials, the difficulty in quantifying the credibility of evidence, and the lack of traceable support for evaluation conclusions, which hinders improvement efforts. It maps key constraints such as evidence materials, version applicability, responsibility and authorization boundaries, time attributes, audit trail information, and the credibility of evidence sources into node and edge attributes of a time-heterogeneous evidence graph. On this graph, a graph neural network is used to represent and learn the relationships between elements and evidence, as well as the credibility of evidence, and output relationship likelihood parameters and credibility likelihood parameters. These likelihood parameters are then introduced into a Bayesian factor graph and confidence propagation is performed to obtain the posterior distribution of element compliance and evidence credibility, thereby achieving global consistency inference and uncertainty quantification output for dispersed evidence. Furthermore, based on the posterior distribution, a three-dimensional scoring result is formed, comprising compliance score, uncertainty score, and coverage score. Evidence chains associated with target elements are extracted based on graph retrieval, outputting traceable evidence chain results containing information such as time, version, responsibility, and audit trail. This ensures that the evaluation conclusions have verifiable evidence and can be directly used to trigger and locate subsequent improvement actions.

[0131] In terms of algorithm structure, this invention does not simply superimpose general scoring rules, but makes targeted improvements to address issues of concern in clinical trial auditing and verification, such as version validity, authorization boundaries, source credibility, and contradictory evidence. Firstly, by incorporating version nodes and version validity periods, responsibility nodes and authorization scope and validity periods, audit trail information of evidence nodes, and edge time attributes into a heterogeneous evidence graph at the same time, the applicability of evidence and attribution of responsibility can be calculably expressed, ensuring the automatic construction of a traceable evidence chain from a data structure perspective. Secondly, by setting evidence source nodes and writing source credibility attributes, factors such as system verification status and data integrity verification results are transformed into inputs for evidence credibility inference, enabling the model to distinguish the credibility differences of the same evidence under different source conditions. Thirdly, by coupling the likelihood parameters output by the graph neural network with Bayesian belief propagation, a combination of local automatic association capabilities and global consistency inference is achieved. Furthermore, conflict factors determined based on time, version, authorization, and audit trail are introduced to suppress the interference of contradictory evidence on conclusions, thereby improving the accuracy, interpretability, and ability to pinpoint improvement targets.

Claims

1. A method for constructing and evaluating a sponsor's drug clinical trial quality management system, characterized in that, include: S1. Obtain data on the quality system elements, evidence materials, and sources of evidence from the sponsor's drug clinical trials, and perform standardized processing. S2. Construct a time-heterogeneous evidence graph based on standardized data, including element nodes, evidence nodes, version nodes, responsibility nodes, and evidence source nodes, as well as the association edges between element nodes and evidence nodes, the version applicability edges between evidence nodes and version nodes, the responsibility attribution edges between evidence nodes and responsibility nodes, and the source association edges between evidence nodes and evidence source nodes. Record time attributes for edges, record version validity attributes for version nodes, record authorization scope and authorization validity attributes for responsibility attribution edges, record audit trail information for evidence nodes, and set source credibility attributes for evidence source nodes; S3. Perform graph neural network representation learning on the time-heterogeneous evidence graph to generate likelihood parameters, including relationship likelihood parameters for the validity of relationships between element nodes and evidence nodes, and evidence credibility likelihood parameters for evidence nodes; S4. Construct a target Bayesian factor graph based on the time-heterogeneous evidence graph and likelihood parameters, including element factors, evidence factors, relationship factors, and conflict factors. Conflict factors are constructed based on the mutual exclusion or contradictory relationships between evidence nodes associated with the same element node. Mutual exclusion or contradictory relationships include conflicts determined based on time attributes, version validity attributes, authorization scope attributes, authorization validity attributes, or audit trail information. The joint probability of conflict is reduced by the factor potential function. S5. Perform confidence propagation operation on the target Bayesian factor graph to obtain the posterior distribution of element compliance for each element node and the posterior distribution of evidence credibility for each evidence node. S6. Calculate compliance score and uncertainty score based on the posterior distribution of element compliance. Calculate coverage score based on predefined evidence requirements and the posterior distribution of evidence credibility. Combine the compliance score, uncertainty score, and coverage score to obtain the three-dimensional scoring result and generate the evaluation conclusion. S7. Based on the evaluation conclusion, determine the target element node, extract the evidence chain associated with the target element node in the time heterogeneous evidence graph, and generate a traceable output associated with the evidence chain.

2. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S1 includes: Obtain the quality system element data, which includes element identifier, element name, element scope of application, corresponding process or activity, corresponding responsible entity, and corresponding evidence requirements. The evidence material data is obtained, which includes evidence identifier, evidence type, evidence content or evidence content index, evidence generation time, version information corresponding to the evidence, and responsibility information corresponding to the evidence, and includes audit trail information; Obtain the evidence source data, which includes a source identifier, source system information, and verification information used to determine the credibility of the source; Convert the above data fields into a preset data format, and establish the correspondence between element identifiers and evidence identifiers, and between source identifiers and evidence identifiers, according to the preset identifier alignment rules; The time field is subjected to timestamp normalization processing, which includes unifying the time format and time zone, and missing time fields are filled or marked.

3. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S2 includes: Element nodes are generated based on the standardized quality system element data; Evidence nodes are generated based on the standardized evidence data, and the audit trail information is written into the evidence nodes; Based on the standardized evidence data, a version node is generated and the version validity period attribute is written; a responsibility node is generated and the role or permission scope information is written. The associated edge, version applicable edge, and responsibility attribution edge are generated based on the identification alignment relationship, and the time attribute is written to each edge. Event nodes are generated based on the event information in the evidence material data, and event occurrence edges are generated between the evidence nodes and the event nodes, and time attributes are recorded for the event occurrence edges.

4. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, In S2, setting the source credibility attribute for evidence source nodes includes: Based on the computerized system verification status, audit trail integrity, and data integrity verification results, source credibility attributes are written to the evidence source nodes according to preset assignment rules, which include standardization and weighted summarization. The source credibility attribute is then written into the evidence node connected to it as the evidence source credibility attribute of the evidence node. When the same evidence node is connected to multiple evidence source nodes, the credibility attributes of the multiple sources are aggregated according to a preset aggregation rule to update the credibility attributes of the evidence source of the evidence node.

5. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S3 include: Node feature vectors are constructed for different types of nodes in the time heterogeneous evidence graph, and the time attributes of the edges and the version validity period attributes are encoded as time features. Based on the node feature vector and the time feature, a graph neural network is used to perform multiple rounds of message passing and aggregation updates on the time heterogeneous evidence graph to obtain the element node representation vector and the evidence node representation vector. The relationship likelihood parameter is calculated based on the element node representation vector and the evidence node representation vector; The likelihood parameter of the evidence credibility is calculated based on the evidence node representation vector and the evidence source credibility attribute of the evidence node.

6. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S4 includes: Set the compliance status variable of each element node in the time heterogeneous evidence graph, and set the credibility status variable of each evidence node in the time heterogeneous evidence graph. Generate element factors based on element compliance status variables, and generate evidence factors based on evidence credibility status variables; For the associated edges, generate relationship factors and connect the corresponding element compliance status variables and evidence credibility status variables; Based on the relationship likelihood parameter and the evidence credibility likelihood parameter, factor potential functions are set for the relationship factor and the evidence factor, respectively. For a set of evidence nodes associated with the same element node, mutual exclusion or contradiction relationships are determined based on the evidence node's time attribute, version validity period attribute, authorization scope attribute, authorization validity period attribute, and audit trail information. Conflict factors are then generated and connected to the corresponding evidence credibility status variables.

7. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S5 include: Set initial marginal distributions for compliance status variables of each element and credibility status variables of each piece of evidence; Iterative message passing is performed in the target Bayesian factor graph, which includes message computation from factor to variable and message computation from variable to factor, and the messages are normalized after each iteration. The iteration stops when the number of iterations reaches a preset threshold or when the change in messages between two adjacent iterations is less than a preset change threshold. The compliance posterior distribution of the element and the credibility posterior distribution of the evidence are calculated based on the message at the time of termination of the iteration.

8. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S6 include: The posterior mean is calculated as the compliance score based on the posterior distribution of the aforementioned elements, and the posterior uncertainty is calculated as the uncertainty score. The posterior uncertainty is determined by the variance of the posterior distribution or by the width of the confidence interval under a preset confidence level. Based on predefined sub-feature mapping relationships, feature nodes are mapped to at least two types of sub-features, and the denominator is determined for each type of sub-feature according to evidence requirements; In the time-heterogeneous evidence graph, evidence nodes that are associated with element nodes and whose posterior distribution of evidence credibility meets a preset credibility threshold are retrieved as coverage evidence, and coverage scores are calculated. The compliance score, uncertainty score, and coverage score are combined to generate a three-dimensional score result.

9. The method for constructing and evaluating a sponsor's drug clinical trial quality management system according to claim 1, characterized in that, S7 includes: Element nodes whose evaluation conclusions indicate that their compliance score is lower than the preset compliance score threshold, their uncertainty score is higher than the preset uncertainty score threshold, or their coverage score is lower than the preset coverage score threshold are identified as target element nodes. In the time-heterogeneous evidence graph, the evidence node set is retrieved along the associated edge starting from the target element node, and the retrieval is further extended along the version application edge, responsibility attribution edge and source association edge to obtain the version node set, responsibility node set and evidence source node set to generate the evidence chain; The identifiers of each node and edge in the evidence chain, along with their time attributes, version validity attributes, source credibility attributes, authorization scope attributes, authorization validity attributes, and audit trail information, are written into the traceable output. The traceable output further includes closed-loop improvement projects generated based on gap types, which include missing evidence, insufficient evidence credibility, evidence not matching version validity, evidence not matching responsibility, or evidence conflict.

Images